The Theory↔︎Computation Cookbook
Graph signal processing and the GNN canon, executed and measured on the jammi engine
This is a runnable book. It bridges applied theory — Stanković et al.’s Data Analytics on Graphs (Foundations & Trends in ML, 2020) and the modern GNN canon — with software and computation: the jammi engine. The signature move of every recipe is to show that one Jammi recipe = one equation in the monograph = one line in the GNN canon, executed and measured.
For the short, “how do I call this verb” reference — dual-language (Rust + Python) and compile-tested — see the How-To Guides in the engine Guide. This book is the long-form, Python, executed-and-measured companion to those recipes.
0.1 How the book is organized
It reads as 4 tiers × 3 rails, echoing AWS Neptune’s “unified yet separate” Database / Analytics / ML split — and adds a fourth tier Neptune structurally lacks: calibrated, provenance-stamped, context-conditioned prediction.
| Tier | Recipe | Theory |
|---|---|---|
| 01 Construct | build_neighbor_graph |
topology from data (monograph Part I) |
| 02 Analyze | propagate_embeddings |
graph signal processing / low-pass filtering = SGC/APPNP (Part II) |
| 03 Learn | fine_tune_graph |
representation learning on graphs (Part III) |
| 04 Predict & Quantify | train_context_predictor + conformal |
context-conditioned posterior + honest coverage (no Neptune analogue) |
The three rails are woven through every tier: provenance (the audit trail on every result), tenancy (the engine’s two isolation layers — catalog-listing and discriminator-column row filtering — and the honest caveat where they stop), and measurement (every recipe ends in a real number). The book is the closed evaluation loop: construct → propagate → learn → MEASURE.
0.2 How it stays honest
Every recipe executes in continuous integration and ends in a real number computed from committed artifacts — never a stand-in figure. The heavy work runs once in the keystone slice and is committed as a small-subset cache; every later chapter reads that cache and asserts its measured verdict against frozen golden metrics within a tolerance. The engine itself is never edited — the book only calls the pinned jammi API, listed in the API reference.