Python sklearn - average classification reports

Question

Using Python's sklearn module,

 from sklearn.metrics import classification_report
 y1_predict = [0, 1, 1, 0]
 y1_dev = [0, 1, 1, 0]
 report_1 = classification_report(y1_dev, y1_predict)
 y2_predict = [1, 0, 1, 0]
 y2_dev = [1, 1, 0, 0]
 report_2 = classification_report(y2_dev, y2_predict)

Is there a way to combine (maybe just an average) report_1 and report_2 I'm looking for an implementation like:

 report_average = average(report_1,report_2)

Or does this have to be done manually? I was hoping that printing the report_average would have average values between the two reports.

Answer 1

It maybe a little bit complicated, since I convert the reports to pandas.DataFrame for calculation. But I think it's worth it, because it works well with two or more report as well. Try below:

import pandas as pd
import numpy as np
from functools import reduce
def report_average(*args):
    report_list = list()
    for report in args:
        splited = [' '.join(x.split()) for x in report.split('\n\n')]
        header = [x for x in splited[0].split(' ')]
        data = np.array(splited[1].split(' ')).reshape(-1, len(header) + 1)
        data = np.delete(data, 0, 1).astype(float)
        avg_total = np.array([x for x in splited[2].split(' ')][3:]).astype(float).reshape(-1, len(header))
        df = pd.DataFrame(np.concatenate((data, avg_total)), columns=header)
        report_list.append(df)
    res = reduce(lambda x, y: x.add(y, fill_value=0), report_list) / len(report_list)
    return res.rename(index={res.index[-1]: 'avg / total'})

output:

report_average  = report_average(report_1, report_2)
print(report_average)
             precision  recall  f1-score  support
0                 0.75    0.75      0.75      2.0
1                 0.75    0.75      0.75      2.0
avg / total       0.75    0.75      0.75      4.0

report_3 = report_2
report_average  = report_average(report_1, report_2,report_3)
print(report_average)
             precision    recall  f1-score  support
0             0.666667  0.666667  0.666667      2.0
1             0.666667  0.666667  0.666667      2.0
avg / total   0.666667  0.666667  0.666667      4.0

Answer 2

Just another way to do this when the reports (as_dict) are passed as list. This will return the result as a dictionary.

def report_average(reports):
    mean_dict = dict()
    for label in reports[0].keys():
        dictionary = dict()

        if label in 'accuracy':
            mean_dict[label] = sum(d[label] for d in reports) / len(reports)
            continue

        for key in reports[0][label].keys():
            dictionary[key] = sum(d[label][key] for d in reports) / len(reports)
        mean_dict[label] = dictionary

    return mean_dict

Answer 3

This is a modification of Rayn's function to fit more sklearn classification report and will have the same format as it

def report_average(*args):
report_list = list()
for report in args:
    splited = [' '.join(x.split()) for x in report.split('\n\n')]
    header = [x for x in splited[0].split(' ')]
    data = np.array(splited[1].split(' ')).reshape(-1, len(header) + 1)
    data = np.delete(data, 0, 1).astype(float)
    rest = splited[2].split(' ')
    accuarcy =np.array([0, 0, rest[1], rest[2]]).astype(float).reshape(-1, len(header))
    macro_avg = np.array([rest[5:9]]).astype(float).reshape(-1, len(header))
    weighted_avg = np.array([rest[11:]]).astype(float).reshape(-1, len(header))
    #avg_total = np.array([x for x in avg]).astype(float).reshape(-1, len(header))
    df = pd.DataFrame(np.concatenate((data, accuarcy,macro_avg,weighted_avg)), columns=header)
    report_list.append(df)
res = reduce(lambda x, y: x.add(y, fill_value=0), report_list) / len(report_list)
return res.rename(index={res.index[-3]: 'accuracy',res.index[-2]: 'macro_avg',res.index[-1]: 'weighted_avg'})