Polyfit
Scikit learn compatible constrained and robust polynomial regression in Python.
Mostly developed for educational purposes, polyfit enables fitting scikit learn compatible polynomial regression models under shape constraints. Under the hood polynomial coefficients are optimized via cvxpy's excellent convex optimizers.
Why?
Often human intuition or prior knowledge gives us an idea that relationships between variables should be monotonic or follow certain asymptotic behaviour. In this example the monotonic fit is visually much more convincing than an unconstrained fit.
Installation:
git clone https://github.com/dschmitz89/Polyfit/
cd Polyfit
pip install .
Usage
Example usage can be found in polyfit/test.py which produces the above image.
Simple example to fit a polynomial of degree 3 which is monotonically increasing for the first feature:
from polyfit import PolynomRegressor, Constraints
polyestimator = PolynomRegressor(deg=3, regularization = None, lam = 0)
monotone_constraint = Constraints(monotonicity='inc')
polyestimator.fit(X, y, loss = 'l2', constraints={0: monotone_constraint})
PolynomRegressor attributes:
Attribute | Possible values |
---|---|
deg | int: Polynomial degree |
regularization | str: "l1" / "l2" |
lam | float: Regularization parameter |
Constraints attributes:
Attribute | Possible values |
---|---|
monotonicity | str: 'inc'/'dec' |
curvature | str: 'convex'/'concave' |
sign | str: 'positive'/'negative' |
constraint_range: | list (lower bound, upper bound): range in which the constraints should be enforced |
gridpoints | int: number of gridpoints for which the constraints are enforced by the optimizer (default: 20). Increase if constraints are not fulfilled. |
PolynomRegressor.fit(X, y) options:
Parameter | Possible values |
---|---|
loss | str: "l2" / "l1" / "huber" (default: "l2") |
interactions | bool: if True, interaction terms are incorporated (default: False) |
m | float: threshold between l1 and l2 loss for huber loss function |
constraints | dict: {feature_1: constraints_1, feature_2: constraints_2, ...}. Dictionary with pairs of feature index and instance of constraints class |
verbose | bool: if True, display optimizer progress (default: False) |
Method
The constraints are enforced by imposing inequality constraints upon the polynomial coefficients. For example, if the resulting one dimensional polynomial is required to be monotonically increasing, its first derivative must be greater than 0. Enforcing this for an interval is not possible but enforcing it for a reasonable number of points within an interval (default: 20) is usually enough to guarantee the monotonicity for this interval. Given the predictor vector x, target vector y and the Vandermonde matrix V the polynomial coefficients p are then estimated by the following optimization problem:
Warning: by default, the polynomial is only monotonic or convex/concave for the interval of the input data!