Overall accuracy of multiclass classification using pandas-CodePudding

I have a data frame as shown below.

id     label     prediction
1      cat       cat
2      dog       cat
3      cow       dog
4      cow       cow
5      dog       cat
6      cat       cat
7      cat       cat
8      dog       dog
9      dog       dog
10     cat       cat

from the above df, I would like to calculate overall accuracy using pandas.

I tried the below code to calculate class-wise accuracy.

class_wise_accuracy = (df.groupby('label')['prediction']
                         .value_counts(normalize=True)
                         .unstack(fill_value=0)
                      )

confusion_matrix = (df.groupby('label')['prediction']
                      .value_counts()
                      .unstack(fill_value=0)
                      .reset_index()
                   )

Expected Output:

overall_accuracy = (4 1 2)/df.shape[0] = 0.7

CodePudding user response：

IIUC, use crosstab and the underlying numpy array:

a = pd.crosstab(df['label'], df['prediction']).to_numpy()

overall_accuracy = a.diagonal().sum()/a.sum()

output: 0.7

intermediates:

pd.crosstab(df['label'], df['prediction'])

prediction  cat  cow  dog
label                    
cat           4    0    0
cow           0    1    1
dog           2    0    2

.tonumpy()

array([[4, 0, 0],
       [0, 1, 1],
       [2, 0, 2]])