K-Centroids Clustering: An Alternative Approach to K-Means Clustering

This text discusses k-centroids clustering as an alternative to k-means clustering. However, the approach proposed in the 2007 SCIENCE paper "Clustering by Passing Messages Between Data Points" differs significantly. This method employs message-passing algorithms where data points iteratively exchange responsibility and availability messages to identify cluster centroids. In implementation, each data point sends messages indicating its suitability as an exemplar (responsibility) and receives messages about other points' preferences (availability), effectively performing unsupervised clustering through distributed computation. The algorithm dynamically updates cluster assignments using matrix operations, typically implemented with similarity matrices and message-passing iterations until convergence. This approach not only handles high-dimensional data effectively but also accommodates complex data structures like images and text data by operating on pairwise similarity measures rather than Euclidean distances. Consequently, it demonstrates broad practical applications in pattern recognition and data mining scenarios where traditional centroid-based methods face limitations.