Learning Individual Behavior Using Sensor Data: The Case of GPS Traces and Taxi Drivers

Abstract

The ubiquitous deployment of mobile and sensor technologies has led to both the capacity to observe human behavior in physical (offline) settings as well as to record it. This provides researchers with a new lens to study and better understand the individual decision processes that were previously unobserved. In this paper, we study decision making behavior of 11,196 taxi drivers in a large Asian city using a rich data set consisting of 10.6 million fine-grained GPS trip records. These records include detailed taxi GPS trajectories, taxi occupancy data (i.e., whether a taxi was occupied with a passenger or was vacant) and taxi drivers’ daily incomes. This capacity to use data where occupancy of the taxi is known is a distinctive feature of our data set and sets this work apart from prior work which has attempted to study driver behavior. The specific decision we focus on pertains to actions drivers take to find new passengers after they have dropped off their current passengers. In particular, we study the role of information derivable from the GPS trace data (e.g., where passengers are dropped off, where passengers are picked up, longitudinal taxicab travel history with fine-grained time stamps) observable by or made available to drivers in enabling them to learn the distribution of demand for their services over space and time. We conduct our study using a heterogeneous Bayesian learning model. We find strong heterogeneity in individual learning behavior and driving decisions, which is significantly associated with individual economic outcomes. Drivers with higher incomes benefit significantly from their ability to learn from not only demand information directly observable in the local market, but also aggregate information on demand flows across markets. Interestingly, our policy simulations indicate information that is noisy at the individual level becomes valuable after being aggregated across various spatial and temporal dimensions. Moreover, the value of information does not increase monotonically with the scale and frequency of information sharing. Finally, our study has important welfare implications in that efficient information sharing leads to an income increase among all drivers, instead of a redistribution of income between different types of drivers. Our work allows us not only to explain driver decision making behavior using these detailed behavioral traces, but also to prescribe information sharing strategy for the firm in order to improve the overall market efficiency.