Create embeddings directly from an adjacency matrix (e.g. numpy.array or scipy.sparse)? about gem HOT 2 OPEN

Each of the methods has a edge_f attribute which can be used to load the graph from a file containing edges, one edge per line. That is actually pretty fast and should work for your case.

from gem.

It looks like .learn_embedding methods with edge_f provided call graph_util.loadGraphFromEdgeListTxt:

This function constructs the complete network in-memory as a networkx object. Creating a networkx graph is the inefficient step that is slow and memory-intensive with the STRING network.

For example, node2vec.learn_embedding ends up reading edge_f to a networkx graph, then writing it to an edgelist. From the code below, it doesn't seem the networkx graph is actually used for anything besides rewriting the edgelist file:

For lle.learn_embedding, it looks like a scipy.sparse matrix is actually what is required for computation, but that a networkx graph is loaded from the edgelist file to create the sparse matrix:

For sdne.learn_embedding, a sparse matrix is also used:

I'd love access the unified API of GEM, but with support for providing an edgelist file or scipy.sparse matrix directly as input, without having this converted to a networkx graph. It seems that several of the methods actually use more efficient data structures than the networkx graph and don't need networkx features at all.

Do you have any suggestions on whether this would be possible? Even if just for one method? Currently, we've used node2vec without GEM to side-step this problem (notebook), but would love the ability to use the implementations in this package!

from gem.