log-reg-dre 0.1.0

Combines logistic regression for estimating density ratios with RU regression to mitigate conditional and covariate shift

LogReg DRE (Logistic Regression for Density Ratio Estimation)

LogReg DRE is a Python package for implementing regression models with a focus on handling covariate shift via density ratio estimation using logistic regression. This approach is particularly useful in scenarios where the training and test data distributions differ significantly.

Features

• Data Generation: Customizable data generation for simulating training and testing datasets with different distributions.
• Density Ratio Estimation: Utilizes logistic regression to estimate density ratios between training and test datasets, which are then used as weights in the loss function to mitigate covariate shift.
• Regression Models: Includes implementation of standard regression models and RU regression models that leverage estimated density ratios.
• Visualization: Functionality to visualize training loss and model predictions to evaluate performance and the impact of covariate shift correction.

Installation

https://github.com/szhang120/log-reg-DRE.git cd log_reg_dre pip install .

Usage

The main components of LogReg DRE are contained within the package directory, which includes modules for data generation, model definition, density ratio estimation, and training routines.

Data Generation

from package.data import generate_x_vals x_train = generate_x_vals(P_TRAIN, NUM_PTS) x_test = generate_x_vals(P_TEST, NUM_PTS)

Density Ratio Estimation

from package.density_ratio_estimation import train_logistic_regression, estimate_density_ratio classifier = train_logistic_regression(x_train, x_test) density_ratios = estimate_density_ratio(x_train, classifier)

Model Training

from package.training import train_standard_model, train_RU_model std_model, std_loss = train_standard_model(X_train, Y_train, classifier) ru_models, ru_loss = train_RU_model(X_train, Y_train, gamma_values, classifier)

Visualization

Refer to the visualization section in main.py for examples on how to plot training losses and model predictions.

License

This project is licensed under the MIT License - see the LICENSE file for details.