【摘要】 针对现有会话推荐采用单一模型无法兼顾全局和局部信息，从而影响推荐性能的问题，提出融合图神经网络和稀疏自注意力的会话推荐模型(SSA-GNN)。模型采用稀疏自注意力构建全局隐向量，以解决无关项的干扰和图神经网络难以表示长距离依赖的问题；采用目标注意图神经网络构建局部隐向量，更深层次的捕获项目间的复杂依赖。最后在预测层将全局和局部隐向量线性连接，有效兼顾了全局和局部信息。模型在Yoochoose1/64数据集上的试验结果比基线模型GC-SAN在评价指标P@20上提高了1.25%,MRR@20上提高了4.59%。更多还原

【Abstract】 Aiming at the problem that the existing session recommendation model cannot take into account the global and local information, which affects the recommendation performance, a session recommendation model(SSA-GNN) combining graph neural network and sparse self-attention is proposed.The model uses sparse self-attention to construct a global hidden vector to solve the interference of irrelevant items and the difficulty of representing long-distance dependencies in graph neural networks; Using a target attention graph neural network to construct local hidden vectors and capture complex dependencies between items at a deeper level.Finally, the global and local hidden vectors are linearly connected in the prediction layer, effectively balancing global and local information.Compared with baseline model GC-SAN,the experimental results on Yoochoose1/64 dataset have increased by 1.25% on P@20 and 4.59% on MRR@20.更多还原