Sparse data augmentation based on encoderforest for brain network classification

Abstract

Brain network classification has attracted increasing attention with the widespread application in the automatic diagnosis of brain diseases. However, limited by the higher cost of detecting and marking for medical imaging, the amount of brain network data is usually small, which largely restricts the performance of current brain network classification models. In this paper, we propose a new sparse data augmentation model (SDAM) based on EncoderForest to effectively enhance the brain network data and improve the classification performance. The EncoderForest based SDAM uses a generator which innovatively encodes the rules of a set of parallel decision trees to generate sparse data with only discriminative connections. The generated data expands the original data set effectively by utilizing the advantages of EncoderForest in learning data feature sparsely and constructing a feature association generation model compactly. In addition, the SDAM is flexible to combine with different classification models, such as random forest, support vector machine, deep neural network, etc. The experimental results on three common brain disease data sets show that our model is able to reasonably augment the brain network data and remarkably improve the performance of various classifiers.