Evaluation of Support Vector Machines SVM for Classification with Imbalanced Datasets

This paper presents the effect of unbalanced data sets on the training of classification models. For this purpose, sensor readings from a wall-following robot dataset available in Kaggle are used. The data was collected as the SCITOS G5 navigated the room using 24 ultrasonic sensors and following the wall clockwise, with 5,456 records distributed unbalanced into four classes. Training is performed by making a 70% to 30% split of the data for training and testing, initially using all the records. The data set is then balanced by sampling and equalizing the records by class. The models are trained with the same percentages for training and testing.