Probability κ-means clustering for neural network architecture

Cluster analysis indices are aimed at classifying the elements used to estimate the quality of the categories based on their similarity. It is a challenging task because, with the same data set, there may be many partitions that fit natural groupings of a given data set. Applications of clustering can include pattern recognition, image analysis and information retrieval. We propose an approach based on the κ-means concept that clustering centers more often have a higher density than their neighbors: then we use a probability κ-means algorithm to achieve fuzzy clustering in continuous form over a relatively large distance from other points with higher densities. Further, in order to follow the mainstream neural network architecture, we define a favorable activation function and corresponding loss function for the clustering iteration. Our method come from the basis of a clustering procedure in which the number of clusters arises intuitively and clusters are achieved regardless of the high dimensions. We demonstrate the result of complete algorithm on several clustering test cases.