A Counterfactual Neural Causal Model for Interactive Recommendation

The survivor effect leads the optimization of recommender systems towards local optima. Existing solutions address the issue via reinforcement learning. These techniques recombine sub-patterns from collected human-system collaborations to obtain higher valuations and, consequently, to mitigate the effect. However, such mitigation from causal perspective is unidentifiable, which makes their estimation unreliable. In this work, we achieve the identification and the estimation by a learnable Neural Causal Model (NCM). The idea is to shift the mitigation towards consistency, a well-defined counterfactual quantity that can be theoretically identified. Specifically, to identify the consistency, we construct the NCM based on the available graphical causal model, which provides a qualitative characterization of preference transitions. Subsequently, we utilize the structural model to quantitatively identify the counterfactual consistency. To estimate the NCM, we introduce the feedforward neural structural functions that leverage the Gumbel-Softmax approximation for differentiation. Additionally, we incorporate multiple types of standard reinforcement learning optimizations into our implementation. Empirical studies further validate the effectiveness of our solution, showcasing improvements of +3.98% in offline experiments and +16.82% in online simulations.