Developing machine learning methods for network anomaly detection

Abstract

Machine learning refers to training of a computer (machine) to be able to acquire knowledge from data (i.e. experience) and improve itself on a given task. The field of machine learning has become a mainstream, improving hundreds of millions of lives. Fraudulent actions in computer networks, credit card transactions and website advertisement traffic might devastate large businesses and cause anually fiscal loss of billions of dollars around the globe. In this thesis, we propose various machine learning methods for three fraud detection problems: network anomaly detection, credit card fraud detection and detection of fraudulent clicks to advertisements on the internet. We design various classifiers such as logistic regression, k-nearest neighbors, decision tree, support vector machine, and ensemble classifiers such as random forest, bagging, stacking and AdaBoost. The hyper-parameters of the classifiers are optimized by performing cross-validation experiments on train sets. In the next step, the models are trained using the optimum hyper-parameter configurations and predictions are computed on test sets. Among the various methods compared the highest accuracy is obtained by ensemble learners.