Semi-Supervised Clustering Based On Mean Field Annealing And Its Application

Clustering is the process of dividing objects into categories or clusters so that objectsfrom the same cluster are more similar to each other while objects from different clusters aremore different. Clustering has been widely used in many fields, including data mining, patternrecognition, image processing, marketing research, and so on. So far, clustering methods aremostly unsupervised, of which the results may be much different from the actual ones. Inrecent years, Semi-Supervised Learning, combining a small amount of labeled data with alarge number of unlabeled data, has become the hotspot of machine learning. Compared tounsupervised clustering, semi-supervised clustering can make effective use of unlabeled datato improve clustering performance.Semi-supervised clustering can make effective use of limited supervised information toimprove clustering performance. Many algorithms, such as those based on seeds set orauxiliary space are easily affected by initial points and noise. A semi-supervised clusteringalgorithm based on Mean Field Annealing is proposed. A simplified Silhouette criterion isused to provide unsupervised information and Rand index is used to provide supervisedinformation. They are combined together to form a new semi-supervised object function. Acontinuous valued Mean Field Annealing algorithm is applied to find out the global optimalsolution.The mean field annealing approach is a new neural network model. Because it usesdeterministic correction rules instead of random rules, it improves simulated annealingapproach greatly. Mean field annealing algorithm is capable of handling large combinatorialoptimization problems. Therefore, we apply it to our semi-supervised clustering problem.Experimental results on UCI Machine Learning Repository show that the algorithm caneffectively improve clustering accuracy.Image segmentation is an important issue in image processing. It has been widelyapplied in many research fields such as computer vision, pattern recognition and medicalimage processing. Clustering-based image segmentation is to convert segmentation problems into unsupervised classification problems and solve the problems in clustering ways. In imagesegmentation, if we can manually provide some pair-wise constraints, we will obtain a bettersegmentation result. That is, to solve the image segmentation problems in a semi-supervisedway. Our algorithm is applied to image segmentation problem. And experimental resultsprove the effectiveness of this algorithm.