Double-Weighting for General Covariate Shift Adaptation

This repository is the official implementation of Double-Weighting for General Covariate Shift Adaptation.

The algorithm proposed in the paper provides efficient learning for the proposed Doble-Weighting for General Covariate Shift (DW-GCS). In particular, we first compute weights $\alpha$ and $\beta$ by solving the Double-Weighting Kernel Mean Matching (DW-KMM). Then, we learn the classifier's parameters by solving the MRC using double-weighting.

Python Code

Requirements

The standard libraries required are listed in the file requirements.txt. To install these libraries using

1. pip

pip install -r requirements.txt

2. conda environment

conda create --name <environment_name> --file requirements.txt

In addition, the implementation of the proposed algorithm utilizes the MOSEK optimizer for which license can be downloaded from here.

Code

The functions in the folder Python_code folder contains experiment scripts required to replicate the experiments of the paper.

• Experiment_Features.py, Experiment_PCA.py and Experiment_20News.py are the main files.
• Experiment_Features.py file perform a experiment for datasets in which we artificially introduce the covariate shift based on the median of the features.
• Experiment_PCA.py file perform a experiment for datasets in which we artificially introduce the covariate shift based on the median of the first principal component of the features.
• Experiment_20News.py file perform a experiment using the “News20groups” dataset that is intrinsically affected by a covariate shift since the training and testing partitions correspond to different times.
• In such files we can modify the feature mapping, and the loss we are going to use. We can also choose if the classifier is deterministic or not.
• The file CovShiftGen.py generate covariate shift in the datasets. In particular, we select training and testing samples with different probabilities based on the medians of the first 3 features, and based on the median of the first principal component of features.
• The folder Auxiliary_Functions contains the function phi.py that calculates the feature mappings using linear, polinomial or random Fourier features (RFF).
• The folder DWGCS contains the functions of the Double-Weighting for General Covariate Shift Adaptation.
• DWGCS.DWKMM computes the estimated weights $\beta$ and $\alpha$ solving the double-weighting kernel mean matching.
• DWGCS.parameters obtains mean vector estimate $\tau$ and confidence vector $\lambda$.
• DWGCS.learning solves the convex MRC optimization problem using double-weighting and obtains the classifier parameters.
• DWGCS.prediction assigns labels to instances and gives the classification error.

Matlab Code

Matlab_code folder contains Matlab scripts required to execute the method:

• run_DWGCS_example1.m and run_DWGCS_example1.m are the main files.
• run_DWGCS_example1.m file perform a experiment for datasets in which we artificially introduce the covariate shift.
• run_DWGCS_example2.m file perform a experiment using the “News20groups” dataset that is intrinsically affected by a covariate shift since the training and testing partitions correspond to different times.
• In such files we can modify the feature mapping, and the loss we are going to use. We can also choose if the classifier is deterministic or not.
• The functions in the folder CovShift_Generation generate covariate shift in the datasets. In particular, we select training and testing samples with different probabilities based on the medians of the first 3 features, and based on the median of the first principal component of features.
• phi.m calculates the feature mappings using linear, polinomial or random Fourier features (RFF).
• DWGCS_weights.m computes the estimated weights $\beta$ and $\alpha$ solving the double-weighting kernel mean matching.
• DWGCS_parameters.m obtains mean vector estimate $\tau$ and confidence vector $\lambda$.
• DWGCS_learning.m solves the convex MRC optimization problem using double-weighting and obtains the classifier parameters.
• DWGCS_prediction.m assigns labels to instances and gives the classification error.

Data

The repository contains multiple datasets to perform experiments as follows -

Dataset	Covariates	Samples	Ratio of majority class	$\sigma$
Blood	3	748	76.20%	0.7491
BreastCancer	9	683	65.01%	1.6064
Haberman	3	306	75.53%	1.3024
Ringnorm	20	7400	50.49%	3.8299
comp vs sci	1000	5309 (tr) / 3534 (te)	55.31%	23.5628
comp vs talk	1000	4888 (tr) / 3256 (te)	60.06%	23.4890
rec vs sci	1000	4762 (tr) / 3169 (te)	50.17%	24.5642
rec vs talk	1000	4341 (tr) / 2891 (te)	55.02%	25.1129
sci vs talk	1000	4325 (tr) / 2880 (te)	54.85%	24.8320

The datasets are available as csv files in this repository in the Datasets and Datasets_20Newsfolders.

Support and Authors

José I. Segovia-Martín

jsegovia@bcamath.org

Santiago Mazuelas

smazuelas@bcamath.org

Anqi Liu

aliu@cs.jhu.edu

Citation

If you find useful the code in your research, please include explicit mention of our work in your publication with the following corresponding entry in your bibliography:

[1] José I. Segovia-Martín, S. Mazuelas, A. Liu "Double-Weighting for Covariate Shift Adaptation". Proceedings of the 40th International Conference on Machine Learning. PMLR, 2023.

The corresponding BiBTeX citation is given below:

@inproceedings{SegMazLiu2023, title={Double-Weighting for Covariate Shift Adaptation}, author={Segovia-Mart'{i}n, Jos'{e} I. and Mazuelas, Santiago and Liu, Anqi}, booktitle={International Conference on Machine Learning}, pages={30439--30457}, year={2023}, organization={PMLR} }