Social network analysis based novel visualizable outlier detection, feature selection and supervised learning methods

Supervised machine learning techniques, which help human decision-making, have become widespread today with the development of data collection devices, the increase in the amount of data and the ease of access to data. Although machine learning models show an increasing success rate, the intelligibility of these models and the interpretability of their classification predictions are declining. Visual models are models suitable for human intuitive understanding. Graphs as a form of visual expression objects, have been able to find application in various ways in machine learning. Within the scope of this thesis various learning techniques suitable for visual interpretation have been developed by first visualization machine learning (training) data through conversion of it to network diagrams, and then processing the resulting graphs through social network analysis and graph theory techniques. The first of these techniques is an outlier data detection system, second one is a feature extraction and selection model, the third one is a visual categorical learner model. These proposed models are applied on various problems in machine learning classification task, and performances are shown in comparison with competing methods. In addition, the contributions of the developed models to the intelligibility of machine learning systems are also discussed within the scope of the thesis. Experimental results show that these proposed systems can be evaluated as new methods that can be preferred in machine learning by both performance and intelligibility criteria.