CN110309427A

CN110309427A - A kind of object recommendation method, apparatus and storage medium

Info

Publication number: CN110309427A
Application number: CN201810553549.1A
Authority: CN
Inventors: 丘志杰; 饶君; 张博; 林乐宇; 冯喆; 陈磊; 胡澜涛; 刘书凯; 刘毅; 孙振龙; 王良栋
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2018-05-31
Filing date: 2018-05-31
Publication date: 2019-10-08
Anticipated expiration: 2038-05-31
Also published as: CN110309427B

Abstract

The embodiment of the present invention provides a kind of object recommendation method, apparatus, storage medium and computer equipment, the present embodiment is based on Recognition with Recurrent Neural Network, the corresponding user's access sequence of multiple sample of users is trained, obtain recommended prediction model, coding calculating is carried out to user's access sequence, similarity calculation is carried out to obtained coding vector and each candidate term vector again, obtains the recommended of the user.It can be seen that, the characteristics of calculating is being encoded based on Recognition with Recurrent Neural Network, so that this model based coding mode of the present embodiment, the long history interest and short-term history interest of user are taken into account, and consider the access order that user accesses object in application platform, it is capable of the interest transition and interest accumulation of more acurrate positioning user, solves the problems, such as that existing ItemCF object recommendation method leads to gained recommended diversity and personalized loss.

Description

A kind of object recommendation method, apparatus and storage medium

Technical field

The present invention relates to technical field of data processing, and in particular to a kind of object recommendation method, apparatus and storage medium.

Background technique

Nowadays, popularizing for internet brings a large amount of information to user, meets user in the information age to information Demand, but with the rapid development of network, network information increases substantially, and when user faces bulk information, is difficult therefrom to obtain The part information actually useful to oneself is obtained, the service efficiency of information is reduced instead.In this regard, technical staff proposes recommendation The interested information of user, product etc. are recommended user, realize individual character by system that is, according to the information requirement of user, interest etc. Change information recommendation, be widely applied to many fields at present, as news is recommended, commercial recommendations, entertainment recommendations, study recommendation, Life recommendation etc..

Currently, common recommended method mainly has based on collaborative filtering (Collaborative Filtering, abbreviation CF) Recommended method, that is, pass through calculate Item-Item (Item can be the objects such as an article or a video) similarity side Formula obtains the object accessed with user, and K most like Item is as recommended.

Wherein, Item-Item similarity calculation is usually that the method for using Item CF is realized, i.e., using Item in user Access object sequence in cooccurrence relation, calculate the similarity of two Item.When different user accesses pair whithin a period of time As Item is identical, but the access order of each access object is different, the recommendation for each user that the recommended method based on Item CF obtains Object will be identical, cannot achieve personalized recommendation, also will affect recommendation accuracy rate, and obtained recommended cannot be considered in terms of use The Long-term Interest and short-term interest at family.

Summary of the invention

The embodiment of the present invention provides a kind of object recommendation method, apparatus, storage medium and computer equipment, may be implemented pair The accurate positionin of user interest transition and interest accumulation, meets different user personalized recommendation demand, and gained recommended is simultaneous The Long-term Interest and short-term interest for having cared for user, improve the accuracy to user's recommended.

To achieve the above object, the embodiment of the present invention provides the following technical solutions:

A kind of object recommendation method, which comprises

User's access sequence is obtained, user's access sequence is the object generation that application platform output is accessed based on user 's；

User's access sequence input recommended prediction model is subjected to coding calculating, obtains user's access pair The coding vector of elephant, the recommended prediction model are based on Recognition with Recurrent Neural Network, user corresponding to multiple sample of users Access sequence training obtains；

Similarity calculation is carried out to the coding vector and each candidate term vector；

Based on similarity calculation as a result, obtaining the recommended of the user.

A kind of object recommendation device, described device include:

Retrieval module, for obtaining user's access sequence, user's access sequence is to access application based on user What the object of platform output generated；

Computation model is encoded, for user's access sequence input recommended prediction model to be carried out coding calculating, The coding vector that the user accesses object is obtained, the recommended prediction model is based on Recognition with Recurrent Neural Network, to multiple The corresponding user's access sequence training of sample of users obtains；

First similarity calculation module, for carrying out similarity calculation to the coding vector and each candidate term vector；

Recommended preference pattern, for based on similarity calculation as a result, obtaining the recommended of the user.

A kind of storage medium is stored thereon with computer program, and the computer program is executed by processor, and realizes as above Each step of the object method.

A kind of computer equipment, the computer equipment include:

Communication interface；

Memory, for storing the computer program for realizing the upper object method；

Processor, for recording and executing the computer program of the memory storage, the computer program is for real Existing following steps:

Based on the above-mentioned technical proposal, a kind of object recommendation method, apparatus, storage medium and meter provided in an embodiment of the present invention Machine equipment is calculated, the present embodiment is based on Recognition with Recurrent Neural Network, is trained, obtains to the corresponding user's access sequence of multiple sample of users To recommended prediction model, coding calculating carried out to user's access sequence, then to obtained coding vector and each candidate word to Amount carries out similarity calculation, obtains the recommended of the user.As it can be seen that the spy calculated in coding based on Recognition with Recurrent Neural Network Point, so that this model based coding mode of the present embodiment, has taken into account the long history interest and short-term history interest of user, and consider The access order that user accesses object in application platform has been arrived, the interest transition and interest product of more acurrate positioning user are capable of It is tired, solve the problems, such as that existing Item CF recommended method leads to gained recommended diversity and personalized loss.

Detailed description of the invention

In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this The embodiment of invention for those of ordinary skill in the art without creative efforts, can also basis The attached drawing of offer obtains other attached drawings.

Fig. 1 is each GRU layers of schematic network structure in a kind of Recognition with Recurrent Neural Network；

Fig. 2 is a kind of flow diagram of object recommendation method provided in an embodiment of the present invention；

Fig. 3 is a kind of network architecture schematic diagram of recommended prediction model provided in an embodiment of the present invention；

Fig. 4 is the flow diagram of another object recommendation method provided in an embodiment of the present invention；

Fig. 5 is the flow diagram of another object recommendation method provided in an embodiment of the present invention；

Fig. 6 is the flow diagram of another object recommendation method provided in an embodiment of the present invention；

Fig. 7 is a kind of keyword sequence generation method schematic diagram provided in an embodiment of the present invention；

Fig. 8 is the flow diagram of another object recommendation method provided in an embodiment of the present invention；

Fig. 9 is the flow diagram of another object recommendation method provided in an embodiment of the present invention；

Figure 10 is the flow diagram of another object recommendation method provided in an embodiment of the present invention；

Figure 11 is a kind of application flow schematic diagram of object recommendation method provided in an embodiment of the present invention；

Figure 12 is a kind of structural schematic diagram of object recommendation device provided in an embodiment of the present invention；

Figure 13 is the structural schematic diagram of another object recommendation device provided in an embodiment of the present invention；

Figure 14 is the structural schematic diagram of another object recommendation device provided in an embodiment of the present invention；

Figure 15 is the structural schematic diagram of another object recommendation device provided in an embodiment of the present invention；

Figure 16 is a kind of hardware structural diagram of computer equipment provided in an embodiment of the present invention.

Specific embodiment

The inventors of the present invention discovered through researches that: have [X, Y, Z] when user A reads article within a certain period of time, user B is read Article have [Y, Z, X], X, Y and Z can be user read article article ID, however, it is not limited to this, obtain the two When the respective recommendation article of user, mainly since the object recommendation method based on Item CF only considers that history is read in article Hold, does not consider that user reads the reading order of article, cause the recommended obtained based on similarity calculation result identical, nothing Method accurately embodies the respective interest transition of the two users and interest accumulation.

Moreover, inventor also found when carrying out similarity calculation, it is difficult control using how many a candidate targets and carries out phase It is calculated like property, if the candidate target of nearest a period of time is selected to carry out similarity calculation, gained recommended can only characterize user Short-term interest；It is not only computationally intensive if the longer candidate target of history is selected to carry out similarity calculation, and what be will lead to pushes away It recommends object is very more, needs further to be ranked up screening to obtained a large amount of recommendeds using sort algorithm, process compares It is cumbersome, affect the efficiency for obtaining the recommended of user.

Based on above-mentioned analysis, inventor proposes a kind of new object recommendation method, can not only consider user in application platform The access order for accessing object (i.e. reading object, viewing video etc.), positions the interest transition and interest of user more accurately Accumulation, and the long history interest and short-term history interest of user has been taken into account, improve the accuracy of the recommendation results of user And efficiency, meet the personalized recommendation demand of different user.

Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other Embodiment shall fall within the protection scope of the present invention.

In order to facilitate understanding recommended method provided in this embodiment, the Recognition with Recurrent Neural Network that the present embodiment is related at this The principle of (Recurrent neural Network, RNN) is briefly described.RNN is a kind of node orientation connection cyclization Artificial neural network, internal state can show dynamic time sequence behavior, and be different from feedforward neural network, Multi-Layer Feedback RNN can use internal memory to handle the list entries of arbitrary sequence, can be easier to handle the hand if not being segmented Write identification, speech recognition etc..

