機器學習 - 特征篩選與降維

特征決定了最優效果的上限，算法與模型只是讓效果更逼近這個上限，所以特征工程與選擇什么樣的特征很重要！

以下是一些特征篩選與降維技巧

# -*- coding:utf-8 -*-
import scipy as sc
import libsvm_file_process as data_process
import numpy as np
from minepy import MINE
from sklearn.feature_selection import VarianceThreshold
from sklearn.feature_selection import SelectKBest
from sklearn.feature_selection import chi2
from sklearn.feature_selection import f_regression
from sklearn.feature_selection import RFE
from sklearn.svm import SVR
from sklearn.linear_model import LogisticRegression
from sklearn.decomposition import PCA
from sklearn.discriminant_analysis import LinearDiscriminantAnalysis


class feature_select:
    """
    特征篩選方式：
        相關鏈接：http://scikit-learn.org/stable/modules/classes.html#module-sklearn.feature_selection
        皮爾遜相關性
        互信息
        單因素 - 卡方判斷，F值，假正率
        方差過濾
        遞歸特征消除法 - 每次消除一個特征，依據是特征前面的系數
        基於模型（LR/GBDT等）的特征選擇 SelectFromModel
            模型（LR/GBDT）必須有feature_importances_ 或 coef_這個屬性
    降維：
        PCA（unsurperised）：一般用於無監督情況下的降維，有監督的時候，也可以小幅降維 去除噪音，然后再使用LDA 降維

        LDA（surperised）:本質上是一個分類器,在使用上，要求降低的維度要小於分類的維度
    """

    def __init__(self):
        self.data_path = "/trainData/libsvm2/"
        self.trainData = ["20180101"]
        # 計算互信息
        self.mine = MINE(alpha=0.6, c=15, est="mic_approx")
        # 方差過濾 一般用於無監督學習
        self.variance_filter = VarianceThreshold(threshold=0.1)
        # chi2 - 卡方檢驗; f_regression - f值; SelectFpr-假正率;等
        self.chi_squared = SelectKBest(f_regression, k=2)
        # 遞歸特征消除
        self.estimator = LogisticRegression()  # SVR(kernel="linear")
        self.selector = RFE(self.estimator, 5, step=1)
        # PCA 降維
        self.pca = PCA(n_components=5)
        # LDA 降維
        self.lda = LinearDiscriminantAnalysis(n_components=2)

    def select(self):
        for i in range(len(self.trainData)):
            generator = data_process.get_data_batch(self.data_path + self.trainData[i] + "/part-00000", 100000)
            labels, features = generator.next()
            # 方差過濾
            filter1 = self.variance_filter.fit_transform(features)
            print filter1.shape, features.shape
            print self.variance_filter.get_support()
            # 卡方檢驗
            filter2 = self.chi_squared.fit_transform(features, labels)
            print filter2.shape
            print self.chi_squared.get_support()
            # 遞歸特征消除（比較耗時 暫時先注釋掉）
            # self.selector.fit(features, labels)
            # print self.selector.support_
            # PCA 降維
            transform1 = self.pca.fit_transform(features)
            print 'transform1:', transform1
            # LDA降維
            self.lda.fit(features, labels)
            transform2 = self.lda.transform(features)
            print 'transform2:', transform2
            for j in range(int(features.shape[1]) - 870):
                features_j = features[0:, j + 870: j + 871]
                self.mine.compute_score(features_j.flatten(), labels.flatten())
                # 計算互信息
                print self.mine.mic()
                # 計算皮爾遜系數
                print j, sc.stats.pearsonr(features_j.reshape(-1, 1), labels.reshape(-1, 1))


if __name__ == '__main__':
    feature_util = feature_select()
    feature_util.select()