Qamro/Prediction-Smartphone-Addiction-Submission-V2

Here's the enhanced version, honest summary of what actually moved the needle:

What improved it:

• Feature engineering was the real driver: missingness indicators for every column (missingness itself carries signal here), plus ratio/interaction features like social_to_screen, sleep_deficit, weekday_weekend_diff, screen_per_age, etc.
• LightGBM with these new features: OOF AUC 0.9628 (up from 0.9620).

Result:

submission_v2.csv, same valid format (296,302 rows matching sample submission), and it should score modestly above your 0.96323 leaderboard result based on the OOF(Out-of-Fold) gain.