A Novel Popularity Extraction Method Applied in Session-Based Recommendation
Yuze Peng,Shengjun Xu,Qingkun Chen,Wenjin Huang,Yihua Huang
DOI: https://doi.org/10.26599/tst.2023.9010061
2024-02-13
Tsinghua Science & Technology
Abstract:Popularity plays a significant role in the recommendation system. Traditional popularity is only defined as a static ratio or metric (e.g., a ratio of users who have rated the item and the box office of a movie) regardless of the previous trends of this ratio or metric and the attribute diversity of items. To solve this problem and reach accurate popularity, we creatively propose to extract the popularity of an item according to the Proportional Integral Differential (PID) idea. Specifically, Integral (I) integrates a physical quantity over a time window, which agrees with the fact that determining the attributes of items also requires a long-term observation. The Differential (D) emphasizes an incremental change of a physical quantity over time, which coincidentally caters to a trend. Moreover, in the Session-Based Recommendation (SBR) community, many methods extract session interests without considering the impact of popularity on interest, leading to suboptimal recommendation results. To further improve recommendation performance, we propose a novel strategy that leverages popularity to enhance the session interest (popularity-aware interest). The proposed popularity by PID is further used to construct the popularity-aware interest, which consistently improves the recommendation performance of the main models in the SBR community. For STAMP, SRGNN, GCSAN, and TAGNN, on Yoochoose1/64, the metric P@20 is relatively improved by 0.93%, 1.84%, 2.02%, and 2.53%, respectively, and MRR@20 is relatively improved by 3.74%, 1.23%, 2.72%, and 3.48%, respectively. On Movieslen-1m, the relative improvements of P@20 are 7.41%, 15.52%, 8.20%, and 20.12%, respectively, and that of MRR@20 are 2.34%, 12.41%, 20.34%, and 19.21%, respectively.
computer science, information systems,engineering, electrical & electronic, software engineering