I wouldn’t say it’s an argument against simple models. Rather, if you understand the fundamental aspects of your problem then the solution will often avoid unnecessary complexity.
00:00 Intro 01:29 Topics to be covered 02:02 XOR problem 04:45 Time-series trick 10:20 Weighted Linear Regression 17:23 Passive-Agressive Algorithms 21:42 Recommender Systems 26:20 Example of Video Games 31:33 Example of Chickens 35:26 Conclusion
This talk makes me fall in love with feature engineering all over again! Today more relevant than ever, with new regulation coming in, keeping our models simple and interpretable is a must.
Not the author, but I may recommend reading "Causal Inference for the Brave and True", with the radial basis functions he was doing a kind of "synthetic control" and maybe finding a resource on splines. For the later portions on machiene learning you can tell he's read "Statistical Rethinking".
This talk is a bit dated. Just because it's easy to understand out-of.the-box, doesn't mean that it is better. I'd rather peer into or interrogate a more accurate deep-learning model to understand how it's working than be satisfied by easy-to-generate plots of a simple and less accurate model.
