# Improving roc auc score when accuracy is good

I have got a binary classification problem with large dataset of dimensions (1155918, 55)

Also dataset is fairly balanced of 67% Class 0 , 33% Class 1.

I am getting test accuracy of 73% in test set and auc score is 50 % Recall is 0.02 for Class 1 I am using a logistic regression and also tried pycaret's classification algorithm

• What methods should I try to improve auc - score Mar 14, 2022 at 9:51
• What is the statistical metric most appropriate to the problem being solved? Accuracy, auroc, some other? AUROC is about rank-ordering vs the confusion matrix which is about tradeoffs. What is the cutoff value used to determine the confusion matrix and is this appropriate to the problem? Mar 14, 2022 at 10:59
• This post could be worth a read.
– Dave
Mar 14, 2022 at 12:22

## 1 Answer

A majority baseline classifier would have 67% accuracy just by predicting every instance as class 0, so 73% accuracy is not especially good. The AUC is a more informative measure, but an AUC of 0.5 is actually the minimum a classifier can do.

And indeed, apparently this classifier is not doing much more than a baseline classifier:

• Recall for class 1 is 0.02, so the number of True Positives (TP) is 67734*0.02=1355.
• Precision for class 1 is 0.36 so the number of predicted positives (TP+FP) is 1355/0.36=3763.
• This means that the classifier predicts only 3763/(184508+67734)=1.5% of the instances as class 1, even though the imbalance is not severe.

So what happens in this: most of the time the classifier doesn't succeed distinguishing the two classes, so it just predicts the majority class 0 (98.5% of the time).

Without any detail it's impossible to know why, maybe the features are not good enough indicators, maybe there is overfitting, maybe logistic regression is not the right approach for this dataset...