Research map / Probabilistic and scientific modeling / Treatment effects
Bi-Level Decision-Focused Causal Learning for Large-Scale Marketing Optimization: Bridging Observational and Experimental Data
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 Bi-DFCL, a bi-level decision-focused causal learning framework that uses RCT data to train a bridge network via an unbiased decision loss while learning the target uplift network on large observational data, solved with implicit differentiation and conjugate gradient.
Problem
Marketing resource allocation for online platforms suffers from prediction-decision misalignment in two-stage ML+OR pipelines and a bias-variance dilemma from combining biased observational data with scarce, unbiased RCT data.
Approach
Assign an unbiased RCT-based primal decision loss (surrogate PPL/PIFD) to the upper level and an MSE prediction loss on observational data with counterfactual pseudo-labels gated by a bridge network to the lower level, and compute the hypergradient via implicit differentiation with conjugate gradient instead of unrolled GD.
Key results
- On CRITEO-UPLIFT v2 (Hybrid), Bi-DFCL-PIFD reaches AUCC 0.7812 (9.37% over TSM-SL), the best among compared methods.
- On Marketing data II, Bi-DFCL-PPL achieves EOM 1.0252 (2.52% improvement over the RCT-only baseline), and on Marketing data I Bi-DFCL-PIFD reaches EOM 1.0297 (2.97%).
- In a four-week online A/B test on 790K shops at Meituan, Bi-DFCL-PIFD achieved a 3.22% improvement in orders over TSM-SL, versus 1.80% for DFCL-PIFD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Bi-DFCL-PIFD instead of LTD-DR | Marketing optimization · CRITEO-UPLIFT v2 (Hybrid) | AUCC: 0.7533 -> 0.7812 | Helped |
| use Bi-DFCL-PIFD instead of AutoDebias | Marketing optimization · Marketing Data I | EOM: 1.0175 -> 1.0297 | Helped |
| use Bi-DFCL-PPL instead of LTD-DR | Marketing optimization · Marketing Data II | EOM: 1.0067 -> 1.0252 | Helped |
| add PPL decision loss (ablation) | Bi-DFCL ablation baseline · Marketing Data I | EOM: 1.0000 -> 1.0167 | Helped |
| add PPL decision loss (ablation) | Bi-DFCL ablation baseline · Marketing Data II | EOM: 1.0000 -> 1.0156 | Helped |
| add bi-level optimization using hybrid RCT and OBS data (ablation) | PPL decision-loss variant · Marketing Data I | EOM: 1.0167 -> 1.0240 | Helped |
| add bi-level optimization using hybrid RCT and OBS data (ablation) | PPL decision-loss variant · Marketing Data II | EOM: 1.0156 -> 1.0175 | Helped |
| add counterfactual labels (ablation) | PPL plus bi-level optimization · Marketing Data I | EOM: 1.0240 -> 1.0248 | Helped |
| add counterfactual labels (ablation) | PPL plus bi-level optimization · Marketing Data II | EOM: 1.0175 -> 1.0213 | Helped |
| add implicit differentiation (ablation) | PPL, bi-level optimization, and counterfactual-label variant · Marketing Data I | EOM: 1.0248 -> 1.0277 | Helped |
| add implicit differentiation (ablation) | PPL, bi-level optimization, and counterfactual-label variant · Marketing Data II | EOM: 1.0213 -> 1.0252 | Helped |
Surprising findings
- Among RCT-only methods on Criteo, the two-stage TSM-SL (AUCC 0.7143) is outperformed by several end-to-end methods, while TSM-CF underperforms TSM-SL on all datasets.
Limitations
- Computational efficiency is noted as a future improvement area.
- Evaluated on one proprietary platform (Meituan) and a public dataset with simulated OBS split.
Experimental setup
Datasets: CRITEO-UPLIFT v2 Meituan Marketing data I Meituan Marketing data II
Benchmarks: CRITEO-UPLIFT v2 (Hybrid) Meituan Marketing data I Meituan Marketing data II
Metrics: AUCC EOM
Baselines: TSM-SL TSM-CF DHCL DFCL-DPL DFCL-DIFD IPS DR-JT CFR-WASS CFR-MMD DragonNet CausE KD-Label
Builds on: DFCL DHCL SPO DragonNet
Closest papers by meaning
- Invariant Deep Uplift Modeling for Incentive Assignment in Online Marketing via Probability of Necessity and Sufficiency (ICML 2025 · similarity 0.85)
- Integrating Inference and Experimental Design for Contextual Behavioral Model Learning (AAAI 2025 · similarity 0.83)
- DFF: Decision-Focused Fine-Tuning for Smarter Predict-Then-Optimize with Limited Data (AAAI 2025 · similarity 0.83)
- On Online Experimentation without Device Identifiers (ICML 2024 · similarity 0.83)
- Rethinking Causal Ranking: A Balanced Perspective on Uplift Model Evaluation (ICML 2025 · similarity 0.83)
- Interference Among First-Price Pacing Equilibria: A Bias and Variance Analysis (ICLR 2025 · similarity 0.82)
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.