Causal Reasoning in Knowledge Graphs¶
Your knowledge graph has hundreds of edges. REQUIRES, SUPPORTS, IMPLEMENTS, CONTRADICTS. They all look the same: source, target, confidence score. But some of those edges are load-bearing walls and others are decorative trim.
This series teaches you to tell the difference. You'll build a causal model from a real knowledge graph, learn which edges carry information and which are redundant, and propagate reward through the structure when rules succeed or fail.
The series¶
| Tutorial | What it covers |
|---|---|
| Not All Edges Are Created Equal | Causal vs correlational edges; the direction map hiding in your relation types |
| The Shape of What You Know | Building a DAG from a messy graph; cycles, breaking them, and what you lose |
| What Can You Learn From Here? | D-separation: which information paths are open, which are blocked, and why conditioning can create dependencies |
| Propagating What Works | Credit assignment through DAG ancestry; connecting reward to the Thompson Sampling bandit |
| The Degradation Chain | The dispatcher's fallback sequence; Pyro-aware fields waiting for Phase 2; where causal meets geometry |
What you need¶
- A qortex knowledge graph (from the ingestion tutorials, or the Chapter 5 fixture data)
pip install qortex(thecausalmodule ships with the core package)- networkx >= 3.3 (for
nx.is_d_separator)
Where this leads¶
The causal layer answers "what leads to what?" The next series, The Geometry of Learning, answers "is the system actually learning, and how fast?" The two bind together exactly where you'd expect: credit assignment outputs tangent vectors on the Fisher manifold. The causal DAG chooses the direction; the manifold determines the distance.
After both series, a roadmap page sketches where this all goes: dynamical systems on curved belief space, Noether's theorem for conservation laws, and the interoception layer that makes the system monitor its own learning.