Research On Adversarial Manifold Propagation With Uncertain Pairwise Constraints

Pairwise constraints,as a type of supervised information,are more generalized than label information and are applied to solve various machine learning problems.However,the initial pairwise constraint relationships available are very limited.In order to obtain more pairwise constraint relationships,many pairwise constraint propagation(PCP)methods have emerged.In existing pairwise constraint propagation methods,it is usually studied how to increase the number of pairwise constraints through propagation learning based on the accurate and reliable initial obtained pairwise constraint relationships.However,with the rapid development of big data and artificial intelligence,the uncertainty of data information becomes more apparent in real scenarios.For pairwise constraints,in addition to precise information,there is often some uncertain information that is also of great significance for the propagation of pairwise constraints.Therefore,how to utilize these uncertain pairwise constraint information to improve the propagation effect of pairwise constraints is a problem that needs to be solved.In response to this issue,the research in this article is as follows:Firstly,in general,there are very few precise pairwise constraint relationships that can be obtained.For example,when the initial pairwise constraint information comes from user feedback,it is difficult to have a full grasp of the provided pairwise constraint information due to the lack of professional knowledge and comprehensive judgment methods.However,it often provides some possibility that two data belong to the same category.A pairwise constraint propagation with uncertain must-link(UMPCP)method is proposed to address the possibility of two sample categories being the same.The main idea is to use a matrix to add some possibilities about uncertain must-links based on the original precise initial pairwise constraints.This possibility is influenced by the adversarial effects of the cannot-link matrix during the propagation process.In addition,there is also an adversarial relationship between the must-link and the cannot-link in the propagation process.These two adversaries are combined to form a new adversarial strength.Thus,it acts on the propagation of manifold regularization with pairwise constraints.Through experiments,it has been proven that the possibility information of must-link effectively acts on the propagation process,improving the propagation effect of pairwise constraints while being more in line with real-world applications.It is worth noting that the addition of uncertain must-links only has a significant impact on the propagation effect of must-links.To improve the overall propagation effect of pairwise constraints,it is necessary to mine uncertain information more comprehensively.Therefore,we also focus on the possibility information of cannot-links and propose an uncertain pairwise constraint propagation(UPCP)method for uncertain pairwise constraints.In this method,by imitating the definition of must-link uncertainty in the UMPCP model,the possibility information of cannot-link is added,and the dual adversarial structure is transformed into a combination of the adversarial between two possibilities and the adversarial between two constraints,so as to minimize the intensity of must-link and cannot-link adversarial in competition.The experimental results show that the UPCP method improves propagation accuracy while overcoming the one-sidedness of UMPCP model in applying uncertain information.