Small-Sample Classification of Economic Impacts from Registered SMEs Using Machine Learning

Abstract

Registered Small and Medium Enterprises (SMEs) serve as a fundamental pillar of Thailand’s economy; however, modeling their economic impact is often constrained by the limitations of small-scale datasets. This study evaluates the performance of seven machine learning algorithms—Decision Tree, Support Vector Machine (SVM), Gradient Boosting, K-Nearest Neighbor (KNN), Naïve Bayes, XGBoost, and Multi-Layer Perceptron (MLP) Neural Network—in classifying the economic contributions of registered SMEs. To mitigate issues arising from class imbalance and limited sample size, the SMOTEENN (Synthetic Minority Over-sampling Technique and Edited Nearest Neighbors) hybrid sampling method was employed. The results indicate that while SMOTEENN significantly improved recall (reaching 1.000 for the majority of models), varying precision levels resulted in moderate F1 scores for the Decision Tree (0.83), SVM (0.87), and Gradient Boosting (0.88) classifiers. Notably, XGBoost and Naïve Bayes achieved peak performance with perfect scores (1.000) across accuracy, F1 score, and AUC metrics, demonstrating a robust capacity for pattern recognition within small-sample contexts, albeit with a recognized risk of overfitting. In contrast, the MLP Neural Network exhibited the lowest efficacy (Accuracy 0.778, AUC 0.438), suggesting that deep learning architectures may require more extensive hyperparameter tuning or larger datasets to remain competitive. These findings suggest that ensemble methods and probabilistic models are most effective for SME classification tasks under data constraints, providing a framework for more precise, data-driven economic policy-making.