Algorithm 2.1.

Stochastic Coordinate Coding with Missing Values

Initialization:
	Samples X = {x1, x2, …, xn}, missing indices Ω = {Ω1, Ω2, …, Ωn}, λ ∈ ℝ, initial dictionary 𝒟0 ∈ ℝm×k, initial combination coeficients Z = {z1, z2, …, zn}, number of iterations T.
1:	ℋ ∈ ℝk×k ← 0
2:	for t = 1 to T do
3:	for i = 1 to n do
4:	Update coefficients via a few iterations of coordinate descent according to (3)
5:	zi ← arg minzi $f_{𝒟}(z_i)\equiv \frac{1}{2}{\Vert {𝒫}_{{\mathrm{\Omega }}_i}(𝒟z_i-x_i)\Vert }_2^2+\lambda {\Vert z_i\Vert }_1$.
6:	Update Hessian matrix
7:	$\mathscr{H}=\mathscr{H}+z_i{z}_i^T$,
8:	Update the dictionary 𝒟i−1 colum by column
9:	for j ∈ {t|1 ≤ t ≤ k, t ∈ ℕ, zi(t) ≠ 0} do
10:	$u_j={𝒫}_{{\mathrm{\Omega }}_i}({𝒟}_{·,j}^{i-1})-\frac{1}{\mathscr{H}[j,j]}{z}_i(j)*{𝒫}_{{\mathrm{\Omega }}_i}({𝒟}^{i-1}*z_i-x_i)$.
11:	${𝒫}_{{\mathrm{\Omega }}_i}({𝒟}_{·,j}^i)\leftarrow u_j$.
12:	${𝒟}_{·j}\leftarrow \frac{1}{\text{max}\{{\Vert D_{·j}\Vert }_2,1\}}{𝒟}_{·j}$.
13:	end for
14:	end for
15:	𝒟0 ← 𝒟n
16:	end for
	Output: 𝒟n.