目的 现有的基于图神经网络(GNN)的推荐方法忽略了会话中有价值用户在项目上的时间驻留信息，无法解决用户无意识点击带来的影响，同时忽略图神经网络中隐藏因素的表达能力，针对以上问题，提出一种融合时间驻留信息的图神经网络会话推荐模型(Graph Neural Network Session-based Recommendation Based on Fusion of Time Resident Information，TRGNN)。方法 首先，对用户在各个项目上的驻留时间信息进行处理，通过时间图神经网络得到时间特征;其次，应用多头注意力机制增强因素的表达能力更好地提取项目特征，TRGNN将时间特征与项目特征进行融合得到最终特征，通过注意力网络得到全局上下文和局部上下文;最后，通过预测层得到最终推荐结果。结果 在Diginetica和Yoochoose两个真实数据集上进行对比实验，实验结果表明:相较于最优基线模型，本模型在Mrr@20评价指标下分别提升了 1。57％和3。30％，在Recall@20指标下分别提升了 1。10％和0。66％。结论 本模型实现了更好的推荐效果，能更好地挖掘隐藏信息，充分应用时间特征和项目隐藏特征来提高推荐准确率，降低用户误触对推荐准确率的影响。
Graph Neural Network Session-based Recommendation Based on Fusion of Time Resident Information
Objective Existing recommendation methods based on graph neural networks(GNN)often overlook the time resident information of valuable users on items within sessions,fail to address the impact of unconscious clicks by users,and neglect the expressive ability of hidden factors in GNNs.To overcome these limitations,this paper proposed a graph neural network session-based recommendation model based on the fusion of time resident information(TRGNN).Methods This method first processed the time resident information of users on various items and extracted time features using a temporal graph neural network.Subsequently,it applied a multi-head attention mechanism to enhance the expressive ability of factors and better extract item features.TRGNN fused time features with item features to obtain final features,captured global and local contexts through an attention network,and ultimately obtained final recommendations through a prediction layer.Results Comparative experiments were conducted using two real datasets,Diginetica and Yoochoose.The experimental results indicated that,compared with the optimal baseline model,the proposed model achieved improvements of 1.57％and 3.30％in the MRR@20 evaluation indicators,and improvements of 1.10％and 0.66％in the Recall@20 indicators,respectively.Conclusion The proposed model demonstrates superior recommendation performance,effectively mining latent information and fully utilizing temporal features and item latent features to enhance recommendation accuracy while reducing the impact of unintended user clicks on recommendation precision.
万方数据
