Generate representations of heterogeneous graphs

[valclemir]

Hey guys! All good? I’m a beginner in the area of ​​graphs, so I soon came across a heterogeneous graph problem. The problem is the following, I need to model the relationship between patient, clinic, company, so that I can generate embeddings to later apply an unsupervised algorithm such as LOF or Isolation Forest to detect any anomaly. Each node, such as a patient, clinic, company, has its specific characteristics. I need to include this when generating embeddings, how do I do it?

This is my problem, can I work using homogeneous graphs and apply GraphSage or node2vec?

[BarclayII]

I would recommend you use heterogeneous graph models. You can try this: dgl/examples/core/rgcn at master · dmlc/dgl · GitHub

But of course, GraphSAGE is also possible, by first running the patient/clinic/company attributes etc. through an MLP dedicated for patients/clinics/companies/etc. that maps them into hidden states of the same size.