DTRP: A Flexible Deep Framework for Travel Route Planning

Abstract

Route planning aims at designing a sightseeing itinerary route for a tourist that includes the popular attractions and fits the tourist’s demands. Most existing route planning strategies only focus on a particular travel route planning scenario but cannot be directly applied to other route planning scenarios. For example, previous next-point recommendation models are usually inapplicable to the must-visiting problem, although both problems are common and closely related in travel route planning. In addition, user preferences, POI properties and historical route data are important auxiliary information to help build a more accurate planning model, but such information are largely ignored by previous studies due to the challenge of lacking an effective way to integrate them. In this paper, we propose a flexible deep route planning model DTRP to effectively incorporate the available tourism data and fit different demands of tourists. Specifically, DTRP includes two stages. In the model learning stage, we introduce a novel multi-input and multi-output deep model to integrate the rich information mentioned above for learning the probability distribution of next POIs to visit; and in the route generation stage, we introduce the beam search strategy to flexibly generate different candidate routes for different traveling scenarios and demands. We extensively evaluate our framework through three travel scenarios (next-point prediction, general route planning and must-visiting planning) on four real datasets. Experimental results demonstrate both the flexibility and the superior performance of DTRP in travel route planning.