- Tags:: 📜Papers
- Author:: Shreya Shankar
- Link:: Predictive Modeling: A Retrospective (shreya-shankar.com) (Google Brain, Facebook)
- Source date:: 2021-01-08
- Finished date:: 2021-06-13
The typical paper that realizes that impact comes from good selection of features working with domain experts. Contradicting what is was said in Range:
To be frank, I don’t think we can benefit from domain expertise too much… It’s very hard to win a competition just by using well known methods. We need more creative solutions.
Also, two modes of experimentation. A very hacky mode, but also, as indicated in How to lead in Data Science:
So I learned the second mode, the slow mode, when I needed to show results to other people. I deploy the pipeline with the DAG scheduler, version everything – data, model, code, artifacts, you name it – and see if I can replicate results multiple times through cross validation. I never trust my results or communicate them to someone else until I’ve seen results of the second mode. I am rigorous in this mode; I put my software engineering hat on and get to work. It is tedious, but if there is anything I have learned about data science, it is that I need to be disciplined and patient in order to repeatedly reap the benefits of this sorcery (p. 14)