Causal Recommendation under Confounder Observability: A Short Survey

Abstract

Causal inference has emerged as a promising paradigm for recommendation systems, moving beyond correlation-based learning to uncover the causal mechanisms behind user–item interactions. Conventional models assume observed interactions directly reflect users’ true preferences, but this assumption often fails in practice: interactions arise from complex system dynamics and various confounding factors, inducing biased feedback loops and spurious correlations that degrade recommendation quality and fairness. Causal inference addresses this by identifying and adjusting for confounding effects, using techniques such as propensity scoring, counterfactual estimation, and causal representation learning to recover the true causal effects behind user–item interactions and enable more accurate, interpretable, and fair recommendations. In this work, we present a focused review of causal recommendation methods, proposing a taxonomy along two dimensions, confounder observability and time-variance, and offering a comparative analysis of existing work alongside directions for future research.