LDA-Driven Classifiers and Stacking Ensembles for Efficient Human Activity Recognition

Betül Uzbaş

doi:10.1109/access.2026.3696768

LDA-Driven Classifiers and Stacking Ensembles for Efficient Human Activity Recognition

2026-01-01

0

PRE

AI

B

Betül Uzbaş *

DOI：10.1109/access.2026.3696768

Original

Original request for help

Save

Abstract

En 中文

Human Activity Recognition (HAR) is a core problem in ambient intelligence, aiming at the accurate classification of human movements using data acquired from wearable sensors. To ensure rigorous scientific validity, the methodology implements a leakage-free pipeline that adheres to the original subject-level separation between training and testing sets. This protocol ensures that participants are strictly partitioned, validating the model's performance on entirely unseen individuals to reflect realistic deployment scenarios. The proposed framework leverages Linear Discriminant Analysis (LDA) to transform the 561-dimensional raw sensor space into a high-utility, five-dimensional manifold, achieving a 99.1% reduction in feature dimensionality while preserving critical discriminative information. Classification performance was benchmarked across a spectrum of architectures, ranging from LDA-driven single classifiers to a complex LDA-augmented stacking ensemble (Stack579). In the Stack579 configuration, a meta-learner integrated diversified base model outputs with the original feature space to maximize spatial context. Empirical results demonstrate that while the Stack579_MetaMLP achieved the peak predictive accuracy of 96.64%, the LDA-SVM and LDA-LR variants followed closely with an accuracy of 96.50%. Critically, a pairwise McNemar test (p > 0.05) was conducted to evaluate the significance of these improvements. The analysis revealed no statistically significant difference between the complex ensemble and the streamlined LDA models. Furthermore, LDA-centric models exhibited an exceptional inference latency of approximately 0.0057 ms per sample, providing a decisive computational advantage. This study concludes that while stacking architectures offer maximum absolute precision, LDA-driven single classifiers represent a highly optimized framework for real-time edge computing and wearable applications by ensuring a superior balance between high-tier accuracy and extreme computational agility.

Keywords:

Dimensionality reduction

human activity recognition

linear discriminant analysis

machine learning

stacking

LDA-Driven Classifiers and Stacking Ensembles for Efficient Human Activity Recognition

Abstract

Journal

Organization

Cited Papers