An Adaptive Framework for Classification of Concept Drift with Limited Supervision

This thesis deals with the problem of classification of data affected by concept drift. In particular, it investigates the area of unsupervised model updating in which a classification model is updated without using information about the changing distributions of the classes.
An adaptive framework that contains an ensemble of classifiers is developed. These can be mature or naive. In particular, only mature classifiers generate decisions, through majority voting, while naive classifiers are candidate to become mature.
The first novelty of the proposed framework is a technique of feedback that combines concepts from ensemble-learning with concepts from self-training. In particular, naive classifiers are trained using unlabelled data and labels generated by mature classifiers over that data, by means of voting. This technique allows updates of the model of the framework in absence of supervision, namely, without using the true classes of the data. The second novelty is a technique that infers the presence of concept drift by measuring the similarity between the decisions of mature classifiers and the decisions of naive classifiers. When concept drift is inferred, a naive classifier is selected to become mature, and a mature classifier is deleted.
A series of experiments are performed. They show that the framework can classify data with Gaussian distribution, and that this capability regards different classification techniques. The experiments also reveal that the framework cannot deal with the concept drift of a uniformly distributed dataset. Moreover, further experiments show that the inference of drift combines quick adapation with low false detections, thus leading to higher classification performance than comparative methods. However, this technique is not able to detect concept drift if the classes are separable.