I am working with data in the form of content that is being reviewed to decide if it violates the policy standards or not. In the review process there are several methods to increase the chances of making the correct decision. Content might be sent for multi-review in which case it is reviewed additional times to help get to a correct decision. We can also check the historical accuracy scores of the humans reviewing the content, higher accuracy scores indicate a higher likelihood of a correct decision. Similarly, we can check some other factors such as length of review time, the type of the original content etc.

There is a separate audit process which evaluates a subset of decisions and decides on the policy correct decision in that instance. We can compare these policy correct decisions to the original decisions to evaluate the marginal lift in accuracy from the multi-review system, reviewer accuracy etc. The goal is to develop a system which incorporates these signals and outputs a score that represents our confidence that the review decision is correct.

I have done analysis comparing the marginal improvement in accuracy based on the individual signals, I have also calculated the margin of error for these scores depending on the number of samples we have. For example, when content gets multi-reviewed it is x% more like to be the correct decision with a margin of error of 0.5%. Some signals have a higher margin of error since we had less samples.

I'm trying to figure out the best way to incorporate all of the signals into a single score. I am not sure if regular classifiers will apply since only a portion of posts get multi-reviewed and we only have historical reviewer accuracy scores for a portion.

The way I have been thinking about it in my head is in the form of a decision tree, if the content has been multi-reviewed yes/no, if we have an accuracy score yes/no. Perhaps a decision tree / random forest model would be suitable. It would be great if the model was interpretable and the output could be explained. I would also be open to a more straight forward solution without the need for complex modelling.

Thanks for any advice you can provide on this :)


1 Answer 1


Combing multiple models together is called typically ensembling.

You are describing hierarchical ensembling. There are simpler methods for ensembling, such as voting and weighted voting.


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.