Identifying and remove the mislabeled training sample of ECG signals using ensemble learning and iterative

摘要

The mislabeled samples in the Electrocardiography(ECG) training set can degrade the classifier's classification accuracy. To mitigate this adverse effect, the ensemble learning and iterative is introduced to identify and remove the mislabeled samples. The basic principle of the proposed method is to form a strong learner through multiple weak learners to make it less affected by noise samples. So the mislabeled samples can be found more accurately. Due to the influence of mislabeled samples in the sub training set, only a small number of mislabeled samples can be detected by only one-time prediction. So we use an iterative process to detect more noise samples and get a cleaner training set. The new training set after denoising is used to train the classifier to obtain higher classification accuracy. Experiments show the effectiveness of this method.