Wherein, RNN can handle the neural network of elongated data as one kind, the historical information of random length can be encoded To a hidden layer (hiddenlayer), the i.e. intermediate output of neural network, certain implicit form of expression of input is characterized, is led to Chang Weiyi vector or matrix.Shown in following formula (1), RNN can compression dimensionality reduction to high latitude data.Especially with close LSTM (Long Short-Term Memory, shot and long term memory network) and GRU (Gated Recurrent are used in several years RNN Unit, gating cycle unit) extensive use, RNN is successfully used to solve natural language processing NLP (Natural Language Process) the technical issues of, such as machine translation, sequence prediction, Speech processing etc..

h_t=g (Wx_t+Uh_t-1) (1)

In formula (1), x_tIt can indicate that characterization vector currently entered (can be access object or visit in the present embodiment Ask that the embedding vector can also be denoted as term vector by the embedding vector of the keyword in object, the present embodiment), h_t-1It can indicate that last moment hidden layer exports, the h of initial time_t-1For a null vector, W and U respectively indicate mapping matrix. In practical applications, this calculation formula can be used to current input carrying out calculating with historical information to merge, and obtain one newly Coding vector (in the present embodiment, which can be encoding vector).

As it can be seen that, at each moment, historical information and current input can be carried out efficient coding in the application of RNN It calculates, obtains a new data representation vector, i.e. coding vector.In the present embodiment, which can be user's visit Ask the history access data (such as reading article, the data for watching video generation etc.) that the object of application platform output generates, when The term vector etc. of the preceding keyword for inputting the term vector or the access object that can be the access object at current time.

It wherein, is by upper a period of time if above-mentioned formula indicates in the coding calculating process of Recognition with Recurrent Neural Network hidden layer The output at quarter obtains the output of subsequent time, by an activation primitive such as tanh function plus the input at this moment with such It pushes away, what is finally exported is the coding vector that the present invention needs.

Since traditional RNN does not consider the increase with the expansion number of plies, network depth can become very deep, lead to reversed ladder Degree, which is propagated, there is exception, as gradient disperse and gradient are exploded.In order to solve this problem, the Recognition with Recurrent Neural Network that the present embodiment uses It can be and increase GRU or LSTM model in traditional RNN, obtained modification Recognition with Recurrent Neural Network, in the circulation of the present embodiment Multiple GRU layers or LSTM layers multiple are generally included in neural network, specific network structure the present embodiment is not described further.The present embodiment Herein only for including multiple GRU layers of Recognition with Recurrent Neural Network, to illustrate to obtain the calculating process of coding vector, about GRU Principle the present embodiment it is not described here in detail.

When calculating hidden layer using GRU, as shown in Figure 1, it is believed that the output of last moment hidden layer is current hide A part of the linear weighted combination of layer output, referring in particular to formula (2):

In formula (2),It can indicate that the candidate of current (i.e. t moment) hides output, i.e., the centre of current hidden layer is defeated Out, it needs to export to be weighted with the hidden layer of last moment and merge, obtain final hidden layer output；Weighted factor Z_tIt can be with H is exported by last moment hidden layer_t-1With current input x_tIt carries out automatic adaptation to be calculated, the following formula of calculation formula (3):

z_t=σ (W_zx_t+U_zh_t-1) (3)

As it can be seen that as last moment hidden layer exports h_t-1The variation inputted with current candidate hidden layer, weighted factor Z_t (i.e. the update door of GRU) all can also change, but weighted factor Z_tBe eventually mapped to (0,1] in section, Z_tBigger expression More value current input information, gives higher weight, the present embodiment is to the Z_tSpecific value be not construed as limiting.Wherein, exist In GUR, the U and W in formula (3) are usually smaller, and σ can be coefficient factor, second, left side σ as shown in figure 1, this implementation Example is not construed as limiting U, W, σ specific value.

In addition, hiding output for the current candidate in above-mentioned formula (2)It can use activation primitive, such as hyperbolic Tangent function tanh is calculated, and can retain the symbolic information and size information being originally inputted, specific formula for calculation (4) is such as Under:

The r in formula (4)_tDoor can be reset, and the Z in above-mentioned formula_tIt can be update door, i.e. two in GRU model Door, the resetting door can be calculated using mode shown in formula (5), and however, it is not limited to this.

r_t=σ (W_rx_t+U_rh_t-1) (5)

σ can be coefficient factor in formula (5), and first, left side σ the present embodiment as shown in figure 1 does not make its specific value It limits.Furthermore, it is desirable to illustrate, the meaning indicated about same letter in the various embodiments described above is identical, is referred to above-mentioned to public affairs The explanation of formula (1) corresponding part.

In conjunction with above-mentioned formula (2)~(5) description and the structure of GRU shown in FIG. 1, it includes two doors, that is, reset Door r_tWith update door Z_t, by resetting door r_tMultiplied by the output h of last moment hidden layer_t-1, whether see will reset or reset much journeys Degree, later with the input x at current time_tSplicing obtains implicit variable h by activation primitive tanh operation_t, then, by upper one The output h at moment_t-1Variable h is implied with this_tLinear combination is carried out, the output at current time is obtained, and so on, it is predicted Required coding vector, wherein carry out the two weight of linear combination and for 1, and the implicit variable h_tWeight be update The output of door, characterization renewal intensity are much.

In the present embodiment practical application, before the recommended for obtaining active user, need first to train to obtain object Recommend prediction model, in conjunction with above-mentioned Recognition with Recurrent Neural Network principles illustrated, it is known that the Recognition with Recurrent Neural Network is encoded to input When calculating, can consider the time sequencing of input, meet the individual requirement to the prediction result of user, and mode input no matter It is longer or compared with short data, the sequence of output result is achieved that during coding, so that thus obtained recommendation results, energy Enough take into account the long history interest and short-term history interest of user.

Based on this, in the present embodiment, the corresponding user's access sequence of available multiple sample of users is followed based on above-mentioned Ring neural network carries out model training to these user's access sequences, obtains recommended prediction model, the present embodiment is to the mould The specific implementation process of type training is not construed as limiting.

Wherein, user's access sequence of sample of users can be by pair of access object of the sample of users in application platform As mark composition, such as according to access time sequence, user's access is generated by the access object ID extracted from history access data Sequence, at this time user's access sequence be properly termed as access object sequence, the present embodiment can using the access object sequence as Training data, implementation model training.

Model is constantly being trained in optimization process, can model prediction result (such as recommended) that this is obtained with The difference of actual access object meets preset condition, i.e., this training pattern meets constraint condition, this training can be obtained Model as recommended prediction model.But the content of the constraint condition is not construed as limiting, and the present embodiment can also limit Optimize number or training data item number, with the training optimization processing of Controlling model, the present embodiment is pre- to how to obtain recommended The implementation method for surveying model is not construed as limiting.

As an alternative embodiment, in order to improve the real-time of recommendation results, the present embodiment is obtaining training data process In, the sequence in each session of available each user is as training data.The session information can be user Current interface flushes to the time window refresh next time before, and internal information is usually all smaller on time difference, makes Model training is carried out with this training data, the recommended prediction model recycled, the recommended energy predicted Enough feedbacks that user is obtained in shorter time interval are more easier to catch user's to guarantee recommendation results real-time Short-term interest.According to this design, the present embodiment can expand the history length of training data, obtain the Long-term Interest of user.

Further, in order to keep training data more abundant, the present embodiment can be for the history access of a sample of users Data constitute a plurality of training data.Assuming that the object identity of object is accessed in application platform by user, obtained access object Sequence is [x1, x2, x3, x4, x5, x6], thus the present embodiment can constitute out following a plurality of training data such as: ([x1], x2), ([x1, x2], x3), ([x1, x2, x3], x4), ([x1, x2, x3, x4], x5) and ([x1, x2, x3, x4, x5], x6), In, the round parentheses left side is currently known access object sequence, and on the right of round parentheses is the target object for needing to predict.Carry out During model training, it can use this plurality of training data and carry out model training, to improve gained recommended prediction model Forecasting accuracy, the present embodiment, to how to be based on Recognition with Recurrent Neural Network, carries out model instruction to above-mentioned a plurality of training data at this Experienced process is not described further.

It is to be appreciated that for the recommended prediction model that the present embodiment obtains sample can be utilized over time User's access sequence that this user updates, optimizes the recommended prediction model, to improve forecasting accuracy, this reality It applies example and its optimization process is not detailed at this.

Referring to Fig. 2, for the embodiment of the invention provides a kind of flow diagram of object recommendation method, this method can be answered For servicing side, i.e. this method can execute realization by server, can specifically include but be not limited to following steps:

Step S101 obtains user's access sequence；

Wherein, the object which can access application platform output based on user generates, the present embodiment pair The specific generating mode of user's access sequence is not construed as limiting, and as needed, the sequential element which includes It can be the object identity of access object, or the keyword of access object, for user's access sequence of different content, Generating mode is often different, and is specifically referred to the description of hereafter corresponding embodiment.

It in the present invention, can be user's access sequence referred to as access object sequence of object identity by sequential element, it will Sequential element is that user's access sequence of keyword is known as keyword sequence, the type of user's access sequence of embodiment not office It is limited to both sequences that the present embodiment is enumerated.

User's access sequence input recommended prediction model is carried out coding calculating, obtains user's visit by step S102 Ask the coding vector of object；

As described above, which can be based on Recognition with Recurrent Neural Network, corresponding to multiple sample of users User's access sequence training obtain, for different training datas, the representation of obtained recommended prediction model can With difference, the meaning that model output data indicates can be different, but the processing logic of the model can be identical, and the present embodiment is to pushing away Object prediction model is recommended not to be described further the coding calculating process of list entries.

Wherein, each sequence when inputting user's access sequence to recommended prediction model, in user's access sequence Element sequentially inputs the recommended prediction model, at this point, an input of recommended prediction model can be accordingly The sequential element at moment.

