Ensemble Learning Approach on MNIST/Fashion MNIST Dataset Using Weighted Majority Voting

Authors : Buğra Hatipoğlu, Hüseyin Aydilek, Mustafa Yasin Erten, Fikret Yalçınkaya, Murat Lüy

Pages : 2291-2310

Doi:10.47495/okufbed.1635821

View : 91 | Download : 177

Publication Date : 2025-12-15

Article Type : Research Paper

Abstract :In deep learning and machine learning, sometimes no matter how complex we make the model, it may fail to achieve the desired results. In fact, an overly complex model can lose its generalization ability and face the problem of overfitting. Ensemble learning approaches, which require the collaboration of multiple models, enhance the generalization capability of the models and prevent overfitting. This study presents an ensemble model created by combining various deep learning models (ResNet, Inception, VGG16) and machine learning algorithms (kNN, Random Forest, SVM) using the weighted majority voting method. The weight of each model is automatically adjusted based on its performance on the training set. This approach has been observed to enhance the overall prediction performance by leveraging the strengths of the models while providing a structure resistant to overfitting. The presented ensemble model was evaluated using experiments on the MNIST handwritten digit dataset and achieved significantly better results compared to the individual models operating alone. In our dataset, which contains 70,000 digit images of 28x28 dimensions, the classification operations have improved the accuracy to as high as 97.88%, surpassing the performance of individual models.
Keywords : Topluluk Öğrenmesi, Ağırlıklı Çoğunluk Oylaması, Makine Öğrenmesi, Derin Öğrenme

ORIGINAL ARTICLE URL