Ordered Incremental Attribute Learning based on mRMR and Neural Networks

Current feature reduction approaches such as feature selection and feature extraction are insufficient for dealing with high-dimensional pattern recognition problems when all features carry similar significance. An applicable method for coping with these problems is incremental attribute learning (IAL) which gradually imports pattern features in one or more size. Hence a new pre-processing called feature ordering should be introduced in pattern classification and regression, and the ordering of imported features should be calculated before recognition. In previous studies, the calculation of feature ordering is similar to wrapper methods. However, such a process is time-consuming in feature selection. In this paper, a substitute approach for feature ordering is presented, where feature ordering is ranked by some metrics such as redundancy and relevance using mRMR criteria. Based on ITID, a neural IAL model is derived. Experimental results verified that feature ordering derived by mRMR can not only save time, but also obtain the best classification rate compared with those in previous studies. In addition, it is also feasible to apply mRMR to calculate feature ordering for regression problems.