The problem. Given a network — nodes and weighted edges — how do you find its communities: groups of nodes more connected to each other than to the rest? On large graphs this is computationally brutal, and most methods didn’t scale.

The idea. Louvain optimizes modularity (how much denser a group’s internal connections are than chance) with a greedy two-phase loop. First, move each node to the neighboring community that most increases modularity. Then collapse each community into a super-node and repeat on the smaller graph. Iterating this “unfolds” a hierarchy of communities in near-linear time, on graphs with millions of nodes.

Why it matters. This is the clustering engine under modern single-cell analysis, even though it was born in network science. Scanpy and Seurat build a k-nearest-neighbor graph of cells and run Louvain (or its successor Leiden) to call cell populations — the clusters I then annotate. Every single-cell figure I’ve read this fortnight has this algorithm somewhere in its methods. It’s another case of a general-purpose idea becoming genomics infrastructure.

Verdict. Foundational and fast, which is exactly why it spread. Its known flaw — it can produce poorly-connected or fragmented communities — is what motivated Leiden, later in this reading list. Read it as the origin of graph-based clustering, then read Leiden to see the fix.