关键词: 推荐系统, 流行度, 长尾效应, 聚类, 负采样

Abstract: Aiming at the issue that existing studies on alleviating the long-tail effect fail to fully consider users' personalized behaviors when dividing popular items and long-tail items, a recommendation method integrating user behavior and improved long-tail algorithm was proposed. Firstly, Bidirectional Encoder Representations from Transformers (BERT)was utilized to encode item attribute information, and items were clustered based on the encoding results. Personalized popular items and long-tail items were redefined according to the user's interaction records with different clusters, thereby integrating personalized user behavior into the process of categorizing popular items. Secondly, the user's popularity sensitivity was evaluated based on interaction records, fully considering the extent to which popularity factors influenced the user. Finally, a novel negative sampling method was proposed, in which different negative sampling approaches were adopted for users with varying popularity sensitivities, and user preference clustering was integrated to select higher-quality negative samples. Experiments were conducted on three publicly available real-world datasets, and the results demonstrate that the proposed personalized categorization method is improved compared to the traditional 80-20 split method in terms of Recall, Hit Rate(HR), and Normalized Discounted Cumulative Gain(NDCG). Specifically, on the NDCG@20 metric, which better reflects recommendation effectiveness, improvements of 0.56, 1.10, and 2.41 percentage points are achieved, respectively. Additionally, the proposed negative sampling method is shown to outperform the optimal negative sampling method in terms of HR and NDCG metrics, with improvements of 3.36, 1.45, and 7.56 percentage points on the NDCG@20 metric, respectively, validating the effectiveness of the proposed methods.

Key words: recommender systems, popularity, long tail, clustering, negative sampling