Select few network features instead of all?

[rachithaiyappa]

Since all topological feaures in creat_numpy_files are obtained using networkx, perhaps it might be useful to remove some of them to make the code faster?

Then we'd need to decide which one's to remove

[skojaku]

Do you identify which feature can be a bottleneck? Shortest paths?

[rachithaiyappa]

Algorithm	Time Complexity	Source
nx.average_clustering(G)
nx.triangles(G)
nx.pagerank(G)
nx.pagerank(G, personalization=hot_vec_copy)
nx.average_neighbor_degree(G)
nx.eigenvector_centrality(G,tol=toler)
nx.resource_allocation_index(G,edge_s)
nx.jaccard_coefficient(G,edge_s)
nx.katz_centrality_numpy(G)
nx.adamic_adar_index(G,edge_s)
nx.preferential_attachment(G,edge_s)
nx.degree_assortativity_coefficient(G)
nx.common_neighbors(G,edge_s[ee][0], edge_s[ee][1])
nx.shortest_path_length(G)
nx.betweenness_centrality(G, normalized=True)
nx.diameter(G)
nx.transitivity(G)

[rachithaiyappa]

Quick update...for a network with ~7k nodes and 50k edges, these are the naive time taken in seconds to calculate some of the features. The bottlenecks become more evident here. I'll make a better plot and post here across a few other networks to decide which feature we can drop since there it is now very evident to me that the script with networkx is inefficient.

I wonder how the authors of the original work managed to use it for their analysis :O

[rachithaiyappa]

Here are plot from 4 of the smaller networks (<50k edges)

[rachithaiyappa]

Here is another way of looking at the same