Event Details

Elif Kozan will present a seminar to the Management Science Department. This is the second speaker of a two speaker session.

Abstract Estimating shortest-path distances in road networks from geometric information is a fundamental problem in transportation, routing, and operational planning. Although Euclidean distance provides a strong baseline, the relationship between geometric proximity and topological distance varies substantially across cities due to differences in road structure and detour behaviour. In particular, detour factor distributions can exhibit pronounced heterogeneity across urban environments, leading to unstable predictions and poor worst-case performance when conventional loss functions and model selection criteria are used.

In this study, we propose a city-adaptive deep learning framework for topological distance estimation in road networks based on a hybrid loss formulation. The proposed approach jointly optimizes absolute distance accuracy, scale-invariant geometric consistency through logarithmic detour factor supervision, and distributional stability via median-based regularization. To account for city-specific structural properties, we introduce lightweight city-level tail and predictability indices that adapt both the training loss and a validation-based model selection score. This score combines average error, tail risk, and ranking consistency in a leakage-free experimental protocol.

Experiments on multiple real-world road networks demonstrate that the proposed method achieves consistently more stable performance across cities, particularly under heterogeneous and heavy-tailed detour behaviour, while maintaining competitive accuracy in well-structured urban environments. The results highlight the importance of distribution-aware objective design and city-adaptive learning strategies for robust distance estimation in operational road network applications.

Supervisor: Burak Boyaci