关键词: 稀疏重构权, 错误标注, 置信度, 检测

Abstract:

The accuracy of data classification depends on the quality and quantity of labeled data. When training data is mislabeled, data classification accuracy is greatly affected. In view of this situation, we propose a detection method based on the sparse reconstruction weights for erroneous labeling data. Firstly, we apply the k-nearest neighbor method to search their neighbor points for the training data that contains wrong labels. Each local sparse reconstruction weight can be calculated by solving the LS model with L1-norm. Secondly, we use parse reconstruction weights to calculate the label confidence level of every labeled data. Finally, by finding the position of the maximum curvature on the confidence curve, this method can adaptively detect the mislabeled data. Experiments on real data demonstrate that the proposed algorithm is effective.

Key words: sparse reconstruction weight, mislabeled, confidence level, detection