基于动态兴趣传播和知识图谱的推荐方法

Recommendation method based on dynamic interest propagation and knowledge graph

知识图谱推荐作为一种信息过滤方法被广泛应用于电子商务和网络社交等领域,然而多数基于知识图谱的推荐方法未采取合适的策略来解决传播过程中实体语义关联性衰减问题,且单维度建模无法利用知识图谱同时丰富用户和项目表示。针对以上问题提出一种基于动态兴趣传播和知识图谱的推荐方法(recommendation method based on dynamic interest propagation and knowledge graph,RDPKG)。首先,通过传播网络挖掘层级用户兴趣生成用户表示,并采用注意力机制区分不同传播层数下用户兴趣的重要性;然后,通过交叉压缩单元提取知识图谱中的有效信息生成项目表示,并采用多任务学习优化推荐单元和知识图谱嵌入单元;最后,将最终的用户表示和项目表示内积获得交互概率。在推荐系统领域的3种公共数据集上进行对比实验,实验结果表明在点击率预测任务中RDPKG的准确率分别达到85.42%、76.09%和69.39%,优于其他对比方法,充分验证了RDPKG方法的有效性。

As an information filtering method,knowledge graph recommendation is widely used in the fields of e-commerce and social networking.However,most knowledge graph-based recommendation methods did not adopt appropriate strategies to solve the problem of entity semantic relevance decay during the propagation process.Additionally,single-dimensional modeling could not utilize knowledge graph to enrich user and item representations at the same time.Therefore,we propose RDPKG,a recommendation method based on dynamic interest propagation and knowledge graph.Specifically,RDPKG employs a propagation network to mine user interests of different layers to generate use representation;and applies an attention mechanism to distinguish the importance of user interests under different propagation layers.Then,RDPKG employs a cross-compression unit to extract valid information in the knowledge graph to generate item representation,and applies multi-task learning to optimize the recommendation unit and the knowledge graph embedding unit.Last,RDPKG takes the inner product of the final user representation and the item representation to obtain the interaction probability.Comparative experiments on three real-world public datasets in the field of recommender systems were carried out.The results demonstrate that the accuracy of RDPKG in the click-through rate prediction task has reached 85.42%,76.09%and 69.39%respectively.RDPKG outperforms other comparison methods,which fully verifies the validity of RDPKG method.