Algorithm 1 The framework of Normal Constrained Active Learning

Input: Two heterogeneous social networks: ${\mathcal{G}}^s$ and ${\mathcal{G}}^t$. Two sets of labeled anchor link: The training set ${\mathcal{A}}_t$ and the validation set ${\mathcal{A}}_v$. The query pool $\mathcal{P}$. The max number of queries $n_q$. The potential entropy computation method $H_p$.

Output: The new training set ${\mathcal{A}}_t$ and the new query pool $\mathcal{P}$ 1:Initialize $n\leftarrow 0$ 2:For each $a(u_i^s,u_j^t)$ in ${\mathcal{A}}_t$ and $\mathcal{P}$, extract four types of features. 3:while $n<n_q$ do 4:    Train an SVM model $\theta$ on ${\mathcal{A}}_t$ according to the training part in MNA; 5:    For each unlabeled anchor link a in $\mathcal{P}$, use $\theta$ to predict the probabilities of its value to be 0 and 1, which is presented as $P_{\theta }(0|a)$ and $P_{\theta }(1|a)$. 6:    if $H_p$ is $H_B$ then for each link a in $\mathcal{P}$, compute $H_p(a)$ by Equation (2) 7:    else 8:        Use $\theta$ as the trained classifier in MNA, and use MNA to predict the labels of all the links in ${\mathcal{A}}_v$. 9:        Compute the probabilities $P_{YY}$ and $P_{NY}$ by Equation (4). 10:        For each a in $\mathcal{P}$, find $R(a)$, and compute $H_p(a)$ by Equation (5). 11:    end if 12:    Select the link $a_h$ which has the highest potential entropy in $\mathcal{P}$, and identify its real label. 13:    if the real label of $a_h$ is “negative” then 14:        $a_h\leftarrow 0$, ${\mathcal{A}}_t\leftarrow {\mathcal{A}}_t\cup \{a_h\}$, $\mathcal{P}\leftarrow \mathcal{P}-\{a_h\}$ 15:    else 16:        $a_h\leftarrow 1$ 17:        Find $R(a_h)$ from $\mathcal{P}$, for each link in $a_r$ in $R(a_h)$, set $a_r\leftarrow 0$ 18:        ${\mathcal{A}}_t\leftarrow {\mathcal{A}}_t\cup \{a_h\}\cup R(a_h)$, $\mathcal{P}\leftarrow \mathcal{P}-\{a_h\}-R(a_h)$ 19:    end if 20:    $n\leftarrow n+1$ 21:end while