Estimating Accessibility of new Demand-ResponsiveTransport Deployments via Transfer Learning

Abstract

Demand-Responsive Transport (DRT) is increasingly introduced to improve mobility in areas where conventional Public Transport (PT) offers limited frequency or coverage. Given the high operational costs of DRT, it is crucial for its further adoption, to rigorously assess its benefits, considering not only traditional operational metrics such as waiting and travel times, but also accessibility. Accessibility measures the ease of reaching surrounding opportunities such as jobs, healthcare, or education. However, data-driven accessibility computation for integrated DRT-PT services remains challenging in newly deployed territories, where historical operational data are unavailable or scarce. This paper proposes a Domain-Adaptive Multi-Task Learning (DAMTL) framework to learn job accessibility patterns from operational data of different “source” territories, in which DRT is already deployed. Through adversarial domain adaptation, such a knowledge is transferred to a target territory with limited data. We apply the framework to three French metropolitan areas: Lille and Strasbourg as source territories, and Orléans as the target. Results show that with approximately 5 weeks of DRT operation in the target territory, the accessibility estimates issued by our method approximate well (70.1%) the estimates achievable over ∼ 6 months. These results show that our method can be successfully employed in the initial phases of its deployment in a new territory (e.g., in a pilot), to estimate, from the beginning, the benefits of DRT to the accessibility of a territory. Such early accessibility estimation can be used to confirm the pertinence of a certain DRT deployment or to diagnose some undesired accessibility distribution outcome and correct it via appropriate service planning intervention.