Gradient Boosting
Gradient Boosting is a Machine Learning result improvement methodology with these characteristics:
The objective is to improve prediction results, that is, to reduce the error gradient ... that is, to boost the results.
Gradient boosting uses an ensemble of models to achieve its result.
Starts with a weak learner, usually a decision tree.
Each model is created sequentially based on the results of the previous model.
Stochastic gradient descent can be used to minimize the loss when adding models.
Python Example
To see details of the GradientBoostingRegressor model, click here.
To download the code below, click here.
""" gradient_boosting_using_scikit_learn.py uses a GradientBoostingRegressor model to demonstrate gradient boosting """ # Import needed libraries import numpy as np import matplotlib.pyplot as plotlib from sklearn import ensemble from sklearn import datasets from sklearn.utils import shuffle from sklearn.metrics import mean_squared_error # Define the number of boosting stages. # Gradient boosting is fairly robust to over-fitting, # so a larger number can result in better performance. number_of_estimators = 500 # Define the maximum number of tree nodes. maximum_depth = 4 # Define the minimum number of samples required to split an internal node. minimum_sample_split = 2 # Define the learning rate. # Learning rate shrinks the contribution of each tree by learning_rate. # There is a trade-off between learning_rate and n_estimators. learning_rate = 0.01 # Define the loss function to be optimized. # In this example, use the least squares regression. loss_function = 'ls' # Define an offset for training and test data. # Needs to be greater than zero and less than one. offset_value = 0.4 # Define a random state value for shuffling data into X and y values. random_state_value = 13 # Load a test dataset. dataset = datasets.load_diabetes() # Shuffle and separate data into X and y values. X, y = shuffle(dataset.data, dataset.target, random_state=random_state_value) print("X:") print(X) print("y:") print(y) # Convert the X data to floating point. X = X.astype(np.float32) # Define an offset based on the offset_value parameter. offset = int(X.shape[0] * offset_value) print("offset:") print(offset) # Create training and test data. X_train, y_train = X[:offset], y[:offset] print("X_train:") print(X_train) X_test, y_test = X[offset:], y[offset:] print("X_test:") print(X_test) # Instantiate a gradient boosting regression model. model = ensemble.GradientBoostingRegressor(n_estimators=number_of_estimators, max_depth=maximum_depth, min_samples_split=minimum_sample_split, learning_rate=learning_rate, loss=loss_function) # Train the model. model.fit(X_train, y_train) # Predict values based on X_test data. X_predictions = model.predict(X_test) # Calculate the mean squared error of X_predictions. mean_squared_error_value = mean_squared_error(y_test, X_predictions) print("X_prediction mean_squared_error:") print(mean_squared_error_value) # Get the training set deviance (loss) score. train_score = model.train_score_ # Get the test set deviance (loss) score. test_score = np.zeros((number_of_estimators,), dtype=np.float64) for prediction, y_pred in enumerate(model.staged_predict(X_test)): test_score[prediction] = model.loss_(y_test, y_pred) # Plot the deviance score curves. plotlib.figure(figsize=(24, 6)) plotlib.plot(np.arange(number_of_estimators) + 1, train_score, 'b-', label='Training Set Deviance (Loss)') plotlib.plot(np.arange(number_of_estimators) + 1, test_score, 'r-', label='Test Set Deviance (Loss)') plotlib.legend(loc='upper right') plotlib.xlabel('Gradient Boosting Iteration Stages') plotlib.ylabel('Deviance (Loss)') # Display the plot. plotlib.show()
Output is shown below:
X:
[[1.50234e+01 0.00000e+00 1.81000e+01 ... 2.02000e+01 3.49480e+02
2.49100e+01]
[5.44114e+00 0.00000e+00 1.81000e+01 ... 2.02000e+01 3.55290e+02
1.77300e+01]
[1.00245e+00 0.00000e+00 8.14000e+00 ... 2.10000e+01 3.80230e+02
1.19800e+01]
...
[7.89600e-02 0.00000e+00 1.28300e+01 ... 1.87000e+01 3.94920e+02
6.78000e+00]
[7.02200e-02 0.00000e+00 4.05000e+00 ... 1.66000e+01 3.93230e+02
1.01100e+01]
[3.30600e-02 0.00000e+00 5.19000e+00 ... 2.02000e+01 3.96140e+02
8.51000e+00]]
y:
[12. 15.2 21. 24. 19.4 22.2 23.3 15.6 20.8 13.8 19.6 27.1 36.5 15.2
11.7 14.1 17.2 16.8 32.9 21.4 32.4 23.5 20.4 13.1 12.6 10.4 50. 23.1
13.4 24.3 25. 7.4 7. 22. 15.3 8.4 16.4 18.1 43.8 8.5 18.6 21.1
50. 11.8 17.4 33.3 14.8 8.8 26.6 16.8 30.1 23.7 50. 19.5 16.1 24.1
20.4 36.4 41.3 21.7 21.7 14. 21.7 20.4 20. 34.7 24.5 11.7 14.3 13.1
17.4 20.1 19.5 21. 30.1 18.4 34.6 20.1 43.5 21.6 18.3 21.4 18.9 13.4
30.8 25. 25.2 8.8 31.1 13.4 48.3 17.8 5.6 12.7 16.1 20.9 19.9 13.9
22.6 21.2 21.2 22.9 20.5 22.8 19.4 21.7 23.1 26.5 18.5 20.2 27.5 50.
