可解释的局部和全局对比多行为推荐算法

Explainable local and global contrastive multi-behavior recommendation algorithms

针对传统的多行为推荐模型无法高效学习异构信息网络的复杂结构以及缺乏可解释性的问题,提出一种多行为推荐模型,即可解释的局部和全局对比多行为推荐。运用一种被广泛应用于提取全局结构的元路径视图从特定的语义角度捕获每个节点之间的特征。设计一个能捕获局部特征和元路径之间的交互信息的超元路径图来捕获多个元路径之间的交互信息,区分不同用户面对不同类别物品的不同行为模式。采用一种异质性可解释对比学习,确定行为类型的重要性,得出更加优质的正负样本进行对比。在两个公共数据集上的实验中,所提模型优于主流先进推荐模型。

A multi-behavior recommendation model,explainable local and global comparative multi-behavioral recommendation,was proposed to address the problems that traditional multi-behavior recommendation models cannot efficiently learn the complex structure of heterogeneous information networks and lack interpretability.A meta-path view,widely used to extract higher-order structures,was employed to capture the features between each node from a specific semantic perspective.A hyper meta-path graph capturing meta-path features and interaction information between meta-paths was designed to capture the interaction information between multiple meta-paths,distinguishing different behavioral patterns of various users facing items of different categories.A heterogeneity-explainable contrastive learning was used to determine the importance of behavior types yielding higher quality positive and negative samples for comparison.The proposed model outperforms mainstream advanced recommendation models in experiments on two public datasets.