By taking recommended prediction model configuration diagram shown in Fig. 3 as an example, to the place of user's access sequence input model Reason process carries out decompression explanation, the recommended prediction model of the present embodiment may include it is GRU layers multiple, i.e. it is multiple in Fig. 3 GRU Cell, in conjunction with the description of above-mentioned GRU principle, an every GRU layers of input is the output of previous moment hidden layer and currently inputs, Output is subsequent time hidden layer status information.

Wherein, in an every GRU layers of calculating process, usually utilize it includes resetting door and update door, realize wait The calculating of hidden layer is selected, and its control is to retain the information of how many previous moment hidden layers, and how many candidates are added in control The information of hidden layer, to be exported.Therefore, it is realized using recommended prediction model provided in this embodiment and user is accessed The coding of sequence calculates, and can flexibly control long short-range Dependency Specification using multiple GRU layers, be suitble to portray sequence data, I.e. remain user in application platform for a long time before reading article simultaneously, and can protrude in the recent period read article so that What the present embodiment obtained is used to predict the coding vector of user's recommended, not only takes into account the Long-term Interest of user and short-term emerging Interest improves forecasting accuracy, meets each personalization of different user and push away it can be considered that user accesses the access order of object Recommend requirement.

It is to be appreciated that the coding vector that step S102 is obtained can be the last layer hidden layer in recommended prediction model The coding vector of output.And above-mentioned recommended prediction model is not limited to configuration diagram shown in Fig. 3, wherein interbed May include it is LSTM layers multiple, can based on the principle of LSTM, realize hidden layer calculate, obtain user access object coding to Amount, the present embodiment are not described further the coding calculating process of this recommended prediction model.

Since LSTM is a kind of time recurrent neural network, it generally is suitable for being spaced and postponing in processing and predicted time sequence Relatively long critical event is usually to introduce three gating devices, to handle memory/forgetting, the input journey of memory unit The problem of degree, output degree, structure is more complicated；And GRU can introduce Reset Gate (resetting door) and Update Gate (updating door), the parameter needed is few, and faster, structure is also relatively easy for training speed, and the present embodiment can select according to actual needs Select the training which kind of Recognition with Recurrent Neural Network to realize recommended model based on, and the coding of user's access sequence to active user It calculates, is only illustrated by taking configuration diagram shown in Fig. 3 as an example herein.

Step S103 carries out similarity calculation to the coding vector and each candidate term vector；

It is to be appreciated that the coding vector is consistent with candidate word vector dimension, and certainly, in above-mentioned coding calculating process, Generated term vector is also identical with finally obtained coding vector dimension, thus guarantee Similarity measures it is normal into Row, the present embodiment are not construed as limiting the particular content and dimension of the dimension of vector.

In different scenes embodiment, which can be different, such as by user's history access pair As the term vector directly obtained, or the term vector etc. obtained by the keyword that user's history accesses, the present embodiment is to word The specific generating mode of vector is not construed as limiting, for example obtains access object or the term vector of keyword etc. using Word2Vec, but It is not limited thereto.

Wherein, Word2Vec can be by distributing a dense vector to each word, with discrete type characteristic processing- One-hot encoding one-hot identification method is contrasted, the semantic dimension information being able to maintain between word and word, the present embodiment pair Embedding Layer in recommended prediction model generates corresponding each access object or crucial calculating The concrete methods of realizing of embedding is not described further.

Optionally, the present embodiment can be real using such as cosine similarity (i.e. Cosine similarity) this similarity algorithm Now to the similarity calculation between two vectors, but it is not limited to a kind of this similarity calculation mode, the present embodiment is only with this For carry out similarity calculation explanation.

Wherein, cosine similarity calculates the following formula of formula (6) used:

In formula (6), u, v respectively indicate the coding vector and candidate term vector that user accesses object, the two vectors Dimension is identical, and u_iIndicate the characteristic value of i-th dimension degree in the coding vector, v_iIndicate the feature of i-th dimension degree in candidate's term vector Value.

Step S104, based on similarity calculation as a result, obtaining the recommended of user；

Optionally, in the present embodiment, for multiple candidate targets of storage, corresponding term vector and encode to The similarity of amount is higher, shows that corresponding candidate object is bigger as the probability of the recommended of the object, i.e. corresponding candidate object A possibility that as prediction recommended, is bigger.As it can be seen that the candidate target that similarity is bigger, user is to its interested possibility It is bigger.

Based on this, the present embodiment can choose the corresponding candidate target of similarity for reaching preset threshold, as the user Recommended；Or select the corresponding candidate target of several highest similarities, the recommended etc. as user, this reality Apply example to how using Similarity measures as a result, the implementation for obtaining the recommended of the user is not construed as limiting.

As another embodiment, if above-mentioned candidate's term vector is the term vector of keyword, similarly, similarity is higher, explanation The user is interested in corresponding keyword, and corresponding keyword is selected as candidate keywords, to determine that the probability of recommended is got over Greatly, thus, in this case, similarity size can indicate user to the interested probability size of corresponding keyword, into And can indicate that the candidate target comprising corresponding keyword becomes the probability size of recommended, the present embodiment can choose similar Several higher keywords are spent, to determine that the interested recommended of user's most probable, specific implementation process are referred to The hereafter description of corresponding embodiment.

In conclusion the present embodiment be based on Recognition with Recurrent Neural Network, to the corresponding user's access sequence of multiple sample of users into Row training, obtains recommended prediction model, and realization carries out coding calculating to user's access sequence, not only takes into account the length of user Phase historical interest and short-term history interest, and the access order that user accesses object in application platform is considered, in this way, base In obtained coding vector and each candidate term vector similarity calculation as a result, obtaining the recommended of the user can accurately determine Position user interest transition and interest accumulation, solve existing Item CF recommended method cause gained recommended diversity and The problem of personalization loss.

It can be by sample in the training process for carrying out recommended prediction model as an alternate embodiment of the present invention User accesses object Item in the above-mentioned history of application platform, is directly becoming access object sequence, is obtained by the access object sequence To the training data for carrying out model training, model training is carried out based on Recognition with Recurrent Neural Network, obtains recommended prediction model.This In the case of planting, the present invention can use object recommendation method shown in Fig. 4, to obtain the recommended of active user, to incite somebody to action It realizes subsequent primary election logic as candidate item, and then the target recommendation pair pushed to user client is obtained by sequence logic As facilitating user fast and accurately to access required object.

The flow diagram of another object recommendation method as shown in Figure 4, this method can also be applied in server side, It can specifically include but be not limited to following steps:

Step S201 obtains a plurality of history access data of user；

In practical applications, user log in application platform, application platform would generally export multiple access objects (such as article, Video, picture etc.), since the content of access object is excessive, the summary or label of each access object are often exported, user is needed Into the display interface of access object, the particular content of the access object can be just shown.

For example, news is read in application platform, it will usually export the title of many news, user can be new according to each item The title of news selects interested news, into the news content display interface, reads the detailed content of this news.It is regarding Frequency plays in application platform, can also show many videos, in order to facilitate user's selection, each video also has corresponding mark Topic or brief description, user select interested video that can just enter the video playing interface, can just play video content.

As it can be seen that the above-mentioned application platform of the present embodiment can be the application platform of the common various APP of user, as audio-video is answered With social applications platforms such as platform, browser application platform, instant messaging application platform etc., correspondingly, the application platform is defeated The access object for user's access out can be a video, an article etc., can be denoted as Item.

Moreover, user carries out in object accesses operating process in application platform, it will corresponding access data are generated, it should Access data can be used as history access data storage into the database of application platform, for showing user in application platform History access behavior.It is to be appreciated that during user operates in application platform, in addition to history access data can be generated Outside, other historical behavior data can also be generated, the present embodiment mainly analyzes history access data, to other behavior numbers According to not being described further.

Wherein, above-mentioned history access data include the letters such as the object identity for accessing object, the content, the title that access object Breath, which shows which object user accesses, which article has such as been read, and has viewed which video etc., that is, is used to area Divide each access object.Therefore, the object identity in the present embodiment can be object ID etc., and the present embodiment accesses data packet to history The content that the content and object identity contained refers to is not construed as limiting.

Optionally, it when application platform storage historical behavior data (it includes history to access data), can be marked according to user Know and carry out classification storage, is i.e. different user is operated in application platform, and generated historical behavior data can be with the use It is stored after the user identifier association at family, so as to the subsequent historical behavior data for quickly finding certain user, but the present embodiment is to going through The specific storage mode of history behavioral data is not construed as limiting, and user identifier may include but be not limited to user account.Based on this, The present embodiment can inquire the associated history access number of user identifier with target user directly from the database of application platform According to, but it is not limited to a kind of this acquisition modes.

Step S202, the object identity separately included by multiple history access data constitute access object sequence；

As described above, each history access data that the present embodiment obtains may include showing current accessed object Object identity, such as object ID can extract each history access data after obtaining a plurality of history access data in effective time The object identity for including, and according to the generation time of each history access data (i.e. the accordingly access time of access object), it is raw At corresponding access object sequence, that is to say, that the element in the object sequence can be each visit of the user in application platform Ask object object identity constitute, and between these elements according to generate Annual distribution, such as according to generation the time sequencing, The object identity of each access object is ranked up, object sequence, specific generating mode of the present embodiment to object sequence are obtained It is not construed as limiting.

