How to choose the optimal threshold for imbalanced classification

threshold_precision_recall_fscore.py

# Calculate the f-score
fscore = (2 * precision * recall) / (precision + recall)

# Find the optimal threshold
index = np.argmax(fscore)
thresholdOpt = round(thresholds[index], ndigits = 4)
fscoreOpt = round(fscore[index], ndigits = 4)
recallOpt = round(recall[index], ndigits = 4)
precisionOpt = round(precision[index], ndigits = 4)
print('Best Threshold: {} with F-Score: {}'.format(thresholdOpt, fscoreOpt))
print('Recall: {}, Precision: {}'.format(recallOpt, precisionOpt))

# Create a data viz
plotnine.options.figure_size = (8, 4.8)
(
    ggplot(data = df_recall_precision)+
    geom_point(aes(x = 'Recall',
                   y = 'Precision'),
               size = 0.4)+
    # Best threshold
    geom_point(aes(x = recallOpt,
                   y = precisionOpt),
               color = '#981220',
               size = 4)+
    geom_line(aes(x = 'Recall',
                  y = 'Precision'))+
    # Annotate the text
    geom_text(aes(x = recallOpt,
                  y = precisionOpt),
              label = 'Optimal threshold \n for class: {}'.format(thresholdOpt),
              nudge_x = 0.18,
              nudge_y = 0,
              size = 10,
              fontstyle = 'italic')+
    labs(title = 'Recall Precision Curve')+
    xlab('Recall')+
    ylab('Precision')+
    theme_minimal()
)