一种基于知识图谱与内容的推荐算法

A Recommendation Algorithm Based on Knowledge Graph and Content

摘要: 推荐系统作为信息筛选工具，通过对用户个人信息记录分析以实现用户兴趣预测，推荐满足用户需求的物品列表，其核心是推荐算法。现有的推荐算法大多需要收集用户大规模的数据训练来实现，存在数据稀疏性的问题，且当用户需要购买新物品时，往往无法进行有效的推荐。针对该问题，本文提出了一种基于知识图谱和内容的推荐算法。首先，解析用户记录并构建知识图谱，设计基于知识图谱的相似度算法，对知识图谱中的实体与关系进行相似度计算；其次，基于文本词嵌入对内容进行向量化表示，并构建基于内容的相似度算法；最后，本文融合了基于知识图谱的相似度算法和基于内容的相似度算法，通过权重计算选取得分高的值作为推荐结果。与其它模型在MovieLens数据集上进行了实验对比，实验结果表明，本文所提出的算法在召回率和准确率方面有明显的提升。

Abstract: As an information screening tool, the recommendation system can predict users’ interests and recommend the list of items that meet users’ needs through the analysis of users’ personal information records. Its core is the recommendation algorithm. Most of the existing recommendation algorithms need to collect users’ large-scale data for training, which has the problem of data sparseness. When a user wants to buy new items, it is often unable to make effective recommendation. In this paper, we propose a novel recommendation algorithm based on knowledge graph and content. Firstly, a content graph was constructed to obtain the similarity between entities and relationships based on graph embedding. Secondly, the content was vectorized based on text word embedding, and a content-based similarity algorithm was constructed. Finally, a fusion method combined graph-based similarity method and content-based similarity method was proposed, and then the entities with high scores would be selected as the recommendation result through weight estimation. The result of comparative experiments on the MovieLens dataset shows that the algorithm proposed has improved the precision and recall on performance.