medoidbased

Medoidbased refers to clustering and related data analysis approaches that use medoids as the central exemplars of clusters, rather than centroids. A medoid is the actual data point within a cluster that minimizes the average dissimilarity to all other points in that cluster. Unlike a centroid, which may lie between observations or be a synthetic mean, a medoid represents a real observed item.

Medoidbased methods are distinguished by their reliance on distance or dissimilarity measures rather than vector means. This makes them robust to outliers and applicable to a wider range of data types, including non-Euclidean spaces and mixed or categorical data where a mean is not meaningful. The objective in medoidbased clustering is typically to minimize the sum of dissimilarities between each point and its assigned medoid.

The most well-known example is k-medoids clustering, including algorithms such as Partitioning Around Medoids (PAM) and its scalable variants CLARA and CLARANS. PAM assigns points to the nearest medoid and then updates medoids by swapping nonmedoid points to reduce the total cost, iterating until convergence. CLARA and CLARANS extend the approach to larger datasets through sampling and randomized search, respectively. These methods are characterized by their emphasis on representative real observations and their flexibility with respect to the distance metric used.

Applications of medoidbased clustering include scenarios where data are non-numeric or contain outliers, where interpretability of cluster centers as actual items is valuable, and where robustness to unusual observations is important. Limitations include potential lower efficiency on very large datasets and sensitivity to the initial choice of medoids.