HMPR: A Hybrid Multi-Paradigm Recommendation Framework for Causal-Aware Academic Service
Year: 2026
Type: conference-paper
Abstract: Existing academic recommendation systems are plagued by selection bias, often conflating item popularity with genuine user interest[1]. We propose HMPR, a causal-enhanced framework that explicitly models the recommendation process as a causal intervention rather than mere correlation fitting. We introduce a Causal Recommender Module utilizing Inverse Propensity Weighting (IPW) and counterfactual reasoning to estimate the true causal effect of recommendations. This approach effectively identifies high-quality yet underexposed potential papers, significantly improving long-tail discovery. To support this mechanism, we construct a robust representation layer synergizing a Multi-Modal Transformer for semantic features[2]and a Heterogeneous Graph Attention Network (HAN) for complex scholarly relationships. Furthermore, we incorporate an Adaptive Interest Evolution Network and Supervised Contrastive Learning to address temporal dynamics and cold-start issues, respectively. Extensive experiments demonstrate that the HMPR framework achieves state-of-the-art performance, reaching a Precision@10 of 80.1% while reducing selection bias by 68%, delivering recommendations that are accurate, diverse, and fair.
Authors: YiDi Ma, YiZe Zhao, LiangBin Yang
Institution: University of International Relations
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7171531726