For example: the object ID of each access object is such as indicated with x1, x2, x3, x4, x5, x6, what the present embodiment generated Access object sequence can be [x1, x2, x3, x4, x5, x6], it is seen then that, can by the sequential element content of the access object sequence To determine that user accesses which object has in effective time in application platform.If access object is article, user can be learnt Which article is read within effective time, and these articles read will be ranked up according to reading order, so as to accurately fixed Interests change and the interest accumulation of position user.Step S203 carries out access object sequence input recommended prediction model Coding calculates, and obtains the coding vector that user accesses object

Wherein, which can be based on Recognition with Recurrent Neural Network, access corresponding to multiple sample of users Object sequence training obtains, and specific training process is referred to the description of foregoing embodiments corresponding portion.

After obtaining recommended prediction model, each sequential element that can include by the access object sequence is successively defeated Enter the recommended prediction model, it is above-mentioned right which is referred to the coding calculating process of list entries The description process of RNN principle, the present embodiment are not described in detail here.

Referring to recommended prediction model configuration diagram shown in Fig. 3, in the present embodiment, input layer input is to use The access object sequence at family, sequential element are the access object Item that the user once accessed in application platform, warp After the dimensionality reduction of Embedding Layer calculates, the corresponding embedding vector (the present embodiment of each access object Item will be generated Term vector can be denoted as).Wherein, the Embedding Layer's of Recognition with Recurrent Neural Network is actually at Data Dimensionality Reduction Layer is managed, the present embodiment specifically can use Word2Vec realization, and concrete methods of realizing is not construed as limiting.

After the obtained corresponding embedding vector of each Item, i.e., each moment corresponding embedding vector can be with The middle layer that multiple GRU Cell are constituted is inputted to utilize such as above-mentioned formula (2), (3), (4) and (5) in every layer of calculating process Coding calculating is carried out, i.e., is believed using last moment hidden layer output (it is usually a coding vector) and the input at current time It ceases (i.e. current time corresponding term vector) and carries out continuous iteration update, the encoding vector finally exported is user's access The coding vector of object.

It should be noted that coding calculating is carried out about to access object sequence, and in a manner of obtaining coding vector, not office It is limited to model framework schematic diagram shown in Fig. 3, also can use multiple LSTM layers of compositions and obtain the realization of recommended prediction model, It is referred to the realization of LSTM principle, this will not be detailed here for the present embodiment.

In addition, needing the embedding vector of each access object Item generation in the present embodiment coding calculating process Dimension and the dimension of encoding vector be consistent, to guarantee subsequent to be able to carry out Similarity measures, and the present embodiment pair The particular content and dimension of the dimension of vector are not construed as limiting.

Step S204 obtains multiple candidate access objects；

Optionally, the high-quality access object that the present embodiment can choose application platform output is constituted and is waited as candidate target Selected works close, wherein high-quality object can be the biggish object of rate of people logging in application platform, the hot topic in nearest social networks The more object of relevant object, nearest a period of time amount of access or at random application platform output object it is collected right As etc., the present embodiment is not construed as limiting the mode for selecting high-quality access object (i.e. the component of candidate collection), not office It is limited to the these types of selection mode that the present embodiment is enumerated.

In practical applications, the candidate target in candidate collection can be with the variation of time, according to selection candidate target Mode, the component of candidate collection is constantly updated, to improve the accuracy of the subsequent recommended for obtaining user accordingly, The present embodiment is not construed as limiting the update mode of candidate collection.

Each candidate access object input language model is obtained corresponding candidate term vector by step S205；

Optionally, which can be Word2Vec, and however, it is not limited to this.If by each of candidate collection Element (i.e. candidate target) is denoted as Xn, and n is positive integer, and specific value size is not construed as limiting, available by these candidate targets Candidate collection can calculate the embedding vector of each candidate target later for [X1, X2, X3 ... Xn] (candidate term vector can be denoted as at this time), implementation method embedding corresponding with each access object Item of above-mentioned calculating to The calculation method of amount may be the same or different, and the present embodiment is not construed as limiting the implementation method for obtaining candidate term vector.

It is to be appreciated that above-mentioned steps S204 and step S205 can be executed in the arbitrary steps before similarity calculation, and It is not limited to the position of the present embodiment description.The present embodiment can store the candidate collection of acquisition, when needing to carry out When similarity calculation, the corresponding candidate term vector of each candidate access object in the candidate collection can be calculated, i.e. the present embodiment is retouched The mode stated.

Certainly, the present embodiment can also precompute the corresponding candidate word of each candidate access object in the candidate collection to Amount directly stores each candidate term vector, in this way, directly acquiring when needing to carry out similarity calculation with coding vector The corresponding candidate term vector of pre-stored each candidate access object, does not need to improve work efficiency in line computation.This feelings Under condition, step S204 and step S205 may be combined in that and obtain the corresponding candidate term vector of multiple candidate access objects, His step is identical, and the present embodiment no longer individually illustrates.

Step S206 carries out similarity calculation to the candidate term vector of the coding vector and each candidate target；

It is to be appreciated that the present embodiment is not construed as limiting the method for how calculating the similarity between two vectors, can adopt With Cosine similarity calculating method described above, the calculation method reciprocal or Pearson's phase of distance can also be used Relationship number calculating method etc., the present embodiment does not do be described in detail one by one herein.

Step S207 screens the recommended of the user based on similarity calculation as a result, from multiple candidate targets.

After the present embodiment obtains the similarity of each candidate access object and prediction result, i.e., each candidate term vector and encode to Similarity between amount, due to the similarity size can characterize user to corresponding candidate access subject interests probability it is big Small, similarity is bigger, and user may more access corresponding candidate subject interests, the corresponding access object of selection access.Therefore, The present embodiment can be directly selected to several highest candidate access objects of similarity to be reached as recommended or similarity To preset threshold candidate access object as recommended etc., the present embodiment to based on similarity calculation as a result, being used The implementation of the recommended at family is not construed as limiting.

For example, the article sequence that user U1 reads article in certain application platform is [x_u1,x_u2,x_u3,x_u4,x_u5, x_u6], the candidate article that the candidate collection of the application platform includes has 10, such as [X1, X2, X3, X4, X5, X6, X7, X8, X9, X10], the article sequence for reading article to user in the manner described above carries out coding calculating, obtains coding vector E_u1, calculate E_u1 Candidate term vector E corresponding with each candidate article_jBetween Cosine similarity, which candidate article j indicate, determines The corresponding candidate article of highest three term vectors of similarity (or similarity reach preset threshold term vector) is X3, X5 and X7, the present embodiment can be using these three candidate articles as article is recommended, and thus constitute that this recalls that logic obtains recalls knot Fruit, so that screening logic subsequent in recommender system therefrom finally recommends article to the target of user's U1 client push.

In conclusion the present embodiment constitutes the training data of model using the access object of user, based on circulation nerve net The recommended prediction model that network training obtains is realized that user accesses the coding calculating of the access object sequence of object, is obtained Coding vector has taken into account the long history interest and short-term history interest of user, and then improves the target finally pushed to user The accuracy of recommended, and due to consideration that user accesses the access order of object, it is capable of the emerging of more acurrate positioning user Interest transition and interest accumulation, the object recommendation method for solving existing Item CF cause to recall result diversity and personalized damage The problem of mistake.

After the present inventor proposes the object recommendation method of above-described embodiment description, though inventor has found this method It so solves the problems, such as object recommendation method of the prior art based on Item CF, but during carrying out Similarity measures, waits Selected works close the candidate access number of objects magnitude for including often all very greatly, especially as current various instant messaging application journeys The access number of objects of sequence, application platform output is ten million rank, and with the increase and access of application platform userbase The extension of object origin, candidate collection only can be increasing, in the future to hundred million ranks even 1,000,000,000 ranks, more or even to 10,000,000,000 Rank, if to carry out similarity calculation, the basic nothing of inline system for the candidate term vector of each candidate access object Method accomplishes to calculate in real time.As it can be seen that candidate collection enormous amount is as its implementation in the object recommendation method that above-described embodiment proposes Important problem.

For this problem, inventor proposes the access object that can be exported to application platform, according to certain sampling plan Several candidate access objects are slightly filtered out, candidate collection is constituted, the present embodiment is not construed as limiting the sampling policy, still, no matter According to the sampling policy of what content, the object recommendation method of above-mentioned alternative embodiment description can all bring the damage of precision of prediction It loses.

Therefore, in order to further increase precision of prediction, the present inventor proposes the improvement to above scheme, is to access object User read article for analyze, it is contemplated that every article be all be made of keyword, although application platform output Article quantity can be more and more, but the total quantity for constituting the keyword of article is a metastable set, usually not Big variation can occur because of the increase of userbase and the extension in article source.So the present embodiment can be by above-mentioned reality Apply example and coding calculating carried out to the access object sequence (can be article sequence herein) of user, be changed to keyword sequence into Row coding calculates, the method that specific implementation process is referred to Examples below description, but is not limited to Examples below Implementation.

It referring to figure 5 and figure 6, is the flow diagram of another object recommendation method provided in an embodiment of the present invention, the party Method still can be applied to server, can specifically include following steps:

Step S301 obtains a plurality of history access data of user；

Step S302 accesses data based on each history, obtains accordingly accessing the corresponding key cluster of object；

In conjunction with the above-mentioned analysis to history access data, the present embodiment can access data based on the history of acquisition, learn Which object user to the access situation of application platform output object, i.e., once had accessed, i.e., each access at past each moment The access order of object.After obtaining title or the content of access object of the user in application platform, can therefrom it extract corresponding Keyword, generate corresponding key cluster, the present embodiment to how from the mode of the corresponding keyword of each access object extraction not It limits, but in the present embodiment, the extracting rule used when extracting the keyword that each access object includes is consistent.

