Research map / Time series / Causal and generative time series
GCF: Estimating Unobserved Demand Using Graph Causal Forecasting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Graph Causal Forecasting (GCF), an RGCN-dilated CNN that estimates counterfactual unobserved product demand using a co-view knowledge graph as synthetic control, plus a B-APP loss for imbalanced demand data, evaluated on Amazon internal data and semi-synthetic METR-LA.
Problem
Forecasting product demand that is suppressed by a treatment (e.g., delivery-time-based search suppression) requires estimating unobserved counterfactual demand, which standard forecasters trained on reduced observed demand underfit.
Approach
Estimate counterfactual demand by splitting it into unaffected local demand (factual dilated CNN forecast) and affected out-of-region demand, modeled by an RGCN-dilated CNN over a heterogeneous product graph whose co-view edges define a synthetic control group; a boundary-aware penalized MSE (B-APP) loss handles class imbalance.
Key results
- On internal data, GCF with co-view graph achieved 75.3% lower MAPE and +10.3% precision versus baseline-2 for counterfactual forecasting.
- On semi-synthetic METR-LA, GCF reached MSE 94.72 vs 134.66 for Temporal GCN, 127.07 for Google Causal Impact and 111.71 for CRN.
- Deployment in 2023 produced +1399 bps products with views and +310 bps products with sales; offline PWAGV rose 61.2%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GCF instead of Temporal GCN for counterfactual forecasting | METR-LA traffic forecasting · METR-LA | MSE: 30% lower MSE | Helped |
| Use GCF instead of CRN for counterfactual forecasting | METR-LA traffic forecasting · METR-LA | MSE: 15% lower MSE | Helped |
| Use GCF instead of Google Causal Impact | METR-LA traffic forecasting · METR-LA | MSE: 25% lower MSE | Helped |
| Use B-APP loss instead of MSE or MAE (ablation) | Dilated CNN factual forecaster · Internal factual forecasting | Precision: 5.3% higher precision than MSE/MAE | Helped |
| Replace co-view edges with price-profile edges (ablation) | RGCN-DCNN counterfactual forecaster · Internal counterfactual forecasting | MAPE: -73.9% vs baseline-2, compared with -75.3% for co-view | Hurt |
| Replace co-view edges with demand-profile edges (ablation) | RGCN-DCNN counterfactual forecaster · Internal counterfactual forecasting | MAPE: -74.3% vs baseline-2, compared with -75.3% for co-view | Hurt |
| Replace co-view edges with product-title edges (ablation) | RGCN-DCNN counterfactual forecaster · Internal counterfactual forecasting | MAPE: -66.4% vs baseline-2, compared with -75.3% for co-view | Hurt |
| Use internal co-view edges instead of co-view edges (ablation) | RGCN-DCNN counterfactual forecaster · Internal counterfactual forecasting | MAPE: -72.6% vs baseline-2, compared with -75.3% for co-view | Hurt |
| Remove current-quarter internal-demand input (ablation) | RGCN-DCNN counterfactual forecaster · Internal counterfactual forecasting | MAPE: -70.1% vs baseline-2, compared with -75.3% for co-view | Hurt |
| Use co-view edges rather than other graph-edge strategies (ablation) | RGCN-DCNN counterfactual forecaster · Internal counterfactual forecasting | MAPE: -75.3% | Helped |
Surprising findings
- Propensity-based control scored worse than baseline on F-score (-27.4%), and DiD gave -74.4% MAPE but only +11.3% F-score, showing metric trade-offs among counterfactual methods.
- Internal co-view graph edges alone (without the co-view-based RGCN) reached a lower F-score than the full co-view model (15.8% vs 20.2%) despite near-equal recall.
Limitations
- Evaluated mainly on proprietary internal Amazon data that cannot be released
- Synthetic control selection relies on hand-chosen domain-knowledge edge types and has hidden-confounder concerns
- Semi-synthetic METR-LA treatment is simulated rather than real
Experimental setup
Datasets: Internal Amazon marketplace product demand data METR-LA
Benchmarks: METR-LA (semi-synthetic treatment) Amazon internal demand forecasting
Metrics: precision recall F-score MAPE MSE PWAGV PWAS
Baselines: Average growth factor Difference-in-differences Propensity matching Catboost LSTM GRU CNN CRN Temporal GCN Google Causal Impact
Builds on: Counterfactual recurrent network TGCN Google Causal Impact RGCN dilated convolution
Closest papers by meaning
- A Cross-View Hierarchical Graph Learning Hypernetwork for Skill Demand-Supply Joint Prediction (AAAI 2024 · similarity 0.81)
- Bi-Level Decision-Focused Causal Learning for Large-Scale Marketing Optimization: Bridging Observational and Experimental Data (NeurIPS 2025 · similarity 0.81)
- Causal Contrastive Learning for Counterfactual Regression Over Time (NeurIPS 2024 · similarity 0.81)
- CoDeR: Counterfactual Demand Reasoning for Sequential Recommendation (AAAI 2025 · similarity 0.80)
- A Recipe for Causal Graph Regression: Confounding Effects Revisited (ICML 2025 · similarity 0.80)
- CausalStock: Deep End-to-end Causal Discovery for News-driven Multi-stock Movement Prediction (NeurIPS 2024 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.