融合知识图谱与用户长短期兴趣的新闻推荐

News Recommendation Integrating Knowledge Graph and Users'Long-Term and Short-Term Interests

针对现有新闻推荐算法研究中忽略了新闻之间知识层面的联系和用户短期偏好等问题,提出一种融合知识图谱和用户长短期兴趣的新闻推荐模型。模型由新闻语义编码器、用户兴趣编码器和点击预测器3部分组成,在新闻语义编码器中,除使用新闻本身的标题、简介、类别信息以学习新闻语义表示外,还利用新闻标题与简介中提及的知识实体并结合WikiData知识图谱构建知识子图,从知识子图中学习新闻之间潜在知识层面的联系。在用户兴趣编码器中,使用注意力机制从用户历史点击新闻序列中提取用户的长期兴趣,并采用GRU网络学习用户的短期偏好,然后结合用户的长期兴趣和短期偏好构建用户综合兴趣表示。在MIND-small数据集上分别进行了对比实验和消融实验,在反映模型准确率的AUC指标上,KGLS模型比最先进的基线模型提高了2.92%。

Aiming at the problem that the existing research on news recommendation systems has ignored the use of external knowledge enti-ties to mine the potential knowledge level relationships between news,and has not combined users′short-term preferences for news recommen-dation,this paper proposes a news recommendation algorithm that combines knowledge graphs and users long-term and short-term interests.The model consists of three parts:a news semantic encoder,a user interest encoder and a click predictor.In the news semantic encoder,in ad-dition to using the news′s own title,introduction,and category information to learn the news semantic representation,it also uses the news ti-tle and the knowledge entities mentioned in the introduction are combined with the WikiData knowledge graph to construct a knowledge sub-graph,and learn the potential knowledge-level connections between news from the knowledge subgraph.In the user interest encoder,the at-tention mechanism is used to extract the user′s long-term interest from the user′s historical click news sequence,and the GRU network is used to learn the user′s short-term preference,and then the user′s long-term interest and short-term preference are combined to construct the us-er′s comprehensive interest representation.Comparative experiments and ablation experiments were carried out on the MIND-small dataset,the KGLS model improved by 2.92%compared with the most advanced baseline model on the AUC index reflecting the accuracy of the model.