Algorithm 1 Generate a Feasible Solution

Input: latent dimension K, knowledge graph G, target predicate p Output: $U^p$, $V^p$, $b^p$   1: Construct bipartite subgraph $G_p$ from G and p   2: $m\leftarrow$ number of subject entities in $G_p$   3: $n\leftarrow$ number of object entities in $G_p$   4: Generate training samples $D_p$, using uniform sampling   5: Initialize $U^p$ ($m\times K$), $V^p$ ($n\times K$), $b^p$ ($n\times 1$) with mean 0 and standard deviation $0.1$   6: for all $(s_p,o_p^{+},o_p^{-})\in D_p$ do   7:  Update $U_s^p$, based on Equation (16)   8: end for   9: return $U^p$, $V^p$, $b^p$