Figure 4:

Results from all 5 datasets (each dataset generated by a different random seed) and parameter choices on highly correlated synthetic datasets. The parentheses contain the best Recovery-F1 scores averaged over all 5 datasets. MCP is shown with γ fixed at 1.5 and 25, and all other choices for γ lie between the shown regions. Our methods and L0Learn outperfom MCP and LASSO in terms of the AUC (left and middle), and better recover the true support (right). L0Learn’s performance heavily overlaps with our methods. Our methods have a computational advantage over L0Learn as shown in the last section.