Comparative Evaluation of Classical Machine Learning Classifiers and Dimensionality Reduction Techniques for American Sign Language Recognition

Abstract

Amid growing demand for efficient and accessible sign language recognition systems, this study evaluates five classical machine learning classifiers—Decision Tree, Random Forest, K-Nearest Neighbors (KNN), Naïve Bayes, and Support Vector Machine (SVM)—for the task of recognizing American Sign Language (ASL) gestures from an MNIST-style image dataset. Following standard preprocessing, Dimensionality reduction techniques such as Principal Component Analysis (PCA) and SelectKBest were applied to assess their effect on classification performance. Each model was trained and tested with and without dimensionality reduction, and hyperparameters were rigorously tuned via grid search. Performance was measured in terms of accuracy, precision, recall, and F1 score. SVM (RBF kernel, C=10, gamma=0.01) emerged as the top performer—achieving 97.83% accuracy and a 0.9753 F1 score—underscoring the value of combining dimensionality reduction with targeted hyperparameter optimization for reliable ASL classification.