Algorithm 1:

Bagging PU learning

Input: positive group $P$, unlabeled group $U$, size of the bootstrap samples $K$, and number of bootstrap samples $B$

Output: a function $p\left(x_u\right)$ to assign a probability of being positive to each $x_u\in U$

$f\left(x_u\right)=0,n\left(x_u\right)=0\forall x_u\in U$//Initialize the accumulators

for $b=1$ to $B$ do:

Draw a bootstrap sample $U_b$ of size $K$ from $U$

Treat $U_b$ as negative and train a classifier $f_b$ to discriminate $P$ from $U_b$

Apply $f_b$ to generate a probability $f_b\left(x_u\right)$ of being positive for all $x_u\in U-U_b$

Update:

$f\left(x_u\right)=f\left(x_u\right)+f_b\left(x_u\right),n\left(x_u\right)=n\left(x_u\right)+1,\forall x_u\in U-U_b$

end for

Return $p\left(x_u\right)=\frac{f\left(x_u\right)}{n\left(x_u\right)}\forall x_u\in U$