Multi-hop sample

[W-rudder]

Hi! When I checked the sampling results of the codes, I found some problems about it. The details are as follows.

1. The first interaction of any couple of nodes will never be sampled. For example, I use wiki for testing. There exists two interactions between node 2 and 8229 at time 36 and 77, but the sample results doesn't include the first one. Here I use batch size 5 for convenient testing. And TGAT config.
   
   
   

2. Is there an error in the sampling logic for multi-hop sampling? For example, when taking a 2-hop subgraph, the new nodes obtained in the first sampling should be re-sampled instead of sampling all nodes. But the logic of the code is to use all nodes, because the root node must be included in the result of the first sampling. Otherwise, repeated one-hop sampling results will be obtained. Same case as above. Got duplicate node 8229 at 77 in second hop sampling
   

• The version of my code is up to date.
• I would appreciate it if you could help me with my question. Thanks a lot!

[tedzhouhk]

Hi, thanks for your interest in our work!

1. Thanks for pointing out this bug! It is due to a boundary issue in the sampler. I will submit a PR, but it might take some time to be merged. For a quick fix, change sampler_core.cpp:298 to for (EdgeIDType k = e_search - 1; k >= std::max(s_search, e_search - neighs); k--).
2. Could you please explain more on this? I follow the GraphSAGE type of sampling in this work. For example, for target node 1, we sample neighbors 2 and 3 in the first layer. Then in the second layer, we sample from nodes 1, 2, and 3, as their hidden features are required in the first layer. In the second layer, nodes 1 can have different sampled neighbors as in the first layer.

[W-rudder]

Hi, thanks for your interest in our work!

1. Thanks for pointing out this bug! It is due to a boundary issue in the sampler. I will submit a PR, but it might take some time to be merged. For a quick fix, change sampler_core.cpp:298 to for (EdgeIDType k = e_search - 1; k >= std::max(s_search, e_search - neighs); k--).
2. Could you please explain more on this? I follow the GraphSAGE type of sampling in this work. For example, for target node 1, we sample neighbors 2 and 3 in the first layer. Then in the second layer, we sample from nodes 1, 2, and 3, as their hidden features are required in the first layer. In the second layer, nodes 1 can have different sampled neighbors as in the first layer.

Thanks for your reply!
About question 2:
I think you are right if it's in GraphSAGE. But in TGAT, it chooses recent neighbors, so in the second layer, node 1 can never have different sampled neighbors as in the first layer. Its neighbors should always be the same as in the first layer.

[W-rudder]

Hi, thanks for your interest in our work!

1. Thanks for pointing out this bug! It is due to a boundary issue in the sampler. I will submit a PR, but it might take some time to be merged. For a quick fix, change sampler_core.cpp:298 to for (EdgeIDType k = e_search - 1; k >= std::max(s_search, e_search - neighs); k--).
2. Could you please explain more on this? I follow the GraphSAGE type of sampling in this work. For example, for target node 1, we sample neighbors 2 and 3 in the first layer. Then in the second layer, we sample from nodes 1, 2, and 3, as their hidden features are required in the first layer. In the second layer, nodes 1 can have different sampled neighbors as in the first layer.

Oh, it's my bad. I think you are right. Thanks again for your reply!