分類法/範例二: Normal and Shrinkage Linear Discriminant Analysis for classification

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

這個範例用來展示scikit-learn 如何使用Linear Discriminant Analysis (LDA) 線性判別分析來達成資料分類的目的

  1. 利用 sklearn.datasets.make_blobs 產生測試資料
  2. 利用自定義函數 generate_data 產生具有數個特徵之資料集,其中僅有一個特徵對於資料分料判斷有意義
  3. 使用LinearDiscriminantAnalysis來達成資料判別
  4. 比較於LDA演算法中,開啟 shrinkage 前後之差異

(一)產生測試資料

從程式碼來看,一開始主要為自定義函數generate_data(n_samples, n_features),這個函數的主要目的為產生一組測試資料,總資料列數為n_samples,每一列共有n_features個特徵。而其中只有第一個特徵得以用來判定資料類別,其他特徵則毫無意義。make_blobs負責產生單一特徵之資料後,利用`np.random.randn` 亂數產生其他`n_features - 1`個特徵,之後利用np.hstack以”水平” (horizontal)方式連接X以及亂數產生之特徵資料。

  1. %matplotlib inline
  2. from __future__ import division
  3. import numpy as np
  4. import matplotlib.pyplot as plt
  5. from sklearn.datasets import make_blobs
  6. from sklearn.discriminant_analysis import LinearDiscriminantAnalysis
  7. n_train = 20 # samples for training
  8. n_test = 200 # samples for testing
  9. n_averages = 50 # how often to repeat classification
  10. n_features_max = 75 # maximum number of features
  11. step = 4 # step size for the calculation
  12. def generate_data(n_samples, n_features):
  13. X, y = make_blobs(n_samples=n_samples, n_features=1, centers=[[-2], [2]])
  14. # add non-discriminative features
  15. if n_features > 1:
  16. X = np.hstack([X, np.random.randn(n_samples, n_features - 1)])
  17. return X, y

我們可以用以下的程式碼來測試自定義函式,結果回傳了X (10x5矩陣)及y(10個元素之向量),我們可以使用pandas.DataFrame套件來觀察資料

  1. X, y = generate_data(10, 5)
  2. import pandas as pd
  3. pd.set_option('precision',2)
  4. df=pd.DataFrame(np.hstack([y.reshape(10,1),X]))
  5. df.columns = ['y', 'X0', 'X1', 'X2', 'X2', 'X4']
  6. print(df)

結果顯示如下。。我們可以看到只有X的第一行特徵資料(X0) 與目標數值 y 有一個明確的對應關係,也就是y為1時,數值較大。

  1. y X0 X1 X2 X2 X4
  2. 0 1 0.38 0.35 0.80 -0.97 -0.68
  3. 1 1 2.41 0.31 -1.47 0.10 -1.39
  4. 2 1 1.65 -0.99 -0.12 -0.38 0.18
  5. 3 0 -4.86 0.14 -0.80 1.13 -1.31
  6. 4 1 -0.06 -1.99 -0.70 -1.26 -1.64
  7. 5 0 -1.51 -1.74 -0.83 0.74 -2.07
  8. 6 0 -2.50 0.44 -0.45 -0.55 -0.42
  9. 7 1 1.55 1.38 0.93 -1.44 0.27
  10. 8 0 -1.95 0.32 -0.28 0.02 0.07
  11. 9 0 -0.58 -0.07 -1.01 0.15 -1.84

(二)改變特徵數量並測試shrinkage之功能

接下來程式碼裏有兩段迴圈,外圈改變特徵數量。內圈則多次嘗試LDA之以求精準度。使用LinearDiscriminantAnalysis來訓練分類器,過程中以shrinkage='auto'以及shrinkage=None來控制shrinkage之開關,將分類器分別以clf1以及clf2儲存。之後再產生新的測試資料將準確度加入score_clf1score_clf2裏,離開內迴圈之後除以總數以求平均。

