杨雨豪_推荐系统的大规模图预训练和微调.pdf

《杨雨豪_推荐系统的大规模图预训练和微调.pdf》由会员分享，可在线阅读，更多相关《杨雨豪_推荐系统的大规模图预训练和微调.pdf（26页珍藏版）》请在三个皮匠报告上搜索。

1、Yuhao Yang,Yuhao Yang,The University of Hong KongThe University of Hong Kong&WeChatWeChatGraphProGraphPro:Graph Pre-training and Prompt Learning for:Graph Pre-training and Prompt Learning for RecommendationRecommendation腾讯微信技术创新奖腾讯微信技术创新奖&犀牛鸟评选第二名犀牛鸟评选第二名GNNs are successful in RecsysGNNs are success

2、ful in RecsysLightGCN,He et al.2020SGL,Wu et al.2022RLMRec,Ren et al.2024GNNs are successful in RecsysGNNs are successful in RecsysSimGCL,Yu et al.2023Question:Are static GNNs enough for recommendation modeling?Question:Are static GNNs enough for recommendation modeling?The static setting differs gr

3、eatly from real-world recommendation problemsModel performances are hardly to be fairly evaluatedChallenges in deploying and translating in real-world applicationsStatic SettingReal-world Problem“We often conceptualize RecSys as the task of predicting missing values in a static user-item interaction

4、 matrix,rather than predicting a users decision on the next interaction within a dynamicdynamic,changingchanging,and application-specific contextapplication-specific context.”Aixin Sun,Beyond Collaborative Filtering:A Relook at Task Formulation in Recommender SystemsQuestion:Are static GNNs enough f

5、or recommendation modeling?Question:Are static GNNs enough for recommendation modeling?Real-world recommendation brings super large graphs with more than 1B edges,which even evolve continuallyHard to learn and update representationsQuestion:Are static GNNs enough for recommendation modeling?Question

6、:Are static GNNs enough for recommendation modeling?Introducing GraphPro:Pretraining and Introducing GraphPro:Pretraining and PromptingPromptingWhat can be a better training and evaluating settingtraining and evaluating setting for recommenders?How to pre-train and promptpre-train and prompt GNNs fo