Deconfounding Representation Learning Based on User Interactions in Recommendation Systems

Representation learning provides an attractive solution to capture users' real intents by modeling user interactions in recommendation systems. However, there exist influencing factors called confounders in the process of user interactions. Most traditional methods might ignore these confounders, resulting in learning inaccurate users' intents. To address the issue, we take a new perspective to develop a deconfounding representation learning model named DRL. Concretely, we infer the unobserved confounders existing in the user-item interactions with an inference network. Then we leverage a generative network to generate users' personalized intents that contain no unobserved confounders. In order to learn comprehensive users' intents, we model the user-user interactions by adopting state-of-the-art GNN with a new aggregating strategy. Thus, the users' real intents we learn not only have their own personalized information but also imply the influence of their friends. The results of two real-world experiments demonstrate that our model can learn accurate and comprehensive representations.