Link prediction questions: train-test, accuracy and prediction

[VoVAllen]

You need to create a complete graph first, calculate the score on each edge and then got the inferenced graph.

@ogggcar

but I may need to score every pais of nodes from this new graph, in order to keep those with a higher score as that which have an edge, right?

Exactly

@ogggcar

@VoVAllen

For the error you met, this might due to the shape of src or pos_idx, which you can use tensor.reshape to change it to the same shape

Any code help for this?

Can you show the the shape of src and pos_idx?

[ogggcar]

@VoVAllen

You need to create a complete graph first, calculate the score on each edge and then got the inferenced graph.

The problem is I already have this graph with its out of the graph embeddings. This graphs are extracted from a text, so its essencial this are real nodes with real information. Graphs used during training were annotated with this 3 types fof links, and now I would like to predict the other 2 types on new no annotated graphs. Sorry if I dont explain myself.

@VoVAllen

Can you show the the shape of src and pos_idx?

This is part of the problem, if a use new real graph without that type of link, pos_idx is obviously messed up:

Graphs: Graph(num_nodes={‘ent’: 122},
num_edges={(‘ent’, ‘link1’, ‘ent’): 0, (‘ent’, ‘link2’, ‘ent’): 486, (‘ent’, ‘link3’, ‘ent’): 0})

pos_idx: torch.Size([0, 1])
src: (14884,)

If i use a graph from the training set, which is annotated with those links, it looks like this:

Graph(num_nodes={‘ent’: 84},
num_edges={(‘ent’, ‘link1’, ‘ent’): 67, (‘ent’, ‘link2’, ‘ent’): 289, (‘ent’, ‘link3’, ‘ent’): 62}

pos_idx: torch.Size([67, 1])
src: (7056,)

I think here you may see its the first graph and not the second one, in which I wanna have inference, being it a real graph from real text with real embeddings I need to use. Problem is, as said, that this new graphs doesnt include any initial link of those types (links 1 and 3), so I dont know how to calculate the score for every pair of nodes.

Thank you so much and sorry to bother.