4 ms·
Data Cascades in Machine Learning
- 6gvONxR4sf7o 5y agoThese kinds of papers are my favorite ML papers (see also: 'machine learning is the high interest credit card of technical debt'). The org design and project aspects of ML projects are some of the most pernicious issues I face, while the modeling and other fun stuff often ends up not being that hard once the right pieces are in place.
- washedup 5y agoGood article. I would suggest following it up with https://pair.withgoogle.com/chapter/data-collection/ https://pair.withgoogle.com/chapter/data-collection/ Monitoring drift in the inputs, predictions, and performance are all crucial for any models in production. Personally, for input/target drift, I prefer using the D-stat from Kolmogorov–Smirnov test to look for any distribution changes.