linear regression

abalone dataset · predicting ring count from physical measurements

regularization

λ 0.01

learning rate 0.01

epochs 25000

train — gradient descent with the λ from the slider. search & train — finds the best λ automatically via 5-fold cross-validation, then retrains with it. this is the correct workflow.

λ min (10^x) -6

λ max (10^x) 1

grid steps 30

cv epochs 5000

R² evolution

train vs validation R² across epochs — gap reveals overfitting

RMSE evolution

train vs validation RMSE in rings — error in original units

predicted vs actual

each dot is one abalone — perfect model sits on the diagonal

feature weights

weight magnitude after training — positive = more rings

weight stability across folds

distribution of each feature's weight across 5 CV folds — wide spread reveals instability

grid search — λ vs validation rmse

5-fold cross-validation across λ values · best λ used to retrain