Question: How is this different from Node Feature Extraction?

[BradKML]

There are Node Feature libraries like:

• https://github.com/benedekrozemberczki/karateclub
• https://github.com/shenweichen/GraphEmbedding

How is RoleSim different?

[abhishekmaha23]

Karate Club is a multi-functional library with implementations of many other algorithms and approaches. The GraphEmbedding implementations appear to be handling a different scenario - my implementation is specifically for role detection and clustering in graph networks.

[BradKML]

@abhishekmaha23 is there a clear way of converting a graph embedding into a Role Detection/Clustering algorithm?

[abhishekmaha23]

That is actually interesting, but I don't think there's a clear answer. I believe it might depend on the form/nature of the embeddings? For example, in places where they're represented by vectors, clustering algorithms like k-means algorithms could work (unsupervised learning).

[BradKML]

@abhishekmaha23 unsupervised learning does sound interesting, but the issue is the vector space itself might have weird topologies, and different algorithms might respond differently. https://www.researchgate.net/figure/a-Barbell-graph-B10-10-b-Roles-identiied-by-RolX-Latent-representations-in-R-2_fig3_316015951 https://www.sciencedirect.com/science/article/abs/pii/S0950705121001350#! and https://ars.els-cdn.com/content/image/1-s2.0-S0950705121001350-gr7.jpg