Clustering is an unsupervised technique that arranges data instances into groups according to inherent similarity, without any predefined labels or categories. Since the groupings emerge from the data itself, clustering is especially useful when the structure of the data is unknown. Streaming variants can update clusters as new instances arrive, so grouping can occur in real time, for instance when segmenting incoming customer behavior or detecting emerging patterns in sensor feeds. The question's twin requirements of grouping in real time and suitability for unknown data are both satisfied.
Classification, by contrast, is a supervised task. It assigns instances to categories that are already defined, and it must first be trained on labeled examples. When the classes in the data are not known beforehand, no labels exist from which to train, so classification is unsuitable.
Partitioning describes a family of clustering strategies, such as k-means, that divide data into a specified number of non-overlapping subsets. It is a method for forming clusters and not a separate algorithm category on the same level. It also generally requires the number of groups to be chosen in advance, which is a hindrance when the data is unfamiliar.
Clustering is therefore the algorithm described. Analysts often follow it with profiling of each group, and sometimes with classification once meaningful labels have been assigned.
Submit