From interests to behavior : hierarchical causal-inspired structure modeling for personalized news recommendation

Abstract

The explosive growth of online news platforms has made personalized news recommendation (NR) essential for helping users navigate overwhelming volumes of content. Unlike conventional recommendation tasks, NR poses unique challenges due to the transient and intent-driven nature of news consumption. Existing methods largely rely on statistical correlations between news items and user behaviors, which often fail to distinguish causally relevant signals from spurious ones. This limitation can result in misaligned or redundant recommendations that do not reflect the underlying decision-making processes of users. In this work, we argue that causal-inspired modeling is crucial for capturing how latent user interests evolve and influence observable behaviors. We identify two complementary types of causality in news consumption: (1) structural dependencies among historical behaviors, and (2) the generative influence of high-level user interests on concrete behaviors. To model these, we propose Hierarchical Causal-Inspired Representation (HCR), a unified framework that explicitly discovers and leverages hierarchical causal structures in user behavior. HCR first disentangles users’ browsing histories into semantically coherent latent interests. It then constructs a two-level causal graph via differentiable causal discovery to model both intra-behavior dependencies and interest-to-behavior influence. The learned structure serves as an effective inductive bias that captures directional and logically meaningful dependencies aligned with user decision patterns. Finally, HCR generates causality-aware user representations that support robust and accurate recommendations. Extensive experiments on real-world datasets demonstrate that our approach consistently surpasses state-of-the-art baselines in both accuracy and diversity.