Causal Representation Learning for Out-of-Distribution Recommendation

Abstract

Modern recommender systems learn user representations from historical interactions, which suffer from the problem of user feature shifts, such as an income increase. Historical interactions will inject out-of-date information into the representation in conflict with the latest user feature, leading to improper recommendations. In this work, we consider the Out-Of-Distribution (OOD) recommendation problem in an OOD environment with user feature shifts. To pursue high fidelity, we set additional objectives for representation learning as: 1) strong OOD generalization and 2) fast OOD adaptation. This work formulates and solves the problem from a causal view. We formulate the user feature shift as an intervention and OOD recommendation as post-intervention inference of the interaction probability. Towards the learning objectives, we embrace causal modeling of the generation procedure from user features to interactions. However, the unobserved user features cannot be ignored, which make the estimation of the interaction probability intractable. We thus devise a new Variational Auto-Encoder for causal modeling by incorporating an encoder to infer unobserved user features from historical interactions. We further perform counterfactual inference to mitigate the effect of out-of-date interactions. Moreover, a decoder is used to model the interaction generation procedure and perform post-intervention inference. Fast adaptation is inherent owing to the reuse of partial user representations. Lastly, we devise an extension to encode fine-grained causal relationships from user features to preference. Empirical results on three datasets validate the strong OOD generalization and fast adaptation abilities of the proposed method.