Scholarlink

<p>Graph-based recommender systems with temporal dynamics offer strong interpretability and adaptability for capturing evolving user preferences. However, their performance is highly sensitive to key parameters, such as time decay functions and short-term session duration, whose values are often set randomly or arbitrarily. In this paper, we propose approaches based on temporal analyses of existing paths between a user and an item at the time preceding their connection in the bipartite interaction graph. To systematically estimate these parameters, we use statistical protocols to adjust time decay functions of graph edge weights according to the evolution of their involvement rate on the paths of length three connecting the new user-item link in the graph. The near-optimal short-term duration is estimated by analyzing the tail of the distribution of this involvement rate as a function of edge duration. To compare our proposed strategies to the Randomized Search Cross-Validation method, experiments are conducted using the Ciao and Epinions datasets, three time decay functions (exponential, power and logistic), the Time Weight Bipartite Graph, the Session-Based Temporal Graph (STG), and the Time Weight STG (TSTG) as recommender graphs, and F1-score@10, Hit Ratio@10, and NDCG@10 as evaluation metrics for Top-10 recommendations. The results obtained for TSTG, which incorporates both short-term duration and time decay functions, show that in 52/54 (96.3%) of cases, the estimated performance is equal to or greater than the average performance of Randomized Search Cross-Validation, and in 41/54 (75.9%) of cases, it is equal to or greater than the maximum performance of Randomized Search Cross-Validation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Dynamic bipartite graph analysis for the systematic estimation of temporal dynamics parameters for graph-based recommender systems

  • 期刊论文

  • 发表日期：

    2026年5月15日

查看全文
  • Armel Jacques Nzekon Nzeko’o,
  • Thomas Messi Nguélé,
  • Xaveria Youh Djam,
  • Saïd Mounpain Ndangouo

摘要

Graph-based recommender systems with temporal dynamics offer strong interpretability and adaptability for capturing evolving user preferences. However, their performance is highly sensitive to key parameters, such as time decay functions and short-term session duration, whose values are often set randomly or arbitrarily. In this paper, we propose approaches based on temporal analyses of existing paths between a user and an item at the time preceding their connection in the bipartite interaction graph. To systematically estimate these parameters, we use statistical protocols to adjust time decay functions of graph edge weights according to the evolution of their involvement rate on the paths of length three connecting the new user-item link in the graph. The near-optimal short-term duration is estimated by analyzing the tail of the distribution of this involvement rate as a function of edge duration. To compare our proposed strategies to the Randomized Search Cross-Validation method, experiments are conducted using the Ciao and Epinions datasets, three time decay functions (exponential, power and logistic), the Time Weight Bipartite Graph, the Session-Based Temporal Graph (STG), and the Time Weight STG (TSTG) as recommender graphs, and F1-score@10, Hit Ratio@10, and NDCG@10 as evaluation metrics for Top-10 recommendations. The results obtained for TSTG, which incorporates both short-term duration and time decay functions, show that in 52/54 (96.3%) of cases, the estimated performance is equal to or greater than the average performance of Randomized Search Cross-Validation, and in 41/54 (75.9%) of cases, it is equal to or greater than the maximum performance of Randomized Search Cross-Validation.