Addressing Unmeasured Confounder for Recommendation with Sensitivity Analysis

Recommender systems should answer the intervention question "if recommending an item to a user, what would the feedback be", calling for estimating the causal effect of a recommendation on user feedback. Generally, this requires blocking the effect of confounders that simultaneously affect the recommendation and feedback. To mitigate the confounding bias, a strategy is incorporating propensity into model learning. However, existing methods forgo possible unmeasured confounders (e.g., user financial status), which can result in biased propensities and hurt recommendation performance. This work combats the risk of unmeasured confounders in recommender systems.

Towards this end, we propose Robust Deconfounder (RD) that accounts for the effect of unmeasured confounders on propensities, under the mild assumption that the effect is bounded. It estimates the bound with sensitivity analysis, learning a recommender model robust to unmeasured confounders within the bound by adversarial learning. However, pursuing robustness within a bound may restrict model accuracy. To avoid the trade-off between robustness and accuracy, we further propose Benchmarked RD (BRD) that incorporates a pre-trained model into the learning as the benchmark. Theoretical analyses prove the stronger robustness of our methods compared to existing propensity-based deconfounders, and also prove the no-harm property of BRD. Our methods are applicable to any propensity-based estimators, where we select three representative ones: IPS, Doubly Robust, and AutoDebias. We