The process of negative sample selection strategy. First, MGCNSS generates centroid vectors |$C_{p}$| and |$C_{u}$| from positive samples and the remaining unlabeled samples, respectively. Then, MGCNSS calculates the CS between each sample in the unlabeled sample set and the centroid vectors |$C_{p}$| and |$C_{u}$|⁠, respectively. Based on the CS, we could divide the unlabeled samples into two groups, Likely Positive Pairs (LP) and Likely Negative Pairs (LN). Next, we update the two centroid vectors using LN and LP. Moreover, MGCNSS adopts ES to repeat these steps until the centroid vectors |$C_{p}$| and |$C_{u}$| converge. Finally, we regard LN as the reliable negative sample set.