Rejection-oriented learning without complete class information

Abstract

Machine Learning is commonly used to support decision-making in numerous, diverse contexts. Its usefulness in this regard is unquestionable: there are complex systems built on the top of machine learning techniques whose descriptive and predictive capabilities go far beyond those of human beings. However, these systems still have limitations, whose analysis enable to estimate their applicability and confidence in various cases. This is interesting considering that abstention from the provision of a response is preferable to make a mistake in doing so. In the context of classification-like tasks, the indication of such inconclusive output is called rejection. The research which culminated in this thesis led to the conception, implementation and evaluation of rejection-oriented learning systems for two distinct tasks: open set recognition and data stream clustering. These system were derived from WiSARD artificial neural network, which had rejection modelling incorporated into its functioning. This text details and discuss such realizations. It also presents experimental results which allow assess the scientific and practical importance of the proposed state-of-the-art methodology.