Growing Neural Gas — a graph that grows into the cloud
Growing Neural Gas starts with two units and, as rows are presented, moves the nearest unit toward the row and its graph neighbors a smaller step. Edges that stop winning are aged out. Every few steps a new unit is inserted near the unit that has accumulated the most error. The graph, not a set of centroids, is the model: clusters are the connected components that remain.
A budget of 10 units on scaled Iris produced 10 components in the published run, silhouette 0.299, adjusted Rand 0.315 against species. Ten units on 150 rows is enough to tile the cloud rather than to recover three species. The node budget is the resolution. A budget of 6 on the same table is a coarser graph; 16 starts to trace local wiggles.
The result depends on presentation order. The seed fixes that order. Without it, a second run grows a different graph.
Settings that change this fit
- max nodes
- Cap on the graph. This is the resolution knob. It is not a cluster count; several units can end up in one component.
- max age
- How long an edge can go without winning before it is deleted. Shorter ages break the graph into more components.