计算机科学与探索

• 学术研究 •    下一篇

融合自适应超图的自监督知识感知推荐模型

周家旋, 柳先辉, 赵晓东, 侯文龙, 赵卫东   

  1. 同济大学 电子与信息工程学院, 上海 201804

Self-Supervised Knowledge-Aware Recommendation Model Integrating Adaptive Hypergraph

ZHOU Jiaxuan,  LIU Xianhui,  ZHAO Xiaodong,  HOU Wenlong,  ZHAO Weidong   

  1. School of Electronic and Information Engineering, Tongji University, Shanghai 201804, China

摘要: 为缓解传统协同过滤推荐系统存在的冷启动问题,知识图谱作为一种辅助知识被引入到推荐系统中。然而,现有的知识图谱推荐模型在充分地建模高阶相互作用方面存在局限性,难以捕获来自高阶邻居的重要信息。此外,监督信号的稀疏性问题也影响着推荐系统性能。为了解决上述问题,提出一种融合自适应超图的自监督知识感知推荐模型。该模型首先使用混合图卷积网络共同学习交互图中低阶交互嵌入与自适应超图中高阶交互嵌入;其次,使用关系感知图注意网络挖掘知识图谱中用户与物品丰富的知识信息;然后,模型在这三种视图基础上构建对比学习任务,通过引入自监督信号来缓解交互数据的稀疏性问题;最后将三种嵌入相结合,用于后续的推荐预测。该模型在多个公开数据集上与KGAT、KGIN、KACL等基准模型进行了对比实验,与7个对比模型中推荐性能最好的模型相比,在Movielens数据集上,Recall@20提升了1.22%,NDCG@20提升了1.17%;在Yelp2018数据集上,Recall@20提升了1.41%,NDCG@20提升了1.60%。实验结果显示该模型的推荐性能优于其他基准模型。

关键词: 推荐系统, 知识图谱, 自适应超图, 自监督学习, 关系感知图注意网络

Abstract: To alleviate the cold-start problem that exists in traditional collaborative filtering recommender systems, knowledge graphs have been introduced as a kind of auxiliary knowledge in recommender systems. However, existing knowledge graph recommendation models have limitations in adequately modeling higher-order interactions, making it difficult to capture important information from higher-order neighbors. In addition, the sparsity problem of supervised signals also affects recommendation system performance. To address the above issues, a self-supervised knowledge-aware recommendation model integrating adaptive hypergraph is proposed. The model first utilizes a hybrid graph convolutional network to jointly learn the low-order interaction embeddings in the interaction graph and the higher-order interaction embeddings in the adaptive hypergraph. Second, it uses a relation-aware graph attention network to mine the rich knowledge information of users and items in the knowledge graph. Then, the model constructs a comparison learning task based on the three views, which mitigates the sparsity problem of the interaction data by introducing the self-supervised signals. Finally, the three kinds of embeddings are combined for subsequent recommendation prediction. The model is experimentally compared with benchmark models such as KGAT, KGIN, and KACL on several publicly available datasets. Compared with the best recommendation performance among the seven compared models, on the Movielens dataset, Recall@20 is improved by 1.22%, NDCG@20 is improved by 1.17%; On the Yelp2018 dataset, Recall@20 is improved by 1.41%, NDCG@20 is improved by 1.60%. The experimental results show that this model outperforms other benchmark models in terms of recommendation performance.

Key words: recommendation system, knowledge graph, adaptive hypergraph, self-supervised learning, relation-aware graph attention network