CIRS: Bursting Filter Bubbles by Counterfactual Interactive Recommender System

While personalization increases the utility of recommender systems, it also brings issue filter bubbles . E.g., if system keeps exposing and recommending items that user is interested in, may make feel bored less satisfied. Existing work studies in static recommendation, where effect overexposure hard to capture. In contrast, we believe more meaningful study interactive recommendation optimize long-term satisfaction. Nevertheless, unrealistic train model online due high cost. As such, have leverage offline training data disentangle causal on To achieve this goal, propose a counterfactual (CIRS) augments reinforcement learning (offline RL) with inference. The basic idea first learn historical capture It then uses learned help planning RL policy. conduct evaluation offline, innovatively create an authentic environment (KuaiEnv) based real-world fully observed rating dataset. experiments show effectiveness CIRS bursting achieving success recommendation. implementation available via https://github.com/chongminggao/CIRS-codes.