Label-Aware Imputation with Cluster Refinement for Smartphone Usage Analytics in Educational Institutions

Abstract

The accurate handling of missing values remains a crucial step in data preprocessing, particularly in behavioral analytics where data incompleteness can distort pattern recognition and predictive modeling. This study presents a novel Label-Aware Imputation with Cluster Refinement (LAICR) framework designed specifically for smartphone usage datasets collected from educational institutions. The method partitions the dataset by usage-level labels (Low, Moderate, High), applies class-specific imputation using iterative reconstruction for numerical data and mode-based filling for categorical data, and refines results through K-Means clustering to improve local consistency.Experiments conducted on school and college datasets demonstrate significant improvements over standard global imputation techniques. The proposed method achieved an RMSE of 0.4575 and R² of 0.7735 for the school dataset, and RMSE of 0.4876 and R² of 0.7636 for the college dataset, outperforming global iterative and statistical baselines. Additionally, classification performance on imputed datasets reached 99.3% accuracy with XGBoost, indicating strong preservation of feature discriminability.The novelty of this work lies in combining label-awareness with intra-class cluster refinement, effectively reducing reconstruction error and preserving behavioral structure. This approach enhances the reliability of smartphone usage analytics, enabling more robust predictive modeling and behavioral interpretation in educational contexts.