Research map / Label-efficient and robust learning / Domain adaptation and generalization

Transfer Learning Through Conditional Quantile Matching

NeurIPS 2026 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Conditional distribution alignment · 0 citations (OpenAlex)

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 TLCQM, which augments a data-scarce regression target domain with synthetic responses from per-source engression generators calibrated to target conditional distributions via quantile matching, and proves excess-risk bounds and quantile-matching convergence rates.

Problem

Transfer learning for regression with heterogeneous source domains where both covariate and conditional response distributions differ, with scarce target data and risk of negative transfer.

Approach

Learn a conditional generative model (engression) per source domain, generate synthetic responses at target covariates, and calibrate them with a linear combination fitted by matching conditional quantiles (squared Mallows metric) to the target response distribution, yielding augmented data without covariate or label shift assumptions.

Key results

• On the Apartment data, TLCQM-augmented NN reduces MSE relative to target-only NN by 99.43% at n0=100 and 87.98% at n0=500; XGBoost gains shrink to -6.80% (worse) at n0=500.
• On the MAU case study (Platform I), TLCQM-augmented XGBoost, NN and KRR achieve MSEs of 1.545, 1.689 and 1.556 versus target-only 4.596, 2.910 and 33.281, and better than TKRR (2.204), CDAR (3.183), DARC (36.380).
• Theory: excess risk of TLCQM ERM bounded by standard generalization error plus importance-weight, quantile matching, distributional learning and transfer bias terms; quantile matching error rate derived.

Surprising findings

• On the Apartment data, XGBoost trained on TLCQM-augmented data performs worse than target-only XGBoost at n0=300 and 500 (-2.28% and -6.80%).

Limitations

• Transfer bias does not vanish unless the target lies in the convex hull of source conditionals; bias can persist.
• Theoretical rates depend on assumptions like bounded densities, bounded generators and Kiefer conditions; density ratio estimation adds error.
• Gains for tree-based models (XGBoost) are smaller and can be negative on real data.

Experimental setup

Datasets: Apartment (UCI rental listings) MAU case study internal data simulated data

Benchmarks: Apartment (UCI) simulated regression MAU prediction case study

Metrics: MSE percentage MSE reduction

Baselines: TKRR CDAR DARC target-only XGBoost target-only KRR target-only NN oracle training

Builds on: engression quantile matching (Mallows metric) kernel mean matching TKRR DARC

Closest papers by meaning

• Wasserstein Transfer Learning (NeurIPS 2025 · similarity 0.86)
• Transfer Learning for Benign Overfitting in High-Dimensional Linear Regression (NeurIPS 2025 · similarity 0.85)
• A High-Dimensional Statistical Method for Optimizing Transfer Quantities in Multi-Source Transfer Learning (NeurIPS 2025 · similarity 0.84)
• Provable Sample-Efficient Transfer Learning Conditional Diffusion Models via Representation Learning (NeurIPS 2025 · similarity 0.84)
• Statistical Matching via Schr\"odinger Bridge beyond Conditional Independence (NeurIPS 2026 · similarity 0.84)
• Rectifying Conformity Scores for Better Conditional Coverage (ICML 2025 · similarity 0.84)