Unbiased Recommendation Through Invariant Representation Learning

Mining user behavior data to align the products with the users’ interests has been a widespread practice in recommender systems. However, the user behavior data is influenced not only by their interests but also by various bias factors, such as product popularity and ranking position. These influences can potentially mislead the recommender systems, steering them in unpredictable directions. Consequently, numerous strategies have been developed to identify and mitigate these biases. Unfortunately, they often rely on human-crafted bias-modeling models or online intervention data, which can be prohibitively expensive or detrimental to the user experience. To this end, we apply an innovative model named the Causal Invariant Recommendation Model (CIRM), which autonomously identifies bias factors in an unsupervised manner and simultaneously disentangles these factors within the recommendation model. Specifically, CIRM employs a dual-tower system, consisting of a causal tower and a spurious tower, to distinctly model users’ interests and inherent biases. We optimize the model adversarially, where the spurious tower partitions training data based on identified spurious features that degrade overall performance. The causal tower, on the other hand, is focused on developing a bias-resistant representation by regularizing the representation invariant across varying data partitions. Extensive experiments on benchmark datasets have validated the effectiveness of CIRM, showcasing its superior performance compared to existing models.