融合视觉信息的协同知识注意力网络推荐模型

Recommendation Model of Collaborative Knowledge Attention Network Fusing Visual Information

推荐系统中将知识图谱作为辅助信息能有效缓解协同过滤算法中的稀疏性和冷启动问题.然而,现有的基于知识图谱的推荐模型往往忽略了视觉信息以及用户和物品历史交互序列中不同物品对当前任务的重要性.由此提出了一种融合视觉信息的协同知识注意力网络推荐模型(CKVI).该模型使用注意力机制动态地捕获用户和物品历史交互数据中蕴含的用户历史偏好信息.同时将知识图谱作为辅助信息,进一步丰富用户和物品的表示,增强模型的可解释性.其次考虑到与物品相关的图像,如电影海报中蕴涵着丰富的视觉信息,设计了一种图像聚合方法,聚合用户的历史行为图像,捕获用户的视觉偏好.最后将几种信息融合,用于推荐.为了验证模型有效性,在MovieLens和Book-crossing两个数据集上进行了实验,结果表明CKVI相比其他对比的模型推荐效果有较大提升.

Knowledge graph auxiliary information in the recommendation system can effectively alleviate sparsity and cold startup problems in collaborative filtering algorithms.However,existing knowledge graph based recommendation models tend to ignore the importance of visual information and the different items in the historical interaction sequence of users and items to the current task.This paper presents a recommendation model of collaborative knowledge attention network fusing visual information(CKVI).This model uses the attention mechanism to dynamically capture the user′s historical preference information contained in the historical interaction data of users and items.At the same time,the knowledge graph is used as auxiliary information to further enrich the representation of users and items and enhance the explanability of the model.Secondly,considering the images related to items,such as movie posters,which contain rich visual information,an image aggregation method is designed to aggregate the historical behavior images of users and capture the users′visual preferences.Finally,several types of information fusion are used for recommendation.In order to verify the validity of the model,experiments were carried out on two datasets,MovieLens and Book-crossing.The results show that CKVI is more effective than other comparative model recommendation.