Causality-driven user modeling for sequential recommendations over time

Abstract

Contemporary sequential recommendation systems predominantly leverage statistical correlations derived from user interaction histories to predict future preferences. However, these correlations often mask implicit challenges. On the one hand, user data is frequently plagued by implicit, noisy feedback, misdirecting users towards items that fail to align with their actual interests, which is magnified in sequential recommendation contexts. On the other hand, prevalent methods tend to over-rely on similarity-based attention mechanisms across item pairs, which are prone to utilizing heuristic shortcuts, thereby leading to suboptimal recommendation. To tackle these issues, we put forward a causality-driven user modeling approach for sequential recommendation, which pivots towards a causal perspective. Specifically, we involves the application of a causal graph to identify confounding factors that give rise to spurious correlations and to isolate conceptual variables that causally encapsulate user preferences. By learning the representation of these disentangled causal variables at the conceptual level, we can distinguish between causal and non-causal associations while preserving the inherent sequential nature of user behaviors. This enables us to ascertain which elements are critical and which may induce unintended biases. The framework of our method can be compatible with various mainstream sequential models, which offers a robust foundation for reconstructing more accurate and meaningful user and item representations driven by causality.