Enhancing the POI data for trip purpose inference using machine learning techniques.

Activity-based modeling has become the backbone behind transportation planning, and trip purposes inferred from large GPS datasets are paving the path to augmenting its accuracy. Since trip purpose inference models are often accompanied by Point of Interest (POI) data, the quality of POI data is highly influential in obtaining outcomes. However, only a few studies attempted to enhance the POI data to achieve higher accurate purpose inference results. Hence, this study focuses on overcoming two common limitations that emerge with POI data in purpose inference utilizing machine learning techniques: (1) Improving the POI categorization with text classification and (2) Identifying the possible entrances for POIs with large land areas from cluster analysis. Support Vector Machines (SVM) predicts the category type of POIs with the highest accuracy from compared classifiers, and 1,656 entrances were identified from the designed methodology to identify entrances. These were tested with a probabilistic trip purpose inference model and evaluated using large-scale taxi GPS data. The experimental evaluation shows that the proposed methods effectively enhance the trip purpose inference compared to its output with the trip purpose proportions of passengers obtained from a travel survey in the study area (i.e. for "Shopping" trips it enhanced from 16.1% to 23.6% which is comparable to 25.2%)