So I’ve been experimenting with Kaggle insurance competition and I am amazed by the speed and acucracy of subsampling data and working through feature engineering process using ML and their feedbacks. Before all I do was creating some features and hyperparameter tuning. Agnostic data science is a real thing and domain knowledge is not always needed!
For example use trees, trees are great when it comes to interpretation and finding good features and interactions of features. Plot them add new stuff and try again, always keep track of your validation scores and improvement. I am now creating a kernel about my journey on this topic. So excited 
Note: It took me just 5 seconds of run time to find the most important feature of the competition which I found out it to be in discussions!
