CIPhy: Causal Intervention with Physical Confounder from IoT Sensor Data for Robust Occupant Information Inference

Occupant information inference with IoT sensor data enables many smart applications, such as patients'/older adults' in-home monitoring. The difficulty of collecting labeled real-world IoT sensor data often leads to reliability and scalability issues for those systems. Extensive prior works (e.g., domain adaptation) focus on the domain shift issues, i.e., the inconsistent data feature and label relationship, and dataset bias is often neglected. Dataset bias is commonly caused by limited and varied accessibility to labeled data for each class, and it is inevitable for real-world datasets. The model trained with a biased dataset fits into the bias, hence cannot further generalize to the testing data for accurate inference.

We propose CIPhy, a causal intervention scheme with physical confounders measured from the sensor data to achieve robust occupant information inference. We model the dataset bias as a confounding problem. There exists a confounder directly impacts both data feature and label, and each class's accessibility to labeled data varies when the confounder's condition changes. The model trained with biased data learns a spurious feature-label correlation conditioned on the confounder's condition in the training data. When testing data has a different condition, i.e., confounding shift, this correlation can not be applied. By using the causal intervention, e.g., backdoor adjustment, the confounding shift's negative impact on the data-driven models can be mitigated. The CIPhy decouples the sensor data to measure the confounder, then conducts the causal intervention for a de-biased occupant information inference. We use a public dataset on occupant identification as a case study, to investigate the feasibility of applying causal intervention to resolve the dataset bias issue. From the experiment, CIPhy achieves up to 11.42% identification accuracy improvement compared to baselines given the biased training data and confounding shift.