Automated feature selection for multimodal human activity recognition using metaheuristic optimization algorithms

Abstract

The purpose of this thesis is to suggest an automated and generalizable feature selection framework for Multimodal Human Activity Recognition (MHAR) settings using metaheuristic optimization. Conventional feature selection methods (Mutual Information, Recursive Feature Elimination, Lasso L1 regularization) have several limitations that make it difficult for researchers to choose the most suitable one for their dataset and configuration. First, these methods require predefining the number of features to be selected and often explore only a limited portion of search space. Moreover, their effectiveness varies widely across datasets, modalities, and possible data corruption. In order to address this inconsistency, the proposed framework uses metaheuristic-based optimization for feature selection that automatically determines the optimal feature subset and explores a larger variation of feature combinations. Fourteen metaheuristic algorithms are evaluated on three benchmark MHAR datasets, UTD-MHAD, CZU-MHAD, and MM-Fit, across a variety of experimental configurations, which include modality sub-combinations and simulated data loss and corruption at 5 different rates. All experiments follow a leave-one-subject-out (LOSO) cross validation protocol to avoid any data leakage. Results show 96.99% LOSO test accuracy on UTD-MHAD, 95.80% on CZU-MHAD, and 98.93% on MM-Fit on main configurations, which outperforms the conventional feature selection baselines and matches the current state-of-the-art result.