Algorithm 2.

Iterative Confidence Preserving Machine

Input: labeled training data ${\{{\mathbf{x}}_i,y_i\}}_{i=1}^n$ with index set 𝒟 = {1, 2, . . . , n}, unlabeled test data ${\{{\mathbf{x}}_i^{te}\}}_{i=1}^m$ with index set 𝒟te = {1, 2, . . . , m}
Output: person-specific classifier wt
1:	ℰ ←𝒟, ℋ←∅;
2:	(w+,w−) ← solve (2);
3:	(ℰ,ℋ) using (1);
4:	repeat
5:	Update relabels ${\eta }_j^{+},{\eta }_j^{-}$, ∀j ∈ ℋ;
6:	(w+,w−) ← solve (2) with fixed ℰ and ℋ;
7:	Estimate virtual labels ${\{{\widehat{y}}_i\}}_{i=1}^m$,
	$${\widehat{y}}_i=\{\begin{array}{ll}1\hfill & {\mathbf{w}}_y^{\top }{\mathbf{x}}_i^{te}>0,\forall y\in \{-1,+1\},\hfill \\ -1\hfill & {\mathbf{w}}_y^{\top }{\mathbf{x}}_i^{te}<0,\forall y\in \{-1,+1\},\hfill \\ 0\hfill & \text{otherwise.}\hfill \end{array}$$
8:	${\mathcal{E}}_t=\{i\in {\mathcal{D}}^{te}\mid \text{sign}({\mathbf{w}}_{+}^{\top }{\mathbf{x}}_i^{te})=\text{sign}({\mathbf{w}}_{-}^{\top }{\mathbf{x}}_i^{te})\}$;
9:	if ∃i, j ∈ ℰt, s.t. ŷi = −1, ŷj = 1 then
10:	wt ← solve (4) given Xte and ${\{{\widehat{y}}_i\}}_{i=1}^m$;
11:	else
12:	${\mathbf{w}}_t=\frac{1}{2}({\mathbf{w}}_{+}+{\mathbf{w}}_{-})$;
13:	end if
14:	Update ${\mathcal{E}}_t=\{i\in {\mathcal{D}}^{te}\mid {\widehat{y}}_i=\text{sign}({\mathbf{w}}_t^{\top }{\mathbf{x}}_i^{te})\}$;
15:	Update (ℰ,ℋ) ← (1);
16:	ℰ ←ℰ ∪ℰt;
17:	until convergence