Regression Models Reference for TryMLEasy

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/Gurneet1928/TryMLEasy/llms.txt

Use this file to discover all available pages before exploring further.

​

Linear Regression

• scikit-learn class: sklearn.linear_model.LinearRegression
• Regularization: None
• Description: Ordinary Least Squares (OLS) regression. Fits a linear model by minimising the sum of squared residuals between predicted and true values. Fast and analytically solved — no iterative optimisation is required.
• Strengths: Extremely fast; fully interpretable coefficients; zero hyperparameters to tune.
• Scaling recommended: Not required for OLS coefficients, but recommended when combining with a decomposition step (e.g., PCA).
• Typical use case: Initial baseline for any regression problem where a linear relationship between features and target is plausible. Inspect the correlation heatmap on the Feature Selection page to assess whether such a relationship exists.

​

Support Vector Regressor

• scikit-learn class: sklearn.svm.SVR
• Regularization: Implicit via the margin (controlled by C and epsilon — defaults used in TryMLEasy)
• Description: Extends Support Vector Machine concepts to regression. Fits the training data within an epsilon-insensitive tube and uses the RBF kernel by default, enabling it to model non-linear feature-target relationships.
• Strengths: Handles non-linear data without explicit feature engineering; robust to moderate outliers within the epsilon tube.
• Scaling recommended: Yes, strongly — the RBF kernel computes distances in feature space, and unscaled features with large numeric ranges will dominate. Always pair SVR with Standard Scaler or Robust Scaler.
• Typical use case: Datasets with complex, non-linear relationships between features and the continuous target variable.

​

Ridge Regression

• scikit-learn class: sklearn.linear_model.Ridge
• Regularization: L2 (adds a penalty proportional to the sum of squared coefficients)
• Description: A regularised extension of Linear Regression. The L2 penalty shrinks coefficient magnitudes toward zero without setting any to exactly zero, reducing overfitting caused by multicollinear or redundant features.
• Strengths: More robust than plain Linear Regression when features are correlated; rarely overfits; analytically solvable.
• Scaling recommended: Yes — L2 penalises coefficients equally, so features must be on a comparable scale for the regularisation to be fair.
• Typical use case: Datasets where features are correlated with each other (visible as bright off-diagonal cells in the correlation heatmap on the Feature Selection page). Ridge is the first upgrade to try when Linear Regression overfits.

​

Least Angle Regression

• scikit-learn class: sklearn.linear_model.LassoLars
• Regularization: L1 (via LASSO) combined with Least Angle Regression path
• Description: Combines the LASSO regularisation (L1 penalty, which drives some coefficients to exactly zero) with the Least Angle Regression algorithm for efficient path computation. Because it can zero out coefficients, it performs implicit feature selection during fitting.
• Strengths: Produces sparse solutions — only the most informative features receive non-zero coefficients; useful for high-dimensional datasets with many potentially irrelevant features.
• Scaling recommended: Yes — L1 regularisation is scale-sensitive; features should be normalised so the penalty is applied consistently.
• Typical use case: Datasets with many features where automatic variable selection is desired, or when model interpretability through sparsity is a priority.

​

Summary Table