The Importance of Feature Engineering in the Age of AutoML

With all the fancy AutoML tools popping up, people think feature engineering is dead. It’s not. I’ve found that understanding your domain and creating meaningful features is still the highest-leverage activity you can do. You can give a mediocre model amazing features, and it will perform decently. But give an amazing model garbage features? It will fail. Talk to domain experts. Read about the problem you’re solving. Sometimes a simple ratio between two existing columns can unlock more accuracy than swapping out your entire neural network architecture.