Groupby class and count missing values in features

Question

I have a problem and I cannot find any solution in the web or documentation, even if I think that it is very trivial.

What do I want to do?

I have a dataframe like this

CLASS FEATURE1 FEATURE2 FEATURE3
  X      A       NaN      NaN
  X     NaN       A       NaN
  B      A        A        A

I want to group by the label(CLASS) and display the number of NaN-Values that are counted in every feature so that it looks like this. The purpose of this is to get a general idea how missing values are distributed over the different classes.

CLASS FEATURE1 FEATURE2 FEATURE3
  X      1        1        2
  B      0        0        0

I know how to recieve the amount of nonnull-Values - df.groupby['CLASS'].count()

Is there something similar for the NaN-Values?

I tried to subtract the count() from the size() but it returned an unformatted output filled with the value NaN

Answer 1

Compute a mask with isna, then group and find the sum:

df.drop('CLASS', 1).isna().groupby(df.CLASS, sort=False).sum().reset_index()

  CLASS  FEATURE1  FEATURE2  FEATURE3
0     X       1.0       1.0       2.0
1     B       0.0       0.0       0.0

---

Another option is to subtract the size from the count using rsub along the 0^th axis for index aligned subtraction:

df.groupby('CLASS').count().rsub(df.groupby('CLASS').size(), axis=0)

Or,

g = df.groupby('CLASS')
g.count().rsub(g.size(), axis=0)

       FEATURE1  FEATURE2  FEATURE3
CLASS                              
B             0         0         0
X             1         1         2

---

There are quite a few good answers, so here are some timeits for your perusal:

df_ = df
df = pd.concat([df_] * 10000)

%timeit df.drop('CLASS', 1).isna().groupby(df.CLASS, sort=False).sum()
%timeit df.set_index('CLASS').isna().sum(level=0)    
%%timeit
g = df.groupby('CLASS')
g.count().rsub(g.size(), axis=0)

11.8 ms ± 108 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
9.47 ms ± 379 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
6.54 ms ± 81.6 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

Actual performance depends on your data and setup, so your mileage may vary.