Graph Traversal & Pathway Interrogation

AA large, densely cross-linked site has more potential connections between articles than any reader can hold in their head. Turning that into an actual graph, rather than leaving it as an implicit similarity table, lets well-understood graph algorithms do the work of finding a route. Shortest-path search doesn't just tell a reader that two articles are related. It gives them the cheapest sequence of intermediate stops and, via the bridge-term interrogation below, what each stop contributed to the journey. The trade-off was the one every road network faces. Capping degree and proximity kept the map legible, but it also meant the "true" nearest neighbour in high-dimensional embedding space was sometimes not directly reachable by a single road, and had to be approached by a short hop through a waystation instead — which is itself a more honest picture of how ideas actually connect than a single straight line would be.

The Local k-NN Spatial Highway Graph Ephemeral / ToolingKnowledge that evolves in months to a year — check for updates

EarlyEarly cartographic web prototypes constructed road lines based on global document hyperlinks or raw Euclidean distance across the map. Long diagonal lines across the canvas created visual clutter and obscured terrain features. To solve this, the SOM atlas built its road network as a Local k-Nearest-Neighbors (k-NN) Spatial Graph:

• Maximum Proximity Bound: Roads were only created between spatial neighbours within a radius of 7.5 hex grid units.
• Degree Capping: Each node was capped at a maximum of 3 outgoing road connections to keep highways clean and readable.
• Terrain Following: Road polylines followed low U-Matrix valleys, avoiding steep mountain ridges wherever possible.

Graph Pathfinding & Route Traversal Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

WhenWhen a reader hovered over a destination node on the map, the frontend pathfinder executed a Breadth-First Search (BFS) or Dijkstra algorithm[2] over the adjacency graph connecting "YOU ARE HERE" to the target node:BFS counts hops Dijkstra adds up weights

[YOU ARE HERE: Risk Management]
        │ (Road 1: Bridge Concepts: "Security", "Compliance")
        ▼
[Waystation 1: Legal Framework in Computing]
        │ (Road 2: Bridge Concepts: "Contracts", "Specification")
        ▼
[TARGET: Project Marking as BDD]

As the algorithm stepped through the road network graph, it accumulated intermediate node titles, road segment bridge terms, and target node concepts into a unified route summary shown in the inspection banner.

SOM Weight Vector Bridge Interrogation Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

BeyondBeyond simple graph traversal, the SOM atlas interrogated the high-dimensional weight vectors at the exact midpoint hex of every road segment. It compared the neuron vector $W_{\text{mid}}$ against the vocabulary term dictionary and extracted the top transitional terms that bridge two distinct articles:a hex with no page still has a weight vector

Interactive Road Polylines Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

InIn SVG rendering, thin 3px road polylines can be difficult to target with a mouse cursor. To ensure effortless user interaction, each road segment is rendered inside an invisible 16px transparent hit area group (<g class="atlas-road-group">). Hovering anywhere near a road line highlights the pathway and displays its bridge concepts in the inspection banner.

This pairing of graph search with mid-edge vector interrogation is what separated the SOM atlas from a conventional "related articles" widget. A similarity list can tell a reader that two pages are related. It can't tell them why, because the relationship exists only as a number, not as anything inspectable along the way. Treating each road as a real edge with a real midpoint — a point that itself sat somewhere in the underlying vector space, with its own nearest vocabulary terms — turned an opaque distance metric into a small, explicit vocabulary of transitional concepts a reader could read before committing to the jump. It was a deliberately cheap technique: no separate explanation-generation step, just a second lookup against the same weight space the layout was already built from. The live atlas replaced it with language-model-generated bridge text.cf. how bridge text is written now

Related Topics

• K-Blades vs. Self-Organizing Maps: How the Live Concept Atlas Works Now — the live atlas's current road network (a minimum spanning tree plus bridge edges) and its LLM-generated "Wayfarer Reading Guidance" text, which replaced this page's k-NN roads and weight-vector bridge terms.
• Triangulating in Complex Concept Spaces — the "where am I, what's nearby, how do I get there" wayfinding frame this pathfinding machinery exists to serve.
• Accessible Navigation — the non-visual fallback that has to state a road segment's proximity and bridge concepts as text rather than showing them spatially.

References

1. Hart, P. E., Nilsson, N. J., and Raphael, B., "A Formal Basis for the Heuristic Determination of Minimum Cost Paths," IEEE Transactions on Systems Science and Cybernetics, vol. 4, no. 2, pp. 100–107, 1968. https://doi.org/10.1109/TSSC.1968.300136
2. Dijkstra, E. W., "A note on two problems in connexion with graphs," Numerische Mathematik, vol. 1, pp. 269–271, 1959. https://doi.org/10.1007/BF01386390