Fig. 10: Improved interpretability of latent motifs over Lyu et al.13.
From: Learning low-rank latent mesoscale structures in networks
We compare subgraphs of CORONAVIRUS PPI that are induced by node sets that we sample using (a) uniformly random k-paths and (b) uniformly random k-walks with k = 10. We also compare the network dictionary with r = 9 latent motifs of CORONAVIRUS PPI that we determine using (c) the NDL algorithm (see Algorithm NDL in the SI) in the present paper to (d) the network dictionary that we determine using the NDL algorithm of Lyu et al.13. We also show the weighted adjacency matrices of the latent motifs. The 10-walks on the network tend to visit the same nodes many times. Consequently, one cannot regard the 10 × 10 mesoscale patches that correspond to those walks as the adjacency matrices of k-node subgraphs of the network. Additionally, the networks in the network dictionary in (d) have clusters of several large-degree nodes, even though the original network does not possess such mesoscale structures.
