Research map / Probabilistic and scientific modeling / Molecules and proteins
Invariant Deep Uplift Modeling for Incentive Assignment in Online Marketing via Probability of Necessity and Sufficiency
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 IDUM, an invariant-learning uplift model that uses probability of necessity and sufficiency (PNS) risk, Gumbel-Softmax feature selection, and IPM balancing to generalize incentive-response estimation to out-of-distribution users.
Problem
Uplift models for incentive assignment are evaluated mainly on in-distribution data, and their performance degrades when user feature distributions shift over time, geography, or context, while observational incentive assignment also suffers selection bias.
Approach
Combines a PNS-based invariant property learning objective (sufficient and necessary causal features, with a monotonicity upper bound and a KL term) with a Gumbel-Softmax top-k feature mask and an IPM balancing regularizer, so the uplift model stays robust under distribution shift.
Key results
- On the Lazada OOD split, IDUM reaches AUUC 0.0274, QINI 0.4633, KENDALL 0.1790, the best AUUC and QINI among compared methods (CFRNet-wass AUUC 0.0269).
- On the Production OOD split, IDUM reaches AUUC 0.2847 and KENDALL 0.2051, outperforming EUEN (AUUC 0.2755, KENDALL 0.1995).
- Ablation on Production OOD shows removing the invariant property learning (IPL) module drops AUUC from 0.2847 to 0.2732 and QINI from 0.3681 to 0.3475.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add balancing discrepancy (BD) (ablation) | IDUM · Production OOD | AUUC: 0.2710 ± 0.0227 -> 0.2847 ± 0.0265 | Helped |
| add balancing discrepancy (BD) (ablation) | IDUM · Production OOD | QINI: 0.3383 ± 0.0264 -> 0.3681 ± 0.0245 | Helped |
| add balancing discrepancy (BD) (ablation) | IDUM · Production OOD | KENDALL: 0.1801 ± 0.0153 -> 0.2051 ± 0.0166 | Helped |
| add invariant property learning (IPL) (ablation) | IDUM · Production OOD | AUUC: 0.2732 ± 0.0235 -> 0.2847 ± 0.0265 | Helped |
| add invariant property learning (IPL) (ablation) | IDUM · Production OOD | QINI: 0.3475 ± 0.0204 -> 0.3681 ± 0.0245 | Helped |
| add invariant property learning (IPL) (ablation) | IDUM · Production OOD | KENDALL: 0.1781 ± 0.0173 -> 0.2051 ± 0.0166 | Helped |
| add Gumbel-Softmax feature selection to IPL (ablation) | IDUM · Production OOD | AUUC: 0.2737 ± 0.0251 -> 0.2847 ± 0.0265 | Helped |
| add Gumbel-Softmax feature selection to IPL (ablation) | IDUM · Production OOD | QINI: 0.3242 ± 0.0253 -> 0.3681 ± 0.0245 | Helped |
| add Gumbel-Softmax feature selection to IPL (ablation) | IDUM · Production OOD | KENDALL: 0.2039 ± 0.0124 -> 0.2051 ± 0.0166 | Helped |
Surprising findings
- On Lazada OOD, several baselines (e.g., UniTE, QINI 0.4499) achieve QINI comparable to IDUM, and IDUM's QINI gain there is small (0.4633 vs 0.4618 for CFRNet-mmd).
Limitations
- Production dataset is proprietary and the OOD split relies on two time points, limiting reproducibility.
- Theoretical guarantees rely on exogeneity and monotonicity assumptions and on a bounded support-shift term that is not empirically verified.
- Hyperparameter sensitivity: performance varies notably with KL weight and IPM weight, requiring careful tuning.
Experimental setup
Datasets: Lazada dataset (voucher distribution, Zhong et al. 2022) Production short-video clarity dataset (proprietary)
Benchmarks: Lazada dataset (ID and OOD) Production dataset (ID and OOD)
Metrics: AUUC QINI KENDALL
Baselines: S-Learner T-Learner TARNet CFRNet-mmd CFRNet-wass DragonNet EUEN UniTE TEED
Builds on: CFRNet CaSN Invariant risk minimization Gumbel-Softmax PNS (Pearl)
Closest papers by meaning
- Rethinking Causal Ranking: A Balanced Perspective on Uplift Model Evaluation (ICML 2025 · similarity 0.88)
- Uplift Modeling with Delayed Feedback: Identifiability and Algorithms (AAAI 2026 · similarity 0.85)
- Bi-Level Decision-Focused Causal Learning for Large-Scale Marketing Optimization: Bridging Observational and Experimental Data (NeurIPS 2025 · similarity 0.85)
- Backdoor Adjustment via Group Adaptation for Debiased Coupon Recommendations (AAAI 2024 · similarity 0.84)
- Causal Customer Churn Analysis with Low-rank Tensor Block Hazard Model (ICML 2024 · similarity 0.82)
- Scaling Reward Modeling without Human Supervision (NeurIPS 2026 · similarity 0.81)
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.