  1. acc_clf1, acc_clf2 = [], []
  2. n_features_range = range(1, n_features_max + 1, step)
  3. for n_features in n_features_range:
  4. score_clf1, score_clf2 = 0, 0
  5. for _ in range(n_averages):
  6. X, y = generate_data(n_train, n_features)
  7. clf1 = LinearDiscriminantAnalysis(solver='lsqr', shrinkage='auto').fit(X, y)
  8. clf2 = LinearDiscriminantAnalysis(solver='lsqr', shrinkage=None).fit(X, y)
  9. X, y = generate_data(n_test, n_features)
  10. score_clf1 += clf1.score(X, y)
  11. score_clf2 += clf2.score(X, y)
  12. acc_clf1.append(score_clf1 / n_averages)
  13. acc_clf2.append(score_clf2 / n_averages)

(三)顯示LDA判別結果

這個範例主要希望能得知shrinkage的功能,因此畫出兩條分類準確度的曲線。縱軸代表平均的分類準確度,而橫軸代表的是features_samples_ratio 顧名思義,它是模擬資料中,特徵數量與訓練資料列數的比例。當特徵數量為75且訓練資料列數僅有20筆時,features_samples_ratio = 3.75 由於資料列數過少,導致準確率下降。而此時shrinkage演算法能有效維持LDA演算法的準確度。

  1. features_samples_ratio = np.array(n_features_range) / n_train
  2. fig = plt.figure(figsize=(10,6), dpi=300)
  3. plt.plot(features_samples_ratio, acc_clf1, linewidth=2,
  4. label="Linear Discriminant Analysis with shrinkage", color='r')
  5. plt.plot(features_samples_ratio, acc_clf2, linewidth=2,
  6. label="Linear Discriminant Analysis", color='g')
  7. plt.xlabel('n_features / n_samples')
  8. plt.ylabel('Classification accuracy')
  9. plt.legend(loc=1, prop={'size': 10})
  10. plt.show()

png

(四)完整程式碼

Python source code: plot_lda.py

  1. from __future__ import division
  2. import numpy as np
  3. import matplotlib.pyplot as plt
  4. from sklearn.datasets import make_blobs
  5. from sklearn.discriminant_analysis import LinearDiscriminantAnalysis
  6. n_train = 20 # samples for training
  7. n_test = 200 # samples for testing
  8. n_averages = 50 # how often to repeat classification
  9. n_features_max = 75 # maximum number of features
  10. step = 4 # step size for the calculation
  11. def generate_data(n_samples, n_features):
  12. """Generate random blob-ish data with noisy features.
  13. This returns an array of input data with shape `(n_samples, n_features)`
  14. and an array of `n_samples` target labels.
  15. Only one feature contains discriminative information, the other features
  16. contain only noise.
  17. """
  18. X, y = make_blobs(n_samples=n_samples, n_features=1, centers=[[-2], [2]])
  19. # add non-discriminative features
  20. if n_features > 1:
  21. X = np.hstack([X, np.random.randn(n_samples, n_features - 1)])
  22. return X, y
  23. acc_clf1, acc_clf2 = [], []
  24. n_features_range = range(1, n_features_max + 1, step)
  25. for n_features in n_features_range:
  26. score_clf1, score_clf2 = 0, 0
  27. for _ in range(n_averages):
  28. X, y = generate_data(n_train, n_features)
  29. clf1 = LinearDiscriminantAnalysis(solver='lsqr', shrinkage='auto').fit(X, y)
  30. clf2 = LinearDiscriminantAnalysis(solver='lsqr', shrinkage=None).fit(X, y)
  31. X, y = generate_data(n_test, n_features)
  32. score_clf1 += clf1.score(X, y)
  33. score_clf2 += clf2.score(X, y)
  34. acc_clf1.append(score_clf1 / n_averages)
  35. acc_clf2.append(score_clf2 / n_averages)
  36. features_samples_ratio = np.array(n_features_range) / n_train
  37. plt.plot(features_samples_ratio, acc_clf1, linewidth=2,
  38. label="Linear Discriminant Analysis with shrinkage", color='r')
  39. plt.plot(features_samples_ratio, acc_clf2, linewidth=2,
  40. label="Linear Discriminant Analysis", color='g')
  41. plt.xlabel('n_features / n_samples')
  42. plt.ylabel('Classification accuracy')
  43. plt.legend(loc=1, prop={'size': 12})
  44. plt.suptitle('Linear Discriminant Analysis vs. \
  45. shrinkage Linear Discriminant Analysis (1 discriminative feature)')
  46. plt.show()