基于知识增强的细粒度个性化新闻推荐用户建模
User Modeling for Fine-Grained Personalized News Recommendation Based on Knowledge Enhancement
熊晓波 1方文涛2
作者信息
- 1. 湖南省高速公路集团有限公司,湖南长沙 410153
- 2. 中联重科股份有限公司,湖南长沙 410013
- 折叠
摘要
传统的新闻推荐用户建模方法难以深入解析新闻的复杂语义和用户的真实需求,为此提出了一种知识增强的新闻建模方法,通过实体表示层、上下文嵌入层和注意力聚合层获取新闻文档表示.在此基础上提出了一种基于知识增强文档的细粒度用户建模方法,利用长文档建模技术将知识增强的新闻文档串联成长文档,通过捕获文档间的词级交互行为得到细粒度用户表示;通过捕获文档内的实体交互行为得到粗粒度用户表示,粗粒度的用户表示和细粒度的用户表示聚合得到最终的用户表示.实验结果显示,提出的新闻建模方法在AUC和NDCG@10指标上优于基线模型,基于此方法的用户建模方法在AUC上至少提升2.51％,在NDCG@10上至少提升4.75％.
Abstract
Traditional user modeling methods for news recommendation struggle to deeply analyze the complex semantics of news and the genuine needs of users.To address this issue,this paper first proposes a knowledge-enhanced news model-ing approach,which obtains news document representations through an entity representation layer,a context embedding lay-er,and an attention aggregation layer.Based on this,a fine-grained user modeling method based on knowledge-enhanced documents is proposed,utilizing long document modeling techniques to concatenate knowledge-enhanced news documents in-to a long document.Fine-grained user representations are obtained by capturing word-level interactions between documents,while coarse-grained user representations are derived from capturing entity interactions within documents.The final user representation is aggregated from both coarse-grained and fine-grained user representations.Experimental results show that the proposed news modeling method outperforms baseline models in terms of AUC and NDCG@10 metrics,with the user modeling method based on this approach achieving at least a 2.51％improvement in AUC and at least a 4.75％improvement in NDCG@10.
关键词
新闻建模/用户建模/知识图谱/细粒度/个性化新闻推荐Key words
news modeling/user modeling/knowledge graph/fine-grained/personalized news recommendation引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据