会话场景下基于特征增强的图神经推荐方法

Feature Augmentation based Graph Neural Recommendation Method in Session Scenarios

基于图神经网络的会话推荐(简称图神经会话推荐)是近年来推荐系统领域的一个研究重点和热点,这主要是因为它们引入了会话图拓扑结构信息来提高物品和会话特征表示的准确性,因而,在一定程度上提升了会话推荐的性能.然而,现有图神经会话推荐方法仍然存在两方面的不足,从而影响其性能:1)它们所构建的会话图中物品间的相关性权重均是在模型训练之前就预先指定并保持固定不变,导致不能准确捕捉物品间的相关性;2)它们只从单个会话的物品序列中产生物品和会话的局部特征,而缺乏从整个会话数据集出发,全局考虑不同物品之间以及不同会话之间的相关性,并由此来生成物品和会话的全局特征,从而充分表示物品和会话的语义特征.为此,本文提出了一种新颖的会话场景下基于特征增强的图神经推荐方法FA-GNR(Feature Augmentation based Graph Neural Recommendation).FA-GNR方法首先基于单个会话构建物品间相关性权重可学习优化的会话图,并借鉴GRU(Gated Recurrent Unit)神经网络来产生物品局部特征,同时基于会话数据集,通过GloVe(Global Vectors)词嵌入方法产生物品全局特征,从而融合物品的局部和全局特征来生成其语义特征.然后,FA-GNR方法基于物品语义特征,利用局部注意力机制来产生会话的局部特征,同时基于物品的全局特征,并通过全局注意力机制来产生会话的全局特征,从而融合会话的局部和全局特征来生成其语义特征.最后,在物品和会话语义特征的基础上,FAGNR方法通过交叉熵损失来学习给定会话下不同物品的点击概率分布.在多个公开数据集上的实验结果表明,FA-GNR方法的推荐性能优于目前主流的方法.

Recommender systems are an effective way to solve the information overload problem.They can filter information from an overwhelming volume of data based on a user’s historical preference to find the content a user is really interested in,and help the user to efficiently obtain the information they want.They have been widely used in various fields,such as news,video,shopping,travel recommendations,etc.Recently,graph neural network based session recommendation(referred to as graph neural session recommendation)has become a research focus and hot in the recommender system community.This is mainly since they introduce the topological structure information of session graph to improve the accuracy of item and session feature representation,and therefore,to a certain extent,can improve the performance of session recommendation.However,existing graph neural session recommendation methods still have two maindrawbacks,which affects the further improvement of performance ofsession recommendation.Firstly,the correlation weights between items in the session graph they construct are all pre-specified before the model training and remain fixed,which leads to the inability to capture the correlations between items accurately.Secondly,they only generate the local features of items and sessions from the item sequence of a single session.And they lack ofthe perspective ofan entire session dataset and the global consideration of the correlations between different items and between different sessions,which can be used to produce global features of items and sessions.This leads to inability to represent the semantic features of items and sessions adequately.To address the above two drawbacks,we innovatively propose FA-GNR(Feature Augmentation based Graph Neural Recommendation method in session scenarios)in this paper.The FA-GNR method first constructs a session graph with learning and optimizing correlation weights between items based on a single session,which is used to generate item local features via GRU(Gated Recurrent Unit)neural network.At the same time,based on the session dataset,item global features are generated by the GloVe(Global Vectors)word embedding method.In this way,item local and global features are fused to produce item semantic features.Then,the FA-GNR method utilizes the local attention mechanism to generate session local features based on item semantic features.Meanwhile,based on item global features,the global attention mechanism is used to generate session global features.In this way,session local and global features are fused to produce session semantic features.Finally,based on item and session semantic features,the FA-GNR method uses cross-entropy loss to learn the click probability distribution of different items under a given session.Experimental results on several public datasets(e.g.Yoochoose1_64,Yoochoose1_4,and Diginetica)indicatethat the recommendation performance of FA-GNR is better than that of existing mainstream methods inthegraph neural session recommendationtask.For example,on the Yoochoose1_4 and Diginetica datasets,the P@20 metric of FA-GNR exceeds the state-of-theart graph-based neural network methods by 2.36% and 3.21% on average,respectively.At the same time,ablation,t-SNE visualization and hyper-parameter experiments further demonstrate the effectiveness of FA-GNR.