Scaffolding type-2 classifier for incremental learning under concept drifts

The proposal of a meta-cognitive learning machine that embodies the three pillars of human learning: what-to-learn, how-to-learn, and when-to-learn, has enriched the landscape of evolving systems. The majority of meta-cognitive learning machines in the literature have not, however, characterized a plug and-play working principle, and thus require supplementary learning modules to be pre-or post-processed. In addition, they still rely on the type-1 neuron, which has problems of uncertainty. This paper proposes the Scaffolding Type-2 Classifier (ST2Class). ST2Class is a novel meta-cognitive scaffolding classifier that operates completely in local and incremental learning modes. It is built upon a multi variable interval type-2 Fuzzy Neural Network (FNN) which is driven by multivariate Gaussian function in the hidden layer and the non-linear wavelet polynomial in the output layer. The what-to-learn module is created by virtue of a novel active learning scenario termed the uncertainty measure; the how-to-learn module is based on the renowned Schema and Scaffolding theories; and the when-to-learn module uses a standard sample reserved strategy. The viability of ST2Class is numerically benchmarked against state-of-the-art classifiers in 12 data streams, and is statistically validated by thorough statistical tests, in which it achieves high accuracy while retaining low complexity. (C) 2016 Elsevier B.V. All rights reserved.