Research And Application Of FCM Initialization Method

The method of Fuzzy C-Means (Fuzzy ISODATA) has been the dominant approach in both theory and practical applications of fuzzy techniques to unsupervised classification. But essentially, it's based on the climbing hill algorithm. So the procedure is very sensitive to the clustering centers. The researches show the behavior of the FCM clustering depends on the quality of the initialization of the parameters strongly because of two fatal disadvantages. One is that the objective function of FCM is nonconvex function and there are many local extremal points. If the parameters are not initialized properly, FCM is likely not to converge to the global optimal solutions. The other is that if the data set is of large amount, the algorithm will waste a long time. Therefore, good initialization is expected. The closer the initial parameter gets to the centers, the more efficient the algorithm will be.By far, the mountain clustering and the subtractive clustering are two ideal methods for initialization, but they all have some disadvantages respectively. A new clustering algorithm, named quick subtractive clustering, is proposed. Based on the mountain clustering and the subtractive clustering, this algorithm remains these advantages of the two algorithms and gets more efficiency than them.However, the initialization method can not promise the number of clusters is the best. The research of clustering validity can solve the problem properly. So the author makes an effort to combine the initialization and the validity checks within the FCM clustering procedure and offered a FCM clustering model with subtractive clustering and validity checks. The simulations show the model can help the FCM clustering to get a better result.One of the developments of FCM is to clustering the time trajectory. Cross-Sectional Fuzzy Clustering Model is one of the significant productions of the research. However, this algorithm is confronted with the same initialization problem as the FCM. So a new clustering algorithm, named Cross-Sectional subtractive clustering, is proposed. This algorithm, which is based on the Cross-Sectional Fuzzy Clustering and the subtractive clustering, can finish the clustering procedure independently, and can be used as an initialization method...