基于图神经网络和注意力机制提取兴趣的会话推荐算法

A Session-Based Recommendation Algorithm for Extracting Interest Based on Graph Neural Network and Attention Mechanism

摘要: 随着用户的隐私意识逐渐增强，平台上存在着大量匿名用户的交互数据，如何为匿名用户进行个性化推荐成为推荐系统备受关注的话题，基于会话的推荐应运而生。会话推荐指的是基于用户当前会话中的交互历史进行推荐，而不依赖用户长期历史行为数据或用户个人属性信息。本文针对现有会话推荐模型难以捕捉用户动态兴趣的问题，提出了一种基于提示词的动态兴趣提取会话推荐模型(Dual-channel Prompt-based Interest Extraction, DPIE)。该模型使用图神经网络作为训练阶段，引入兴趣提示词，通过设计滑动注意力窗口机制，对会话中的项目进行细粒度的表示学习，自适应地捕捉用户兴趣的变化，结合提示词增强项目的语义表示，得到动态兴趣表示。同时，引入位置编码，结合长期兴趣提示词学习用户的长期兴趣表示。这种设计使得模型能够更好地建模用户兴趣的动态变化过程，解耦动态兴趣和长期兴趣，从而提供更精准的推荐，在多个真实数据集上的实验结果验证了该方法的有效性。

Abstract: With heightened user privacy awareness, anonymous user interaction data is prevalent on platforms. Personalizing recommendations for anonymous users has become a focal point in recommender systems, leading to the rise of session-based recommendation. Session-based recommendation focuses on leveraging interaction history within the current session, rather than long-term user history or profile attributes. Addressing the challenge of existing session-based models in capturing dynamic user interests, this paper proposes a Dual-channel Prompt-based Interest Extraction (DPIE) model. DPIE employs graph neural networks in the training phase and introduces interest prompts. By designing a sliding attention window mechanism, it enables fine-grained representation learning of items within a session, adaptively capturing shifts in user interests, and enhances item semantic representations with prompts to derive dynamic interest representations. Furthermore, position encoding is incorporated, combined with long-term interest prompts to learn long-term user interest representations. This design allows the model to better model the dynamic evolution of user interests, decoupling dynamic and long-term interests for more precise recommendations. Experimental results on multiple real-world datasets validate the effectiveness of the proposed approach.