A Comparative Analysis of Distance Metrics for Multilabel Land-Use Classification Using Mobile Phone Data in Trento

Mobile-network traces provide a complementary view of land use, yet the role of distance functions in multilabel neighborhood-based inference remains underexplored in heterogeneous regions. We compare eleven pointwise, complexity-invariant, and elastic distances across four multilabel k-nearest-neighbor (k-NN) algorithms using weekly profiles of incoming calls, outgoing calls, and internet activity from 241 manually labeled cells in Trentino, Italy. Four non-k-NN classifiers with fixed settings constitute the reference baselines. We evaluate performance through nine different metrics under the same iterative stratified 10-fold cross-validation procedure. The distances belonging to the Canberra family perform well in several classifier–channel pairs, including an average precision of 0.905 and a ranking loss of 0.125 in ML-localkNN for incoming calls. The experiment also highlights the inconsistent effects of complexity correction and varying sensitivities to distance selection across classifiers and channels. The study contributes a systematic empirical assessment of distance families, a consistent multilabel evaluation framework, and spatial diagnostics linking numerical performance to land-use interpretation.