Abstract
ML-proxkNN combines feature-space locality with Moore-neighbourhood relationships in the spatial grid to improve multi-label place identification from call detail records. Experiments on Milan and Trento compare eight multi-label classifiers under standard and spatial block cross-validation. The method achieves strong average precision and F1 scores while retaining its advantage when spatial autocorrelation is controlled, indicating that grid adjacency can improve locally adaptive neighbourhood selection across contrasting geographic contexts.
Citation
Mendoza-Hurtado, M. (2026). A new local proximity-based k-nearest neighbours method for multi-label population patterns identification using mobile phone data. Manuscript under review.