Research On Semi-supervised Learning And Its Application

With the development of computer technology, the ability of data collection and storage has been improved greately. Not only in science research, but also in daily life plenty of data has been gained. How to analyse these data got and mine the useful information hidden in the data, which is a critical requirement in every field. In the traditional machine learning, just the unlabeled data or the labeled data were considered. Both of them are existing in many instances, how to use the combined informations, then semi-supervised learning is proposed.Semi-supervised learning is one of the most important research fields in pattern recognition and machine learning. It is widely applied in classification and clustering. In this paper only semi-supervised clustering is mentioned.In this paper, firstly, the study status on semi-supervised clustering is reviewed. Secondly, we introduce the theory of the non-supervised clustering and supervised clustering, and get the reasons why semi-supervised learning is widely attended. We also introduce the general method and algorithm in semi-supervised clustering. Lastly, we present what we have done about semi-supervised clustering which can be generalized two things.(1) We propose a modified differential evolution algorithm for semi-supervised fuzzy clustering. On the bases of traditional fcm algorithm and evolutionary algorithm, we introduce inertia-weighted coefficient by considering inertia-weighted idea of particle swarm algorithm, which keeps diversity of individual at early stages and quickens convergent speed at later stages, and at the same time improves the performance of the algorithm. Experimental results for remote sensing data indicate the efficiency.(2) We propose a semi-supervised clustering algorithm based on modified pairwise constraints. We adjust the old few pairwise constraints to get more information at first, then utilize new supervision to integrate dimensional reduction. We use pairwise constraints k-means algorithm to cluster in the subset in which the closures are changed by closures center. This new algorithm solves the problem of violating pairwise constraints, also improves the performance of clustering. The feasibility is proved on UCI database.