新颖性会话推荐机制研究

Novel Session-Based Recommendation Mechanism

摘要: 目前会话推荐模型大多通过计算项目物品与用户会话的准确性匹配度，从而决定用户的推荐列表。但是，仅依靠准确性匹配度进行推荐导致存在推荐新颖性不足、推荐结果倾向热门化问题。为解决上述问题，本文提出了一种能够提升会话推荐列表新颖性的新颖性机制，将其运用在会话推荐模型中以观察效果。本文首先将用户与会话推荐系统的交互行为数据构建为有向会话图，使用图神经网络的门控更新机制融合邻域信息与全局偏好，为每一个项目生成向量信息；然后，将会话的局部兴趣与全局偏好的注意力加权融合，计算会话的向量，并通过会话的向量与项目进行相似性计算，得出每个项目的推荐得分；最后，引入基于流行度的物品新颖度指标，可以通过修改参数x平衡新颖性与准确性，得出每个物品的综合得分，取出综合得分排名前列的物品作为推荐列表。在调整参数x的过程中，本文根据准确性和新颖性实验结果，将新颖化推荐分为三种模式：精准推荐模式、新颖性推荐模式、新颖性优先模式。系统管理者可以根据不同的场景灵活设置参数x以完善用户体验。本文为提升会话推荐系统的新颖性推荐能力提供了一种新的技术路径。

Abstract: Currently, most session-based recommendation models determine the user’s recommendation list by calculating the accuracy matching degree between items and the user session. However, relying solely on accuracy matching for recommendations leads to issues such as insufficient recommendation novelty and a tendency for results to favor popular items. To address these problems, this paper proposes a novel mechanism capable of enhancing the novelty of session-based recommendation lists and applies it to session-based recommendation models to observe its effects. This paper first constructs the interaction data between users and the session-based recommendation system as a directed session graph. It then employs the gated update mechanism of a graph neural network to fuse neighborhood information and global preferences, generating vector representations for each item. Next, the local interests of the session and the global preferences are fused via attention-weighted aggregation to compute the session vector. The similarity between the session vector and items is calculated to derive the recommendation score for each item. Finally, a popularity-based item novelty metric is introduced, which allows balancing novelty and accuracy by adjusting parameter x, resulting in a comprehensive score for each item. The top-ranked items based on this comprehensive score are selected as the recommendation list. During the process of adjusting parameter x, this paper categorizes novelty-enhanced recommendations into three modes based on experimental results of accuracy and novelty: Accuracy-First Mode, Novelty-First Mode, and Novelty-Priority Mode. System administrators can flexibly set parameter x according to different scenarios to improve the user experience. This paper provides a new technical approach for enhancing the novelty recommendation capability of session-based recommendation systems.