Could DGL-KE be applied to undirected graph? about dgl-ke HOT 7 OPEN

For DGL, an undirected graph is stored as a bi-directional graph. If there is an edge from A to B, there will be an edge from B to A. In DGL-KE, we use DGL to store DGL to store knowledge graphs. If your graph is undirected, I think you can store it as a bi-directional graph. You can store a triplet (A, r, B) and (B, r', A) for every edge. I think you can create a separate relation for the reverse edge.

from dgl-ke.

The reason why to change the undirected graph to stored as bi-directional graph is that the graph now contains two types of relation, concretely, inverse relation, right?

from dgl-ke.

I think so. But it is also up to the model. Some of the KGE models, such as DistMult, are symmetric. For the symmetric models, I guess one direction is enough.

from dgl-ke.

We know that DGL-KE have provided knowledge graph embedding algorithm like TransE and many others.
I see that PyG and some other similar system like GraphVite both have implements of random walk node embedding algorithm like Node2Vec, DeepWalk, LINE, etc, does DLG-KE plan to develop this feature?

from dgl-ke.

Yes, we plan to implement them. We have a prototype of DeepWalk and Metapath2vec, but haven't fully optimized them yet.

from dgl-ke.

That's great.
Now we have two choices of embedding models, one is node embedding like DeepWalk, Node2Vec, while another knowledage graph embedding like TransE, DistMult.
As the first question of this issue we talked, actually, KGE models could be used for homogeneous graph where DeepWalk, Node2Vec is applied. But this does not mean that KGE models could replace those node embedding models, right? Otherwise, there is no need to develop and optimize DeepWalk.
I still don't know the performance differences between the two, do you ever consider this problem. As now u developed the two models, please provide some advice on how to choose the two while user of DGL-KE tried to do node embedding with homogeneous graph, thx.

from dgl-ke.

This is an interesting question. We never compare the two kinds of algorithms. KGE is supervised by link prediction while DeepWalk is random walk sequence. When we apply KGE on a homogeneous graph, it should be similar to simple matrix factorization. DeepWalk should outperform matrix factorization, such as SVD?

from dgl-ke.