多层级用户兴趣与多意图融合的下一篮推荐算法  

Multi-level User Interest and Multi-intent Fusion for Next Basket Recommendation

作　　者：魏楚元 袁保杰 王昌栋[2] WEI Chuyuan;YUAN Baojie;WANG Changdong(School of Electrical and Information Engineering,Beijing University of Civil Engineering and Architecture,Beijing 100044,China;School of Computer Science and Engineering,Sun Yat-sen University,Guangzhou 510275,China)

机构地区：[1]北京建筑大学电气与信息工程学院,北京100044 [2]中山大学计算机学院,广州510275

出　　处：《计算机科学与探索》2025年第3期749-763,共15页Journal of Frontiers of Computer Science and Technology

基　　金：科技部“十四五”重点研发计划项目-子课题(2022YFB3305602);教育部人文社会科学研究规划一般项目(22YJAZH110);北京市教育科学“十四五”规划重点课题(CHAA22061)。

摘　　要：下一篮推荐旨在根据用户历史交互的篮子序列,为用户推荐下一篮可能感兴趣的商品。针对现有下一篮推荐算法未能较好解离篮子内的多意图以及仅从单一层面考虑用户的兴趣或意图,导致推荐效果受限等问题,提出了一种多层级用户兴趣与多意图融合的下一篮推荐模型(MLIMI),从多个层级分别考虑用户兴趣与多意图,构建全局级的用户-项目交互图。考虑到用户行为会随时间发生变化,设计一种长短期时间衰减权重平衡交互项的重要性,通过图卷积网络学习用户的动态兴趣;构建局部级篮子-项目图,通过图解离网络学习解离化的篮子内多意图,随后通过一个多头自注意力层对多意图进行编码,得到最终的意图表示。设计一个跨层级的对比学习范式,结合来自不同层级的项目表示,以增强不同层级项目之间的语义信息。在预测层中融合来自不同层级的用户兴趣和意图,进行下一篮预测。在两个公共基准数据集Ta Feng和Dunnhumby上与MITGNN、TAIW、MINN等主流模型进行了对比实验,结果表明MLIMI的性能优于当前许多基线模型。Next basket recommendation aims to recommend the next basket of items that users may be interested in based on the basket sequence of user historical interactions.The existing next-basket recommendation algorithms fail to adequately disentangle multi-intents within baskets and consider user interests or intents from only a single level,thereby resulting in suboptimal recommendation performance.To address the limitations,this paper proposes a multi-level user interest and multi-intent fusion model(MLIMI)for next-basket recommendation.This model separately considers user interests and multi-intents from multiple levels.Firstly,a global-level user-item interaction graph is constructed.Considering that user behavior changes over time,a long and short-term time decay weight is designed to balance the importance of the interaction items,and then the user’s dynamic interests are learnt through graph convolution networks.Secondly,a local-level basket-item graph is constructed to learn the disentangled multi-intents within baskets via a graph disentangled network,and subsequently the multi-intents are encoded via a multi-head self-attention layer to obtain the final intent rep-resentations.A cross-level contrastive learning paradigm is also designed to combine item representations from different levels in order to enhance the semantic information between items at different levels.Finally,user interests and intents from different levels are fused in the predict layer for the next basket of predictions.Comparative experiments with main-stream models such as MITGNN,TAIW and MINN on two public benchmark datasets,TaFeng and Dunnhumby,show that MLIMI outperforms many current baseline models.

关 键 词：下一篮推荐 图解离网络 多意图学习 对比学习 多头注意力机制 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]