Research map / Time series / Causal and generative time series

Learning Deployable Causal Action Geometry under Temporal Non-Stationarity

NeurIPS 2026 · Time series › Causal and generative time series · approach: Dynamical system reconstruction

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.

Defines deployable causal action geometry, an anchored, link-scale, amplitude-profiled action-response shape that persists under temporal drift, and estimates it with a two-stage orthogonal honest profiled selection procedure, validated on synthetic data, Dominick's, and a Taobao pricing A/B test.

Problem

Learning a continuous-action response law from non-stationary observational time series, where pooled regression conflates action effects with drifting baselines and can reverse action rankings for future deployment.

Approach

Target only the anchored action-response shape on a selected monotone link (identity or log) after profiling out positive amplitude, and estimate it via cross-fitted orthogonal pilots followed by honest profiled selection over link, representation, and shape classes, with guarantees under approximate geometry persistence.

Key results

• In the Taobao 26-day online A/B test on about 280k items, GMV improved +1.16% (p=0.037) and buy counts +1.70% (p=0.004).
• Offline on Taobao, the method achieved RMAE 0.103 and MAPE 12.75 versus 0.138 and 16.33 for fixed log link, and DR policy value index 128.32 versus 122.36 for fixed log and 121.72 for NPM.
• In synthetic regime with shifted history distribution, the method's NRegret (0.1385) was below plug-in selection (0.1412), fixed identity (0.1416) and fixed log (0.1422).

Surprising findings

• Pooled regression can reverse the deployable action ranking under drift-aligned logging (ranking-reversal construction).
• Log-linked contrast was preferred in only four of six rolling splits on Taobao, indicating time-varying preference between link scales.

Limitations

• Validity relies on one-step exchangeability, overlap, a finite link family, and approximate geometry persistence.
• Guarantees are criterion-level within a finite admissible library and do not imply unique latent representation recovery.
• Target is a one-step proximal response, not long-horizon dynamic treatment effects.

Experimental setup

Datasets: Taobao item-day pricing panel Dominick's retail dataset synthetic DGP regimes R1-R4

Benchmarks: Dominick's retail dataset Taobao offline rolling-origin split synthetic regimes R1-R4

Metrics: SIMSE regret RMAE MAPE stability loss DR policy value index GMV lift buy count lift commission lift

Baselines: NPM VCNet Causal Transformer Pooled MLP Time-aware pooled MLP Fixed identity link Fixed log link

Builds on: orthogonal/debiased learning VCNet Causal Transformer invariant risk minimization

Closest papers by meaning

• Observable Neural ODEs for Identifiable Causal Forecasting in Continuous Time (NeurIPS 2026 · similarity 0.87)
• Learning Structural Causal Models from Ordering: Identifiable Flow Models (AAAI 2025 · similarity 0.86)
• Temporal Difference Flows (ICML 2025 · similarity 0.86)
• Learning Time-Aware Causal Representation for Model Generalization in Evolving Domains (ICML 2025 · similarity 0.86)
• A Dynamical View of the Question of Why (ICLR 2024 · similarity 0.85)
• Sim-to-Real Causal Transfer: A Metric Learning Approach to Causally-Aware Interaction Representations (CVPR 2025 · similarity 0.84)