分類法/範例四: Classifier comparison

這個範例的主要目的

  • 比較各種分類器
  • 利用圖示法觀察各種分類器的分類邊界及區域

(一)引入函式並準備分類器

  • 將分類器引入之後存放入一個list裏
  • 這邊要注意 sklearn.discriminant_analysis 必需要 sklearn 0.17以上才能執行
  1. import numpy as np
  2. import matplotlib.pyplot as plt
  3. from matplotlib.colors import ListedColormap
  4. from sklearn.cross_validation import train_test_split
  5. from sklearn.preprocessing import StandardScaler
  6. from sklearn.datasets import make_moons, make_circles, make_classification
  7. from sklearn.neighbors import KNeighborsClassifier
  8. from sklearn.svm import SVC
  9. from sklearn.tree import DecisionTreeClassifier
  10. from sklearn.ensemble import RandomForestClassifier, AdaBoostClassifier
  11. from sklearn.naive_bayes import GaussianNB
  12. from sklearn.discriminant_analysis import LinearDiscriminantAnalysis
  13. from sklearn.discriminant_analysis import QuadraticDiscriminantAnalysis
  15. h = .02  # step size in the mesh
  17. names = ["Nearest Neighbors", "Linear SVM", "RBF SVM", "Decision Tree",
  18.          "Random Forest", "AdaBoost", "Naive Bayes", "Linear Discriminant Ana.",
  19.          "Quadratic Discriminant Ana."]
  20. classifiers = [
  21.     KNeighborsClassifier(3),
  22.     SVC(kernel="linear", C=0.025),
  23.     SVC(gamma=2, C=1),
  24.     DecisionTreeClassifier(max_depth=5),
  25.     RandomForestClassifier(max_depth=5, n_estimators=10, max_features=1),
  26.     AdaBoostClassifier(),
  27.     GaussianNB(),
  28.     LinearDiscriminantAnalysis(),
  29.     QuadraticDiscriminantAnalysis()]

(二)準備測試資料

  • 利用make_classification產生分類資料，n_features=2表示共有兩個特徵, n_informative=2 代表有兩個類別
  • 所產生之 X: 100 x 2矩陣，y: 100 元素之向量，y的數值僅有0或是1用來代表兩種類別
  • 利用X += 2 * rng.uniform(size=X.shape)加入適度的雜訊後將(X,y)資料集命名為linear_separable
  • 最後利用make_moon()及make_circles()產生空間中月亮形狀及圓形之數據分佈後，一併存入datasets變數
  1. X, y = make_classification(n_features=2, n_redundant=0, n_informative=2,
  2.                            random_state=1, n_clusters_per_class=1)
  3. rng = np.random.RandomState(2)
  4. X += 2 * rng.uniform(size=X.shape)
  5. linearly_separable = (X, y)
  7. datasets = [make_moons(noise=0.3, random_state=0),
  8.             make_circles(noise=0.2, factor=0.5, random_state=1),
  9.             linearly_separable
  10.             ]

(三)測試分類器並作圖

接下來這段程式碼有兩個for 迴圈，外迴圈走過三個的dataset，內迴圈則走過所有的分類器。
為求簡要說明，我們將程式碼簡略如下：

  1. 外迴圈：資料迴圈。首先畫出資料分佈，接著將資料傳入分類器迴圈
    1.  for ds in datasets:
    2.      X, y = ds
    3.      #調整特徵值大小使其在特定範圍
    4.      X = StandardScaler().fit_transform(X)
    5.      #利用train_test_split將資料分成訓練集以及測試集
    6.      X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=.4)
    7.      #產生資料網格來大範圍測試分類器，範例EX 3有詳述該用法
    8.      xx, yy = np.meshgrid(..........省略)
    9.      # 畫出訓練資料點
    10.      ax.scatter(X_train[:, 0], X_train[:, 1], c=y_train, cmap=cm_bright)
    11.      # 畫出測試資料點，用alpha=0.6將測試資料點畫的"淡"一些
    12.      ax.scatter(X_test[:, 0], X_test[:, 1], c=y_test, cmap=cm_bright, alpha=0.6)

  2. 內迴圈：分類器迴圈。測試分類準確度並繪製分類邊界及區域

    1.  for name, clf in zip(names, classifiers):
    2.      clf.fit(X_train, y_train)
    3.      score = clf.score(X_test, y_test)
    5.      # Plot the decision boundary. For that, we will assign a color to each
    6.      # point in the mesh [x_min, m_max]x[y_min, y_max].
    7.      if hasattr(clf, "decision_function"):
    8.          Z = clf.decision_function(np.c_[xx.ravel(), yy.ravel()])
    9.      else:
    10.          Z = clf.predict_proba(np.c_[xx.ravel(), yy.ravel()])[:, 1]
    12.      # Put the result into a color plot
    13.      Z = Z.reshape(xx.shape)
    14.      ax.contourf(xx, yy, Z, cmap=cm, alpha=.8)

