How to choose the optimal threshold for imbalanced classification

threshold_roc_gmean.py

# Calculate the G-mean
gmean = np.sqrt(tpr * (1 - fpr))

# Find the optimal threshold
index = np.argmax(gmean)
thresholdOpt = round(thresholds[index], ndigits = 4)
gmeanOpt = round(gmean[index], ndigits = 4)
fprOpt = round(fpr[index], ndigits = 4)
tprOpt = round(tpr[index], ndigits = 4)
print('Best Threshold: {} with G-Mean: {}'.format(thresholdOpt, gmeanOpt))
print('FPR: {}, TPR: {}'.format(fprOpt, tprOpt))

# Create data viz
plotnine.options.figure_size = (8, 4.8)
(
    ggplot(data = df_fpr_tpr)+
    geom_point(aes(x = 'FPR',
                   y = 'TPR'),
               size = 0.4)+
    # Best threshold
    geom_point(aes(x = fprOpt,
                   y = tprOpt),
               color = '#981220',
               size = 4)+
    geom_line(aes(x = 'FPR',
                  y = 'TPR'))+
    geom_text(aes(x = fprOpt,
                  y = tprOpt),
              label = 'Optimal threshold \n for class: {}'.format(thresholdOpt),
              nudge_x = 0.14,
              nudge_y = -0.10,
              size = 10,
              fontstyle = 'italic')+
    labs(title = 'ROC Curve')+
    xlab('False Positive Rate (FPR)')+
    ylab('True Positive Rate (TPR)')+
    theme_minimal()
)