Unsupervised and Semi-supervised Clustering: a Brief Survey

ia that provide significant distinctions between clustering methods and can help selecting appropriate candidate methods for one's problem: . Objective of clustering. Many methods aim at finding a single partition of the collection of items into clusters. However, obtaining a hierarchy of clusters can provide more flexibility and other methods rather focus on this. A partition of the data can be obtained from a hierarchy by cutting the tree of clusters at some level. . Nature of the data items. Most clustering methods were developed for numerical data, but some can deal with categorical data or with both numerical and categorical data. . Nature of the available information. Many methods rely on rich representations of the data (e.g. vectorial) that let one define prototypes, data distributions, multidimensional intervals, etc., beside computing (dis)similarities. Other methods only require the evaluation of pairwise (dis)similarities between data items; while imposing less restrict