文章编号:1673-5196(2023)01-0094-09基于图神经网络特征交叉的协同过滤算法王燕*,赵妮妮,范林(兰州理工大学计算机与通信学院,甘肃兰州730050)摘要:学习用户和项目有效的向量表示是推荐系统的核心目标,现有的推荐模型大多通过深度神经网络或专门设计的特征交叉,来学习用户-项目间的特征交叉生成用户(项目)向量表示,但并未将用户(项目)特征间的交叉信息编码到嵌入向量中充分利用特征交叉信息,且多个特征交叉信息对于生成最终的用户(项目)向量表示的影响不同.基于此,构建两个图神经网络模块,学习用户(项目)特征间的交叉信息、用户-项目之间的特征交叉信息,并通过计算注意力分数对特征交叉信息进行加权,得到用户(项目)的特征信息;然后通过门控循环神经网络(GRU)聚合原始的特征信息和网络层学习到的特征交叉信息,得到最终的用户(项目)向量表达;最后通过用户向量与项目向量的元素积得到最终的推荐结果.在数据集MovieLens1M、Book-Crossing和Taobao上验证了模型的有效性.关键词:协同过滤;图神经网络;GRU;双线性特征交叉;注意力机制中图分类号:TP391.3文献标志码:ACollaborativefilteringalgorithmbasedongraphneuralnetworkcross-featureWANGYan,ZHAONi-ni,FANLin(SchoolofComputerandCommunication,LanzhouUniv.ofTech.,Lanzhou730050,China)Abstract:Learningeffectivevectorrepresentationsofusersanditemsarethecoregoaloftherecommen-dationsystem.Mostoftheexistingrecommendationmodelsusedeepneuralnetworksorspeciallydesignedfeaturecrossingtolearnthefeaturecrossingbetweenusersandprojectstogenerateuser(item)vectorrep-resentations,butthecrossinformationbetweenuser(item)featuresisnotcodedintotheembeddingvec-tortomakefulluseofthefeaturecrossinformation,andmultiplefeaturecrossinformationhasdifferenteffectsonthegenerationofthefinaluser(item)vectorrepresentation.Basedonthis,twographneuralnetworkmodulesareconstructedtolearntheintersectioninformationbetweenuser(item)featuresandthefeatureintersectioninformationbetweenuseranditem,followingthefeaturecrossinformationweigh-tedbycalculatingtheattentionscoretoobtaintheuser(item)featureinformation.Thentheoriginalfea-tureinformationandthefeaturecrossinformationlearnedfromthenetworklayerareaggregatedthroughthegatedrecurrentneuralnetwork(GRU)...
