优达机器学习：特征缩放

练习：最小值/最大值重缩放器编码

import numpy as np

def featureScaling(arr):
    arr = np.array(arr)
    max = np.max(arr)
    min = np.min(arr)
    res = []
    for item in arr:
        data = float(item-min)/(max-min)
        res.append(data)
    return res

# tests of your feature scaler--line below is input data
data = [115, 140, 175]
print featureScaling(data)

练习：需要重缩放的算法练习

• 使用 RBF 核函数的 SVN
• K-均值聚类

练习：缩放类型

• MinMaxScaler

练习：计算重缩放特征

• salary : 0.17962406631
• stock : 0.0290205889347

import numpy as np
stocklist = []
for item in data_dict:
    stock = data_dict[item]['exercised_stock_options']
    if stock != 'NaN':
        stocklist.append( stock )

stocklist = np.array(stocklist)

print (1000000.0 - np.min(stocklist)) / (np.max(stocklist) - np.min(stocklist))

salarylist = []
for item in data_dict:
    salary = data_dict[item]['salary']
    if salary != 'NaN':
        salarylist.append( salary )

print (200000.0 - np.min(salarylist)) / (np.max(salarylist) - np.min(salarylist))

很奇怪的是，使用sklearn里面的MinMaxScaler的缩放率计算有误差，代码如下

import numpy as np
stocklist = []
for item in data_dict:
    stock = data_dict[item]['exercised_stock_options']
    if stock != 'NaN':
        stocklist.append( stock )

stocklist = np.array(stocklist)

from sklearn import preprocessing
min_max_scaler = preprocessing.MinMaxScaler()
min_max_scaler.fit_transform(stocklist)
print 1000000 * min_max_scaler.scale_

salarylist = []
for item in data_dict:
    salary = data_dict[item]['salary']
    if salary != 'NaN':
        salarylist.append( salary )

salarylist = np.array(salarylist)

min_max_scaler.fit_transform(salarylist)
print 200000 * min_max_scaler.scale_
print np.max(salarylist)
print np.min(salarylist)

练习：何时部署特征缩放

• 重要