21.9 23.4 32.7 14.9 15.6 20.3 11.9 30.5 31.6 21.9 25. 23. 17. 7.2
44.8 16. 38.7 20.4 22.5 21.7 12.7 5. 21.4 23.7 21. 19.5 20.1 24.6
36.1 23. 18.5 32.5 19.1 23.3 18.5 21.5 19.3 26.4 31. 22.3 13.3 7.
22.5 27.5 30.1 10.2 20. 25. 17.8 13.8 32. 23.7 23.8 16.7 23.8 18.8
22. 29. 21.2 33.1 30.7 24.8 21.7 37.6 23.1 22.9 13.6 14.6 18.9 22.6
31.7 19.4 12.7 20.1 30.3 18.8 8.1 20.6 33.2 21.1 31.5 20. 8.5 22.2
24.8 50. 21.7 24. 15. 13.5 9.7 18.9 22.6 29.6 20.6 24.3 16.2 19.6
35.1 17.5 12.5 22.2 22.9 34.9 28. 17.4 7.2 23.4 21.2 27.9 20.2 22.7
26.2 50. 32. 20.7 15. 17.2 23.1 10.9 21.5 17.2 35.2 10.9 23.8 17.8
25. 10.5 26.6 11.9 20.3 21.9 14.5 10.8 23.1 25. 14.9 6.3 24.2 13.2
24.7 19.8 18.5 23.9 29.6 18.7 29.1 10.5 32.2 50. 35.4 7.5 16.3 25.
25.3 19.1 28.7 14.3 23.1 19.8 17.5 20. 8.3 23.2 26.7 17.8 19.3 18.
10.2 22.2 28.4 21.2 11. 34.9 36.2 19.7 22.5 18.7 29. 13.5 22.4 18.4
36.2 28.6 14.1 33. 50. 19.1 24.7 24.5 19. 23.3 22.9 28.2 24.1 26.4
50. 48.5 11.3 29.4 10.2 13.6 13. 24.4 15.6 9.6 22.3 19.9 46.7 19.2
20.7 24.4 5. 22.8 19.1 29.8 13.8 18.2 46. 18.3 29.8 14.2 21.4 19.6
19.3 20. 24.8 37.9 24.8 24.6 22.6 16.1 10.4 14.1 23.9 50. 25. 19.6
18.6 16.5 33.4 19.4 20.6 15.4 20.5 22.4 28.7 20.5 18.2 19.3 24.4 22.
13.8 14.5 50. 41.7 22. 20.8 12.3 42.8 23.6 23.9 23. 14.4 22.8 50.
16.6 19.9 20.1 24.7 22.1 12.1 42.3 17.1 24.4 29.9 17.1 22. 20.6 35.4
33.4 19. 34.9 15.1 22. 33.8 8.7 27.9 33.2 37.3 7.2 19.7 31.6 50.
12.8 22.7 23.3 13.3 20.3 24.5 19.6 16.6 11.8 50. 13.9 20.8 19.5 33.1
14.4 19.3 16.2 13.1 23.9 19.2 20.6 21.8 20.3 23.6 28.7 26.6 44. 43.1
14.6 27.5 16.7 37. 19.8 29.1 27.5 23.2 13.3 50. 50. 16.5 23.7 14.9
48.8 17.3 23.2 22.2 9.5 18.7 20.9 15.6 28.4 28.1 31.2 13.1 37.2 22.
11.5 13.8 39.8 28.5 15.2 23.8 19.4 27.1 18.9 17.9 45.4 15.6 21.6 21.4
19.9 17.8 23. 15.4 8.3 27. 36. 22.8 17.1 22.6 23.9 17.7 31.5 8.4
14.5 13.4 15.7 17.5 15. 21.8 18.4 25.1 19.4 17.6 18.2 24.3 23.1 24.1
23.2 20.6]
offset:
202
X_train:
[[1.50234e+01 0.00000e+00 1.81000e+01 ... 2.02000e+01 3.49480e+02
2.49100e+01]
[5.44114e+00 0.00000e+00 1.81000e+01 ... 2.02000e+01 3.55290e+02
1.77300e+01]
[1.00245e+00 0.00000e+00 8.14000e+00 ... 2.10000e+01 3.80230e+02
1.19800e+01]
...
[3.30450e-01 0.00000e+00 6.20000e+00 ... 1.74000e+01 3.76750e+02
1.08800e+01]
[1.96091e+01 0.00000e+00 1.81000e+01 ... 2.02000e+01 3.96900e+02
1.34400e+01]
[1.61282e+00 0.00000e+00 8.14000e+00 ... 2.10000e+01 2.48310e+02
2.03400e+01]]
X_test:
[[1.15779e+01 0.00000e+00 1.81000e+01 ... 2.02000e+01 3.96900e+02
2.56800e+01]
[1.70040e-01 1.25000e+01 7.87000e+00 ... 1.52000e+01 3.86710e+02
1.71000e+01]
[4.26131e+00 0.00000e+00 1.81000e+01 ... 2.02000e+01 3.90740e+02
1.26700e+01]
...
[7.89600e-02 0.00000e+00 1.28300e+01 ... 1.87000e+01 3.94920e+02
6.78000e+00]
[7.02200e-02 0.00000e+00 4.05000e+00 ... 1.66000e+01 3.93230e+02
1.01100e+01]
[3.30600e-02 0.00000e+00 5.19000e+00 ... 2.02000e+01 3.96140e+02
8.51000e+00]]
X_prediction mean_squared_error:
19.748551548039462