Travel itinerary recommendation using interaction-based augmented data

Itinerary planning is complicated for travelers because the traveling content, including places to visit,acceptable times, and distances, can be diverse. Travel recommender systems (TRSs) recommend the mostrelevant itineraries for a traveler. In this paper, we propose an interactive framework that allows users toedit itineraries directly on a map to suit their preferences better. The proposed framework collects feedbackdata by recording user itinerary modifications, infers positive and negative preferences, and fine-tunes therecommender models with our ranking-based loss function. This interaction-based data augmentation approachaddresses data sparsity issues due to personalization by capturing a variety of travel item combinations. Inour experiments, we evaluate multiple combinations of models and itinerary generation methods to show theeffectiveness of integrating interaction data into TRSs. Our experimental evaluations demonstrate that ourinteractive TRS can provide itineraries that align with users' preferences more in terms of point-set-wise andrank-wise accuracy; the integration consistently improves the accuracy for all combinations of the components,and particularly, the improvement is large for small backbone models