The Lasso is a popular tool for variable selection because it is simple and often works well in practice. However, its selection stability suffers when predictors are highly correlated, leading to inconsistent choices across similar datasets. Existing fixes have limitations, which motivated the authors of a recent arXiv preprint to propose a new approach.

The method integrates a weighting scheme directly into the Lasso penalty. Each predictor's weight is an increasing function of a correlation-adjusted ranking that reflects its predictive power. This down-weights unstable or redundant predictors while preserving the Lasso's sparsity-inducing behavior.

In experiments on both simulated and real-world datasets, the proposed technique improved selection stability compared with standard Lasso. Additional results suggest the weighting scheme can be applied to other regularization-based selection methods, pointing toward a general-purpose solution for a known weakness in penalized regression.