Graph validation: type constraints, cardinality and quality gates

Last updated: 5 Oct 20265 min read

tutorial

A graph load should be rejected when its claims violate declared relation types, identity rules or required evidence fields.

Validate shape rules

A teaches relation can require a lesson subject and concept object. A published lesson may need exactly one canonical subject, while a concept can have many teaching lessons. Express required properties, allowed types and cardinality in a versioned schema. Some constraints are closed-world editorial rules; others should not be mistaken for universal facts about the world.

Check contradictions

Two prerequisites may form a cycle even if each claim is individually well typed. A cycle might be permitted for loose “related to” edges but invalid for strict learning order. Identify which predicates require acyclicity and run those checks after ingestion. Relation direction affects both validation and downstream traversal.

Measure quality by source

Report missing evidence, duplicate claim IDs, invalid targets, orphan entities and unresolved aliases by ingestion source. A high overall pass rate can hide one newly added feed whose claims are nearly all wrong. Quarantine bad claims rather than silently dropping them; preserve the reason and record count for reconciliation with input batches.

Test publication

Insert a teaches claim from a learner node to a concept node, a duplicate prerequisite ID and a strict-prerequisite cycle among three lessons. The type and duplicate checks should fail before publication; the cycle check should isolate the loop. A clean batch should switch into the serving snapshot only after counts and required relationships reconcile.

Implementation

def invalid_typed_claims(claim_rows, allowed_relations):
    return [claim["claim_id"] for claim in claim_rows
            if allowed_relations.get(claim["predicate"])
            != (claim["subject_type"], claim["object_type"])]

Performance and operating cost

Type validation costs O(C) expected time for C claims. Cycle checks on a selected directed subgraph cost O(V + E) time and space; run them only for predicates where a cycle is actually forbidden.

Common Mistakes

Do not treat every relation as if it had the same cardinality.
Do not publish a batch with silent invalid-claim loss.
Do not apply prerequisite cycle rules to ordinary related-to links.