For example: assuming that the object sequence of user U2 is [x1, x2, x3], i.e., in effective time, the user U2 of acquisition There are three access objects in application platform, the object sequence can be constituted by corresponding object ID, from each object ID The keyword extracted in corresponding access object can be denoted as Tagnn, and n is integer, in practical applications, if object is article, At least one keyword for including can be extracted from the title of article, naturally it is also possible to extract at least one from article content Keyword etc., and the extraction of keyword can be realized using natural language processing technique, implement this implementation of process Example is not described further.

Access pair as shown in table 1 below can be generated after obtaining the corresponding key cluster of each access object in the present embodiment As-keyword mapping relations, so as to subsequent direct inquiry.As it can be seen that above-mentioned steps S302 can use each history access data, obtain To the object identity of each access object, later, direct queried access object-keyword mapping relations obtain corresponding access object The keyword for including obtains the corresponding key cluster of each access object.It is to be appreciated that access object-keyword mapping relations Representation is not limited to form shown in the following table 1.

Table 1 accesses object-keyword mapping relations

Object identity	Keyword
		x1	Tag11, Tag12
x2	Tag21, Tag22, Tag23
		x3	Tag31, Tag32

Step S303 constitutes keyword sequence by the corresponding key cluster of each access object；

Optionally, in order to preferably characterize the boundary information for accessing object, the present embodiment can be in each access object pair In the key cluster answered, virtual keyword head and tail portion are added, is denoted as head-tag and tail-tag respectively, wherein After each content for accessing addition head-tag and tail-tag in the corresponding key cluster of object can be identical, and its content determines It can immobilize, the present embodiment is not construed as limiting the head-tag and the tail-tag content respectively indicated.

It is still to be visited shown in [x1, x2, x3] and table 1 with the access object sequence of above-mentioned user U2 referring to Fig. 7 based on this Ask and be illustrated for object-keyword mapping relations, the key cluster of corresponding each access object be respectively [Tag11, Tag12]；[Tag21,Tag22,Tag23]；[Tag31, Tag32], each key cluster can be understood as an article Tag sequence, i.e., each key cluster may map at least one access object, access object 1 as shown in Figure 7, access pair As 2 etc..The present embodiment can be according to the access time of the corresponding access object of each key cluster, by the key cluster of each access object As sequential element, generate keyword sequence, i.e., [head-tag, Tag11, Tag12, tail-tag, head-tag, Tag21, Tag22,Tag23,tail-tag,head-tag,Tag31,Tag32,tail-tag]。

It can be seen that accessing data, the obtained sequence for input model in the history of application platform for user U2 By [x1, x2, x3], become [head-tagTag11, Tag12, tail-tag, head-tagTag21, Tag22, Tag23, tail-tag,head-tag,Tag31,Tag32,tail-tag]。

Keyword sequence input recommended prediction model is carried out coding calculating, obtains user access by step S304 The coding vector of object；

The corresponding keyword sequence of the available multiple sample of users of the present embodiment, obtains training number used in model training According to obtain recommendation pair used in step S304 to these training datas progress model training based on Recognition with Recurrent Neural Network As prediction model, model training process is referred to the description of above-described embodiment corresponding portion.

Wherein, by the coding process of keyword sequence input recommended prediction model, with above-mentioned to object sequence Coding process it is similar, difference be to input the access object by above-described embodiment, become access object keyword, The present embodiment is not detailed cataloged procedure.Therefore, the present embodiment, which obtains coding vector, can be used to predict that user is interested The keyword for accessing object, can not directly predict the interested access object of user.

Step S305, candidate's term vector corresponding with each candidate keywords to the coding vector carry out similarity calculation；

Unlike the step S206 of above-mentioned alternative embodiment, the present embodiment and coding vector carry out similarity calculation It is the corresponding candidate term vector of candidate keywords, rather than the corresponding candidate term vector of candidate access object, at the same time, this reality Applying the element in the candidate collection in example no longer is candidate access object, and becomes candidate keywords.

Wherein, the present embodiment generates candidate collection, the candidate keywords of selection can be in application platform access times compared with High keyword, keyword relevant to social networks hot spot theme etc., the present embodiment is to the side for how selecting candidate keywords Formula is without limitation.And the similarity calculating method class about the similarity calculating method between vector, with above-described embodiment description Seemingly, such as cosine similarity calculation method, however, it is not limited to this, and the present embodiment is no longer superfluous to the specific implementation process of step S305 It states.

Step S306 obtains the corresponding candidate keywords of the candidate term vector of highest first quantity of similarity；

The present embodiment is not construed as limiting the specific value of the first quantity, can rule of thumb or test setting.Moreover, It screens candidate keywords, the Measurement of Similarity of screening can also be preset, i.e., preset similarity threshold, then, step S306 can become obtaining the corresponding candidate keywords of candidate term vector that similarity reaches default similarity threshold, may be used also certainly To realize the screening of candidate keywords using other modes, it is not limited to mode given herein.

Step S307 constitutes candidate keywords cluster by the first quantity candidate keywords obtained；

It is to be appreciated that step S307 be object recommendation method provided in this embodiment is described for convenience, so, this In embodiment practical application, subsequent step directly can be carried out using the candidate keywords filtered out, be not necessarily to execute step Rapid S307, it is implicit to perform step S307 in other words during executing step S306.

Step S308, the candidate keywords for including using candidate keywords cluster, in recommended prediction model upper one The output of moment hidden layer is inputted as current time hidden layer, is continued coding and is calculated；

Above-mentioned similarity calculation process can be thought decoding process by the present embodiment, and the present embodiment can be using repeatedly decoding Mode, to obtain multiple candidate key word sequences, for obtaining the recommended of user.

Optionally, the present embodiment can use seq2seq technology and realize repeatedly decoding.The seq2seq is actually one The Recognition with Recurrent Neural Network of Encoder-Decoder (i.e. coder-decoder) structure, input is a sequence, and output is also One sequence.Wherein, the signal sequence of a variable-length can be become the vector table of regular length by Encoder layers of processing It reaches, i.e., by the input coding of input at a vector, Decoder layers of processing can become the vector of this regular length can The signal sequence of the target of elongated degree, that is, combine coding generate the possible output object of vector forecasting, in the present embodiment for A sequence is write by the coding vector prediction of user, specific implementation process the present embodiment is not detailed.

Based on this, after the present embodiment obtains candidate keywords cluster, one of candidate keywords can be arbitrarily selected, or A candidate keywords, or the selection highest candidate keywords of similarity etc. are selected in sequence, are continued as mode input Coding calculating is carried out, i.e., using the candidate keywords selected as current input, utilizes a period of time upper in recommended prediction model Hidden layer output is carved, continues coding and calculates, obtain new coding vector, with above-described embodiment description by each sequential element After sequentially inputting recommended prediction model, the calculation of hidden layer is similar, and this will not be detailed here for the present embodiment.

Step S309 accesses the coding vector of object, return step S305 using obtained new coding vector as user；

Step S310, detection similarity calculation number reach the second quantity, utilize the second quantity candidate key of composition With dimension candidate keywords in word cluster, the first quantity candidate key word sequence is generated；

The present embodiment is not construed as limiting multiple coding number, i.e., is not construed as limiting to the numerical value of the second quantity, can be according to warp The numerical value of the second quantity of setting is tested or tested, can also be that similarity size is true according to the candidate keywords that this coding obtains It is fixed etc.；And need to continue whether coding can also meet preset condition by judging the candidate keywords cluster of composition, such as Whether the similarity of the candidate keywords in candidate keywords cluster reaches certain threshold value, or the phase of the candidate keywords obtained every time It is realized like the degree conditions such as tend towards stability, how the present embodiment is to determine whether to continue the condition of coding without limitation, not It is confined to the mode of the present embodiment description.

Such as above-mentioned analysis, in the present embodiment practical application, it is assumed that decoding calculates every time, i.e., every time by similarity meter It calculates, filters out the highest K candidate keywords of similarity, is i.e. the first quantity is K, realizes T decoding according to above description mode Afterwards, i.e. the second quantity be T, will obtain T length for K candidate keywords cluster, later, can from these key clusters, according to The secondary candidate keywords extracted with dimension, generate corresponding candidate key word sequence, the K candidate key that will such as obtain every time A column of the word sequence as matrix, so obtain T column data, take the candidate keywords of same a line to constitute one and predict obtained time Keyword sequence is selected, [Tagm1, Tagm2, Tagm3 ..., TagmT], m=1,2,3 ..., K are denoted as, it is seen then that the present embodiment can be with Obtain the candidate key word sequence that K length is T.

It is waited it is to be appreciated that the candidate key word sequence generating mode of the present embodiment is not limited to above-described same latitude Keyword extraction mode is selected, a candidate keywords can also be arbitrarily extracted from each candidate keywords cluster and are generated, in this way A large amount of candidate key word sequences will be obtained, the processing mode of other subsequent steps is identical, and the present embodiment no longer individually describes.

Step S311 obtains the recommendation of user based on the candidate keywords that the first quantity candidate key word sequence includes Object.

Optionally, the present embodiment can pre-establish the keyword of each object that application platform can export to access object Inverted index, specific construction method the present embodiment is not described further.Wherein, inverted index is properly termed as reverse indexing, merging shelves Case or reversed archives can be used to be stored under full-text search, storage of some word in a document or one group of document The mapping of position, the present embodiment storage is then the mapping of keyword to access object.

