修订:
如果我有一个包含5列Col1
&Col2
&Col3
&Col4
&Col5
的pandas DataFrame,我需要通过考虑值来获得(Col2
,Col3
)和(Col2
,Col4
)和(Col2
,Col5
)之间的最大Pearson相关系数在Col1
通过下一个公式得到的Col2
的修改值:
df['Col1']=np.power((df['Col1']),B)
df['Col2']=df['Col2']*df['Col1']
其中B
是变化的变量(单个值),以获得最大Pearson之间的相关系数(Col2
,Col3
的新值)和(Col2
,Col4
的新值)和(Col2
,Col5
的新值)。
更新:
上表中包含5列,如上所述,(Col2
,Col3
)和(Col2
,Col4
)和(Col2
,Col5
)之间的系数之间的相关性如下表所示。
我需要根据两个提到的方程式改变Col2
的值,其中变化的值是B
。
所以问题是如何获得B
的最佳值,使得新的相关系数大于或等于其对应物(旧)?
更新2:
COL1,col2的,COL3,COL4,COL5
2,0.051361397,2618,1453,1099
4,0.053507779,306,153,150
2,0.041236151,39,54,34
6,0.094526419,2755,2209,1947
4,0.079773397,2313,1261,1022
4,0.083891415,3528,2502,2029
6,0.090737243,3594,2781,2508
2,0.069552772,370,234,246
2,0.052401789,690,402,280
2,0.039930675,1218,846,631
4,0.065952096,1706,523,453
2,0.053064126,314,197,123
6,0.076847486,4019,1675,1452
2,0.044881545,604,402,356
2,0.073102611,2214,1263,1050
0,0.046998526,938,648,572
不是很优雅,但有效;随意使这更通用:
import pandas as pd
from scipy.optimize import minimize
def minimize_me(b, df):
# we want to maximize, so we have to multiply by -1
return -1 * df['Col3'].corr(df['Col2'] * df['Col1'] ** b )
# read your dataframe from somehwere, e.g. csv
df = pd.read_clipboard(sep=',')
# B is greater than 0 for now
bnds = [(0, None)]
res = minimize(minimize_me, (1), args=(df,), bounds=bnds)
if res.success:
# that's the optimal B
print(res.x[0])
# that's the highest correlation you can get
print(-1 * res.fun)
else:
print("Sorry, the optimization was not successful. Try with another initial"
" guess or optimization method")
这将打印:
0.9020784246026575 # your B
0.7614993786787415 # highest correlation for corr(col2, col3)
我现在从clipboard
读取,用你的.csv
文件替换它。然后你应该避免列的硬编码;上面的代码仅用于演示目的,以便您了解如何设置优化问题本身。
如果您对总和感兴趣,可以使用(其余代码未经修改):
def minimize_me(b, df):
col_mod = df['Col2'] * df['Col1'] ** b
# we want to maximize, so we have to multiply by -1
return -1 * (df['Col3'].corr(col_mod) +
df['Col4'].corr(col_mod) +
df['Col5'].corr(col_mod))
这将打印:
1.0452394748131613
2.3428368479642137