LLM-Based News Recommendation System With Multi-Granularity News Content Fusion and Dual-View User Interest Perception

Abstract

Owing to its formidable semantic comprehension capabilities, the Large Language Model (LLM) is extensively employed to enhance the accuracy and personalization of news recommendation. However, two issues restrict the performance of news recommendation systems based on LLM: 1) The single-prompt-based LLM fails to capture the multi-scale and multi-perspective perception of users when processing news content, leading to inaccurate and less diverse news representation; 2) The imbalanced distribution of user interests in historical news browsing leads to calibration issues in news recommendation scenarios, preventing the accurate identification of the full spectrum of user interests. To address these issues, we propose the LLM-Based News Recommendation System with Multi-granularity News Content Fusion and Dual-View User Interest Perception (LLMNR-MCFDIP). Specifically, LLMNR-MCFDIP constructs a unified representation of news content and dynamically adjusts user interest models through multi-granularity news content fusion and dual-view user interest perception. It employs a Multi-granularity Fusion News Encoder to generate intermediate representations of news and then fuses them into a unified representation in an adaptive manner. Meanwhile, by forming a weighted combination of user interests through dual views (autocorrelation among historically clicked news and correlation with candidate news), it avoids interest bias and enhances the accuracy and balance of recommendations. Extensive experimental results on MIND datasets demonstrate our LLMNR-MCFDIP significantly outperforms existing techniques.