Inferring the purposes of taxi trips using GPS and POI data considering the destination context

Semantic enrichment of the taxi trajectories with trip purposes has been a moderately discussed subject area in travel behavior analysis. Two data sources, (a) Point of Interest (POI) and (b) GPS records of taxi trips, are often combined in predicting the trip purposes of taxi passengers. Three methods (rule-based, probabilistic, and machine learning) have been emerged to infer the trip purpose using these two data sources. However, choosing a model has become a challenge. As rule-based and probabilistic methods are the most widely used in this context, this study considered (1) closest assignment, (2) gravity model, and (3) Bayes theorem-based model accordingly. This study aims to compare and identify the best-performing model among those three models. POI data from Google Places API and GPS records from a taxi service provider in Colombo District, Sri Lanka is utilized for this evaluation. Comparing all these three models it is concluded that Bayes theorem-based model produces more consistent results. The possibility to use crowdsourced user reviews to measure the attractiveness of POIs in probabilistic models is also assessed. Finally, the limitations of the evaluated models and the possibilities of machine learning-based models to overcome those are presented.