Analyzing Bias in Machine Learning Models: Insights from the AlexNet Architecture on Balanced and Imbalanced Data

Machine learning is being applied in almost every field nowadays. Considering its abilities, this wide usage and rapid extension so many domains is not surprising. Machine learning and artificial intelligence applications, especially deep learning models are developing rapidly and starting to play an effective role in different fields. This wide usage of these methods in various domains gives rise to some concerns and problems along with positive effects on our lives. With the beginning of its use in selection and decision-making mechanisms in social life, some concerns related to its fair and equal work have emerged. The fairness of the algorithms and data used in this field has started to be the subject of some controversy. With this study, the relationship of bias with the cases of balanced and unbalanced data sets, one of the characteristics of data sets, has been considered. In order to create a test environment suitable for the study, MNIST data sets were used as a balanced dataset and MNIST data sets as an unbalanced data set. The AlexNet algorithm was used to work with these data sets and the obtained mathematical results were examined. The recognition performance of AlexNet for each class in these datasets is evaluated and its behavior is analyzed in detail. As a result of these studies, apart from the data sets used, it has been understood that the only reason for bias formation in a model is not only the datasets but also the algorithm used can cause bias formation. Finally, it is observed that additional to the data imbalance, the dataset content and the complexity of the classes should be considered carefully to be able to avoid bias in machine learning methods. It has been determined that the compliance of the learning phase with the samples in the dataset plays an important role in preventing bias.