基于图神经网络和注意力机制的会话推荐

Session recommendation based on graph neural network and attention mechanism

为解决基于循环神经网络会话推荐方法全局偏好表示不准确,以及欠考虑目标项目与所有项目相关性的问题,提出一种基于图神经网络和注意力机制的会话推荐方法。利用图神经网络捕捉会话项目间的依赖关系,得到项目嵌入;通过多头注意力生成全局嵌入准确表示全局偏好,根据目标注意力生成目标嵌入激活目标项目相关性;融合当前嵌入,得到会话嵌入,预测下一次点击。在公共数据集上进行对比实验,实验结果表明,相较最优基准模型,P@20最高达到了71.74%,提高超过0.3个百分点,MRR@20最高达到了35.20%,提高超过3个百分点,验证了该方法的有效性。

To solve the problem that the representation of global preferences based on the recurrent neural network is inaccurate in session recommendation,and the relevance of the target items with all items is neglected to consider,a session recommendation based on graph neural network and attention mechanism method was proposed.The graph neural network was applied to capture dependencies between items to obtain item embedding.Multi-head attention accurately represented global preferences,the global embedding was obtained,target attention activated the relevance of the target items with all items,target embedding was obtained.By fusing the local embedding,the ultimate session embedding was gained,and the next click was predicted.The comparison experiments were carried out in the public data sets.Experimental results show that,the proposed method is superior to the optimal benchmark method,P@20 reaches up to 71.74%and increases more than 0.3 percentage point,MRR@20 reaches up to 35.20%and increases more than 3 percentage point,verifying the effectiveness of the proposed method.