Research on User Correlation Behavior Analysis Based on Graph Neural Networks
基于图神经网络的用户关联行为分析研究

摘要

针对传统用户行为分析方法难以刻画用户之间高阶关联、时序依赖与关系传播的问题，提出一种基于图神经网络 的用户关联行为分析方法。首先，将用户属性、浏览点击、收藏购买及交互时间统一表示为多关系图，分别构建用户 —行为二部图、用户关联图和时序行为图；其次，引入关系消息传播、时间衰减权重与注意力聚合机制，学习用户节 点的结构化表示；最后，通过关联解码器计算用户间潜在关联概率，并支持群体发现、异常关系识别和个性化运营。 在包含640个用户、2932条正关联边的可控仿真数据上进行验证，结果表明，所提方法的F1值和AUC分别达到0.755 和0.814，较多层感知机分别提高3.5和5.8个百分点。

Abstract

To address the difficulty of traditional user behavior analysis methods in characterizing high-order correlations, temporal dependencies, and relationship propagation among users, this paper proposes a user correlation behavior analysis method based on graph neural networks. First, user attributes, browsing and clicking, favoriting and purchasing, and interaction time are uniformly represented as a multi-relational graph, and a user-behavior bipartite graph, a user correlation graph, and a temporal behavior graph are constructed respectively. Second, relation-based message propagation, time-decay weights, and attention aggregation mechanisms are introduced to learn structured representations of user nodes. Finally, a correlation decoder is used to calculate the potential association probability between users and support group discovery, abnormal relationship identification, and personalized operations. The method is validated on controllable simulation data containing 640 users and 2,932 positive association edges. The results show that the proposed method achieves F1 and AUC values of 0.755 and 0.814, respectively, representing improvements of 3.5 and 5.8 percentage points over the multilayer perceptron model.