I have a graph with many disjoint subgraphs that are not connected to each other. Essentially these subgraphs could represent different clusters. What is a general process to figure out node embeddings so that when I process a new incoming node, the resulting embedding will be near other nodes that belong to the same cluster that the inference time node would belong to. At inference time I am just receiving the node with its features but none of the structural information.
I have been thinking about whether one could create an embedding space that contains individual nodes as well as entire clusters(these disjoint clusters are essentially graphs themselves so cluster embeddings would be the same as global graph embeddings of the subgraphs). Essentially if I have a node A, then the embedding for that node would be similar to the embedding of the entire cluster that Node A resides in(global cluster embedding). However I do not know if this is the right way to go about it.
Overall my question how would someone create this node/cluster embedding space or if this is not possible, how would someone train a model that could take a new node with only its features at inference time and classify it to an already existing cluster?
