Let’s say I have a classification model. And my job is to predict the correct class out of 30 different classes. The current accuracy is 60%.

The thing is: I have to consume another team’s classification result which is 80% accurate. So I’m using their prediction result as a feature. I’ll call it “golden feature”. Let’s say I’m aiming >80% accuracy with the golden feature.

Here is my current approach:

(I’m using Deep Learning.) I have several features and each feature has its own weight. I also create a weight vector for one hot vector (1 by 30) of “golden feature” and train all weights together. However the result doesn’t seem to provide much.

I thought about the reason why and realized that the learned vector (30 by 30) won’t be that meaningful. They would be just positive numbers. (Please yell at me if my reasoning is wrong!)

Has anyone faced the similar problem? Any suggestion will be highly appreciated! The method that you suggest doesn’t have to be Deep Learning approach.

  • $\begingroup$ What is consume? Pseudo label? $\endgroup$
    – Aditya
    Commented Jul 23, 2019 at 2:29
  • $\begingroup$ @Aditya NOT pseudo label. Predicted label by another pretty accurate model. $\endgroup$
    – aerin
    Commented Jul 23, 2019 at 2:58
  • $\begingroup$ You are leaking information by doing so.(if these aren't the oof's) $\endgroup$
    – Aditya
    Commented Jul 23, 2019 at 10:05
  • $\begingroup$ After you fed the output of the 80% accurate model as input to your model the resulting accuracy was still near 60% or it got near 80% but somewhat lower? $\endgroup$
    – Khanis Rok
    Commented Sep 30, 2021 at 0:34

1 Answer 1


You are describing ensembling, combing a collection of models.

The most common ensembling design patterns that could be applied in your situation are:

  • Stacking - The output of model becomes the input of another model.

  • Bagging - Each model votes for final result.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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