通用範例/範例四: Imputing missing values before building an estimator

http://scikit-learn.org/stable/auto_examples/missing_values.htm

在這範例說明有時補充缺少的數據(missing values)，可以得到更好的結果。但仍然需要進行交叉驗證。來驗證填充是否合適
。而missing values可以用均值、中位值，或者頻繁出現的值代替。中位值對大數據之機器學習來說是比較穩定的估計值。

(一)引入函式庫及內建測試資料庫

引入之函式庫如下

  1. sklearn.ensemble.RandomForestRegressor: 隨機森林回歸
  2. sklearn.pipeline.Pipeline: 串聯估計器
  3. sklearn.preprocessing.Imputer: 缺失值填充
  4. sklearn.cross_validation import cross_val_score:交叉驗證

(二)引入內建測試資料庫(boston房產資料)

使用 datasets.load_boston() 將資料存入， boston 為一個dict型別資料，我們看一下資料的內容。

n_samples 為樣本數

n_features 為特徵數

  1. dataset = load_boston()
  2. X_full, y_full = dataset.data, dataset.target
  3. n_samples = X_full.shape[0]
  4. n_features = X_full.shape[1]
顯示 說明
(‘data’, (506, 13)) 機器學習數據
(‘feature_names’, (13,)) 房地產相關特徵
(‘target’, (506,)) 回歸目標
DESCR 資料之描述

共有506筆資料及13個特徵(‘CRIM’, ‘ZN’, ‘INDUS’, ‘CHAS’, ‘NOX’, ‘RM’, ‘AGE’, ‘DIS’, ‘RAD’,’TAX’, ‘PTRATIO’, ‘B’, ‘LSTAT’)用來描述房地產的週邊狀況，如CRIM (per capita crime rate by town)跟該區域之犯罪率有關。而迴歸目標為房地產的價格，以1000美元為單位。也就是說這個範例希望以房地產的週遭客觀數據來預測房地產的價格。

(三)利用整個數據集來預測

全部的資料使用隨機森林回歸函數進行交叉驗證，得到一個分數。

Score with the entire dataset = 0.56

  1. estimator = RandomForestRegressor(random_state=0, n_estimators=100)
  2. score = cross_val_score(estimator, X_full, y_full).mean()
  3. print("Score with the entire dataset = %.2f" % score)

(四)模擬資料損失時之預測情形

設定損失比例，並估計移除missing values後的得分
損失比例75%，損失樣本數為379筆，剩餘樣本為127筆。

將127筆資料進行隨機森林回歸函數進行交叉驗證，並得到一個分數。

Score without the samples containing missing values = 0.49

  1. missing_rate = 0.75
  2. n_missing_samples = np.floor(n_samples * missing_rate)
  3. missing_samples = np.hstack((np.zeros(n_samples - n_missing_samples,
  4.                                       dtype=np.bool),
  5.                              np.ones(n_missing_samples,
  6.                                      dtype=np.bool)))
  7. rng.shuffle(missing_samples)
  8. missing_features = rng.randint(0, n_features, n_missing_samples)
  10. X_filtered = X_full[~missing_samples, :]
  11. y_filtered = y_full[~missing_samples]
  12. estimator = RandomForestRegressor(random_state=0, n_estimators=100)
  13. score = cross_val_score(estimator, X_filtered, y_filtered).mean()
  14. print("Score without the samples containing missing values = %.2f" % score)

(五)填充missing values，估計填充後的得分

每一筆樣本資料都在13個特徵中隨機遺失一個特徵資料，

使用sklearn.preprocessing.Imputer進行missing values的填充。

  1. class sklearn.preprocessing.Imputer(missing_values='NaN', strategy='mean', axis=0, verbose=0, copy=True)

填充後進行隨機森林回歸函數進行交叉驗證，獲得填充後分數。

  1. X_missing = X_full.copy()
  2. X_missing[np.where(missing_samples)[0], missing_features] = 0
  3. y_missing = y_full.copy()
  4. estimator = Pipeline([("imputer", Imputer(missing_values=0,
  5.                                           strategy="mean",
  6.                                           axis=0)),
  7.                       ("forest", RandomForestRegressor(random_state=0,
  8.                                                        n_estimators=100))])
  9. score = cross_val_score(estimator, X_missing, y_missing).mean()
  10. print("Score after imputation of the missing values = %.2f" % score)

利用數據填充後的迴歸函數，去測試填充前的資料，預測的準確率獲得提升。

Score after imputation of the missing values = 0.57

(六)完整程式碼

Python source code: missing_values.py

http://scikit-learn.org/stable/auto_examples/missing_values.html#example-missing-values-py

  1. import numpy as np
  3. from sklearn.datasets import load_boston
  4. from sklearn.ensemble import RandomForestRegressor
  5. from sklearn.pipeline import Pipeline
  6. from sklearn.preprocessing import Imputer
  7. from sklearn.cross_validation import cross_val_score
  9. rng = np.random.RandomState(0)
  11. dataset = load_boston()
  12. X_full, y_full = dataset.data, dataset.target
  13. n_samples = X_full.shape[0]
  14. n_features = X_full.shape[1]
  16. # Estimate the score on the entire dataset, with no missing values
  17. estimator = RandomForestRegressor(random_state=0, n_estimators=100)
  18. score = cross_val_score(estimator, X_full, y_full).mean()
  19. print("Score with the entire dataset = %.2f" % score)
  21. # Add missing values in 75% of the lines
  22. missing_rate = 0.75
  23. n_missing_samples = np.floor(n_samples * missing_rate)
  24. missing_samples = np.hstack((np.zeros(n_samples - n_missing_samples,
  25.                                       dtype=np.bool),
  26.                              np.ones(n_missing_samples,
  27.                                      dtype=np.bool)))
  28. rng.shuffle(missing_samples)
  29. missing_features = rng.randint(0, n_features, n_missing_samples)
  31. # Estimate the score without the lines containing missing values
  32. X_filtered = X_full[~missing_samples, :]
  33. y_filtered = y_full[~missing_samples]
  34. estimator = RandomForestRegressor(random_state=0, n_estimators=100)
  35. score = cross_val_score(estimator, X_filtered, y_filtered).mean()
  36. print("Score without the samples containing missing values = %.2f" % score)
  38. # Estimate the score after imputation of the missing values
  39. X_missing = X_full.copy()
  40. X_missing[np.where(missing_samples)[0], missing_features] = 0
  41. y_missing = y_full.copy()
  42. estimator = Pipeline([("imputer", Imputer(missing_values=0,
  43.                                           strategy="mean",
  44.                                           axis=0)),
  45.                       ("forest", RandomForestRegressor(random_state=0,
  46.                                                        n_estimators=100))])
  47. score = cross_val_score(estimator, X_missing, y_missing).mean()
  48. print("Score after imputation of the missing values = %.2f" % score)
  1. results:
  2. Score with the entire dataset = 0.56
  3. Score without the samples containing missing values = 0.48
  4. Score after imputation of the missing values = 0.55