    為了顯示方便，我將原始碼的內圈改為 for name, clf in zip(names[0:4], classifiers[0:4]):只跑過前四個分類器。

  1. %matplotlib inline
  3. figure = plt.figure(figsize=(30,20), dpi=300)
  4. i = 1
  5. # iterate over datasets
  6. for ds in datasets:
  7.     # preprocess dataset, split into training and test part
  8.     X, y = ds
  9.     X = StandardScaler().fit_transform(X)
  10.     X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=.4)
  12.     x_min, x_max = X[:, 0].min() - .5, X[:, 0].max() + .5
  13.     y_min, y_max = X[:, 1].min() - .5, X[:, 1].max() + .5
  14.     xx, yy = np.meshgrid(np.arange(x_min, x_max, h),
  15.                          np.arange(y_min, y_max, h))
  17.     # just plot the dataset first
  18.     cm = plt.cm.RdBu
  19.     cm_bright = ListedColormap(['#FF0000', '#0000FF'])
  20.     ax = plt.subplot(len(datasets), (len(classifiers) + 1)//2, i)
  21.     # Plot the training points
  22.     ax.scatter(X_train[:, 0], X_train[:, 1], c=y_train, cmap=cm_bright)
  23.     # and testing points
  24.     ax.scatter(X_test[:, 0], X_test[:, 1], c=y_test, cmap=cm_bright, alpha=0.6)
  25.     ax.set_xlim(xx.min(), xx.max())
  26.     ax.set_ylim(yy.min(), yy.max())
  27.     ax.set_xticks(())
  28.     ax.set_yticks(())
  29.     i += 1
  31.     # iterate over classifiers
  32.     for name, clf in zip(names[0:4], classifiers[0:4]):
  33.         ax = plt.subplot(len(datasets), (len(classifiers) + 1)//2, i)
  34.         clf.fit(X_train, y_train)
  35.         score = clf.score(X_test, y_test)
  37.         # Plot the decision boundary. For that, we will assign a color to each
  38.         # point in the mesh [x_min, m_max]x[y_min, y_max].
  39.         if hasattr(clf, "decision_function"):
  40.             Z = clf.decision_function(np.c_[xx.ravel(), yy.ravel()])
  41.         else:
  42.             Z = clf.predict_proba(np.c_[xx.ravel(), yy.ravel()])[:, 1]
  44.         # Put the result into a color plot
  45.         Z = Z.reshape(xx.shape)
  46.         ax.contourf(xx, yy, Z, cmap=cm, alpha=.8)
  48.         # Plot also the training points
  49.         ax.scatter(X_train[:, 0], X_train[:, 1], c=y_train, cmap=cm_bright)
  50.         # and testing points
  51.         ax.scatter(X_test[:, 0], X_test[:, 1], c=y_test, cmap=cm_bright,
  52.                    alpha=0.6)
  54.         ax.set_xlim(xx.min(), xx.max())
  55.         ax.set_ylim(yy.min(), yy.max())
  56.         ax.set_xticks(())
  57.         ax.set_yticks(())
  58.         ax.set_title(name,fontsize=28)
  59.         ax.text(xx.max() - .3, yy.min() + .3, ('%.2f' % score).lstrip('0'),
  60.                 size=30, horizontalalignment='right')
  61.         i += 1
  63. figure.subplots_adjust(left=.02, right=.98)
  64. plt.show()

png

(四) 原始碼列表

Python source code: plot_classifier_comparison.py

http://scikit-learn.org/stable/auto_examples/classification/plot_classifier_comparison.html

