The starter code I provided is a basic DNN structure for modelling complex non-linear relationships between five soil properties and 3000+ predictors.
In practice, I found that some of the properties require even more complex DNN structure to achieve better predictive accuracy. The 50-50-50 setup is a very solid starting point for the readers to begin their own experiments.
Thank you. How did you come up with the 50-50-50 setup, or was it purely empirical ? Did you try something simpler first, and how did that simpler method perform vis-a-vis this DNN ? Congratulations on topping the leaderboard.
Thanks! Yes, I always start with much simpler networks like 10, 10-10, 10-10-10. Unfortunately, the regression problems here are quite complex hence bigger networks are required (well, it wouldn't be on Kaggle otherwise).
In practice, I found that some of the properties require even more complex DNN structure to achieve better predictive accuracy. The 50-50-50 setup is a very solid starting point for the readers to begin their own experiments.