Paweł Zyblewski, Ph.D., from our Faculty won the Polish Artificial Intelligence Society award for the best Ph.D. Dissertation in Artificial Intelligence (11th Edition, 2021).

The jury appreciated his work titled Classifier selection for imbalanced data stream classification, supervised by Prof. Michał Woźniak from the Department of Systems and Computer Networks.

– The subject of my doctoral dissertation is closely related to the "Algorithms for classification of unbalanced data streams" project implemented at Wrocław University of Science and Technology and financed by the National Science Centre – says Dr. Paweł Zyblewski. – It was aimed at combining two important research trends in the field of pattern recognition and related to the analysis of the so-called difficult data, i.e. data stream classification and imbalanced data classification – he adds.

Imbalanced data is characterized by a different number of objects belonging to particular problem classes, thus making the classification task significantly more difficult when standard algorithms are used, as these tend to prioritize the majority class which is often less important.

In turn, data streams, prone to the so-called concept drift, require continuous improvement of the models used. The phenomenon of concept drift may lead to a shift in the decision boundary of the problem and, consequently, to the degeneration of the classifier's decision-making ability during its operation. These data disturbances, which are disadvantageous in practice, have a negative impact on the forecasting of many important economic processes, including supporting medical, logistic and consumer decisions.

Our researcher decided to focus on this issue as it has been so far addressed by a relatively small number of works, despite the fact that real data streams can often show a high and dynamically changing degree of imbalance. In effect, there is a growing demand for algorithms capable of dealing with these difficult conditions.

– To deal with these problems, I proposed methods based on dynamic classifier selection, which select the classifier ensemble depending on the value of the competence measures for a given object – explains Dr. Paweł Zyblewski. – These approaches have been supplemented with data pre-processing techniques aimed at balancing the class numbers of analysed problems in order to reduce bias in relation to the majority class – he adds.

The study was aimed at using the natural ability of classifier selection algorithms to deal with data imbalance and at proposing new, effective solutions to the problem of classifying highly imbalanced data streams, which is an issue rarely discussed in the literature.

Based on these assumptions, the dissertation formulated and substantiated the hypothesis that there are methods employing both data pre-processing and classifier selection methods that exceed the prediction quality of the difficult data classification methods known in the literature.

The doctoral dissertation is available at the Lower Silesian Digital Library.

Complete results of the contest can be found at the PSSI website.

Paweł Zyblewski, Ph.D. is currently one of the members of the team implementing the SWAROG project, the aim of which is to develop a system that detects sources of deliberate disinformation.

In the near future, he will also go on a one-month research internship to the University of Birmingham in Great Britain in order to cooperate with the team of Prof. Leandro Minku. The internship will concern the application of semi-supervised learning in the task of classifying data streams.