What does the index of the embedding mean?

[vadhoob]

Hi,

I am currently generating embedding using g.embed method using the following code block:

g = node2vec.PreComp(p=in_parameter, q=out_parameter, workers=4, verbose=True)
g.read_edg(edges_path, delimiter=',', weighted=False, directed=False)

g.preprocess_transition_probs()

emb = g.embed(dim=num_dim, num_walks=number_of_walks, walk_length=lenght_walk, window_size=size_window, epochs=num_epochs)

I then use for example emb[0] to get the embedding associated with node 1. Is this the correct interpretation of what g.emd produces? I am using these embeddings for a downstream classification task where I join these embeddings with other features based on index key.

Many thanks!

[RemyLau]

Hi @vadhoob, thanks for your interest in using PecanPy! In this case, the index of emd will be consistent with the graph g, i.e., the first row in the emd matrix correspond to the embedding vector of the first node in g.

The relevant code is as follows

PecanPy/src/pecanpy/pecanpy.py

Line 285 in a2b5f7d

return w2v.wv[self.nodes]

where w2v is the trained word2vec object, and w2v.wv is the corresponding keyedvectors. This keyedvectors can be indexed by the corresponding words (i.e., node IDs in the graph). Thus, using the node IDs self.nodes of the graph g via w2v.wv[self.nodes] produces an embedding matrix whose node order is consistent with the graph.

[vadhoob]

Thanks @RemyLau, based on your reply this is what I have done:

node_list = [int(x) for x in g.nodes]

and then:

embedding_with_key = dict(zip(node_list, emb))

This is now gives me a dictionary containing the node id and its corresponding embedding. Is this correct? I am trying to avoid doing any loop operation as my graph is quite big.

Many thanks!

[RemyLau]

Yes, this looks correct to me. Maybe if you only need the embedding_with_key but not the node_list, you can simply do the following

embedding_with_str_key = dict(zip(g.nodes, emb))

# Or map the nodes to int if that's what you wanted as well
embedding_with_int_key = dict(zip(map(int, g.nodes), emb))

[vadhoob]

okay thank you!