Ensemble learning¶

In [1]:
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt

from sklearn.model_selection import train_test_split
from sklearn import datasets
from sklearn.tree import DecisionTreeClassifier
from sklearn.ensemble import BaggingClassifier, RandomForestClassifier, GradientBoostingClassifier

Classification by optical recognition of handwritten digits¶

The digits dataset has 1797 labeled images of hand-written digits.

  • $X$ = digits.data has shape (1797, 64).
    • Each image $\mathbf{x}_i$ is represented as the $i$th row of 64 pixel values in the 2D digits.data array that corresponds to an 8x8 photo of a handwritten digit.
    • We can revover the $i$th 8x8 matrix via digits.data[i].reshape(8, 8).
  • $y$ = digits.target has shape (1797,). Each $y_i$ is a number from 0 to 9 indicating the handwritten digit that was photographed and stored in $\mathbf{x}_i$.
    • So this is a multiclass classification problem with $C = 10$ classes.

First we explore the data.

In [2]:
digits = datasets.load_digits()
In [3]:
print(f'digits.data.shape={digits.data.shape}')
print(f'digits.target.shape={digits.target.shape}')
row_0 = digits.data[0, :]
print(f'The first image is digits.data[0] (shown here as 8x8 instead of 1x64):')
print(f'The first target is y={digits.target[0]}.')
print(f'{row_0.reshape(8, 8)}')

digits.data.shape=(1797, 64)
digits.target.shape=(1797,)
The first image is digits.data[0] (shown here as 8x8 instead of 1x64):
The first target is y=0.
[[ 0.  0.  5. 13.  9.  1.  0.  0.]
 [ 0.  0. 13. 15. 10. 15.  5.  0.]
 [ 0.  3. 15.  2.  0. 11.  8.  0.]
 [ 0.  4. 12.  0.  0.  8.  8.  0.]
 [ 0.  5.  8.  0.  0.  9.  8.  0.]
 [ 0.  4. 11.  0.  1. 12.  7.  0.]
 [ 0.  2. 14.  5. 10. 12.  0.  0.]
 [ 0.  0.  6. 13. 10.  0.  0.  0.]]
In [4]:
_ = plt.imshow(digits.data[0].reshape(8, 8), cmap='binary') # 'binary' gives greyscale
In [5]:
# now show the first image again in the context of the first 40 images:
fig = plt.figure(figsize=(10, 4)) # new blank figure
gs = fig.add_gridspec(nrows=4, ncols=10) # grid of plot axes
for i in np.arange(40): # i goes from 0 to 39; or if we consider it as a 2-digit number, from 00 to 39
    # In the next line, the row is i's first digit and the column is i's second digit.
    # e.g. Image 23 goes in row 2 and column 3 of the 10x4 plot.
    ax = fig.add_subplot(gs[i // 10, i % 10])
    ax.text(x=0, y=1, s=digits.target[i], color='red')
    ax.imshow(digits.data[i].reshape(8, 8), cmap='binary')

We can see from digits.target that the first three rows are repeats of 0 through 9, but the fourth row seems to be random digits:

In [6]:
first_40_y_values = digits.target[0:40].copy()
first_40_y_values.reshape(4, 10)
Out[6]:
array([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
       [0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
       [0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
       [0, 9, 5, 5, 6, 5, 0, 9, 8, 9]])

Here we split the digits data into training, validation, and test sets:

In [7]:
X = digits.data
y = digits.target
# split 80% training data, 20% "_tmp" for validation & test
X_train, X_tmp, y_train, y_tmp = train_test_split(X, y, test_size=.2,
                                                  random_state=0, stratify=y)
# of remaining 20%, split in half to get 10% validation, 10% test
X_valid, X_test, y_valid, y_test = train_test_split(X_tmp, y_tmp, test_size=.5,
  random_state=0, stratify=y_tmp)

Here we try ensemble learning.¶

We will try each of bagging, random forest, and gradient boosting to see whether they improve performance over a basic decision tree.

In [8]:
# Here is a basic ID3 DecisionTree from long ago in unit 3.
clf = DecisionTreeClassifier(criterion='entropy', max_depth=8)
clf.fit(X_train, y_train)
validation_score = clf.score(X_valid, y_valid)
print(f'DecisionTree validation_score={validation_score:.3}')

# Here is Bagging with the DecisionTree and some default values.
clf = BaggingClassifier(estimator=DecisionTreeClassifier(max_depth=8),
                        n_estimators=100, random_state=0)
clf.fit(X_train, y_train)
validation_score = clf.score(X_valid, y_valid)
print(f'Bagging validation_score={validation_score:.3}')

# Here is a RandomForest. I tuned the max_depth=8 by hand, starting
# with 2, getting lousy results, and finding 8 was best of 2 through 9.
clf = RandomForestClassifier(max_depth=8, random_state=0)
clf.fit(X_train, y_train)
validation_score = clf.score(X_valid, y_valid)
print(f'RandomForest validation_score={validation_score:.3}')

# Here is GradientBoosting with default values.
clf = GradientBoostingClassifier(n_estimators=100, learning_rate=1.00,
    max_depth=1, random_state=0)
clf.fit(X_train, y_train)
validation_score = clf.score(X_valid, y_valid)
print(f'GradientBoosting validation_score={validation_score:.3}')

DecisionTree validation_score=0.872
Bagging validation_score=0.956
RandomForest validation_score=0.972
GradientBoosting validation_score=0.183

With very little effort, Bagging and RandomForest improved upon the basic DecisionTree. Hyperparameter tuning (via GridSearchCV() or RandomizedSearchCV()) might help further.

This run of GradientBoosting is a disaster! The problem, I think (from code experiments in class) is the learning rate $\alpha = 1.0$ is too large. I may add code later to do hyperparameter tuning to find a good value and see that GradientBoosting can help too.

(In fact $\alpha = 0.25$ yields 0.956, but I will leave the learning_rate=1.0 code here to show the failure with $\alpha=1.0$.)

From the previous four models, choose the best (on validation data) and evaluate it on test data.¶

In [9]:
clf = RandomForestClassifier(max_depth=8, random_state=0)
clf.fit(X_train, y_train)
validation_score = clf.score(X_valid, y_valid)
print(f'RandomForest validation_score={validation_score:.3}')
test_score = clf.score(X_test, y_test)
print(f'RandomForest test_score={test_score:.3}')

RandomForest validation_score=0.972
RandomForest test_score=0.956