Overview Lasso implements linear regression with L1 regularization. The L1 penalty can shrink some coefficients to exactly zero, performing automatic feature selection. Lasso uses coordinate descent optimization.Constructor import { Lasso } from "bun-scikit" ; const model = new Lasso ( options );
Parameters Configuration options for the model Regularization strength. Higher values specify stronger regularization:
alpha = 0: Equivalent to ordinary least squares (not recommended)alpha > 0: Adds L1 penalty (sum of absolute values of coefficients) Whether to calculate the intercept for this model
Maximum number of coordinate descent iterations
Tolerance for optimization convergence
Methods fit Fit the lasso regression model using training data.
fit ( X : Matrix , y : Vector , sampleWeight ?: Vector ): this
Parameters:
X: Training data matrix of shape [nSamples, nFeatures]y: Target values vector of length nSamplessampleWeight: Optional sample weights (not currently used)
Returns: The fitted model instanceExample: const X = [[ 0 ], [ 1 ], [ 2 ], [ 3 ], [ 4 ], [ 5 ]]; const y = [ 0 , 1 , 2 , 3 , 4 , 5 ]; const lasso = new Lasso ({ alpha: 0.01 , maxIter: 2000 , tolerance: 1e-8 }); lasso . fit ( X , y );
predict Predict using the lasso regression model.
predict ( X : Matrix ): Vector
Parameters:
X: Samples matrix of shape [nSamples, nFeatures]
Returns: Predicted values vector of length nSamplesExample: const predictions = lasso . predict ([[ 2.5 ]]); console . log ( predictions [ 0 ]); // Predicted value
score Return the coefficient of determination (R² score) of the prediction.
score ( X : Matrix , y : Vector ): number
Parameters:
X: Test samples matrixy: True target values
Returns: R² score (coefficient of determination)Attributes After fitting, the following attributes are available:
Estimated coefficients for the lasso regression problem. Some coefficients may be exactly zero.
Independent term in the linear model
Number of iterations run by the coordinate descent solver
Complete Example import { Lasso } from "bun-scikit" ; // Training data const X = [[ 0 ], [ 1 ], [ 2 ], [ 3 ], [ 4 ], [ 5 ]]; const y = [ 0 , 1 , 2 , 3 , 4 , 5 ]; // Create and fit the model const lasso = new Lasso ({ alpha: 0.01 , maxIter: 2000 , tolerance: 1e-8 }); lasso . fit ( X , y ); // Evaluate the model const r2 = lasso . score ( X , y ); console . log ( "R² score:" , r2 ); // >0.98 // Check convergence console . log ( "Iterations:" , lasso . nIter_ ); // Inspect coefficients console . log ( "Coefficients:" , lasso . coef_ ); console . log ( "Non-zero coefficients:" , lasso . coef_ . filter ( c => c !== 0 ). length );
Feature Selection Example import { Lasso } from "bun-scikit" ; // Data with 5 features, but only 2 are relevant const X = [ [ 1 , 2 , 0.1 , - 0.2 , 0.05 ], [ 2 , 4 , - 0.15 , 0.1 , - 0.03 ], [ 3 , 6 , 0.2 , 0.15 , 0.08 ], [ 4 , 8 , - 0.1 , - 0.05 , 0.02 ], [ 5 , 10 , 0.05 , 0.2 , - 0.06 ], ]; const y = [ 3 , 6 , 9 , 12 , 15 ]; // y ≈ x₀ + 2×x₁ // Lasso with strong regularization const lasso = new Lasso ({ alpha: 0.1 , maxIter: 2000 , tolerance: 1e-8 }); lasso . fit ( X , y ); console . log ( "Coefficients:" , lasso . coef_ ); // Features 0 and 1 should have large coefficients // Features 2, 3, 4 should be close to or exactly zero const selectedFeatures = lasso . coef_ . map (( coef , idx ) => ({ idx , coef: Math . abs ( coef ) })) . filter ( f => f . coef > 0.01 ) . map ( f => f . idx ); console . log ( "Selected features:" , selectedFeatures ); // [0, 1]
Notes
Lasso can set coefficients exactly to zero, making it useful for feature selection.
The coordinate descent algorithm is efficient for high-dimensional sparse problems.
If the solution doesn’t converge, try increasing maxIter or adjusting alpha.
For a mix of L1 and L2 regularization, see ElasticNet .
For automatic selection of the best alpha value, see LassoCV.