4
$\begingroup$

Training and testing data have around 1% positives, but the model predicts only around 0.1% as positives.

The model is an xgboost classifier.

I’ve tried calibration but it didn’t improve much. I also don’t want to pick thresholds since the final goal is to output probabilities.

What I want is for the model to have a number of classified positives similar to the number of positives in the actual data.

$\endgroup$
  • $\begingroup$ What method of calibration did you try? $\endgroup$ – Ben Reiniger Jan 31 at 17:40
  • $\begingroup$ @BenReiniger isotonic and sigmoid (sklearn's CalibratedClassifierCV) $\endgroup$ – Rodrigo Nader Jan 31 at 18:54
  • $\begingroup$ "...the model predicts only around 0.1% as positives" with the default cutoff? Have you looked at a calibration plot? How do other metrics look? Maybe the model is well-calibrated, but not very discriminative: then predictions might only barely breach 50%, but that's an honest assessment of the likelihood of being in the positive class. $\endgroup$ – Ben Reiniger Jan 31 at 19:02
  • 1
    $\begingroup$ Yes, the calibration plot is very distorted (off the middle line), and when I apply a calibrator it helps, but still, only few samples are classified as positives. Precision and recall seem fine, the model seems to be performing well, the problem is related to the probability distribution. Too few samples are getting a probability above 50%. $\endgroup$ – Rodrigo Nader Jan 31 at 19:08
1
$\begingroup$

The first (and easiest) option is to make sure that your model is calibrated in probabilites. In Python, it means that you should pass the option binary:logistic in your fitting method.

The alternative is to transform the output of your model into probabilities. There are different approaches for that.

This could be achieved with some sort of regression techniques to find the relationship between probabilities and your output. Python's isotonic regression should work for that purpose. However, without more information on your score ditribution it is possible it doesn't work well.

This can also be achieved with platt scaling : transforming your output into binary prediction (0 and 1) with a threshold, then calibrate a logistic regression on those new variables. It is relatively easy to do, but in my experience doesn't necssarily work well with unbalanced problems with non-linear relationships.

Finally, there are some approaches that just correct the output depending on your model. For logistic regression that would mean changing your bias variable so that the overall predicted proportion match the one of your data-set. This also can be used to counter the effects of rare events (see this). I have found this to work wuite well with logistic regression. However, I am not sure if it would directly be appliable to XGBoost, but it could be worth a try.

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.