Classification method for uncertain data based on sparse denoising autoencoder neural network

Abstract

Due to its importance in machine learning, pattern recognition, and many other applications, uncertain data mining has attracted much attention. This paper proposes a classification method for uncertain data based on a sparse de-noising auto-encoder neural network. Firstly, a hyperellipsoid convex model is used to describe the uncertain interval vector, and give an approach for uncertain data classification based on an interval uncertainty support vector machine. Secondly, this paper introduces a sparse de-noising auto encoder neural network, which can convert highdimension data into low-dimensional characteristic space. Finally, this paper establishes a three layered auto-encoder neural network, and whose deep structure is tuned with stochastic gradient descent parameters fine-tuned by layer greedy pre-training and back propagation. Experimental results show that this proposed method creats better classification accuracy, and has stronger robustness for noise parameter, so it is effective for uncertain data classification.