Closed marvin-hansen closed 2 years ago
Totally agree with the above, this project started before the rapid development of the graph learning field and should move towards becoming a tooling library for interfacing with external graph ML projects.
James,
Take a look at Pilosa
Any hypergraph can be represented as a NxM matrix and Pilosa does just that so that might be an option if performance and latency at scale matters and you can afford a bit of engineering work.
Arango already comes with a ready to use graph deep learning solution for it's graph. It's marvelous and immediately production ready although no type system and no rule based reasoning.
If you aren't bound to TypeDB, Arango is the way to go. If you are invested in TypeDB, well, good luck with patching ML together as your are essentially all on your own.
On Fri, Apr 8, 2022, 18:15 James Fletcher @.***> wrote:
Totally agree with the above, this project started before the rapid development of the graph learning field and should move towards becoming a tooling library for interfacing with external graph ML projects.
— Reply to this email directly, view it on GitHub https://github.com/vaticle/kglib/issues/157#issuecomment-1092704534, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFYR7XHQ6WYYK2DRHMSBN7LVEABLBANCNFSM5Q3Y777Q . You are receiving this because you modified the open/close state.Message ID: @.***>
Hadn't heard of Pilosa, does look interesting. We moved away from solutions that required the whole knowledge graph to be ingested due to scalability concerns. We thought it better to take a subgraph sampling approach for that reason. Their indexing approach sounds promising though.
As an employee at Vaticle and author of this repo I'll be staying with TypeDB! :)
Have you considered DGL as underlying ML engine?
Imho, it's way easier to convert a typeDB query result into a networkX graph which than can be processed by the many state of the art algorithms in DGL instead of reinventing the wheel. This would substantially boost the usefulness of this project and actually give people a strong reason to build on typeDB in the first place.
https://www.dgl.ai
https://github.com/awslabs/dgl-ke
Any thoughts on this?