MULTI-TASK RECOMMENDATION MODEL COMBINING SELF-ATTENTION MECHANISM AND KNOWLEDGE GRAPH
-
摘要: 借助知识图谱提供辅助信息以提升推荐系统性能愈加受到研究者的关注。针对基于知识图谱的推荐算法用户表示较为单一，无法充分挖掘隐藏信息的问题，提出一种融合自注意力机制和知识图谱的推荐模型KSMR。通过自注意力捕获用户交互序列的上下文信息，得到融合兴趣转移的用户向量，采用文本卷积网络实现特征修正与再提取；交替训练推荐任务和知识图谱嵌入任务，达到协同优化的目的。在真实数据集MovieLens-1M与Last.FM上的实验结果表明，模型的点击率预测（CTR）性能相较于对比算法均有明显提升。Abstract: Researchers have got increasingly attention to obtain auxiliary information with the help of knowledge graph. Aimed at the problem that recommendation algorithms based on knowledge graph have single user representation and cannot fully mine hidden information, a recommendation model combining self-attention mechanism and knowledge graph (KSMR) is proposed. The context information of user interaction sequence was captured by self-attention mechanism to obtain the user vector fused with interest transfer, and the feature correction and re-extraction were realized by text CNNs. Alternating training was used to combine the knowledge graph embedding task and recommendation task to achieve the purpose of collaborative optimization. Experimental results on real datasets MovieLens-1M and Last.FM show that, the CTR (Click Through Rate) prediction performance of the model has obvious advantages over the comparison algorithms.
-
计量
- 文章访问数: 11
- HTML全文浏览量: 1
- PDF下载量: 14
下载:
