LAUSR

The objective of this paper is to make an improvement on ensemble learning for imbalanced problem. Multi-matrices approach and nearest entropy are introduced into model of base classifier for the sake of utilizing spatial information of data and geometric relation between instances. Our method utilizes the variety of matrix to mine the potential information in the data and constructs regularization term that measures the neighboring relationship among instances with entropy to enhance the stability of decision boundary. The different shapes of matrix contain distinct spatial information. As a result, the origin vector-oriented data are reorganized into multiple shapes of matrix to expand the different spatial information. The nearest entropy is used to measure the local certainty of instances so that the stable instances can be selected to train by the new regularization term. In order to compare the advantages of introducing the multi-matrices and entropy, several ensemble learning methods that have similar ensemble strategy and variants of linear classification models are selected to implement experiments, based on 55 binary classification datasets of KEEL benchmark.