Assessing the Potential of Google Location History (GLH) Data for Travel Behavior Research in the Context of Developing Country

Using passive data to investigate travel behavior is becoming increasingly prevalent, owing to its convenient data acquisition process. This study seeks to evaluate the feasibility of leveraging Google Location History (GLH) data for analyzing travel behavior within the context of a developing nation like Bangladesh, characterized by high population density, diverse land use, and heterogeneous traffic patterns, including a significant presence of non-motorized vehicles, and relatively low motorized vehicle speeds. A group of 60 individuals willing to share their GLH data stored in the Google Maps application was recruited to accomplish this. A dedicated mobile phone application named Trip Tracker was developed to facilitate the collection of ground truth data. Validation of the GLH data was carried out through a three-step procedure. Initially, the identification of home and work locations from GLH, based on visit frequency and duration, was cross-verified against user-provided inputs, demonstrating 100% accuracy. Subsequently, the accuracy of day-to-day travel data, including arrival and departure times and locations, was assessed against GLH information, yielding a spatial and temporal matching accuracy of 82%. Thirdly, the modes of transportation extracted from ground truth data were compared with those provided by GLH, revealing a mode prediction accuracy of 53 % for GLH data. This discrepancy was attributed to the intricate nature of Dhaka's traffic system and the prevalence of non-motorized transportation modes like rickshaws. Additionally, GLH tends to aggregate multimodal trips, revealing only the high-speed mode and neglecting the mode(s) used for the last/first-mile connection. Consequently, two predictive models were developed utilizing Random Forest (RF), a tree-based machine learning (ML) algorithm, and a long short-term memory neural network (LSTM-based NN) to refine the GLH-predicted travel mode information. The RF and LSTM models achieved mode prediction accuracies of 86% and 68%, respectively, representing a notable improvement over GLH predictions. Further enhancements in accuracy can be anticipated by increasing the sample size.