Knowledge-Based Multi-Behavior Recommendation with Factor Disentanglement
Tingyu Liu, Ying Li · 2022
Traditional recommendation models usually consider only one type of user behavior, some recommendation models focus on user behaviors that frequently occur (e.g. click, view), which have more interaction data but does not accurately reflect the users’ interests, and another types of recommendation models focus on user behaviors which reflect a stronger user preference (e.g. purchase), but has problem of data sparcity. Therefore, in order to improve the performance of the model, multi-behavior-based recommendation is a suitable solution. However, early efforts towards multi-behavior recommendation either do not utilize external knowledge to encode explicit relationships, or cannot fully explore the semantic information of user behavior. Hence, in this work, we solve these problems by constructing a heterogeneous interaction graph and proposing a new multi-behavior recommendation model named KMBFD(short for Knowledge-based Multi-Behavior Recommendation with Factor Disentanglement), which encodes knowledge-based information by multi-relational item graph and user graph and captures semantic information of multi-behavior interaction by a feature aggregation module with factor disentanglement method that is used to disentangle multi-dimensional latent factors hidden in the training data, which can enhance expressiveness and interpretability.