Towards travel recommendation interpretability: Disentangling tourist decision-making process via knowledge graph

Abstract

Understanding tourists’ decision-making processes, in which many factors ranging from functional attributes to geographical configurations are highly intertwined, has long been a crux for tourism management. Existing studies are typically based on manual surveys that extract the intricate psychological or behavioural mechanisms, but the huge expense of the required samplings limits the generalization and comprehensiveness of the findings. This study proposes a novel explainable recommendation method—Knowledge-Graph-aware Disentangled Auto-Encoder (KGDAE)—to automatically unravel the tourists’ decision processes from massive historical behaviour data. Based on the constructed tourism-KG that integrates multidimensional factors into 23 types of entities corresponding to 37 semantic and geographic relationships, KGDAE realizes a macro-micro supervised disentangled learning for the interaction of multiple determinants. Macroscopically, the hierarchical attention mechanisms are designed to distinguish the dominance of either functional or geographical factors, and capture the effect of the residential environment; microscopically, the preference-propagation-based technique is introduced to infer the fine-grained characteristics and relations of tourist interests on the tourism-KG. Extensive experiments show that KGDAE can effectively restore tourists’ decision processes according to two empirical studies while boosting the recommendation performance compared to multiple state-of-the-art methods with an increase of 1∼19%. Furthermore, the advantaged interpretability also guarantees the robustness of sparse recommendation scenario to achieve the lowest degradation at 7.8%.