Onboard occupancy reconstruction in a public transport network through fraud modeling by combining AFC and APC data

Abstract

Accurate estimation of ridership and fare evasion is essential for monitoring public transport networks. The combination of Automated Fare Collection and Automated Passenger Counting (APC) data delivers good results in fraud detection but its estimation on unequipped vehicles remains challenging. This paper proposes a new method to estimate fraudulent boardings and alightings on vehicles that are not equipped with APC, using transfer learning from equipped vehicles. Constrained linear regressions are fitted at multiple spatio-temporal granularities. To address data sparsity, a K-Nearest Neighbors approach transfers information across close spatio-temporal granularities. The method is evaluated on four public transport networks showing a significant improvement in onboard occupancy reconstruction compared to other existing approaches, and highlighting the importance of incorporating temporal dynamics in fraud estimation.