Ad conversion

How the model was chosen

Five models and a baseline that tells everyone "won't buy" took the same test: the data was split into 10 parts with the share of buyers kept, each part was held out once, and this was repeated 5 times. The table shows the mean over the 50 tests.

Cross-validation results; ± standard deviation for ROC AUC.
Model ROC AUC F1 Precision Recall Accuracy

Why this model

Why accuracy misleads

ROC curves of all models

    Why gender is not in the model

    We also tested the same models with gender added. ROC AUC did not rise meaningfully for any model. Gender was left out so the model does not use information that adds nothing and that could lead to discrimination in ad targeting.

    ROC AUC without and with gender.
    Model Without gender With gender Difference

    Model card