Ensemble self-training algorithm based on nearest neighbor density and semi-supervised k nearest neighbor (SSKNN)

Abstract

The invention discloses an ensemble self-training algorithm based on nearest neighbor density and semi-supervised k nearest neighbor (SSKNN). A nearest neighbor density algorithm is used to select initializing labeled samples, adding k nearest neighbor samples around the labeled sample into a labeled candidate set is avoided, the initializing labeled samples are enabled to be as decentralized as possible to better reflect an original spatial structure of the samples, and at the same time, the samples with highest densities are selected, in the labeled sample candidate set, as the labeled samples. In order to improve the performance of data editing, the SSKNN is used to replace WKNN (weighted k nearest neighbor), the problem that when the WKNN is used to perform data editing, only influences of the labeled samples on to-be-measured sample categories are taken into account, and unlabeled samples around to-be-measured samples are not utilized is offset, and comparison experiment on UCI datasets verifies the validity of the provided algorithm.

Description

technical field [0001] The invention relates to the field of computer information technology, in particular to an integrated self-training method based on neighbor density and semi-supervised KNN. Background technique [0002] Integrated self-training algorithm [1] is integrated learning [2] and self-training [3，4] combined semi-supervised learning [5] frame. Compared with other semi-supervised learning methods, it does not require harsh assumptions, so it is favored by many scholars, but how to select reliable samples to add to the training set in the integrated self-training algorithm has always been a hot issue in semi-supervised learning. . [0003] Some scholars select reliable samples by taking the method of confidence, MFA Hady [6] A Co-Trainingby Committee integrated self-training learning framework is proposed. This method integrates multiple classifiers for self-training learning, where the confidence is the average posterior probability of multiple classifie...

Images