Classification method for sequencing and transferring by utilizing privilege information

Abstract

The invention discloses a classification method for sorting and transferring by utilizing privilege information. Considering the condition that positive labeled samples, unlabeled samples and privilege information of the positive labeled samples and the unlabeled samples exist at the same time, positive samples PS, negative samples NS, the unlabeled samples and similar weights of the unlabeled samples are combined into a learning model of a sorting support vector machine to obtain an extended first sorting support vector machine model, and the extended first sorting support vector machine model is trained on a training set only with the privilege information. The actual distance between common training samples (not containing privilege information) by utilizing a classifier trained is calculated by the privilege information. Finally, the constant interval 1 is replaced with the actual distance, and a second sorting support vector machine model is trained on the common training sample set. The actual sample spacing calculated by using privilege information is more accurate than the constant spacing 1, and the classifier can be more accurate by using the actual spacing to learn.

Description

technical field [0001] The invention relates to the technical field of machine learning, in particular to a classification method for sorting and transferring by using privileged information. Background technique [0002] In traditional supervised learning, only labeled training samples are learned, so as to build a model for predicting the labels of unknown samples. With the rapid development of data collection and storage technology, it is quite easy to collect a large number of unlabeled samples, and it is quite difficult to obtain a large number of labeled samples because obtaining these labels can be resource-intensive. For example, mark exceptions. Since unlabeled samples are readily available, learning on both positively labeled and unlabeled data (PU learning) has attracted much attention. There are many studies on PU learning in existing works, which show that unlabeled samples are more likely to be located near the decision boundary and play a crucial role in the...

AI Technical Summary

Problems solved by technology

[0005] For PU learning problems, when the quality of negative samples extracted from unlabeled samples is not good, the accuracy of the classifier is often unsatisfactory. The purpose of the present invention is to overcome this deficiency and provide a sorting transfer method using The classification method combines positive samples PS, negative samples NS, unlabeled samples and their similarity weights into the learning model of sorting support vector machine, and then in the training set with only privileged information (x * , y) for training, use the privileged information to explore the actual spacing of the training samples, and then use the actual spacing to replace the constant spacing 1, and train on the training set (x, y)

Images