  1. print(__doc__)
  4. # Code source: Gaël Varoquaux
  5. #              Andreas Müller
  6. # Modified for documentation by Jaques Grobler
  7. # License: BSD 3 clause
  9. import numpy as np
  10. import matplotlib.pyplot as plt
  11. from matplotlib.colors import ListedColormap
  12. from sklearn.cross_validation import train_test_split
  13. from sklearn.preprocessing import StandardScaler
  14. from sklearn.datasets import make_moons, make_circles, make_classification
  15. from sklearn.neighbors import KNeighborsClassifier
  16. from sklearn.svm import SVC
  17. from sklearn.tree import DecisionTreeClassifier
  18. from sklearn.ensemble import RandomForestClassifier, AdaBoostClassifier
  19. from sklearn.naive_bayes import GaussianNB
  20. from sklearn.discriminant_analysis import LinearDiscriminantAnalysis
  21. from sklearn.discriminant_analysis import QuadraticDiscriminantAnalysis
  23. h = .02  # step size in the mesh
  25. names = ["Nearest Neighbors", "Linear SVM", "RBF SVM", "Decision Tree",
  26.          "Random Forest", "AdaBoost", "Naive Bayes", "Linear Discriminant Analysis",
  27.          "Quadratic Discriminant Analysis"]
  28. classifiers = [
  29.     KNeighborsClassifier(3),
  30.     SVC(kernel="linear", C=0.025),
  31.     SVC(gamma=2, C=1),
  32.     DecisionTreeClassifier(max_depth=5),
  33.     RandomForestClassifier(max_depth=5, n_estimators=10, max_features=1),
  34.     AdaBoostClassifier(),
  35.     GaussianNB(),
  36.     LinearDiscriminantAnalysis(),
  37.     QuadraticDiscriminantAnalysis()]
  39. X, y = make_classification(n_features=2, n_redundant=0, n_informative=2,
  40.                            random_state=1, n_clusters_per_class=1)
  41. rng = np.random.RandomState(2)
  42. X += 2 * rng.uniform(size=X.shape)
  43. linearly_separable = (X, y)
  45. datasets = [make_moons(noise=0.3, random_state=0),
  46.             make_circles(noise=0.2, factor=0.5, random_state=1),
  47.             linearly_separable
  48.             ]
  50. figure = plt.figure(figsize=(27, 9))
  51. i = 1
  52. # iterate over datasets
  53. for ds in datasets:
  54.     # preprocess dataset, split into training and test part
  55.     X, y = ds
  56.     X = StandardScaler().fit_transform(X)
  57.     X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=.4)
  59.     x_min, x_max = X[:, 0].min() - .5, X[:, 0].max() + .5
  60.     y_min, y_max = X[:, 1].min() - .5, X[:, 1].max() + .5
  61.     xx, yy = np.meshgrid(np.arange(x_min, x_max, h),
  62.                          np.arange(y_min, y_max, h))
  64.     # just plot the dataset first
  65.     cm = plt.cm.RdBu
  66.     cm_bright = ListedColormap(['#FF0000', '#0000FF'])
  67.     ax = plt.subplot(len(datasets), len(classifiers) + 1, i)
  68.     # Plot the training points
  69.     ax.scatter(X_train[:, 0], X_train[:, 1], c=y_train, cmap=cm_bright)
  70.     # and testing points
  71.     ax.scatter(X_test[:, 0], X_test[:, 1], c=y_test, cmap=cm_bright, alpha=0.6)
  72.     ax.set_xlim(xx.min(), xx.max())
  73.     ax.set_ylim(yy.min(), yy.max())
  74.     ax.set_xticks(())
  75.     ax.set_yticks(())
  76.     i += 1
  78.     # iterate over classifiers
  79.     for name, clf in zip(names, classifiers):
  80.         ax = plt.subplot(len(datasets), len(classifiers) + 1, i)
  81.         clf.fit(X_train, y_train)
  82.         score = clf.score(X_test, y_test)
  84.         # Plot the decision boundary. For that, we will assign a color to each
  85.         # point in the mesh [x_min, m_max]x[y_min, y_max].
  86.         if hasattr(clf, "decision_function"):
  87.             Z = clf.decision_function(np.c_[xx.ravel(), yy.ravel()])
  88.         else:
  89.             Z = clf.predict_proba(np.c_[xx.ravel(), yy.ravel()])[:, 1]
  91.         # Put the result into a color plot
  92.         Z = Z.reshape(xx.shape)
  93.         ax.contourf(xx, yy, Z, cmap=cm, alpha=.8)
  95.         # Plot also the training points
  96.         ax.scatter(X_train[:, 0], X_train[:, 1], c=y_train, cmap=cm_bright)
  97.         # and testing points
  98.         ax.scatter(X_test[:, 0], X_test[:, 1], c=y_test, cmap=cm_bright,
  99.                    alpha=0.6)
  101.         ax.set_xlim(xx.min(), xx.max())
  102.         ax.set_ylim(yy.min(), yy.max())
  103.         ax.set_xticks(())
  104.         ax.set_yticks(())
  105.         ax.set_title(name)
  106.         ax.text(xx.max() - .3, yy.min() + .3, ('%.2f' % score).lstrip('0'),
  107.                 size=15, horizontalalignment='right')
  108.         i += 1
  110. figure.subplots_adjust(left=.02, right=.98)
  111. plt.show()