利用门控网络构建用户动态兴趣的序列推荐模型

在推荐系统中，现有多数序列推荐方法将用户行为视为一个时间有序的序列进行用户兴趣建模，用户兴趣的动态变化导致模型难以从用户行为序列中捕捉准确的用户兴趣信息。针对该问题，同时考虑到项目间成对的共现模式应作为交互信息的补充，提出利用门控网络构建用户动态兴趣的序列推荐模型DCGN。使用门控线性单元捕获交互序列中的用户兴趣，利用带有注意力权重的门控循环网络学习用户的动态兴趣。在此基础上，对用户交互项目间的共现模式进行建模，与用户兴趣信息以及用户信息进行融合后输入深度神经网络，得到最终推荐结果。在ML100K、Amazon 5-Elect、Retailrocket 3个公开数据集上进行实验，使用精确率、归一化折损累积增益和命中率进行性能评估，结果表明，DCGN模型较NARM、GRU4Rec、NLR等主流序列推荐模型性能更优，其归一化折损增益和精确率在Retailrocket数据集上平均提升1.9%和1.22%，在Amazon 5-Elect数据集上平均提升0.82%和1.05%，在ML100K数据集上平均提升0.36%和0.31%。

Sequential Recommendation Model Using Gated Network to Construct User's Dynamic Interest

In recommendation systems, most existing sequence recommendation methods regard user behavior as a time-ordered sequence to model user interest.The dynamic change of user interest, however, makes it difficult for the model to capture accurate user interest information from the user behavior sequence.Aiming at solving this problem, and considering that the paired co-occurrence pattern between items should be used as a supplement to interaction information, a sequential recommendation model DCGN is proposed to construct user's dynamic interests by using gated network.A Gated Linear Unit(GLU) is used to capture user interests in interaction sequences, and a gated recurrent network with attention weights is used to learn the user's dynamic interests.In this way the co-occurrence pattern between user interaction items is modeled, and the user interest information and user information are integrated into the deep neural network to obtain the final recommendation result.Experiments are carried out on three public datasets, ML100K, Amazon 5-Elect, and Retailrocket, and the performance is evaluated in terms of the following three metrics:precision, Normalized Depreciation Cumulative Gain(NDCG) and hit rate.The results show that the DCGN model has better performance compared with the mainstream sequence recommendation models such as NARM, GRU4Rec, NLR, et al.Its NDCG and precision increase of 1.9% and 1.22% on Retailrocket dataset, 0.82% and 1.05% on Amazon 5-Elect dataset and 0.36% and 0.31% on ML100K dataset, in average.