Model selection with Lasso-Zero and a robust extension with an application to the problem of missing covariates

We propose a new model selection technique based on the limit of the lasso path as the penalty parameter tends to zero. The method provably guarantees model selection under a weaker condition than the lasso and performs better empirically in terms of false discovery rate (FDR). We extend the method to the situation of missing covariates.

• Speaker: Sylvain Sardy, Université de Genève
• Friday 01 February 2019, 16:00–17:00
• Venue: MR12.
• Series: Statistics; organiser: Dr Sergio Bacallado.