多元线性回归scikit-learn和statsmodel

时间:2016-06-21 19:36:53

标签: python scikit-learn statsmodels

我尝试使用scikit-learn在数据集上使用多元线性回归,但我在获取正确的系数时遇到问题。我使用的是休伦湖数据,可以在这里找到:

https://vincentarelbundock.github.io/Rdatasets/datasets.html

在转换之后,我有以下一组值:

         x1        x2        y
0  0.202165  1.706366  0.840567
1  1.706366  0.840567  0.694768
2  0.840567  0.694768 -0.291031
3  0.694768 -0.291031  0.333170
4 -0.291031  0.333170  0.387371
5  0.333170  0.387371  0.811572
6  0.387371  0.811572  1.415773
7  0.811572  1.415773  1.359974
8  1.415773  1.359974  1.504176
9  1.359974  1.504176  1.768377
...  ...       ...       ...

使用

df = pd.DataFrame(nvalues, columns=("x1", "x2", "y"))
result = sm.ols(formula="y ~ x2 + x1", data=df).fit()

print(result.params)

产量

Intercept   -0.007852
y2           1.002137
y1          -0.283798

这是正确的值,但如果我最终使用scikit-learn,我会得到:

a = np.array([nvalues["x1"], nvalues["x2"]])
b = np.array(nvalues["y"])

a = a.reshape(len(nvalues["x1"]), 2)
b = b.reshape(len(nvalues["y"]), 1)

clf = linear_model.LinearRegression()
clf.fit(a, b)

print(clf.coef_)

我得到[[-0.18260922 0.08101687]]

为了完整我的代码

from sklearn import linear_model

import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import statsmodels.formula.api as sm

def Main():
    location = r"~/Documents/Time Series/LakeHuron.csv"
    ts = pd.read_csv(location, sep=",", parse_dates=[0], header=0)

    #### initializes the data ####
    ts.drop("Unnamed: 0", axis=1, inplace=True)

    x = ts["time"].values
    y = ts["LakeHuron"].values

    x = x.reshape(len(ts), 1)
    y = y.reshape(len(ts), 1)

    regr = linear_model.LinearRegression()
    regr.fit(x, y)

    diff = []
    for i in range(0, len(ts)):
        diff.append(float(ts["LakeHuron"][i]-regr.predict(x)[i]))

    ts[3] = diff

    nvalues = {"x1": [], "x2": [], "y": []}

    for i in range(0, len(ts)-2):
        nvalues["x1"].append(float(ts[3][i]))
        nvalues["x2"].append(float(ts[3][i+1]))
        nvalues["y"].append(float(ts[3][i+2]))

    df = pd.DataFrame(nvalues, columns=("x1", "x2", "y"))
    result = sm.ols(formula="y ~ x2 + x1", data=df).fit()

    print(result.params)

    #### using scikit-learn ####
    a = np.array([nvalues["x1"], nvalues["x2"]])
    b = np.array(nvalues["y"])

    a = a.reshape(len(nvalues["x1"]), 2)
    b = b.reshape(len(nvalues["y"]), 1)

    clf = linear_model.LinearRegression()
    clf.fit(a, b)

    print(clf.coef_)

if __name__ == "__main__":
    Main()

2 个答案:

答案 0 :(得分:2)

问题在于

a = np.array([nvalues["x1"], nvalues["x2"]])

因为它没有按照您的意图对数据进行排序。相反,它将生成一个数据集

x1_new    x2_new
-----------------
 x1[0]     x1[1]
 x1[2]     x1[3]
[...]
 x1[94]    x1[95]
 x2[0]     x2[1]
[...]

尝试改为

ax1 = np.array(nvalues["x1"])
ax2 = np.array(nvalues["x2"])
ax1 = ax1.reshape(len(nvalues["x1"]), 1)
ax2 = ax2.reshape(len(nvalues["x2"]), 1)
a = np.hstack([ax1,ax2])

可能有一种更干净的方法,但这种方式有效。回归现在也给出了正确的结果。

编辑: 更简洁的方法是使用transpose()

a = a.transpose()

答案 1 :(得分:0)

根据@Orange建议,我已将代码更改为更有效的东西:

#### using scikit-learn ####
a = []
for i in range(0, len(nvalues["x1"])):
    a.append([nvalues["x1"][i], nvalues["x2"][i]])

a = np.array(a)
b = np.array(nvalues["y"])

a = a.reshape(len(a), 2)
b = b.reshape(len(nvalues["y"]), 1)

clf = linear_model.LinearRegression()
clf.fit(a, b)

print(clf.coef_) 

这与简单回归的scikit-learn网站上的示例类似