Based on this, multiple subsequent key word sequences are being obtained, i.e. [Tagm1, Tagm2, Tagm3 ..., TagmT], m=1, 2,3 ..., K can successively pull the time comprising each candidate keywords in the access object that application platform can export Choosing access object, that is, obtain at least one candidate access object that each candidate keywords are mapped to, to generate each candidate key The corresponding Inverted List of word such as pulls the candidate access object x2 comprising Tagm1, at this point, candidate access object x2 may be also It can include other candidate keywords, that is to say, that candidate keywords may be mapped to the same candidate access object, later, The number being mapped by calculating each candidate access object, according to the number size, from the candidate access object being mapped to, sieve The recommended of user is selected, specific implementation process is without limitation.

Certainly, if candidate key word sequence remains word order, the present embodiment can also be pressed these candidate key word sequences Spliced according to word order, later, can use text similarity calculation, calculates spliced candidate key word sequence and each Similarity between the corresponding keyword sequence of candidate access object, and then select the highest preset quantity of similarity is candidate to visit Ask recommended of the object as user.

It, can be with to obtain the recommended of user as it can be seen that the implementation of step S311 is not limited to a certain mode It is realized using any mode given above, and is not limited to above-described two kinds of implementations.

In conclusion the present embodiment accesses the keyword in object using user, the training data of model is constituted, based on following The recommended prediction model that ring neural metwork training obtains realizes the coding meter for accessing user the keyword sequence of object The characteristics of calculating, obtaining the coding vector of the prediction interested keyword of user, be based on Recognition with Recurrent Neural Network, this coding mode is simultaneous The long history interest and short-term history interest of user have been cared for, and then has improved the target recommended finally pushed to user Accuracy, and due to consideration that user accesses the access order of object is capable of the interest transition of more acurrate positioning user and emerging Interest accumulation solves the problems, such as that the object recommendation method of existing Item CF causes to recall result diversity and personalized loss.

Moreover, because the present embodiment is in the term vector that the candidate term vector of similarity calculation is keyword, relative to access The term vector of object, it is quantitatively many less, and with the variation of time, calculative candidate's term vector there will not be very Big variation, reduces calculation amount, which raises the accuracy of object recommendation and stability.

Optionally, following two mode is provided for the implementation method of the step S311 of above-mentioned alternative embodiment, but simultaneously It is not limited to two kinds of implementations described below:

Mode one:

It may comprise steps of referring to Fig. 6 and flow diagram shown in Fig. 8, this method:

Step A1 obtains the inverted index that the keyword constructed is mapped to object；

The present embodiment does not limit the building mode for the inverted index that the keyword in current application platform is mapped to object It is fixed.

Step A2 inquires the inverted index, obtains the row of falling of each of each candidate key word sequence candidate keywords List；

Wherein, the Inverted List is for characterizing at least one Candidate Recommendation object that the candidate keywords are mapped to.As above The mode of text description can retract at least from the access object that application platform can export for each candidate keywords Comprising the candidate access object to candidate keywords, or perhaps obtain the candidate access pair that each candidate keywords are respectively mapped to As can quickly be obtained by the Inverted List crucial comprising corresponding candidate to generate the Inverted List of each candidate keywords Which the candidate access object of word has.

Step A3 counts each time in each candidate key word sequence based on the Inverted List of obtained each candidate keywords Select the number that recommended is mapped；

In conjunction with above-mentioned analysis it is found that for the candidate access object that the present embodiment retracts, it includes candidate keywords (i.e. Candidate keywords in each candidate key word sequence obtained above) quantity is more, i.e. time that is pulled of the candidate access object Number is more, illustrates that the candidate access object may more be recommended to user, the probability for becoming recommended is bigger.So this The number that embodiment can be pulled based on each candidate access object determines that the candidate access object is corresponding according to certain rule Score value, later, according to the score value size of each candidate access object, Lai Shixian subsequent step.In the present embodiment, candidate target Score value it is bigger, illustrate that user is bigger to the probability of the candidate access subject interests, gets over as the probability of recommended Greatly.

Optionally, in practical applications, the present embodiment can also count each candidate access object to each candidate key word order The keyword coverage rate of column, the keyword coverage rate is bigger, illustrates that corresponding candidate access object is got over as the probability of recommended Greatly, the present embodiment can directly carry out subsequent processing on this basis, can also with it is reported that obtain the score value of each subsequent access object, Indicate that corresponding candidate access object becomes the probability etc. of recommended by score value size, the present embodiment covers the keyword The circular of rate is not construed as limiting.

Step A4 screens the recommendation pair of the user from multiple Candidate Recommendation objects of acquisition based on statistical result As.

In the present embodiment, due to be mapped the more Candidate Recommendation object of number be screened for recommended probability more Greatly, that is to say, that the corresponding candidate keywords quantity of candidate access object is more, illustrates that it is recommended to the probability of user and gets over Greatly, therefore, the present embodiment can screen recommendation pair of the p most candidate target of corresponding candidate keywords quantity as user As.

Specifically, the number that the present embodiment can be mapped each candidate access object is ranked up, it is secondary according to being mapped The sequence of number from big to small, selects p candidate access object as recommended, but be not limited to this implementation.

Mode two:

It may comprise steps of referring to Fig. 6 and flow diagram shown in Fig. 9, this method:

Step B1, according to the word order of each candidate keywords in each candidate key word sequence, to the first quantity candidate key The candidate keywords that word sequence includes are spliced；

Step B2 obtains the corresponding keyword sequence of multiple candidate access objects, and the candidate access object is at least Include the candidate keywords in any candidate key word sequence；

It is to be appreciated that step B2 can be executed before step B1 in the present embodiment practical application, it is not limited to this This sequence of steps of embodiment.

Optionally, when the keyword sequence of candidate access object is excessive, in order to reduce similarity calculation workload, this reality Candidate access object can be filtered using Inverted List mode by applying example, such as filtered out and contained at least one candidate keywords The candidate access object of candidate keywords in sequence, for completing subsequent similarity calculation.

Step B3 carries out the candidate key word sequence keyword sequence corresponding with each candidate access object that splicing obtains Text similarity computing；

In the present embodiment practical application, a text may be considered for the keyword sequence obtained after splicing, because This, the present embodiment can use Text similarity computing mode, and it is corresponding with each candidate access object to calculate candidate key word sequence Keyword sequence between similarity, to filter out recommended.The present embodiment is specific to Text similarity computing method Realization process is not described further.

Step B4 screens the recommendation pair of user based on Text similarity computing as a result, from multiple candidate access objects As.

Wherein, the higher candidate target of the similarity be screened for the probability of recommended it is bigger, therefore, the present embodiment Can be from candidate access object, the screening highest preset quantity candidate access object of similarity is recommended, can also be with The candidate access object that screening similarity reaches preset threshold is recommended etc., specific implementation of the present embodiment to step B4 Process is not construed as limiting.

Recommended based on the user that the various embodiments described above obtain can push away in practical applications as target Recommend the candidate item of object, that is to say, that after obtaining the recommended of user, can also do to it further using some logics Screening, and the effect of the step of object recommendation method of the various embodiments described above description in practical applications may is that according to specific The portrait information of user carries out data pull according to dimensions such as various accurate personalized, general personalizations, temperatures, that is, pulls user Possible interested recommended, the quantity of the recommended usually obtained at this time is relatively more, can further screen, specific real Existing method is referred to flow diagram shown in Fig. 10, but is not limited to method shown in Fig. 10.

It as shown in Figure 10, is the flow diagram of another object recommendation method provided in an embodiment of the present invention, in this method The recommended for obtaining user realizes process, is referred to the description of the various embodiments described above, and details are not described herein for the present embodiment, This is only described the treatment process obtained after the recommended of user, and therefore, this method may also comprise the following steps::

Step S401 carries out preliminary screening to the recommended of user, obtains primary election recommended according to ad hoc rules；

Wherein, ad hoc rules can be the factor that user accesses object dependencies, timeliness, region, diversity etc. It determines, the particular content that the present embodiment includes to the ad hoc rules is without limitation.

The schematic diagram of the recommender system of application scenarios as shown in figure 11, the application scenarios specifically can be certain instant messaging Information display platform is to the recommendation scene of its output information in client, and as shown in figure 11, which may include recalling Multiple functional modes such as logic, primary election logic and sequence logic (i.e. Rank in Figure 11), wherein recall logic and be accomplished that The acquisition process of the recommended of the user of each embodiment description is stated, primary election logic is used to realize the realization process of step S401, And sequence logic is used to realize the subsequent ordering process to primary election recommended.

Step S402 obtains the coding vector of the corresponding coding vector of each primary election recommended and current accessed object；

It is similar between primary election recommended and current browsing object in addition to directly calculating in the present embodiment practical application Degree, in such a way that similarity size is ranked up each primary election recommended outside, the present embodiment can also use above description Similarity calculation mode to get to the corresponding coding vector of access object, thus by the similarity between vector, realize pair The sequence of each primary election recommended.

Based on this, the present embodiment can according to manner described above, obtain each primary election recommended it is corresponding encode to Amount generates access object sequence such as using each primary election recommended as sequential element, is sequentially input recommended prediction mould Type carries out coding calculating, obtains corresponding coding vector, but is not limited to a kind of this coding mode, can also use above-mentioned pass The coding mode of keyword sequence, specific implementation process are referred to the generation step of the coding vector of each embodiment description above.

Step S403 carries out phase to the corresponding coding vector of each primary election recommended and the coding vector of current accessed object It is calculated like degree；

Optionally, the present embodiment can use cosine similarity calculation method, obtain the similarity between vector, specific real Existing process is referred to the description of foregoing embodiments corresponding portion, but the similarity calculating method between vector is not limited to this Kind implementation method.

