Reasoning & analytics

TL;DR

The payoff for all that construction: once knowledge is a graph, whole families of questions become computable that tables and documents can't express at any price. Five families cover nearly everything on a product roadmap. Traversal follows chains: exposure, lineage, "how are these two connected?" Centrality finds which nodes matter most — the PageRank family. Community detection finds which nodes cluster — fraud rings, customer segments, duplicate candidates. Link prediction finds which edges are missing or likely next — recommendations, next-best-action. Rule-based inference derives facts that follow logically from stated facts — "subsidiary of a sanctioned entity is sanctioned-adjacent." Each family maps to a product capability you can name in a roadmap review, and each has a different trust profile. Traversal and inference produce explainable answers you can show a regulator. Predictions produce probabilistic ones you should ship as suggestions, not facts. Knowing which family a proposed feature sits in tells you its cost, its latency, and how much explanation your UX owes the user.

🎯 For the product leader

Why it matters — "Graph analytics" sounds like a data-science luxury. It's actually the menu of what your graph investment can ship: the fraud feature, the recommendation engine, the risk-exposure dashboard are all entries on this menu. If you can't name the families, you can't spot which roadmap items the graph makes cheap. You also can't be appropriately suspicious when a pitch calls a prediction a fact.

What it changes in your decisions — Roadmap items get tagged with their family, which sets three things at once: compute pattern (real-time query vs. batch job), explainability (deterministic vs. probabilistic), and eval burden (correctness vs. precision/recall on a golden set).

Ask yourself — "For this graph-powered feature: is the answer traversed, inferred, or predicted — and does the UX honestly reflect which one it is?"

Risk if ignored — A link prediction ships rendered like a stored fact. A customer asks "why does your product say we work with X?" The honest answer — "an algorithm guessed" — lands in a complaint, or a headline.

The five families, as product capabilities

Traversal is the workhorse — the multi-hop questions from lesson 1, plus the underrated pathfinding form: "show every chain connecting this customer to this sanctioned entity." The answer is the path itself — evidence you can put in front of an auditor. It runs live, in the request path, if depth-capped.

Centrality ranks importance by structure. PageRank and its relatives find the supplier whose failure cascades furthest, the person the org actually routes through (rarely the org chart's answer), and the account whose loss would unravel a network. This is batch computation, refreshed on a schedule, and consumed as a score.

Community detection finds dense clusters no one labeled: accounts that share devices, addresses, and payment methods (a fraud ring, structurally, before any individual account misbehaves — the reason graph analytics is standard kit in financial crime), customers that cluster by actual usage rather than firmographics, and — a nice dogfooding touch — probable duplicates the resolution pipeline missed.

Link prediction guesses edges that are missing or coming: customers who look like this one also need that (recommendations), this person likely knows that person, this record is probably the same entity as that one. The modern machinery is graph embeddings — compressing each node's neighborhood into a vector so "structurally similar" becomes computable — and increasingly graph neural networks. It is powerful, and probabilistic to the bone: these are suggestions wearing math.

Rule-based inference derives facts logically: part-of chains roll up, ownership percolates through corporate trees, "handles EU personal data" propagates to every system downstream of one that does. It is deterministic and auditable, and it is the reason formal ontologies exist. In regulated domains, this family is the product.

The trust gradient

The families differ most in what kind of answer they produce — and your UX, evals, and compliance posture must track it:

The gradient also sets where mistakes hurt. A wrong stored fact corrupts every family downstream — this is why construction quality outranks algorithm choice. A wrong prediction, honestly framed as a suggestion, costs a shrug. The catastrophic combination is a prediction stored back into the graph as a fact: the guess launders itself into evidence, and six months later nobody remembers it was a guess. If predictions must be persisted, they must carry their confidence and provenance forever.

Operational shape

Two compute patterns, two cost profiles. Query-time work (traversal, inference on demand) lives in the request path: milliseconds, depth caps, timeouts. Feature latency is graph latency. Batch work (centrality, communities, embeddings, prediction scoring) runs on schedules and ships scores. It's cheap to serve, but stale by design — the fraud ring detected nightly is invisible for up to 24 hours, and whether that's fine is a product decision, not an infrastructure one. Most platforms bundle these algorithms into libraries, so the engineering lift is usually integration, not invention. The real work is choosing thresholds and owning what the scores mean.

Failure modes

• Predictions cosplaying as facts — the link-prediction output rendered in the same UI as curated knowledge. Trust in the whole graph dies with the first bad guess.
• Garbage in, insight out — centrality and community algorithms run happily on a graph full of false merges. The "key supplier" is two companies fused by a resolution bug.
• Celebrity-node blowups — the mega-hub (the bank everyone uses, the component in everything) makes traversals explode and communities meaningless. Dense nodes need modeling care, caps, or exclusion rules.
• Unstable scores presented as truth — community boundaries shift run to run. Yesterday's "Segment 7" is today's "Segment 12," and downstream teams built process on the labels.
• Batch cadence mismatch — a real-time decision (payment approval) fed by a nightly score. The ring that formed at 9 a.m. clears payments all day.

Practitioner checklist

• Is every graph-powered roadmap item tagged with its family — and therefore its compute pattern, explainability, and eval plan?
• Do predictive features look predictive in the UX (suggestion framing, why-shown), and are they never silently persisted as facts?
• For path-and-rule features: can the product show its work — the path or the rule — to a user, an auditor, or a regulator?
• Are batch cadences matched to decision speed — and is someone on record accepting the staleness window?
• Have we identified our celebrity nodes and decided how each family handles them?
• Do algorithm outputs carry enough lineage that a wrong score can be traced to the facts (or bugs) that produced it?

Related lessons

• Building the graph — analytics amplify whatever quality the pipeline delivers, garbage included.
• Knowledge graphs & LLMs — the newest consumer of all five families.
• Evals — the measurement discipline the predictive family borrows wholesale.