NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·2h ago
AI Research
arXiv CS.AI·May 25·8 min read
Expressive Power of Deep Homomorphism Networks over Relational Databases
TL;DR
Graph Neural Networks (GNNs) struggle to express complex queries over relational databases. Researchers analyzed Deep Homomorphism Networks (DHNs)—a neural architecture designed to mirror SQL conjunctive queries—by mapping them to logical fragments.
✦ Why It Matters
Engineers can now predict which DHN aggregation function suits specific database query patterns before implementation.
Key Takeaways
- DHNs with sum aggregation handle more complex logical patterns than max or mean variants.
- Map your database query requirements to logical fragments to select appropriate DHN architecture.
- Static analysis problems (emptiness, subsumption) are decidable for DHN variants—enabling verification tools.
- Expressiveness differences in logic directly correlate with real-world prediction task performance.
Full Summary
Message-passing Graph Neural Networks have known limitations in expressing complex relationships, motivating research into more powerful architectures for database learning tasks. Deep Homomorphism Networks (DHNs) are proposed as particularly suitable for relational databases because they structurally align with SQL conjunctive queries (database queries selecting data matching multiple conditions).
The authors established formal connections between DHNs and fragments of first-order logic (FO)—a mathematical system for expressing logical statements. They proved that DHNs using max, sum, and mean aggregations correspond to specific logical fragments: unary negation fragment (UNFO), extensions with counting quantifiers, and quantifier alternation variants.
Through these mappings, the work illuminates which database queries different DHN variants can and cannot compute. The team also studied decidability of static analysis problems like emptiness and subsumption checking.
Experimental validation confirmed that theoretical expressiveness differences translate to measurable performance gaps on prediction tasks.