Achieving semi-supervised incremental learning with Learn++ and simple recycled selection
Frederico Damasceno Bortoloti, Patrick Marques Ciarelli · 2016
In many real-world tasks a lot of unlabeled data are collected over time and, although they may be useful to improve the quality of classification models, they are usually ignored. Semi-supervised learning techniques combine unlabeled and labeled data to capture more useful information about a particular task. On the other hand, an incremental learning technique can incorporate new information to an existing model, so that it can dynamically adapt its structure to follow the environment changes. In order to unify the characteristics of both approaches, in this paper is proposed an incremental semi-supervised learning method called SSLearn++, which is based on the techniques Simple Recycled Selection (SRS) (semi-supervised learning), and Learn++ (incremental learning). To treat multi-class problems, SRS was combined with Sequential Forward Selection (SFS) to generate cascades of binary classifiers. Experiments were conducted using publicly available benchmark data sets. Results show the proposed method is promising.