CoAI

Main package with tools to make any predictive model cost-aware. To install, run python setup.py install from the main directory; python setup.py install --user for local install. Installation should take only a few minutes.

For more examples, see experiments, in particular experiments/ED Trauma.ipynb and experiments/NHANES.ipynb.

Basic Usage:

from coai import CoAIOptimizer
import numpy as np
import shap
import xgboost as xgb

# Random data
Xtrain = np.random.random((100,10))
Xtest = np.random.random((100,10))
ytrain = np.random.choice([0,1],size=100)
ytest = np.random.choice([0,1],size=100)

# Features fall into two groups, each with a randmly assigned cost
costs = np.hstack(([np.random.random()]*5,[np.random.random()]*5))
groups = np.hstack(([1]*5,[2]*5))

# Train models at three cost thresholds
thresholds = np.array([0.5,1.25,1.75])

# Base model and explainer
bst = xgb.XGBClassifier(tree_method='approx',silent=1)
explainer = shap.TreeExplainer

# Train CoAI
CoAI = CoAIOptimizer(bst,explainer,scale_ints=10000)
CoAI.fit(Xtrain,ytrain,
    feature_costs=costs,
    feature_groups=groups,
    thresholds=thresholds)

# Generate predictions with budget 1.25
deployment_threshold=1.25
preds = CoAI.predict(Xtest,max_cost=deployment_threshold)

Dependencies (and versions known to work):

• Python 3 (3.6)
• Numpy (1.17.0)
• Pandas (0.25.1)
• Scikit-Learn (0.23)
• SHAP (0.35.0)
• LightGBM (2.3.0) -- required for coai.cegb module wrapping Cost-Efficient Gradient Boosting
• Matplotlib (3.1.1)
• TQDM (4.49.0)

The coai.cwcf module wrapping the RL-based Classification With Costly Features requires successful installation of the CWCF repository

CoAI should work on any operating system but has been tested so far on Ubuntu and CentOS.