Research On Fuzzy Clustering Algorithm With Supervision

Learning algorithm includes two types, which are supervised and unsupervised algorithm. It has an important role in pattern recognition theory, and has been widely applied in many fields, such as computer vision, data mining, target recognition and machine learning. The supervised learning algorithm employs the labeled samples to learn recognition system, so that the identification system can classify and discriminate the unknown samples. Unsupervised algorithm reveals the natural attributes and the structure of the sample set which is based on some similarity between samples as the basis for classification, Traditional fuzzy C-means (FCM, Fuzzy c-Means) algorithm is one of unsupervised clustering methods, and it has a wide range of application and attention because it could be converted into an optimization problem and designed easily. In this paper, we employ feature transformation and absorption technology, extend the traditional unsupervised FCM algorithm to supervised learning, and establish the supervised FCM algorithm, which not only can reveal the internal structure and the law of sample data, but also realize supervised learning and testing.Whether clustering is effective or not has two principles: (1) small distance between inner classes; (2) large distance between inter-classed. Current self-learning algorithms classify and recognition patterns by the principle with small distance between inner classes, which do not consider that degree between the two different centers couble be enough small and it is sometimes large. A feature tranformation algorithm is presented in the paper. And it makes the degree of different centers approximately close to 0, enlarges the distance between classes, and thus effectively improves the performance of the FCM algorithm.Considering that the density of training sample is very large and there is flaw of information, a feature weighted FCM algorithm based on absorption is proposed, which can be sparse on the training smple through cluster validity function, and expected that flaw of sample set is overcomed to some degree. Additionally, the algorithm also cuts the training samples.As some isolated samples will be produced in the learning process of the supervised FCM algorithm, which is deviated from the other samples, there will be a bad bargin and result in over-fitting phenomenon if we emphasis their roles. If we play down their roles or remove these samples, there will be large "blind spot" when the algorithm is used to identify test samples in future. That is because the isolated samples have more important information, though they are alone. Similar to the role of absorb, we introduce feature transformance method according to the characteristics of"univerisal gravity", and establish absorption based on univerisal gravity and absorb these isolated samples through classified data set.Training samples have some flaws and paradoxical phenomenon. Existing learning algorithms consider the training samples as correct by default. The paper utlizes the FCM algorithm based on weighted feature to train smples at different levels and establish three types of samples: homogeneous cluster, homogeneous cluster with few samples and heterogeneous cluster with few samples, determine the best training class number according to the cluster validity function. Finally, the latter two types of cluster are processed through the feature transformation, and cluster centers are obtained. Then unknown samples are identified through these cluster centers and centers of homogeneous cluster. Compared to the traditional self-learning system, the algorithm presented in the paper is more practical due to increasing heterogeneous clusters, and experimental results show that the algorithm is feasible.