Dynamic User Interest Evolution Modeling for Personalized Recommendation Systems Based on Multi-Scale Temporal Awareness and Attention Mechanisms

Abstract

This paper addresses the problem of dynamic user interest changes in recommendation systems by proposing a deep recommendation method based on time awareness and interest evolution modeling. The method introduces multi-scale time windows, time decay factors, and attention mechanisms to hierarchically model user behavior sequences. It extracts both short-term and long-term interest features and integrates them into a dynamic interest representation. The model considers not only the sequential nature of behaviors but also incorporates time intervals, behavior frequency, and the weighting of key interaction segments to improve the accuracy of interest modeling. In the experimental section, multiple evaluations are conducted on the MovieLens 1M dataset. The results show that the proposed method significantly outperforms existing mainstream models in terms of Precision@10, Recall@10, and NDCG@10. In addition, sensitivity experiments are conducted on learning rate, optimizer, and interest evolution speed. These experiments confirm the model's stability and robustness to parameter changes. Overall, this study builds a recommendation framework that performs well in dynamic environments. It effectively improves the accuracy and practicality of personalized recommendations.