【摘要】 针对目前基于图神经网络会话推荐的意图难以捕捉和新项目表示不充分问题，提出一种使用意图推理网络进行基于会话的新项目推荐模型。该模型从会话角度捕捉项目间的多级关系生成多级用户意图，同时捕获用户在分类数据上的偏好，将两者结合得到用户意图表示。通过基于元学习的新项目表示层学习旧项目获取元知识，并根据新项目属性信息表示新项目嵌入。将用户意图嵌入与新项目嵌入的点积进行归一化得到每个新项目的得分，将得分较高的新项目推荐给用户。实验结果表明所提出模型在Amazon G&GF和Yelpsmall两个数据集上P@20和MRR@20比最优基线算法分别提升22.6%、76.1%和14.5%、8.5%。更多还原

【Abstract】 To address the challenges of difficult intent capturing and inadequate representation of new items in current sessionbased recommendations using graph neural networks, a session-based new item recommendation model utilizing an intent reasoning graph neural network was proposed. Multi-hop relations among items were captured from a session perspective to generate multi-level user intents, while user preferences on categorical data were simultaneously captured. These two components were combined to derive the user intent representation. Additionally, a meta-learning-based new item representation layer was employed to extract meta-knowledge from historical items, which enables the model to generate embeddings for new items based on their attribute information. Finally, the dot product between the user intent embedding and the new item embedding was normalized to compute a score for each new item. New items with the highest scores were then recommended to the user. Experimental results demonstrate that the proposed model achieves significant improvements over the state-of-the-art baseline algorithms, with P@20 and MRR@20 increasing by 22. 6% and 76. 1% on the Amazon G&GF dataset, and by 14. 5% and 8. 5% on the Yelpsmall dataset, respectively.更多还原