Linear Feature Sensibility for Output Partitioning in Ordered Neural Incremental Attribute Learning

Abstract

Feature Ordering is a special training preprocessing for Incremental Attribute Learning IAL, where features are trained one after another. Since most feature ordering calculation methods, compute feature ordering in one batch, no matter, this study presents a novel approach combining input feature ordered training and output partitioning for IAL to compute feature ordering with considering whether the output of the classification problem is univariate or multivariate. New metric called feature's Single Sensibility SS is proposed to individually calculate features' discrimination ability for each output. Finally, experimental benchmark results based on neural networks in IAL show that SS is applicable to calculates feature's discrimination ability. Furthermore, combined output partitioning can also improve further the final classification performance effectively.