Experimental evaluation of online classification models
Πειραματική αποτίμηση μοντέλων άμεσης κατηγοριοποίησης
View/ Open
Keywords
Machine learning ; Online learning ; Experimental evaluation ; Linear methods ; Kernel methods ; SVM ; Perceptron ; Logistic regression ; WinnowAbstract
In this thesis we compare experimentally several online machine learning methods, for classification. The term "online" refers to the fact that the classifier examines the training examples one by one and not as a whole dataset, and additionally the classifier's decision is irrevocable. We evaluate their performance on three different datasets. Specifically, the online learning methods used are Perceptron, Support Vector Machines (SVMs), Winnow, Logistic Regression (LR), along with kernel methods for Perceptron and SVM. The datasets used cover real-life application fields. The first dataset is used to predict spam e-mails, the second for cancellation of shipment and the third for constant surveillance of people suffering epilepsy. After transforming the datasets to the desired form, we implement the said methods and algorithms from scratch using Python, evaluate the models created and present our results. Our findings are summarized as follows. First, the kernel methods do not outperform the non-kernel ones, due to the lack of quantity of training examples. Second, the Winnow performance was poor, probably due to its lack of negative correlation. Finally, we find that the deviation between the models we created, and the corresponding static models created by Python's sk-learn, was insubstantial, therefore proving our models' performance was satisfactory.