Content-Aware Anomaly Detection with Network Representation Learning

Given a network with node descriptions or labels, how to identify anomalous nodes and anomalous links in it? Existing methods (e.g., non-negative matrix factorization) mostly focuses on structural anomalies, without taking node descriptions or labels seriously into account. However, such information is obviously valuable for detecting anomalous nodes and links. On the other hand, network representation learning aims to represent the nodes and links in a network as low-dimensional, real-valued and dense vectors, so that the resulting vectors have representation and reasoning ability. It is straightforward that the reconstruction errors between normal node representations are small, while anomaly ones are not. Therefore, we propose a novel Content-Aware Anomaly Detection (CAAD) method based on network representation learning and encoder-decoder. The CAAD method learns structural and content representations with convolutional neural networks. By using the learned low-dimensional node representations, an encoder-decoder model is trained to perform anomaly detection in terms of reconstruction errors. Experiments on two synthetic datasets and one real-world dataset demonstrate that CAAD consistently outperforms the existing baseline methods. For more information and source codes of this study please visit https://github.com/lizhong2613/CAAD.