Q&A — Phase 26.1 ConceptGraph: Design Decisions & Usage

[web3guru888]

Frequently Asked Questions — ConceptGraph

Q1: Why combine semantic networks with prototype theory?

A: Classical semantic networks (Quillian 1967) represent concepts as nodes with labeled edges, giving us inheritance and relational reasoning. But they treat category membership as binary — something either IS_A bird or not. Rosch's prototype theory adds graded membership via the typicality field. A robin (typicality=0.95) is a more typical bird than a penguin (typicality=0.3). This lets us model fuzzy boundaries without abandoning the graph structure.

Q2: How does property inheritance handle conflicts in diamond hierarchies?

A: When a concept has multiple parents via IS_A (diamond inheritance), properties from all parents are collected via depth-first traversal. If the same property name appears with different values, the own (most specific) property always shadows inherited ones. For conflicts between equally-distant ancestors, we use topological ordering — the first parent in parent_ids takes precedence.

Q3: What is the computational complexity of spreading activation?

A: Worst-case O(V + E) where V is nodes and E is edges — essentially a BFS. In practice, the threshold parameter prunes most of the graph. With decay=0.8 and threshold=0.1, activation dies after ~10 hops (0.8^10 ≈ 0.107). The actual runtime is proportional to the activated subgraph, which is typically much smaller than the full graph.

Q4: How does the similarity metric work when embeddings are missing?

A: The similarity function uses a weighted combination of three signals:

• Wu-Palmer path similarity (weight 0.4) — always available via graph structure
• Property Jaccard coefficient (weight 0.3) — always available via propagated properties
• Cosine embedding similarity (weight 0.3) — only if both concepts have embeddings

When embeddings are missing, the weight is redistributed proportionally: Wu-Palmer gets 0.57, Jaccard gets 0.43.

Q5: Can ConceptGraph handle cycles (e.g., A causes B, B causes A)?

A: Yes. IS_A relations are enforced to be acyclic (DAG) — add_relation rejects cycles for IS_A edges. But other relation types (CAUSES, ENABLES, INHIBITS, SIMILAR_TO) allow cycles, representing feedback loops and mutual relationships. Spreading activation handles cycles by tracking visited nodes and only updating when the new activation exceeds the existing one.

Q6: How does ConceptGraph integrate with WorldModel (13.1)?

A: WorldModel stores entities with properties and relations. ConceptGraph provides the type hierarchy — WorldModel entities are instances of ConceptGraph concepts. When WorldModel encounters a new entity type, it queries ConceptGraph for default properties via propagate_properties(). When an entity's behavior deviates from its concept's expectations, SurpriseDetector (13.4) is triggered.

Q7: What testing strategy validates graph correctness?

A: We test three layers:

1. Structural correctness: add/remove operations maintain graph invariants (no IS_A cycles, bidirectional adjacency, referential integrity)
2. Semantic correctness: subsumption transitivity, property inheritance shadows, LCA computation on diamond graphs
3. Algorithmic correctness: spreading activation decay matches decay^hop, similarity bounds [0,1], threshold pruning prevents over-activation

---

Issue: #578 | Planning: #577