Transfer Learning Through Conditional Quantile Matching (TLCQM)

This repository contains the Python implementation for high-quality data augmentation and synthetic data generation in transfer learning tasks.

Paper Reference: Y. Zhang, S. Wilkins-Reeves, W. Lee, and A. Hofleitner. Transfer Learning Through Conditional Quantile Matching. (2026+).

Overview

TLCQM is a framework that addresses both covariate shift and concept shift between source and target domains. It leverages conditional quantile matching to calibrate generated samples from multiple source domains to match the target domain distribution, enabling effective transfer learning with limited labeled target data.

Requirements

• Python >= 3.10 (earlier versions might be applicable)
• NumPy
• scikit-learn
• PyTorch (for neural network models and auto-differentiation)
• engression
• cvxopt
• Optional: pandas and Matplotlib (for data processing and plotting)

File Descriptions

Core Modules

File	Description
TLCQM.py	Main implementation of the TLCQM framework with the fit_TLCQM() function
quantile_match.py	Quantile matching estimator via an iterative procedure
covariate_shift.py	Utilities for handling covariate shift between domains
utils.py	Utility functions for data simulation

Simulation Scripts

File	Description
Sim_TLCQM.py	Simulation code I for TLCQM
Sim_TLCQM_Ratio.py	Simulation code II for TLCQM
Sim_Compare.py	Comparison code I with baseline methods
Sim_Compare_Ratio.py	Comparison code I with baseline methods
Syn_Sim_Res.py	Synthetic simulation result processing

Real-World Data Experiments

File	Description
Apartment_TLCQM.py	TLCQM experiments on apartment rental data
Apartment_Compare.py	Baseline comparisons on apartment rental data

Visualization

File	Description
Figure_Plotting.ipynb	Jupyter notebook for generating publication figures

SLURM Batch Scripts

Each .py experiment file has a corresponding .sbatch file for HPC cluster submission.

Usage

Basic Example

import numpy as np
import torch
from TLCQM import fit_TLCQM
from utils import sim_data

np.random.seed(123)

# Generate synthetic data with covariate and concept shift
dat_source, dat_target, dat0_full, dat_test0 = sim_data(
    n_s=1000,           # samples per source
    n_0=50,             # labeled target samples
    n_test=5000,        # test samples
    sig=0.5,            # noise std
    mu_s=np.ones(5),    # source covariate mean
    mu_t=np.zeros(5),   # target covariate mean
    Sigma=np.eye(5),    # covariate covariance
    beta1=1/np.arange(1, 6)  # response coefficients
)

# Fit TLCQM model
Y_matched, beta_hat = fit_TLCQM(
    dat_source=dat_source,
    dat_target=dat_target,
    X_dat_tensor=None,
    n_sampler=3000,
    random_state=42,
    # Engression model hyperparameters
    eng_num_layer=2,
    eng_hidden_dim=100,
    eng_noise_dim=5,
    eng_lr=0.001,
    eng_num_epochs=1000,
    eng_pred_sample_size=500,
    # Quantile matching hyperparameters
    qm_stop_eps=1e-8,
    qm_max_iter=1000,
    qm_positive=False,
    qm_verbose=False
)

# Y_matched: calibrated responses for X_dat_tensor
# beta_hat: estimated quantile matching coefficients

Running Experiments on HPC

# Run simulation experiment
sbatch Sim_TLCQM.sbatch

# Run apartment data experiment
sbatch Apartment_TLCQM.sbatch

Data

The data/ directory contains:

• apartments_for_rent_classified_100K.csv: Real-world apartment rental dataset for experiments

Contribute

See the CONTRIBUTING file for how to help out.

License

TLCQM is MIT licensed, as found in the LICENSE file.