Evaluation of the bias effect in feature selection methods in terms of classification accuracy

In today's world, data is everywhere, abundant and easily accessible. Data is abundant, easy to obtain, but due to its continuously increasing structure, it is becoming increasingly difficult to process and make it meaningful. Especially in big data studies, image-based studies, data stream-based studies, it is important to reduce attributes into meaningful subsets. Without feature selection, processing power is used intensively, the training time of classifiers is prolonged and this makes some data unprocessable. Feature selection in machine learning is a field of study that has received much attention. Feature selection aims to avoid the curse of dimensionality by reducing the number of features in the data. For this purpose, it examines the data with various approaches and selects the most meaningful attributes using various decision-making mechanisms. When selecting attributes, a bias effect may occur due to the inherent effects of the data. The bias effect negatively affects attribute selection. One of the important issues in feature selection is that the iterative change of the training, validation and test set we use can introduce bias effects. The relationship between the number of samples and the variation of the selected attributes is also an important issue. When we select attributes when the number of instances is large, we expect a similar subset of attributes to be selected each time. In order to avoid the bias effect, using various cross validation methods is a positive way to reduce the effect. Using different cross validation methods for feature selection gives us different similarity metric ratios and gives us information about which method has less bias effect. Research on this topic continues with intense interest. In this thesis, three different datasets and three different feature selection methods were used for feature selection. These feature selection methods were also run with three different cross validation methods and three different classifiers. In this way, eighty-one different runs were performed. Bias effect was observed by using two different similarity metrics for the studies. According to the results obtained, the cross validation method with the least bias effect was determined independently of the dataset and feature selection method.