supply_chain_resilience: dual-mode GNN delay predictor (train from scratch or load pre-created)#92
Open
cafzal wants to merge 6 commits into
Open
supply_chain_resilience: dual-mode GNN delay predictor (train from scratch or load pre-created)#92cafzal wants to merge 6 commits into
cafzal wants to merge 6 commits into
Conversation
…ratch or load pre-created) The template's supplier delay risk now comes from an actual GNN, not a hand-authored table, with a TRAIN_GNN toggle: train from scratch (GPU + Snowflake experiment schema) or run on the bundled, GNN-produced predictions (default; fast, no GPU). - data/generate_corpus.py + the multi-year labelled corpus it produces (shipment_corpus.csv + temporal train/val/test splits): the bundled 262 shipments are too sparse to train a GNN. Delay risk combines supplier reliability, a recurring seasonal surge, and one-hop upstream propagation through the supply graph — a high-own-reliability shipper like B004 is risky via its upstream B003, signal a per-supplier model misses and message-passing recovers. - supply_chain_resilience_predictive.py: the dual-mode GNN stage (Shipment->Supplier->upstream-Supplier graph, leakage-controlled features), aggregating per-shipment lateness to a per-supplier delay probability written to data/delay_prediction.csv. - runbook step 5b documenting the optional from-scratch training. Draft: the bundled delay_prediction.csv is still the prior table; the from-scratch GNN run that regenerates it is the pending verification step.
…+ regenerated predictions Reworked the predictive stage to the proven CSV-backed (smoker_local) shape after a heterogeneous graph hit EMPTY_TABLE on local model.data secondary node tables: homogeneous Shipment nodes + a relatedness edge list (co-supplier + upstream links). Supplier reliability is denormalized as a feature; the graph corrects it via label propagation. Verified end-to-end (TRAIN_GNN=true, seed 42): non-degenerate, differentiated 0.19-0.52, B003 #1 with B014/B017 on top. B004 (own-reliability 0.90) lifts to the HIGH tier (0.315) above its reliability peers (~0.19-0.25) purely via the upstream-propagation edges (B004<-B003) — the graph tell. data/delay_prediction.csv regenerated from this run (model gnn_v3.0); generator now emits supplier_reliability + shipment_edges.csv.
Collaborator
Author
|
Verified the from-scratch path end-to-end ( |
Hoist the collections import to the top and split semicolon-joined statements (ruff E402/I001/E702 — the CI lint failure). Sort the upstream-set iteration before the seeded random.choice: set iteration order is hash-randomized per process, so shipment_edges.csv was non-deterministic across runs despite the fixed SEED. Now generate_corpus.py reproduces a byte-identical corpus + edges every run (corpus/labels were already deterministic — worst_up uses max()).
…easoner chain Makes the GNN delay predictor a full stage in the multi-reasoner chain instead of a separate dual-mode script plus a data load. The chain is now Stage 0 Reachability -> 1 Graph -> 2 Predictive -> 3 Rules -> 4 Prescriptive; the GNN's per-supplier delay risk (DelayPrediction.predicted_delay_prob) is the hand-off the rules stage consumes. Combined script: fold the standalone GNN trainer in as Stage 2 (dual-mode -- default loads the bundled predictions, TRAIN_GNN=true retrains on the multi-year corpus in its own Model); renumber Rules->3 and Prescriptive->4; delete supply_chain_resilience_predictive.py. The predictive import is deferred into the train branch so the default chain runs on base relationalai with no GPU stack (matching datacenter_compute_allocation); pyproject stays relationalai==1.11.0. Runbook + README: add Predictive to the chain ASCII, reasoner overview, reasoning_types, and How-it-works; renumber; reword the single-model claim since the GNN trains on a separate corpus Model. Static-verified (py_compile, ruff); the TRAIN_GNN=true run and end-to-end chain numbers need a live RAI engine to confirm.
…ve reasoning type Auto-generated gallery indexes (generate_version_indexes.py) reflecting the new reasoning_types + description; the verify CI check enforces they stay in sync.
… the GNN discriminates and the demo holds Running the chain on the engine surfaced that the bundled GNN predictions clustered near the base rate (all 23 suppliers above the old 0.15 threshold), which flagged 15 watch suppliers and broke the documented chain ($8,545 baseline, Watch->Avoid +3609%). Sharpened the corpus risk model (raise the eff_unrel cap/amplification, trim the seasonal/lead noise that blurred per-supplier rates) so genuinely-risky suppliers separate (B003 0.76, B014 0.68, B017 0.63; reliable ~0.10); retrained the GNN (val roc_auc ~0.68 -> predictions span 0.84 down to <0.13) and set DELAY_PROB_THRESHOLD=0.50. The chain now reproduces the documented $1,865 baseline / S004-offline +88.5% / Watch->Avoid +0.0% -- verified end-to-end on the engine (logic + GPU predictive + prescriptive). Also fixed the from-scratch path: gnn.fit() raised [Ambiguous model] because Stage 2 created a second Model; it now trains on the single main model (the portfolio pattern). Reconciled runbook + README to the actual run (rules counts 2->5, threshold, GNN predicted spread).
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Adds a predictive stage to the
supply_chain_resiliencetemplate so the supplier delay risk it feeds into the Rules → Prescriptive chain comes from an actual GNN, not a hand-authoreddelay_prediction.csv. Dual-mode via aTRAIN_GNNtoggle:TRAIN_GNNunset (default) — downstream stages use the bundled, GNN-produceddata/delay_prediction.csv. Fast, no GPU, no Snowflake experiment schema.TRAIN_GNN=true— train the GNN from scratch on the bundled corpus and regenerate the predictions.Why a GNN
Delay risk propagates through the supply graph: a shipper with high own reliability is still risky when its upstream supplier is unreliable (e.g. B004 ← B003). A per-supplier or flat tabular model misses that; message-passing over
Shipment → Supplier → upstream-Supplieredges recovers it.Changes
data/generate_corpus.py+ the multi-year labelled corpus it produces (shipment_corpus.csv+ temporalshipment_{train,val,test}.csv). The bundled 262 shipments are far too sparse to train a GNN; the corpus combines supplier reliability, a recurring seasonal surge, and one-hop upstream propagation. Baseline check: per-shipment test roc_auc ≈ 0.63, B004 elevated to 0.40 despite 0.90 own-reliability.supply_chain_resilience_predictive.py— the dual-mode GNN stage (leakage-controlled features; per-shipment lateness aggregated to a per-supplier delay probability).runbook.md— step 5b for the optional from-scratch training.Draft / pending verification
The bundled
delay_prediction.csvis still the prior table. The pending step is a realTRAIN_GNN=trueGPU run to (1) confirm the GNN trains non-degenerate and recovers the graph-aware ranking, and (2) regenerate the shippeddelay_prediction.csvfrom it. Marking ready once that run is green.