基于知识图谱的多任务推荐算法

Multi⁃task recommendation algorithm based on knowledge graph

基于知识图谱的推荐算法可以丰富物品特征,挖掘用户兴趣,有效解决传统推荐算法存在的冷启动和数据稀疏性问题,然而现有基于知识图谱的推荐算法常忽略用户交互中协同信息对图谱训练的正向作用,在图谱缺失度较高的情况下难以挖掘物品的深层特征。为此,本文提出一种基于知识图谱的多任务推荐方法,联合训练推荐任务与图谱补全任务。该算法首先构建用户-物品连通图和物品知识图谱,利用图卷积神经网络分别扩充用户物品的交互表征与实体关系的结构表征,传播协同信息和图谱信息;同时,采用两层注意力结构分别建模同阶邻域的重要性差异和异阶邻域的信息衰减,自适应聚合信息;最后交叉共享物品与实体的高阶表征,学习来自对方任务的知识。该算法充分刻画物品和实体表征,在提高图谱完备性的基础上提高推荐效率。在三个公开数据集和一个自建数据集上与基准算法进行对比实验,结果表明本文算法在AUC、F1等指标上有明显提高。

The knowledge graph⁃based recommendation algorithm can enrich the characteristics of items,explore user interests,and effectively solve the cold start and data sparsity issues of traditional recommendation algorithms.However,existing knowledge graph⁃based recommendation algorithms often overlook the positive effect of collaborative information in user interactions on graph training,making it difficult to explore the deep features of items when the graph has a high degree of missingness.Therefore,in this article,a multi⁃task recommendation method called MRGC based on the knowledge graph is proposed,which jointly trains the recommendation task and the graph completion task.Firstly,the algorithm constructs a user⁃item connectivity graph and an item⁃related knowledge graph.It utilizes graph convolutional neural networks to expand the interaction representation of users and items and the structural representation of entity relationships,propagating collaborative information and graph information.At the same time,a two⁃layer attention structure is used to model the importance differences of same⁃order neighborhoods and the information decay of different⁃order neighborhoods,adaptively aggregating information.Finally,high⁃order representations of items and entities are cross⁃shared to learn knowledge from the other task.This algorithm fully characterizes item and entity representations,improving recommendation efficiency based on improving graph completeness.Comparative experiments are conducted with benchmark algorithms on three public datasets and one self⁃built dataset,the results show that the MRGC algorithm significantly improves metrics such as AUC and F1.