ICPOptimize

The Iterative Constrained Pathways Optimizer

ICP is a constrained linear model optimizer built with a focus on memory efficiency, flexibility, and solution interpretability.

Description

This repository contains implementations of the Iterative Constrained Pathways (ICP) optimization method, the ICP Rule Ensemble (ICPRE), linear classifier, regressor, and other methods. Currently, hinge, least-squares, and absolute-value loss modes are supported, with support for other loss functions planned. Coefficients can be constrained by sign or by arbitrary intervals. L1 & L2 norm constraints are also supported.

Further discussion about and motivation for the methods can be found on my blog:

nogilnick.github.io/Posts/63.html

Features

• Linear Classification using Hinge Loss
• Regression Support using L1 and L2 Penalties
• Arbitrary Interval Constraints
• L1 and L2 Coefficient Norm Constraints
• Useful Default Settings
• Support for DataFrames and Sparse Matrices

Installation

Install via PyPi:

pip install ICPOptimize

PyPi Project:

https://pypi.org/project/ICPOptimize/

Examples

Rule Ensemble Classifier

from ICP.Models import ICPRuleEnsemble

...

IRE = ICPRuleEnsemble().fit(A[trn], Y[trn])
YP  = IRE.predict_proba(A)

Linear Model

from ICP.Models import ICPLinearRegressor

...
# Fit linear regressor with absolute loss and L_1 norm <= 10
ILR = ICPLinearRegressor(p='l1', L1=10.0).fit(A[trn], Y[trn])
YP  = ILR.predict(A)

Further examples are available on the ICPExamples GitHub page:

https://github.com/nogilnick/ICPExamples