MARLIN-X: Margin-Based Feature Selection and Rule Distillation using Mosaic Patch Migration Optimization for Academic Performance Prediction

Objectives: This study aims to develop an optimization based framework for predicting student academic performance using smartphone behavioural indicators. Method: The proposed Margin-based Rule Learning Integrated Network with eXplainable Optimization (MARLIN-X) framework integrates Tri-Label Margin Feature Selection, DCAR-Net classification and Rule-Set Distillation under a novel Mosaic Patch Migration Optimizer. The optimizer operates using patch based population structuring, migration, fragmentation and elite gene preservation. Feature subsets, hyperparameters and rule masks are optimized within a unified search process. Findings: The proposed method achieves 97.8% accuracy on the School dataset and 98.3% on the College dataset. Feature dimensionality is reduced by more than 50% while increasing tri-label margin values. The distilled rule set achieves 96.5% fidelity with respect to deep model predictions. Balanced accuracy and Macro-F1 scores show consistent improvement over recent methods. Novelty: The study introduces a Mosaic Patch Migration Optimizer that performs patch-level structured search rather than individual-level movement. Keywords: Academic Performance Prediction, Smartphone Usage Analytics, Feature Selection, Nature-Inspired Optimization, Deep Learning, Rule Distillation, Educational Data Mining