README.md

Minimax-Risk-Classifier

Different variations of minimax risk classifiers(MRC) using different loss functions and uncertainity set of distributions.

The variants available here are -

1. MRC with 0-1 loss (MRC.py)
2. MRC with log loss (MRC.py)
3. MRC with 0-1 loss and fixed instances' marginals (CMRC.py)
4. MRC with log loss and fixed instances' marginals (CMRC.py)

Requirements

We will need have installed the following libraries:

• numpy
• sklearn
• cvxpy

We can install the requirements directly from the file "requirements.txt"

pip install -r requirements.txt

Usage

Fitting the models

MRC with 0-1 loss

clf = MRC(r=r, loss='0-1').fit(X, Y)

MRC with log loss

clf = MRC(r=r, loss='log').fit(X, Y)

MRC with 0-1 loss and fixed instances' marginals

clf = CMRC(r=r, loss='0-1').fit(X, Y)

MRC with log loss and fixed instances' marginals

clf = CMRC(r=r, loss='log').fit(X, Y)

A small training example of MRC

from MRC import MRC
from datasets import load_mammographic

X, Y = load_mammographic(return_X_y=True)
r = len(np.unique(Y))
clf = MRC(r=r).fit(X, Y)
y_pred = clf.predict(X[:2,:])

Bounds on the error for MRC

clf = MRC(r=r).fit(X, Y)
upper_bound = clf.upper
lower_bound = clf.getLowerBound()

Only available for the MRC class

Setting the interval estimates while fitting

Using instances X_ and Y_ to calculate tau and lambda

clf = MRC(r=r).fit(X, Y, X_, Y_)

By passing the values for tau and lambda

clf = MRC(r=r).fit(X, Y, _tau=0.5, _lambda=0.1)