Step S404 selects the highest preset quantity primary election recommended of similarity as target recommended；

It is to be appreciated that the present embodiment from multiple primary election recommendeds, screens the implementation method of target recommended not It is confined to the mode of the present embodiment description, with above-described embodiment description from multiple candidate access objects, screens the recommendation of user The Method type of object, the present embodiment do not enumerate herein.

The client that the target recommended is sent to user is shown by step S405.

In conclusion the present embodiment is in such a way that above-described embodiment describes, the recommended of obtained user can not only The interest transition and interest accumulation of user are enough accurately positioned out, additionally it is possible to the Long-term Interest and short-term interest for taking into account user, from this The target recommended of the user filtered out in the recommended of sample can be more in line with the current demand of user, so that using this The application server of this object recommendation method of embodiment, preferably can provide recommendation service for user.

For applying recommender system shown in above-mentioned Figure 11, field is carried out to object recommendation method provided by the above embodiment Illustrating under scape, the article that the present embodiment is only exported using accessing object as application platform, and based on object sequence as instruction Practice data instance to be illustrated, for the server of the application, the text of multiple users in the available application platform Zhang Xulie, and training pattern is thus obtained, model training is carried out based on Recognition with Recurrent Neural Network, obtains recommended prediction model, Later, server can be denoted as target user, can obtain in the manner described above for any user using the application The article sequence of target user is sequentially input the recommended prediction model, is obtained for predicting that user may be interested Article coding vector, and the term vector of itself and preset multiple candidate articles is subjected to similarity calculation, to select phase Like spending the recommendation article of highest several candidate articles as target user, i.e. what logic was recalled in Figure 11 execution recalls result. Certainly, during model training, it also can be used the corresponding keyword sequence of multiple users, composing training data, in this way, Predict target user recommendation article when, acquisition be then target user keyword sequence, be inputted recommended prediction After model, the coding vector of output can predict user may interested keyword, later, calculate separately the coding vector with Similarity (may be considered a decoding process) between the term vector of each candidate keywords selects K similarity highest Candidate keywords generate the keyword sequence that multiple length are T, later, according to two kinds of texts given above after repeatedly decoding The processing mode for recalling logic obtains the recommendation article of target user such as inverted index or Text similarity computing mode, i.e., should The article of target user recalls result.

Later, it can use primary election logic to screen the multiple recommendation articles recalled, obtain multiple primary election articles, then Multiple primary election articles are ranked up using Rank logic, also can use the above-described side for obtaining and recommending article at this time Formula is ranked up primary election article, obtains target and recommends article, and is shown and use client display interface in target user On.

In practical applications, above-mentioned treatment process can be realized by server is lower online, when certain user is connected using client When logical server, the associated target of the user identifier directly can be recommended article according to the user identifier of the user by server Client displaying is fed back to, but is not limited to this implementation.

Referring to Fig.1 2, it is a kind of structural schematic diagram of object recommendation device provided in this embodiment, which can apply In server, the apparatus may include but be not limited to consisting of structure:

Retrieval module 11, for obtaining user's access sequence；

Wherein, which is to access the object that application platform exports based on user to generate；

Computation model 12 is encoded, by carrying out user's access sequence input recommended prediction model based on coding It calculates, obtains the coding vector that the user accesses object；

Wherein, which is to be visited based on Recognition with Recurrent Neural Network the corresponding user of multiple sample of users Ask that sequence training obtains, and the Recognition with Recurrent Neural Network includes multiple gating cycle elementary layers or multiple shot and long term memory networks Layer, specific network structure and principle are referred to the description of above method embodiment corresponding portion.

First similarity calculation module 13, for carrying out similarity calculation to the coding vector and each candidate term vector；

Recommended selecting module 14, for based on similarity calculation as a result, obtaining the recommended of the user.

Optionally, referring to Fig.1 shown in 3, training data content needed for training recommended prediction model is different, acquisition The sequential element content of user's access sequence will be different, and then the mode for obtaining user's recommended also can accordingly change.

Based on this, above-mentioned retrieval module 11 may include:

First data capture unit 1110, a plurality of history for obtaining user access data, and the history accesses data It is to be generated based on access operation of the user to application platform output object；

First ray Component units 1111, the object identity for being separately included by a plurality of history access data, structure At access object sequence.

In this case, as shown in figure 13, device provided in this embodiment can also include:

Candidate access object acquisition module 15, for obtaining multiple candidate access objects；

First term vector obtains module 16, for obtaining corresponding candidate for each candidate access object input language model Term vector

Correspondingly, above-mentioned first similarity calculation module 13 specifically can be used for the coding vector and candidate access pair As corresponding candidate term vector carries out similarity calculation.

Wherein, retouching for embodiment of the method corresponding portion above is referred to about the similarity calculating method between vector It states.

Recommended selecting module 14 specifically can be used for selecting the highest preset quantity candidate access object of similarity, Recommended as user；Or similarity is selected to reach recommended of the candidate access object of preset threshold as user Etc., how the present embodiment is to screening the mode of the recommended of user not from multiple candidate access objects using similarity It limits.

As another embodiment of the application, as shown in figure 14, above-mentioned retrieval module 11 may include:

Second data capture unit 1120, a plurality of history for obtaining user access data；

Key cluster acquiring unit 1121 obtains accordingly accessing the corresponding pass of object for accessing data using each history Keyword cluster；

Second Sequence composition unit 1122, for constituting keyword sequence by the corresponding key cluster of each access object.

In the present embodiment, the realization process for obtaining keyword sequence is referred to retouching for embodiment of the method corresponding portion above It states.

At this point, above-mentioned apparatus can also include:

Candidate keywords obtain module, for obtaining multiple candidate keywords；

Second term vector obtains module, for the multiple candidate keywords to be distinguished input language model, obtains corresponding Candidate term vector.

Correspondingly, above-mentioned first similarity calculation module 13 specifically can be used for closing the coding vector and each candidate visit The corresponding candidate term vector of keyword carries out similarity calculation.

In practical applications, due to application platform output access object keyword quantity be it is metastable, not Can rapidly it increase with the user of application platform and the increase of output object origin, this makes the present embodiment similarity calculation Calculation amount it is smaller, and the recommended of thus obtained user is also relatively stable.

Optionally, in the case where candidate collection is candidate keywords, as described in Figure 14, above-mentioned recommended selecting module 14 may include:

Candidate keywords acquiring unit 141, it is corresponding for obtaining the candidate term vector of highest first quantity of similarity Candidate keywords；

Key cluster generation unit 142 constitutes candidate keywords for the first quantity candidate keywords by obtaining Cluster；

Encode computing unit 143, a candidate keywords and the recommendation for including using the candidate keywords cluster The output of last moment hidden layer continues coding calculating in object prediction model；

Similarity calculation terminates unit 144, for accessing the coding of object using obtained new coding vector as user Vector, execution is described to carry out Similarity measures step to the coding vector and each candidate term vector, until similarity calculation is secondary Number reaches the second quantity, or the candidate keywords cluster constituted meets preset condition；

Keyword sequence generation unit 145, for utilizing same dimension in the second quantity candidate keywords cluster constituted Candidate keywords generate the first quantity candidate key word sequence；

Recommended acquiring unit 146, the candidate pass for including based on the first quantity candidate key word sequence Keyword obtains the recommended of the user.

Optionally, above-mentioned recommended acquiring unit 146 can specifically include:

Inverted index obtains subelement, the inverted index for being mapped to object for obtaining the keyword constructed；

Subelement is inquired, for inquiring the inverted index, obtains the candidate pass of each of each candidate key word sequence The Inverted List of keyword；

Wherein, the Inverted List is for characterizing at least one Candidate Recommendation object that the candidate keywords are mapped to.

It counts subelement and counts each candidate key word sequence for the Inverted List based on obtained each candidate keywords In the number that is mapped of each Candidate Recommendation object；

First screening subelement, for screening the use from multiple Candidate Recommendation objects of acquisition based on statistical result The recommended at family.

Wherein, be mapped the more Candidate Recommendation object of number be screened for the probability of recommended it is bigger.

As another alternative embodiment, above-mentioned recommended acquiring unit 146 also may include:

Splice subelement, for the word order according to each candidate keywords in each candidate key word sequence, to first number The candidate keywords that a candidate key word sequence includes are measured to be spliced；

Keyword sequence obtains subelement, for obtaining the corresponding keyword sequence of multiple candidate access objects, institute Candidate access object is stated including at least the candidate keywords in any candidate key word sequence；

Similarity calculation subelement, the candidate key word sequence for obtaining to splicing are corresponding with each candidate access object Keyword sequence carries out Text similarity computing；

Second screening subelement, for being based on Text similarity computing as a result, from the multiple candidate access object, sieve Select the recommended of the user.

Wherein, the higher candidate access object of the similarity be screened for the probability of recommended it is bigger.

It is to be appreciated that realizing process about each functional module in above-mentioned apparatus embodiment or the function of unit, it is referred to The description of above method embodiment corresponding portion.

Optionally, on the basis of the various embodiments described above, as shown in figure 15, above-mentioned apparatus can also include:

Preliminary screening module 17, for carrying out preliminary screening to the recommended of the user, obtaining according to ad hoc rules Primary election recommended；

Coding vector obtains module 18, for obtaining the corresponding coding vector of each primary election recommended and current accessed pair The coding vector of elephant；

Second similarity calculation module 19, for the corresponding coding vector of each primary election recommended and the current accessed The coding vector of object carries out similarity calculation；

Target recommended selecting module 120, for selecting the highest preset quantity primary election recommended of similarity to make For target recommended；

Target recommended sending module 121, for the target recommended to be sent to the client of the user It is shown.

