The contents of this repository cover the experiments described in the paper:

A Probabilistic Theory of Supervised Similarity Learning for Pointwise ROC Curve Optimization. Robin Vogel, Aurélien Bellet, Stéphan Clémençon ; Proceedings of the 35th International Conference on Machine Learning, PMLR 80:5062-5071, 2018.

Three experiments are described in the paper:

• Experiment 1: Pointwise ROC optimization - Describes how we can optimize for a specific point of the ROC curve, with guarantees derived from our analysis, in a very simple special case.
• Experiment 2: Fast Rates - Shows that the fast rates can be illustrated with simples distributions, when we satisfy the assumptions of a Mammen-Tsybakov type assumption.
• Experiment 3: Scalability by sampling - Shows that for the MMC algorithm, which is a metric learning objective which formulation is very close to our problem, subsampling very agressively the negative pairs does not hinder learning.

• Experiment 1: numpy, matplotlib.
• Experiment 2: numpy, matplotlib, pandas, scipy.stats
• Experiment 3: numpy, matplotlib, scikit-learn, autograd, configargparse.

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