Disentangling User Interest and Popularity Bias for Recommendation with Causal Embedding.
Y. Zheng, Chen Gao, Xiang Li, Xiangnan He, Yong Li, Depeng Jin · arXiv (Cornell University) · 2020
Recommendation models are usually trained on observational data. However, observational data exhibits various bias, such as popularity bias. Biased data results in a gap between online environment and offline evaluation, which makes it difficult to measure the real effect of recommendation. To bridge this gap, most existing methods exploit causal inference to eliminate the bias in historical observational data, e.g., by re-weighting training samples or leveraging a small fraction of unbiased data. However, different causes of an interaction are bundled together as a unified representation in these algorithms. In this paper, we present DICE, a general framework that learns representations where user interest and item popularity are structurally disentangled. We leverage separate embeddings for different causes, and make each embedding capture only one cause by training with cause-specific samples and imposing direct disentanglement supervision. Our algorithm outperforms state-of-the-art baselines with remarkable improvements on two real-world datasets with various backbone models. We further demonstrate that the learned embeddings successfully capture the desired causes.