Computing Full Conformal Prediction Set with Approximate Homotopy.

This package implements a computation of conformal prediction set based on approximate solution of empirical risk minimization when the input observations changes sequentially. We develop a new homotopy continuation technique so that only a finite number of computation are needed to approximate infinitely many solution on a given interval [ymin, ymax].

Installation & Requirements

This package has the following requirements:

• numpy
• scipy
• scikit-learn
• python-intervals

We recommend to install or update anaconda to the latest version and use Python 3.

Reproducibility

• Figure 1 is generated by visualization_ridge_conf.py
• Table 1: the left figure is generated by bench_homotopy_time.py and the right figure by benchmark.py with (method = "lasso", dataset = "climate", alpha = 0.1)
• Figure 2 is generated by benchmark.py with the range of coverage level alphas = np.arange(1, 10) / 10. The figure (a) corresponds to (method = "lasso", dataset = "diabetes") and the figure (b) corresponds to (method = "logcosh", dataset = "boston").
• Table 2 is generated with alpha = 0.1 by benchmark.py with (method = "logcosh", dataset = "boston") and (method = "linex", dataset = "diabetes").