Abstract

Incremental Attribute Learning (IAL) is a novel supervised machine learning approach, which sequentially trains features one by one. Thus feature ordering is very important to IAL. Previous studies on feature ordering only concentrated on the contribution of each feature to different outputs. However, besides contribution, correlations among input features and output categories are also very important to the final classification result, which has not yet been researched in feature ordering but has confirmed in multivariate statistics. This study aims to find out the relations between feature ordering and feature correlations. This paper presents a new method for feature ordering calculation which is based on correlations between input features and outputs. Experimental results confirm that correlation-based feature ordering can produce better classification results than contribution-based approaches, feature orderings with theoriginal sequence sorted in the database, and conventional methods where all features are trained in one batch.