基于知识表示学习的协同矩阵分解方法

Collective Matrix Factorization Based on Knowledge Representation Learning

针对协同过滤算法中用户反馈数据的稀疏性问题,提出一种基于知识库的协同矩阵分解方法.该方法从物品的知识图谱中学习其向量表示,并在此基础上联合地分解反馈矩阵和物品关联度矩阵,两种矩阵共享物品向量,利用物品的语义信息弥补反馈数据的缺失.实验结果表明,该方法显著地提升了矩阵分解模型的推荐效果,在一定程度上解决了协同过滤的冷启动问题.

In order to solve the problem of user feedback data sparseness existed in collaborative filtering method,a collective matrix factorization method was proposed based on knowledge graph.The method was arranged to make up for the scarce of the user feedback data with additional item sematic information.Learning item embeddings from items’knowledge graph,the method was designed to jointly factorize a user feedback matrix and an item relatedness matrix with the same item embeddings.Experimental results on two datasets show that the proposed method can significantly improve the performance of matrix factorization models,and it can solve the cold start problem to some extent.