Fig. 2.

Schematic illustration of the proposed deep weighted subclass-based sparse multi-task learning for feature selection. $f({\tilde{\mathbf{Y}}}^{\left(h\right)},{\tilde{\mathbf{X}}}^{\left(h\right)},{\delta }^{\left(h\right)})={\Vert {\tilde{\mathbf{Y}}}^{\left(h\right)}-{\tilde{\mathbf{X}}}^{\left(h\right)}{\mathbf{W}}^{\left(h\right)}\Vert }_2^2+{\lambda }^{\left(h\right)}{\Vert {\Delta }^{\left(h\right)}\odot {\mathbf{W}}^{\left(h\right)}\Vert }_{2,1}$ denotes an objective function in Eq. (7), δ^(h) is defined by Eq. (6), and a^(h) (a⁽⁰⁾ = 0) and ${\mathbb{F}}^{\left(h\right)}$ denote, respectively, the validation accuracy and a set of the selected features at the h-th hierarchy