A label query and change method based on active learning

Abstract

The invention relates to a label query and change method based on active learning, belonging to the active learning field. The method comprises the following steps: firstly, training a plurality of classification models by using a 10-fold cross-validation method and a strong-weak classification algorithm; Then calculating the residual matrix and confidence matrix; Secondly, splicing the two new matrices into one matrix horizontally as the eigenmatrix. using the K-Means algorithm to cluster the generated feature matrices, and obtaining four types of clusters A, B, C and D according to the threshold; Finally, relabeling some samples by active learning method for the suspected noise clusters (type A and type C), and obtaining the final noise clusters according to the labeling results. The sample labels of noise clusters are modified to correct labels, and the modified A, C samples and B, D samples are merged into the final training sample set. The invention has excellent effect of screening out label noise, can reduce the cost of manual labeling, and can obtain higher classification accuracy rate with very small active learning cost.

Description

technical field [0001] The invention belongs to the field of active learning and relates to a label noise query system based on active learning. Background technique [0002] In traditional supervised learning, the learning algorithm trains the model based on the labeled samples as the training set. Generally speaking, the more samples in the training set and the higher the quality, the stronger the classification ability and generalization ability of the obtained model. However, in real tasks, a large number of samples are unlabeled, and it is relatively difficult to obtain labeled samples. It usually requires domain experts to manually label, so it requires high labor costs. The emergence of the crowdsourcing model has solved the problem of data labeling to a certain extent. Witkey platforms such as Amazon Mechanical Turk allow task publishers to publish various tasks (such as data annotation, filling out questionnaires, etc.) on the platform, and attract users from all o...

AI Technical Summary

Problems solved by technology

With the continuous development of crowdsourcing technology, the cost of obtaining labels is gradually reduced. However, due to the inability to effectively control the labeling process, it is inevitable that the obtained labels will be mixed with several noise labels, resulting in a decline in the quality of the training set samples. affect the learning effect

Images