What does `scores_arr > np.mean(scores_arr)` mean here?

[Isuxiz]

What does scores_arr > np.mean(scores_arr) mean here?
Doesn't scores_arr store the probability of this sample belongs to positive examples?
Why not just use scores_arr >= 0.5 ?

CVAR/main.py

Line 78 in d97942d

return roc_auc_score(labels, scores), f1_score(labels, (scores_arr > np.mean(scores_arr)).astype(np.float32).tolist())

[BestActionNow]

The distribution of the predicted scores depends on the datasets you chose. On a dataset with higher posterior ctr, the score distribution will be larger overall compared to a dataset with lower ctr. Therefore the threshold you mentioned usually needs tuning to adapt the dataset. Here we simply use scores_arr > np.mean(scores_arr) and omit the hyperparameter tuning process.