Abstract: Online purchase prediction aims to predict a users’ purchase behavior, which can generate considerable commercial value for shopping websites. Currently, most purchase prediction studies ignore the exploration of implicit preferences. Thus, this paper proposes a two-stage dynamic interest recognition model for online purchase prediction. First, the first stage of the model constructs a click frequency graph of user-item pairs and utilizes a light graph convolutional neural network to learn contextual characteristics as their static interest’s representation. Then, the second stage invokes Bi-GRU with attention mechanism to explore the transformation process of user preference. Finally, aiming at the potential high-dimensional features, a purchase prediction model integrating dynamic interest and implicit features is built. This paper conducts extensive experiments on two real e-commerce datasets. Moreover, the results validate that the proposed model can greatly improve the performance of purchase prediction from accuracy and F1 score.

摘要： 在线购买预测旨在预测用户的购买行为，从而为购物网站带来可观的商业价值。目前，大多数购买预测研究忽略了对隐含偏好的挖掘。因此，本文提出基于两阶段动态兴趣识别的购买行为预测模型预测用户购买商品的概率。首先，模型的第一阶段构建用户-项目的点击频率图，并利用轻量图卷积神经网络(Light-GCN, Light-Graph Convolutional Neural Network)学习图的上下文特征作为用户的静态兴趣表征。其次，第二阶段采用带有注意力机制的双向门控递归单元神经网络(Bidirectional Gated Recurrent Unit，Bi-GRU)探索用户偏好的转化过程。最后，针对潜在的高维特征，建立了一个融合动态兴趣和隐含特征的购买预测模型。本文在两个真实电子商务数据集上进行大量实验，结果表明提出的模型可以从准确率和 F1 分数上大大提高购买预测的性能。