In summary, the present embodiment be based on Recognition with Recurrent Neural Network, to the corresponding user's access sequence of multiple sample of users into Row training, obtains recommended prediction model, and realization carries out coding calculating to user's access sequence, not only takes into account the length of user Phase historical interest and short-term history interest, and the access order that user accesses object in application platform is considered, in this way, base In obtained coding vector and each candidate term vector similarity calculation as a result, obtaining the recommended of the user can accurately determine Position user interest transition and interest accumulation, solve existing Item CF recommended method cause gained recommended diversity and The problem of personalization loss.

Wherein, according to different needs, the content of training data used in model training can be carried out with flexible choice, that is, is used The content of family access sequence improves the flexibility of object recommendation method.

The embodiment of the present invention also provides a kind of computer equipment, and the hardware configuration of the computer equipment can be such as Figure 16, should The hardware configuration of computer equipment may include: communication interface 1, memory 2 and processor 3；

In embodiments of the present invention, communication interface 1, memory 2, processor 3 can be realized mutual by communication bus Communication, and the communication interface 1, memory 2, processor 3 and communication bus quantity can be at least one.

Optionally, communication interface 1 can be the interface of communication module, such as the interface of gsm module；

Processor 3 may be a central processor CPU or specific integrated circuit ASIC (Application Specific Integrated Circuit), or be arranged to implement the integrated electricity of one or more of the embodiment of the present invention Road.

Memory 2 may include high speed RAM memory, it is also possible to further include nonvolatile memory (non-volatile Memory), a for example, at least magnetic disk storage.

Wherein, memory 2 is stored with computer program, the computer program that processor 3 calls memory 2 to be stored, with Realize each step of the above-mentioned object recommendation method applied to computer equipment；

Optionally, which is primarily useful for:

In the present embodiment practical application, above-mentioned computer equipment can be application server, such as various instant messaging visitors Corresponding application server in family end etc..

The embodiment of the present invention also provides a kind of storage medium, the storage medium recorded processor having suitable for computer equipment The computer program of execution, to realize each step of the above-mentioned object recommendation method applied to computer equipment, the object recommendation The realization process of method is referred to the description of above method embodiment corresponding portion.

Each embodiment in this specification is described in a progressive manner, the highlights of each of the examples are with other The difference of embodiment, the same or similar parts in each embodiment may refer to each other.For device disclosed in embodiment, For computer equipment, since it is corresponded to the methods disclosed in the examples, so be described relatively simple, related place referring to Method part illustration.

Professional further appreciates that, unit described in conjunction with the examples disclosed in the embodiments of the present disclosure And algorithm steps, can be realized with electronic hardware, computer software, or a combination of the two, in order to clearly demonstrate hardware and The interchangeability of software generally describes each exemplary composition and step according to function in the above description.These Function is implemented in hardware or software actually, the specific application and design constraint depending on technical solution.Profession Technical staff can use different methods to achieve the described function each specific application, but this realization is not answered Think beyond the scope of this invention.

The step of method described in conjunction with the examples disclosed in this document or algorithm, can directly be held with hardware, processor The combination of capable software module or the two is implemented.Software module can be placed in random access memory (RAM), memory, read-only deposit Reservoir (ROM), electrically programmable ROM, electrically erasable ROM, register, hard disk, moveable magnetic disc, CD-ROM or technology In any other form of storage medium well known in field.

The foregoing description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, as defined herein General Principle can be realized in other embodiments in the case where not departing from core of the invention thought or scope.Therefore, originally Invention is not intended to be limited to the embodiments shown herein, and is to fit to and the principles and novel features disclosed herein Consistent widest scope.

Claims

1. a kind of object recommendation method, which is characterized in that the described method includes:

User's access sequence is obtained, user's access sequence is to access the object that application platform exports based on user to generate；

User's access sequence input recommended prediction model is subjected to coding calculating, the user is obtained and accesses object Coding vector, the recommended prediction model are to be accessed based on Recognition with Recurrent Neural Network the corresponding user of multiple sample of users Sequence training obtains；

Similarity calculation is carried out to the coding vector and each candidate term vector, candidate's term vector corresponds in candidate collection Candidate target or candidate keywords；

2. the method according to claim 1, wherein the Recognition with Recurrent Neural Network includes multiple gating cycle lists First layer or multiple shot and long term memory network layers.

3. method according to claim 1 or 2, which is characterized in that user's access sequence is access object sequence, institute Stating acquisition user's access sequence includes:

The a plurality of history for obtaining user accesses data, and the history access data are to export object to application platform based on user What access operation generated；

The object identity separately included by a plurality of history access data, constitutes access object sequence.

4. according to the method described in claim 3, it is characterized in that, the method also includes:

Obtain multiple candidate access objects；

By each candidate access object input language model, corresponding candidate term vector is obtained.

5. according to method described in 1 or 2, which is characterized in that user's access sequence is keyword sequence, the acquisition user Access sequence, comprising:

The a plurality of history for obtaining user accesses data；

Data are accessed using each history, obtain accordingly accessing the corresponding key cluster of object；

By the corresponding key cluster of each access object, keyword sequence is constituted.

6. according to the method described in claim 5, it is characterized in that, the method also includes:

Obtain multiple candidate keywords；

The multiple candidate keywords are distinguished into input language model, obtain corresponding candidate term vector.

7. according to the method described in claim 5, it is characterized in that, the similarity calculation that is based on is as a result, obtain the user Recommended, comprising:

Obtain the corresponding candidate keywords of the candidate term vector of highest first quantity of similarity；

By the first quantity candidate keywords obtained, candidate keywords cluster is constituted；

Last moment is hidden in the candidate keywords for including using the candidate keywords cluster, with the recommended prediction model Layer output is hidden as current time hidden layer input in the recommended prediction model, continues coding and calculates；

The coding vector of object is accessed using obtained new coding vector as user, is executed described to the coding vector and each Candidate term vector carries out Similarity measures step, until similarity calculation number reaches the second quantity, or the candidate key constituted Word cluster meets preset condition；

Using, with dimension candidate keywords, generating in the second quantity candidate keywords cluster of composition, the first quantity is candidate to be closed Keyword sequence；

Based on the candidate keywords that the first quantity candidate key word sequence includes, the recommended of the user is obtained.

8. the method according to the description of claim 7 is characterized in that described be based on the first quantity candidate key word sequence The candidate keywords for including obtain the recommended of the user, comprising:

Obtain the inverted index that the keyword constructed is mapped to object；

The inverted index is inquired, the Inverted List of each of each candidate key word sequence candidate keywords is obtained, it is described Inverted List is for characterizing at least one Candidate Recommendation object that the candidate keywords are mapped to；

Based on the Inverted List of obtained each candidate keywords, each Candidate Recommendation object quilt in each candidate key word sequence is counted The number of mapping；

The recommended of the user is screened, wherein reflected from multiple Candidate Recommendation objects of acquisition based on statistical result Penetrate the more Candidate Recommendation object of number be screened for the probability of recommended it is bigger.

9. the method according to the description of claim 7 is characterized in that described be based on the first quantity candidate key word sequence The candidate keywords for including obtain the recommended of the user, comprising:

According to the word order of each candidate keywords in each candidate key word sequence, to the first quantity candidate key word sequence packet The candidate keywords contained are spliced；

The corresponding keyword sequence of multiple candidate access objects is obtained, the candidate access object includes at least any candidate A candidate keywords in keyword sequence；

Text similarity is carried out to the candidate key word sequence keyword sequence corresponding with each candidate access object that splicing obtains It calculates；

Based on Text similarity computing as a result, screening the recommended of the user from the multiple candidate access object, In, the higher candidate access object of the similarity be screened for the probability of recommended it is bigger.

10. method according to claim 1 or 2, which is characterized in that the method also includes:

According to ad hoc rules, preliminary screening is carried out to the recommended of the user, obtains primary election recommended；

Obtain the coding vector of the corresponding coding vector of each primary election recommended and current accessed object；

Similarity calculation is carried out to the coding vector of the corresponding coding vector of each primary election recommended and the current accessed object；

Select the highest preset quantity primary election recommended of similarity as target recommended；

The client that the target recommended is sent to the user is shown.

11. a kind of object recommendation device, which is characterized in that described device includes:

Retrieval module, for obtaining user's access sequence, user's access sequence is to access application platform based on user What the object of output generated；

Computation model is encoded, for user's access sequence input recommended prediction model to be carried out coding calculating, is obtained The user accesses the coding vector of object, and the recommended prediction model is based on Recognition with Recurrent Neural Network, to multiple samples The corresponding user's access sequence training of user obtains；

12. device according to claim 11, which is characterized in that the retrieval module includes:

First data capture unit, a plurality of history for obtaining user access data, and the history access data are based on use Family generates the access operation of application platform output object；

First ray Component units, the object identity for being separately included by a plurality of history access data, constitute access pair As sequence.

13. device according to claim 11, which is characterized in that the retrieval module includes:

Second data capture unit, a plurality of history for obtaining user access data；

Key cluster acquiring unit obtains accordingly accessing the corresponding key cluster of object for accessing data using each history；

Second Sequence composition unit, for constituting keyword sequence by the corresponding key cluster of each access object.

14. a kind of storage medium, is stored thereon with computer program, which is characterized in that the computer program is held by processor Row realizes each step of the object method as described in claim 1-10 any one.

15. a kind of computer equipment, which is characterized in that the computer equipment includes:

Communication interface；

Memory, for storing the computer program for realizing the object method as described in claim 1-10 any one；

Processor, for recording and executing the computer program of memory storage, the computer program for realizing with Lower step: