This is an approach that I have used when doing attribution. Given error signals of a larger system, I couldn’t get great performance to attribute the errors to

This is an approach that I have used when doing attribution. Given error signals of a larger system, I couldn’t get great performance to attribute the errors to a particular broken component in the system. However, when I broke down that component into its set of particular issues and built a classifier per issue, I was able to get great performance. With the light weight models we used, it was straight forward to automate most of the training / validation of these component-issue specific models and decom them when the issue no longer existed (a fix was put in).