基于双编码器的会话型推荐模型

A Session Recommendation Model Based on Dual Encoders

针对现有会话型推荐模型难以准确捕获物品间全局依赖的问题,提出了一种基于双编码器的会话型推荐模型(SR-BE)。该双编码器由基于自注意力网络的全局编码器和基于图神经网络的局部编码器组成,无论被浏览物品之间的时间间隔长还是短,全局编码器都能够利用注意力机制自适应地捕获被浏览物品之间的全局依赖,并将其编码为全局隐向量。为弥补自注意力网络没有结构信息而难以捕获邻近物品间局部依赖的不足,在局部编码器中,首先将会话序列构建成会话图,然后通过图神经网络在会话图上捕获邻近物品间的局部依赖,并将其编码为局部隐向量。最后将从双编码器得到的全局隐向量和局部隐向量线性组合为会话表示,再通过预测层解码会话表示得到每个候选物品被点击的概率。实验结果表明:将基于自注意力网络的全局编码器与基于图神经网络的局部编码器结合在一起,比单一地使用全局编码器或局部编码器在命中率上分别提高了3.11%和6.55%。通过与同类模型客观定量比较,SR-BE模型在两个公开数据集上取得了突出的效果,表明该模型有效、可行。

A session recommendation model based on dual encoders(SR-BE)is proposed to solve the problem that it is difficult to accurately capture the global dependencies between items in the existing session recommendation models.The dual encoder consists of a global encoder based on self-attention network and a local encoder based on graph neural network.The global encoder can adaptively capture the global dependencies among all items by using the attention mechanism regardless of their time intervals and encode them into global latent vectors.In order to make up for the lack of local dependencies in self-attention network,the local encoder first constructs the session sequences into a session graphs,then the graph neural networks are used to capture the local dependencies among adjacent items and encode them into local latent vectors.Finally,the linear combination of global and local latent vectors obtained from the dual encoder is used to obtain the session representations and the click probability of each candidate item is obtained by decoding the session representions through the prediction layer.Experimental results and comparisons with a single global encoder and a single local encoder show that the combination of the global encoder based on self-attention networks and the local encoder based on graph neural networks improves the hit rate by 3.11%and 6.55%,respectively.Comparisons with state-of-the-art methods on two public datasets validate the effectiveness and feasibility of the proposed method.