Community detection

Community detection finds groups, or communities, of nodes in a network that connect to each other more densely than they connect to nodes outside the group. In a social graph this might surface friend groups or interest clusters; in a transaction graph it might surface groups of accounts that transact mostly among themselves.

It is a form of clustering applied specifically to graph-structured data, using connection patterns rather than a table of feature values as the basis for grouping. Common algorithms include modularity-based methods, which search for a partition that maximizes the difference between actual and expected within-group connections, and label-propagation methods, which spread a group label along edges until the assignment stabilizes. This differs from centrality measures, which rank individual nodes; community detection instead partitions the whole graph into groups.

Community detection matters for use cases from marketing segmentation and social listening to fraud rings and organizational network analysis, wherever the structure of connections, not just individual attributes, defines a meaningful group. A key pitfall is that most algorithms are sensitive to a resolution parameter or similar setting, and the same network can be validly partitioned at different granularities — a handful of large communities or many small ones — with no single answer being objectively correct. link prediction is often run on the same graph to complement the partitions this produces.