DataFrame correlation produces NaN although its values are all integers

问题

I have a dataframe df:

df   = pandas.DataFrame(pd.read_csv(loggerfile, header = 2))

values = df.as_matrix()

df2 = pd.DataFrame.from_records(values, index = datetimeIdx, columns = Columns) 

EDIT:

Now reading the data this way as suggested:

df2 = pd.read_csv(loggerfile, header = None, skiprows = [0,1,2])

Sample:

                         0              1       2   3   4   5   6   7   8   \
0  2014-03-19T12:44:32.695Z  1395233072695  703425   0   2   1  13   5  21   
1  2014-03-19T12:44:32.727Z  1395233072727  703425   0   2   1  13   5  21   

   9   10  11   12  13   14  15  16  
0  25   0  25  209   0  145   0   0  
1  25   0  25  209   0  146   0   0

The columns are all type int (except the first one):

print df2.dtypes

0     object
1      int64
2      int64
3      int64
4      int64
5      int64
6      int64
7      int64
8      int64
9      int64
10     int64
11     int64
12     int64
13     int64
14     int64
15     int64
16     int64

But in my correlation, some columns seem to be NaN.

df2.corr()

     1          2    3          4           5   6   7            8           ...    
1    1.000000   NaN  0.018752   -0.550307   NaN NaN 0.075191     0.775725
2    NaN        NaN  NaN         NaN        NaN NaN NaN          NaN
3    0.018752   NaN  1.000000   -0.067293   NaN NaN -0.579651    0.004593 
...

回答1:

Those columns do not change in value right now, yes

As, Joris points out you would expected NaN if the values do not vary. To see why take a look at correlation formula:

cor(i,j) = cov(i,j)/[stdev(i)*stdev(j)]

If the values of the ith or jth variable do not vary, then the respective standard deviation will be zero and so will the denominator of the fraction. Thus, the correlation will be NaN.

来源：https://stackoverflow.com/questions/22655667/dataframe-correlation-produces-nan-although-its-values-are-all-integers