Scikit-learn:拟合模型中的错误 - 输入包含NaN,无穷大或者对于float64来说太大的值

Question

我正在使用Python scikit-learn对从csv获得的数据进行简单的线性回归.

reader = pandas.io.parsers.read_csv("data/all-stocks-cleaned.csv")
stock = np.array(reader)

openingPrice = stock[:, 1]
closingPrice = stock[:, 5]

print((np.min(openingPrice)))
print((np.min(closingPrice)))
print((np.max(openingPrice)))
print((np.max(closingPrice)))

peningPriceTrain, openingPriceTest, closingPriceTrain, closingPriceTest = \
    train_test_split(openingPrice, closingPrice, test_size=0.25, random_state=42)


openingPriceTrain = np.reshape(openingPriceTrain,(openingPriceTrain.size,1))

openingPriceTrain = openingPriceTrain.astype(np.float64, copy=False)
# openingPriceTrain = np.arange(openingPriceTrain, dtype=np.float64)

closingPriceTrain = np.reshape(closingPriceTrain,(closingPriceTrain.size,1))
closingPriceTrain = closingPriceTrain.astype(np.float64, copy=False)

openingPriceTest = np.reshape(openingPriceTest,(openingPriceTest.size,1))
closingPriceTest = np.reshape(closingPriceTest,(closingPriceTest.size,1))

regression = linear_model.LinearRegression()

regression.fit(openingPriceTrain, closingPriceTrain)

predicted = regression.predict(openingPriceTest)

最小值和最大值显示为0.0 0.6 41998.0 2593.9

然而,我收到此错误ValueError: Input contains NaN, infinity or a value too large for dtype('float64').

我该如何删除此错误？因为从上面的结果来看,它确实不包含无限或纳米值.

这是什么解决方案？

编辑:所有股票clean.csv是avaliabale在http://www.sharecsv.com/s/cb31790afc9b9e33c5919cdc562630f3/all-stocks-cleaned.csv

Answer 1

你的回归问题是不知何故NaN已经潜入你的数据.可以使用以下代码段轻松检查:

import pandas as pd
import numpy as np
from  sklearn import linear_model
from sklearn.cross_validation import train_test_split

reader = pd.io.parsers.read_csv("./data/all-stocks-cleaned.csv")
stock = np.array(reader)

openingPrice = stock[:, 1]
closingPrice = stock[:, 5]

openingPriceTrain, openingPriceTest, closingPriceTrain, closingPriceTest = \
    train_test_split(openingPrice, closingPrice, test_size=0.25, random_state=42)

openingPriceTrain = openingPriceTrain.reshape(openingPriceTrain.size,1)
openingPriceTrain = openingPriceTrain.astype(np.float64, copy=False)

closingPriceTrain = closingPriceTrain.reshape(closingPriceTrain.size,1)
closingPriceTrain = closingPriceTrain.astype(np.float64, copy=False)

openingPriceTest = openingPriceTest.reshape(openingPriceTest.size,1)
openingPriceTest = openingPriceTest.astype(np.float64, copy=False)

np.isnan(openingPriceTrain).any(), np.isnan(closingPriceTrain).any(), np.isnan(openingPriceTest).any()

(True, True, True)

如果您尝试输入如下所示的缺失值:

openingPriceTrain[np.isnan(openingPriceTrain)] = np.median(openingPriceTrain[~np.isnan(openingPriceTrain)])
closingPriceTrain[np.isnan(closingPriceTrain)] = np.median(closingPriceTrain[~np.isnan(closingPriceTrain)])
openingPriceTest[np.isnan(openingPriceTest)] = np.median(openingPriceTest[~np.isnan(openingPriceTest)])

您的回归将顺利运行而不会出现问题:

regression = linear_model.LinearRegression()

regression.fit(openingPriceTrain, closingPriceTrain)

predicted = regression.predict(openingPriceTest)

predicted[:5]

array([[ 13598.74748173],
       [ 53281.04442146],
       [ 18305.4272186 ],
       [ 50753.50958453],
       [ 14937.65782778]])

简而言之:您的数据中缺少值,如错误消息所示.

编辑 ::

或许更简单,更直接的方法是在用pandas读取数据后立即检查是否有任何缺失数据:

data = pd.read_csv('./data/all-stocks-cleaned.csv')
data.isnull().any()
Date                    False
Open                     True
High                     True
Low                      True
Last                     True
Close                    True
Total Trade Quantity     True
Turnover (Lacs)          True

然后使用以下两行中的任何一行来估算数据:

data = data.fillna(lambda x: x.median())

要么

data = data.fillna(method='ffill')