Handwritten Digit Recognition using Machine Learning

Handwritten Digit Recognition (HDR) remains a fundamental benchmark in pattern recognition and machine learning due to its practical applications and inherent classification challenges posed by diverse handwriting styles. This study investigates and compares two classical statistical classifiers—Gaussian Naive Bayes (GNB) and Linear Discriminant Analysis (LDA)—to recognize the digits from the MNIST dataset. Both models assume underlying normality in feature distributions and offer computational efficiency, making them suitable for high-dimensional input such as image pixels. Using 60,000 training and 10,000 test samples, we evaluate model performance through accuracy, precision, recall, F1 score, and confusion matrices. The results reveal that while GNB achieves moderate accuracy (55.58\%), LDA significantly outperforms it with an accuracy of 87.30\%, demonstrating superior capability in distinguishing visually similar digits. Our analysis further highlights the limitations of GNB’s independence assumption and underscores LDA’s strength in capturing shared variance across classes. These findings reinforce the effectiveness of LDA as a robust baseline for HDR tasks, especially when interpretability and computational simplicity are desired.