Regression with an Abalone Dataset

Kaggle competition: Playground Series - Season 4, Episode 4

GOAL: predict the age of abalone from various physical measurements.
The evaluation metric for this competition is Root Mean Squared Logarithmic Error.
PRIVATE SCORE: 0.14605 (Top 20% of leaderboard)

---

Architecture

EDA

The features:

1. Sex (categoacal): M, F, and I (infant)
2. Length (continuous) - Longest shell measurement, mm
3. Diameter (continuous) - perpendicular to length, mm
4. Height (continuous) - with meat in shell, mm
5. Whole weight (continuous) - whole abalone, grams
6. Shucked weight (continuous) - weight of meat, grams
7. Viscera weight (continuous) - gut weight (after bleeding), grams
8. Shell weight (continuous) - after being dried, grams
9. Rings (integer): +1.5 gives the age in years

Generated data was used for the competition. The training set was supplemented with original data.

Choosing the default model

	Train RMSLE	Test RMSLE	Test R2
Linear Regression	0.049806	0.049200	0.667240
Ridge Regression	0.049806	0.049200	0.667239
Random Forest Regression	0.017691	0.045806	0.716091
Bagging Regression	0.020856	0.047807	0.690532
XGBoost Regression	0.041198	0.045048	0.724207
LightGBM Regression	0.044126	0.044750	0.727181
CatBoost Regression	0.042644	0.044591	0.730209

CatBoost Regression with default parameters shows the best metrics on train data.

VotingRegressor with Optuna

Creating an ensemble of different multihead-models: LightGBM x3, XGBoost Regression x3, CatBoost x3 are tuned with Optuna.
Than VotingRegressor is tuned by Optuna too.