Understanding LinkNeighborLoader

[pstrybol]

Hi,

I am having some difficulties understanding all the arguments and return attributes of LinkNeighborLoader . All following questions are related to the example in this google collab: https://colab.research.google.com/drive/1r_FWLSFf9iL0OWeHeD31d_Opt031P1Nq?usp=sharing

1. Why does the attribute num_sampeld_nodes equal 3 even if we specify num_neighbors=[20, 10] so for 2 iterations? Because the attribute num_sampled_edges does equal to 2?
   1.1 does num_neigbhors sample 20 neigbhors for each node type and edge type?
2. When specifying disjoint_train_ratio=0 we expect the number of edges in edge_index and edge_label_index to be the same, why is this not the case here?
3. As pointed out in discussion questions about e_id and node_id on LinkNeighborLoader #7233 the attribute node_id is provided by the user and n_id is made by pyG to map the edge indices in the batch back to the original indices. But in the notebook these are the same? How can I map the batch edge indices back to the original indices?

4. Why is has the reverse edge type less edges sampled than the normal edge type?
5. Let's say there is a second edge type e.g. 'movie', 'country_of_origin', 'country'. If this edge type is omitted from the edge types in RandomLinkSplit and LinkNeighborLoader is this information still used to update the embeddings of the movie nodes?

Thanks in advance!

[rusty1s]

1. There are 3 types of nodes we include when sampling for 2 hops: (1) the seed nodes, (2) the nodes in the first hop, (2) the nodes in the second hop
   1.1 Yes.
2. edge_label_index will remove reverse edges. Might this be the case?
3. Manually setting node_id is no longer necessary, and is now automatically added by us via n_id. You can map indices back via n_id[edge_index].
4. This may always happen since you will sample edges for each edge type in every hop.
5. Yes, as long as it part of the returned mini-batch, it will be used for message passing.

[rusty1s]

1. Yes. Regarding 2-hops, I mean the second iteration of sampling first order neighbors from sampled neighbors.
2. Keep me posted :)

[pstrybol]

Regarding 2, the edge_label_index also includes negative samples which are not included in the edge_index. I overlooked this.

Would it be possible to define the splits yourself? Let's say I have 2 edge types and I want to train to predict 1 specific edge type. Is it as easy as just defining the edge_label_index and edge_label yourself? Also, I would not need the functionalities of LinkNeighborLoader so can I just use the basic DataLoader?

Thanks!

[rusty1s]

In general, this should be pretty straightforward. No one is forced to use RandomLinkSplit as it may not even fit every use-case. If you know the edges you want to predict on, use these as edge_label_index, and all the remaining edges as edge_index.

I don't see how you could use the basic DataLoader here though. If you are working on a single graph, you either can use no loader at all (if the whole data fits into GPU), or have to use mini-batch sampling via our neighbor loaders.

[pstrybol]

It works as expected, thanks! Maybe a little off-topic but because we use 2 different embeddings layers does is still make sense to compare distances between 2 different node types? For example, I want to see which user nodes are closest to a certain movie node in the embedding space? Can I assume that the space in which both node types are embedded are the same?

[rusty1s]

Yes, if you train the embeddings via inner product, this is not a problem.

[songsong0425]

Hi, sorry for updating the completed question. Does the numbers in num_neighbors=[n_1, ..., n_k] mean that sample the k-hop neighbors from seed nodes? Until now, I thought that n_k meant the number of neighbors per source and target node. Then, if I run the 3-layered GCN, should I set the num_neighbors=[n_1, n_2, n_3]?

[rusty1s]

num_neighbors=[n_1, n_2, n_3] means that you sample n_1 neighbors for every seed node in the first hop, you sample n_2 neighbors in the second hop for each neighbor sampled in the first hop, and you sample n_3 neighbors in the third hop for each neighbor sampled in the second hop.

[songsong0425]

I understand. Thank you for your kind reply!