Analysis
Use standard graph mode. Generate Graph first, then Run Analysis in the right panel. Metrics are computed on the server graph, then projected onto the display sample.
Explain next to each block calls the Workbench LLM with the current hubs / communities. It is commentary, not a substitute for the numeric job.
Centrality
Tick the measures you need. Defaults in the UI are all on.
| Measure | What it highlights |
|---|---|
| PageRank | Importance from quantity and quality of inbound links |
| Betweenness | Nodes that sit on many shortest paths (brokers) |
| Closeness | How few hops to everyone else |
| Degree | Raw connection count |
| Eigenvector | Connected-to-important-nodes |
| Katz | Attenuated walks of all lengths |
| Harmonic | Closeness that tolerates disconnected graphs |
| Eccentricity | Longest shortest-path from the node |
| Clustering coefficient | How clustered the node’s neighbors are |
After a run, set Size nodes by (often PageRank) and Color nodes by (often Community) on the canvas toolbar.
Workbench2 uses NetworkX for these scores (Python port of the former Java graphtheory API).
Community detection
Enable Community Detection, then choose:
| Algorithm | Notes |
|---|---|
| Louvain modularity | Default. Fast partitions; modularity is reported in global metrics |
| Girvan–Newman | Edge-betweenness splits; slower on large graphs |
Optional number of communities is a target for algorithms that accept a count; Louvain still follows modularity.
Coloring by Community is the usual way to read the result on the canvas.
Shortest path
Enable Shortest Path, enter source and target node ids, then run analysis (or the dedicated path call). The UI draws the path when it exists. Backend supports BFS / Dijkstra depending on weights.
Use Explain on the path block after ids are filled.
After a run
- Node details show per-node scores
- Results table lists ranked nodes
- Optional degree distribution chart in the analysis drawer
- Writeback (left panel) can merge scores into Mongo for later Runtime lookups or dashboards
Heavy algorithms may skip or simplify above ~100 000 edges; the hard processing ceiling is 2 000 000 edges.
Not the Runtime bandit
Dynamic Engagement Network Analysis (approach in campaign properties) ranks offers from an acceptance network at score time. That is documented under algorithms. This page is the Workbench graph lab.