playground-series-s6e3

[merry spruce]

Hii! Soo the competition just began, and I wanted to do an EDA on the data, cuz that's something which I haven't done a lot in the past and wanted to try my hand there. I think it also might be beneficial to know the features and how they interact in the data.
Here's the notebook for that: https://www.kaggle.com/code/rupankarmajumdar/predict-customer-churn-s6e3-eda-baseline
Please do let me know if I need to add anything for making the EDA more impactful. Also gave a baseline with LightGBM.

[twin ruin]

explore and feedback is welcome ! https://www.kaggle.com/code/suhanigupta04/predict-customer-churn-playground-problem

[analog bramble]

Any feedback, please? Thanks!! https://www.kaggle.com/code/sangeetha007/trial-3-churn-logistic-regression

[rare folio]

I'd consider using CatBoost given the dataset is dominated by category values.

You may also consider using the SVM but if you do use my HVRT repository to reduce samples down to 50k to make it feasible https://github.com/jpeaceau/hvrt. Also, determine irreducible error, investigate the residuals of the SVM, logistical regression model and CatBoost to evaluate a potential ensemble approach.

Do not use random sampling, with notable class imbalance of features, it will likely hurt performance.

Also hyperparameter optimization (HPO) with cross-validation for logistical regression and CatBoost is appropriate. You can for the SVM as well, though I'd use multiprocessing as even while reducing down to 50k samples to make it usable, the SVM is quite slow

[late thistle]

sure

[late thistle]

nice

[coarse trout]

I recently worked on the Playground Series S6E3 (Churn Prediction) competition and would really appreciate feedback from you on my approach. here is notebook link https://www.kaggle.com/code/mahmudulmadu/s6e3-adv-fe-xgboost-lightgbm-catboost-ensemble

[idle terrace]

Hi, I'm Doing RealMLP + CatBoost + XGB, is that a good approach? I've gotten a score of only 0.91702.

can someone please provide feedback?

[zenith mesa]

Yes, that is a good approach. A score of .91702 is very good!

[zenith mesa]

You could try adding LightGBM to your ensemble and/or experimenting with weighting strategies

[idle terrace]

hey, LGBM, even with Extensive HPO actually reduced the score, I'm trying Hill Climb based weighting now.

[idle terrace]

That Shake up was crazy, I climbed almost 200 spots.

[grand vessel]

I finished 21st in PS S6E3 and published my writeup.
It focuses on final blend selection with Ridge and Nelder-Mead.
If it helps anyone, here is the link:
https://www.kaggle.com/competitions/playground-series-s6e3/writeups/21st-place-solution-final-blend-selection-with-ri

[rotund briar]

check my bio 😁

[grand ginkgo]

check my bio 😁

[queen vapor]

Check my bio 😁

[shell ridge]

No!