Addressing Confounding Feature Issue for Causal Recommendation

In recommender systems, some features directly affect whether an interaction would happen, making the happened interactions not necessarily indicate user preference. For instance, short videos are objectively easier to finish even though may like video. We term such feature as confounding , and video length is a in recommendation. If we fit model on data, just done by most data-driven will be biased recommend more, deviate from actual requirement. This work formulates addresses problem causal perspective. Assuming there factors affecting both other item features, e.g., creator, find opens backdoor path behind user-item matching introduces spurious correlation. To remove effect of path, propose framework named Deconfounding Causal Recommendation (DCR) which performs intervened inference with do-calculus . Nevertheless, evaluating requires sum over prediction all possible values feature, significantly increasing time cost. address efficiency challenge, further mixture-of-experts (MoE) architecture, modeling each value separate expert module. Through this way, retain expressiveness few additional costs. demonstrate DCR backbone neural factorization machine (NFM) showing that leads more accurate preference small release our code at: https://github.com/zyang1580/DCR