面向社区团购的深度图神经网络推荐系统

Deep Graph Neural Network Recommendation System for Community Group Buying

由于我国农业生产的熟人社会属性与各种社区团购日益普及,基于社交关系的推荐系统已成为农产品社区团购等新兴互联网应用中不可或缺的组成部分。然而,现有研究大多没法探索、量化用户偏好和社会关系间的相关性,忽略了可能影响某些社交关系拓扑的农产品特征间的相关性(同好者交友现象),因此提出一种基于深度图神经网络的社交推荐模型(GNN-R4A)。首先,该模型将用户与农产品特征空间抽象为两个图网络,分别通过图神经网络方法进行编码;然后,将两个编码空间嵌入矩阵分解的两个隐因子中完成用户—农产品评分矩阵中缺失的评分值;最后,在3组数据集上开展实验,使用均方根误差、均方误差、归一化折损累计增益作为评估指标,结合消融实验验证推荐方法的有效性。实验表明,GNN-R4A相较于已有方法推荐效果更优,可为面向社区团购的推荐系统提供参考与借鉴。

Due to the social attributes of acquaintances in agricultural production in China and the increasing popularity of various community group buying,recommendation systems based on social relationships have become an indispensable component of emerging internet applica⁃tions such as agricultural product community group buying.However,most existing research attempts to explore and quantify the correlation between user preferences and social relationships,neglecting the correlation between agricultural product features that may affect certain so⁃cial relationship topologies(the phenomenon of like-minded friends).Therefore,a social recommendation model based on deep graph neural networks(GNN-R4A)has been proposed.Firstly,the model abstracts the user and agricultural product feature space into two graph net⁃works,which are encoded using graph neural network methods;Then,embed the two encoding spaces into the two hidden factors of the matrix decomposition to complete the missing scoring values in the user agricultural product scoring matrix;Finally,experiments were conducted on three datasets,using root mean square error,mean square error,and normalized cumulative loss gain as evaluation indicators to validate the effectiveness of the recommended method through ablation experiments.Experiments have shown that GNN-R4A has better recommendation performance compared to existing methods,and can provide reference for recommendation systems for community group buying.