Network Clustering Analytics

The question

In a dense network — think wireless sensors or IoT devices — 600 nodes drawn on a screen overlap into a blur. Can unsupervised learning find the structure that the eye can't? Built for an Artificial Intelligence module (CMPG 313).

How it works

• A realistic network — a stochastic block model with seven communities: dense connections inside each (probability 0.27–0.32), sparse ones between them (0.05).
• Layout — a spring layout, compressed per community, with small random jitter so groups touch instead of separating perfectly.
• Energy — each node gets a battery level from 10 to 100, deliberately independent of its position.
• K-Means (scikit-learn) groups the nodes by position, run with k = 7 and k = 3, ten initialisations and a fixed random seed so results reproduce.

What it found

• With k = 7, clusters of 75–95 nodes; with k = 3, about 200 each — the trade-off between detail and a high-level view.
• Low-energy nodes (under 20) appear in every cluster, so position alone can't find the nodes that need attention: energy needs its own analysis.
• Results are shown as 2-D network plots and an interpolated 3-D energy surface.