Multi Agent-Learner based Online Feature Selection system

Online Feature Selection (OFS) is an important technique in pattern recognition and machine learning. Our challenge is how to enhance the classification performance in real contexts where the large-scale training data arrive sequentially with a big number of features. The major problem is how to choose the best accurate and efficient state-of-the art OFS method that can select the relevant features or if we do a combination between these methods can we improve the classification performance? In this paper, we propose a framework of OFS using the characteristics of multi-agent systems (MAS) to overcome this challenge. We propose firstly a new OFS model; Agent-Learner based OFS (ALOFS) which represents each agent in our MAS. ALOFS is a generalization of first-order and second-order online learning methods based feature selection. Secondly, we propose the Multi Agent-Learner based OFS (MALOFS) system which is our MAS. MALOFS uses two levels of selection: the first level aims to select the more confident learners and the second level has as object to select the relevant features using a proposed negotiation method (MANOFS). MALOFS is applicable to different domains successfully and achieves highly accuracy with some real world applications.