基于图神经网络和改进自注意网络的会话推荐

Session recommendation based on graph neural network and improved self-attention network

基于会话推荐旨在根据用户当前会话和历史会话预测用户的下一次点击。现有的会话推荐系统大多数基于当前会话建立局部偏好来预测用户行为,而低估了会话全局序列蕴含的信息。同时多数推荐系统忽略了会话交互序列的相对位置关系。针对这些问题,提出了一种基于图神经网络与改进自注意力网络融合的会话推荐模型(GNN-SAP)。GNN-SAP通过GNN与注意力机制来提取当前会话节点的局部偏好,通过改进自注意网络来捕获会话节点的全局偏好;同时在会话节点中加入可学习的位置嵌入,来更好地把握用户兴趣变化的过程。最终,通过线性融合全局偏好和局部偏好的方式来预测行为。通过大量的实验验证了GNN-SAP模型在常用的稀疏、密集数据集和不同评价指标上都优于现有的会话推荐方法,并且通过对GNN-SAP不同组件的消融实验验证了通过将基于GNN短期偏好和基于改进自注意力的全局偏好融合的有效性。

Session-based recommendation aims to predict the user’s next click based on the user’s current and historical sessions.Most existing session recommendation systems establish local preferences based on current sessions to predict users’ behaviors,but underestimate the information contained in the global sequence of sessions.They usually ignore the relative positional relationship of session interaction sequences that contain rich information.In this regard,this paper proposes a session recommendation model,based on the fusion of graph neural network and the improved self-attention network,named GNN-SAP.GNN-SAP extracts the local preferences of current session nodes through GNN and the attention mechanism,and the global preferences of session nodes through the improved self-attention network.It also integrates learnable positional embeddings to the nodes to better grasp the process of user interest changes.The final behavior prediciton is completed by linear transformation over the concatenation of the local and global preferences.A large number of experiments demonstrate that GNN-SAP is superior to the existing conversational recommendation methods on commonly used sparse and dense datasets and different evaluation indicators,and the ablation experiments on different components of GNN-SAP proves the effectiveness of the fusion of the short-term preference based on GNN and the global preference based on improved self-attention network.