One of the unfortunate consequences of Kaggle's inability to host tabular data competitions any more will be that the fine art of feature engineering will slowly fade away. Feature engineering is rarely, if ever, covered in ML courses and textbooks. 1/
There is very little formal research on it, especially on how to come up with domain-specific nontrivial features. These features are often far more important for all aspects of the modeling pipeline than improved algorithms. 2/
I certainly would have never realized any of this were it not for tabular Kaggle competitions. There, over many years, a community treasure trove of incredible tricks and insights had accumulated. Most of them unique. 3/
Now, regrettably, that all seems to be coming to an end. It is still possible to find many stellar tricks and insights in the old competitions. But there has been very little, if any, new development. 4/
Many legendary Kagglers have now for all practical purposes stopped Kaggling. I really miss seeing them around. I have looked up to them and learned so much from them over the years. 5/
• • •
Missing some Tweet in this thread? You can try to
force a refresh
