Trip purpose prediction using travel survey data with POI information via gradient boosting decision trees

At present, data obtained from the Global Positioning System (GPS) is significantly valuable in mobility research. However, GPS-based data lacks include trip purpose information. Consequently, many researchers have endeavoured to predict or impute these missing attributes. Existing studies have focused on constructing more features to improve prediction accuracy, but paid less attention to the model's applicability and transferability. In this study, five trip purposes are extracted, including education, recreation, personal, shopping, and transportation, from Chengdu Household Travel Survey (HTS) data. The individual and trip characteristics that are common and can be easily derived from GPS data are carefully selected and extracted. Point of Interest (POI) data of the trip destination are also collected to enhance input characteristics. To obtain more accurate results, an ensemble learning model, Gradient Boosting Decision Trees (GBDT), is employed to predict trip purposes. grid search and cross-validation techniques are used to optimize the hyper-parameters. Empirical results show that the proposed model achieves 0.788 accuracy, which is 22.17%, 14.53%, 10.36%, and 6.77% higher than Multinominal Logit (MNL), Artificial Neural Network (ANN), Random Forest (RF), and Deep Belief Network (DBN), respectively. It is also found that although increasing trip features improve the model's accuracy, it simultaneously impairs model's transferability and generalizability. This study explores the trip purpose prediction problem using 2016 Household Travel Survey (HTS) data and POI data in Chengdu, China.image