An incremental self-trained ensemble algorithm

Incremental learning has boosted the speed of Data Mining algorithms without sacrificing much, or sometimes none, predictive accuracy. Instead, by saving computational resources, combination of such kind of algorithms with iterative procedures that improve the learned hypothesis utilizing vast amounts of available unlabeled data could be achieved efficiently, in contrast to supervised scenario where all this information is rejected because no exploitation mechanism exists. The scope of this work is to examine the ability of a learning scheme that operates under shortage of labeled data for classification tasks, based on an incrementally updated ensemble algorithm. Comparisons against 30 state-of-the art Semi-supervised methods over 50 publicly available datasets are provided, supporting our assumptions about the learning quality of the proposed algorithm.

I. Introduction

Despite the fact that the power of cutting edge computer systems has been increased all these years with a stable trend, the corresponding increase rate of the volume of produced data is dramatically sharper, leading to vast amounts of datasets while, simultaneously, rising issues of efficient data manipulation. Taking also into consideration the complex cases of “concept drift” [1] – where the inputs to general Machine Learning (ML) recognition systems change dynamically - and fast evolving systems [2] which occur mainly due to the nature of the tackled application and the aspects of online learning - it is easily perceived that more and more applications need to be satisfied under both time and capacity restrictions, avoiding to mention big data era. Some usual ways of coping with such kind of tasks are: i) to increase the computational power of employed working systems [3], ii) design more efficient algorithms that could operate under such environments [4] and iii) convert effective algorithms to parallel versions [5].