An important factor to guarantee a responsible use of data-driven recommendation systems is that we should be able to communicate their uncertainty to decision makers. This can be accomplished by constructing prediction intervals, which provide an intuitive measure of the limits of predictive performance.
This package contains a Python implementation of Conformalized quantile regression (CQR) [1] methodology for constructing marginal distribusion-free prediction intervals. It also implements the equalized coverage framework [2] that builds valid group-conditional prediction intervals.
CQR is a technique for constructing prediction intervals that attain valid coverage in finite samples, without making distributional assumptions. It combines the statistical efficiency of quantile regression with the distribution-free coverage guarantee of conformal prediction. On one hand, CQR is flexible in that it can wrap around any algorithm for quantile regression, including random forests and deep neural networks. On the other hand, a key strength of CQR is its rigorous control of the miscoverage rate, independent of the underlying regression algorithm.
[1] Yaniv Romano, Evan Patterson, and Emmanuel J. Candes, “Conformalized quantile regression.” 2019.
To support equitable treatment, the equalized coverage methodology forces the construction of the prediction intervals to be unbiased in the sense that their coverage must be equal across all protected groups of interest. Similar to CQR and conformal inference, equalized coverage offers rigorous distribution-free guarantees that hold in finite samples. This methodology can also be viewed as a wrapper around any predictive algorithm.
[2] Y. Romano, R. F. Barber, C. Sabbatti and E. J. Candès, “With malice towards none: Assessing uncertainty via equalized coverage.” 2019.
This package is self-contained and implemented in python.
Part of the code is a taken from the nonconformist package available at https://github.com/donlnz/nonconformist. One may refer to the nonconformist repository to view other applications of conformal prediction.
- python
- numpy
- scipy
- scikit-learn
- scikit-garden
- pytorch
- pandas
The development version is available here on github:
git clone https://github.com/yromano/cqr.gitPlease refer to cqr_real_data_example.ipynb for basic usage. Comparisons to competitive methods and additional usage examples of this package can be found in cqr_synthetic_data_example_1.ipynb and cqr_synthetic_data_example_2.ipynb.
The notebook detect_prediction_bias_example.ipynb performs simple data analysis for MEPS 21 data set and detects bias in the prediction. The notebook equalized_coverage_example.ipynb illustrates how to run the methods proposed in [2] and construct prediction intervals with equal coverage across groups.
The code available under /reproducible_experiments/ in the repository replicates the experimental results in [1] and [2].
-
Blog: BlogFeedback data set.
-
Bio: Physicochemical properties of protein tertiary structure data set.
-
Bike: Bike sharing dataset data set.
-
Community: Communities and crime data set.
-
STAR: C.M. Achilles, Helen Pate Bain, Fred Bellott, Jayne Boyd-Zaharias, Jeremy Finn, John Folger, John Johnston, and Elizabeth Word. Tennessee’s Student Teacher Achievement Ratio (STAR) project, 2008.
-
Concrete: Concrete compressive strength data set.
-
Facebook Variant 1 and Variant 2: Facebook comment volume data set.
The Medical Expenditure Panel Survey (MPES) data can be downloaded using the code in the folder /get_meps_data/ under this repository. It is based on this explanation (code provided by IBM's AIF360).
-
MEPS_19: Medical expenditure panel survey, panel 19.
-
MEPS_20: Medical expenditure panel survey, panel 20.
-
MEPS_21: Medical expenditure panel survey, panel 21.
This project is licensed under the MIT License - see the LICENSE file for details.
