The proposal of dynamic thresholds in an immune algorithm for fuzzy clustering

Abstract

Most datasets obtained in real-world applications are typically unlabeled, requiring a manual labor of classifying a sample of such data or the application of unsupervised learning. Clustering is typically used to devise how data are grouped together before sampling the data to be labeled. Most clustering algorithms often assumes that the number of clusters is known and that a given instance from the dataset belongs to only one cluster. The Fainet algorithm is a bioinspired fuzzy clustering algorithm that finds fuzzy partitions and dynamically estimates the number of clusters. The results from the literature showed that, given a correct parameters set, this algorithm can outperform most clustering methods from the literature. However, in order to obtain such optimal set, a typical user should first acquire a knowledge of the dataset being studied. This work proposes dynamic rules to finetune the parameters set on-the-fly. The advantages of the proposed method is that the parameters not only adapts to the dataset characteristics but also to how close the solutions are from the optima. The results show that the method greatly improves the prototypes representativeness while optimizing the estimated number of clusters.