An inductive non-negative projection semi-supervised data classification method and system

Abstract

The method of the present invention explicitly combines the semi-supervised data representation and classification error with the existing projective non-negative matrix factorization framework for joint minimization learning, thereby applying the weight coefficient construction and label propagation process to the projective non-negative matrix factorization, which can effectively Avoid the negative impact of noise, corruption or heterogeneity that may be contained in the original data on the similarity measure and label prediction results. In addition, the above joint minimization process can also effectively preserve the neighborhood information and spatial structure during the projected non-negative matrix factorization process, resulting in more accurate data representation results. In addition, the weight construction and inductive learning are integrated into a unified model, and an adaptive weight coefficient matrix can be obtained, thereby avoiding the difficulty of selecting the best neighbor in traditional algorithms. The method of the invention is an inductive model, which can complete the induction and prediction of out-of-sample data without introducing an additional reconstruction process, and has good scalability.

Description

technical field [0001] The invention relates to an inductive non-negative projection semi-supervised data classification method and system, belonging to the technical fields of pattern recognition and data mining. Background technique [0002] Graph-based semi-supervised learning has been an important topic in the fields of data mining and pattern recognition. Since the model can learn from a small amount of labeled data and a large amount of unlabeled data, it is very suitable for the characteristics of practical application data. Existing models can be broadly categorized into transductive learning and inductive learning based on whether they can be efficiently extended to out-of-sample new data. [0003] Label propagation, as a typical classification model, has attracted considerable attention and interest in academia in recent years. Existing typical transductive label propagation algorithms include Gaussian field and harmonic functions, local and global consensus lear...

AI Technical Summary

Problems solved by technology

Although the existing inductive label propagation methods effectively solve the out-of-sample problem through embedding, there are still obvious shortcomings. Most of the data in practical applications usually contain redundancy or various noises, which have a great impact on classification performance.

Images