融合相似性负采样和短期偏好的图卷积推荐模型

Graph convolution recommendation model combining similarity negative sampling and short-term preference

大多数基于知识图谱的推荐算法在进行知识图谱学习任务时多采用随机替换的方式对负样本进行采样,不能帮助系统高效地学习样本特征;此外,在学习用户的兴趣时忽略了时间因素对用户偏好的影响。针对以上两点,提出了一种融合相似性负采样和用户短期偏好的推荐模型(SPKG)。首先,使用TransE将实体嵌入到向量空间,采用K-means聚类算法将实体进行聚类,通过同簇实体的相互替换可获得高质量的负三元组;然后,采用基于注意力机制的双向门控循环网络从用户近期交互的物品序列中提取用户的短期偏好;最后,通过用户的短期偏好对用户进行推荐。在3个数据集上对模型的性能进行验证,结果表明,相比于基线模型,SPKG在AUC、召回率和F1指标上都有所改善。

Most of the knowledge graph-based recommendation algorithms mostly use random substitution to sample negative samples when performing knowledge graph learning tasks,which does not help the system to learn sample features efficiently.In addition,the influence of temporal factors on user preferences is ignored when learning users’interests.To address these two points,a recommendation model(SPKG)that incorporates negative sampling of similarity and short-term user preference is proposed.Firstly,entities are embedded into the vector space using TransE and clustered using the K-means clustering algorithm,and high quality negative triads can be obtained by mutual replacement of entities in the same cluster;then,a two-way gated recurrent network based on an attention mechanism is used to extract users’short-term preference from the sequence of items that users have recently interacted with;finally,users are recommended by their short-term preference.The performance of the model is validated on three datasets,and the results show that SPKG improves on AUC,recall and F1 metrics compared to the baseline model.