MSE example - test-train split

In this example we consider a polynomial linear regression, splitting the data into two sets. We measure the quality of the regression by comparing the predictions with the output on the test set.

from __future__ import annotations

import matplotlib.pyplot as plt
from gemseo import configure_logger
from gemseo import create_dataset
from gemseo.datasets.io_dataset import IODataset
from gemseo.mlearning import create_regression_model
from gemseo.mlearning.quality_measures.mse_measure import MSEMeasure
from numpy import arange
from numpy import argmin
from numpy import hstack
from numpy import linspace
from numpy import sort
from numpy.random import choice
from numpy.random import normal
from numpy.random import seed

configure_logger()

<RootLogger root (INFO)>

Define parameters

seed(12345)
n_samples = 10
noise = 0.3**2
max_pow = 5
amount_train = 0.8

Construct data

We construct a parabola with added noise, on the interval [0, 1].

def f(x):
    return -4 * (x - 0.5) ** 2 + 3


x = linspace(0, 1, n_samples)
y = f(x) + normal(0, noise, n_samples)

Indices for test-train split

samples = arange(n_samples)
n_train = int(amount_train * n_samples)
n_test = n_samples - n_train
train = sort(choice(samples, n_train, False))
test = sort([sample for sample in samples if sample not in train])
print("Train:", train)
print("Test:", test)

Train: [1 3 4 5 6 7 8 9]
Test: [0 2]

Build datasets

data = hstack([x[:, None], y[:, None]])
variables = ["x", "y"]
groups = {"x": IODataset.INPUT_GROUP, "y": IODataset.OUTPUT_GROUP}
dataset = create_dataset(
    "synthetic_data", data[train], variables, variable_names_to_group_names=groups
)
dataset_test = create_dataset(
    "synthetic_data", data[test], variables, variable_names_to_group_names=groups
)

Build regression model

model = create_regression_model("PolynomialRegressor", dataset, degree=max_pow)
print(model)

PolynomialRegressor(degree=5, fit_intercept=True, l2_penalty_ratio=1.0, penalty_level=0.0)
   based on the scikit-learn library

Predictions errors

measure = MSEMeasure(model)

mse_train = measure.evaluate_learn()
mse_test = measure.evaluate_test(dataset_test)

print("Training error:", mse_train)
print("Test error:", mse_test)

Training error: [0.0003947]
Test error: [2.29565983]

Compute predictions

measure = MSEMeasure(model)
model.learn()

n_refined = 1000
x_refined = linspace(0, 1, n_refined)
y_refined = model.predict({"x": x_refined[:, None]})["y"].flatten()

Plot data points

plt.plot(x_refined, f(x_refined), label="Exact function")
plt.scatter(x, y, label="Data points")
plt.legend()
plt.show()

Plot predictions

plt.plot(x_refined, y_refined, label=f"Prediction (x^{max_pow})")
plt.scatter(x[train], y[train], label="Train")
plt.scatter(x[test], y[test], color="r", label="Test")
plt.legend()
plt.show()

Compare different parameters

powers = [1, 2, 3, 4, 5, 7]
test_errors = []
for power in powers:
    model = create_regression_model("PolynomialRegressor", dataset, degree=power)
    measure = MSEMeasure(model)

    test_mse = measure.evaluate_test(dataset_test)
    test_errors += [test_mse]

    y_refined = model.predict({"x": x_refined[:, None]})["y"].flatten()

    plt.plot(x_refined, y_refined, label=f"x^{power}")

plt.scatter(x[train], y[train], label="Train")
plt.scatter(x[test], y[test], color="r", label="Test")
plt.legend()
plt.show()

Grid search

print(test_errors)
print("Power for minimal test error:", argmin(test_errors))

[array([0.54513687]), array([0.00518409]), array([0.00584647]), array([0.06387849]), array([2.29565983]), array([1.16961302])]
Power for minimal test error: 1