Reweighting Semi-supervised Classification For Noisy Labels

This paper proposes a semi-supervised classification algorithm based on importance re-weighting for a two-class problem,where only a few data contains noisy labels.The Bayesian formula and unconstrained least squares fitting are used to estimate the noise rate.BP neural network is then used to solve the weighted optimization problem step by step.The experimental results on multiple benchmark sets show that the proposed method can reduce the impact on classification accuracy originated from the label insufficiency and noise.The structure of this paper is as follows: The first chapter is the introduction,which introduces the importance of noise data and semi-supervising and clarifies the main work of this paper.The second chapter is preliminary knowledge,which systematically summarizes the importance re-weighting method and conditional probability density estimation.The third chapter proposes our semi-supervised importance re-weighting algorithm.Experiments are proceeded in the fourth chapter,where we show that the proposed semi-supervised importance re-weighting algorithm has competitive classification accuracy and application potential.