Abstract
This study compares eleven distance metrics across four multi-label k-nearest-neighbour classifiers using weekly profiles of incoming calls, outgoing calls and internet activity in Milan. The task assigns overlapping home, work and density labels to grid cells. Canberra ranks highest overall among the tested k-nearest-neighbour configurations, while complexity-invariant dynamic time warping performs poorly despite its greater computational cost. With outgoing calls, the local-k classifier using Canberra achieves a ranking loss of 0.095 and an accuracy of 0.742.
Citation
Mendoza-Hurtado, M., Cuevas-Muñoz, J. M., & Ortiz-Boyer, D. (2026). Distance metrics for multi-label urban function and density classification from mobile phone data: A Milan benchmark. GeoInformatica, 30(2), Article 37. https://doi.org/10.1007/s10707-026-00598-8