Crime Prediction using Adaptive Quadtrees

ICLR 2026 Conference Submission18844 Authors

Keywords: Scalable Hierarchical Crime Prediction, Adaptive Quadtree, Regression, Ensemble Modelling, Clustering

Abstract: Urban crime prediction demands scalable methods for large, skewed spatio-temporal data. We introduce SMART-CARE, an adaptive quadtree-based hierarchical framework that dynamically partitions urban spaces and refines local predictors. Given $\mathcal{D}=\{(\mathbf{x}_i,t_i,c_i)\}_{i=1}^N$, SMART-CARE learns $f:(\mathbf{x},t)\mapsto\hat{c}$ through: (i) variance-driven median splitting with adaptive capacity $T_{\max}$ and depth $L_{\max}$, (ii) periodic local re-tuning with leaf merging to prevent over-fragmentation, and (iii) parent→child knowledge transfer for model fine-tuning. Experiments on NYC and Chicago crime data show SMART-CARE outperforms uniform grids, static quadtrees, and standard baselines in accuracy and efficiency while enabling fine-grained localized forecasts.

Primary Area: learning on time series and dynamical systems

Submission Number: 18844