Influence of Various ML-Based Binary Classifiers on the Performance on Handwritten Digit Recognition

The handwritten dataset from MNIST is a well-known and reliable tool for the study on various classification solutions. The aim of this study is to compare four competing Machine Learning classifiers, including a boosting algorithm, when applied to the MNIST dataset for predicting 0-Images. Accuracy (ACC) of Support Vector Machine (SVM), Stochastic Gradient Descent (SGD), Adaboost and Random Forest models are evaluated through 3-fold and 10-fold Cross Validation. The work has achieved precisely 98.8% of F1 score and 99.9% of AUC for one digit classification without use of deep learning algorithm. When tested on a test dataset (unseen data), it shows accuracy of 99.74% after being tuned. Models are also compared according to run-time and resource utilization (memory and CPU) during training and prediction. Based on this comparison, suitable model can further be fine-tuned for implementation in small and medium-sized enterprises (SMEs).