1. 8

  2. 6

    Actually a really good rundown. It’s about 50% data collection and cleaning, 10% training and running a model, and 40% analysis of “why didn’t this work well”. These ratios are how you know someone is an actual practitioner and not just a poseur.

    1. 1

      I’m very confused why XGBoost was used here with seemingly no rationale. Modeling failure rates is a pretty old problem, and there are many models in this space that are used for everything from factory equipment failure to server failure. I appreciate talking about data cleaning and choosing training sets, but I still don’t understand why XGBoost was used and not something like a Cox Hazards Model.