Query2Trip: Dual-Debiased Learning for Neural Trip Recommendation

Peipei Wang, Lin Li, Ru Wang, Xiaohui Tao

Abstract: Trip recommendation (TripRec) seeks to recommend a trip that consists of an ordered sequence of points-of-interest (POIs) for a tourist through a user-specific query. Recent neural TripRec methods with sequence-to-sequence models have achieved remarkable performance. However, alongside the exposure bias in general recommender systems, the selection bias caused by the lack of explicit feedback (e.g., ratings) from the trip data exacerbates the tendency toward users’ unsatisfactory experience in TripRec. To this end, a novel debiased representation learning method for neural TripRec is proposed to fulfill sequence generation from Query to Trip named Query2Trip. It develops dual-debiased learning to mitigate selection bias and exposure bias in TripRec. The former happens as the visit by a user does not necessarily mean the user exhibits a positive preference for the visit. Benefiting from the query provided by a user, Query2Trip designs a debiased adversarial learning module by conditional guidance to alleviate this selection bias from positives (visited). The latter happens as unvisited is not equivalent to negative. Query2Trip devises a debiased contrastive learning module by negative weighting to mitigate this exposure bias from negatives (unvisited). Experiments conducted on eight real-world datasets empirically demonstrate the superior performance of Query2Trip compared to the state-of-the-art baselines.