Kaggle - National Data Science Bowl 2019
| Public Leeaderboard(QWK) | Private Leaderboard(QWK) |
|---|---|
| 593(0.551) | 62(0.550) |
Features
This repository contains my solution for the National Data Science Bowl 2019 held in Kaggle over the last 3 months. The solution in itself is pretty straightforward but I manually engineered a number of features which I found very interesting.
- Individual
game/activityperformance features:- for each of the
gameandactivityI looked into mainly 3 things (they are different for each of the games and activities but kind of follow the below pattern): a.weight: Assign weights to each round of the game as per their difficulty (I played the games - so the weights were subjective) b.round: Number of rounds completed c.misses: Number of misses each round
- for each of the
Then the accuracy/performance of the game/activity was measured by something like:
where a penalty was assigned for skipping rounds. The equations for penalty are subjective and I played the games over and over to tune the numbers so it's not a good approach but worked out pretty well at the end.
- Path Efficiency:
I calculated the efficiency of the
pathtaken by a child to arrive at the assignment. It was done by a very similar process to what we use for full backward tree search in dynamic programming for a discrete time Markov Reward Process (MRP) with a discounting. This was driven by the fact that child who skips or deviates from the sequence of games designed by PBS KIDS, often does a poorly in an assignment.
The above two features helped the model quite a bit.
Apart from that past accuracy, count and a few onehot features were used. In total there were only ~35-40 features in my final model.
Model
The final model is a plain single LGB bagged 10 times. Each of the bag takes around ~2 minutes including the CV in my 32-core machine to run. I didn't get much time to tune the hyperparameters so used the one from a few I tried out manually. In whole the complete solution runs in ~30 minutes.