codes problem about graphsage HOT 2 CLOSED

Hi yes, this part of the code is extremely hard to parse, though it is doing what it needs to.. The main complicating thing (which made directly implementing the psuedo-code exactly as written impossible in Tensorflow) was TensorFlow's (at the time) lack of support for dynamic compute graphs. The concatenation in line 5 of algorithm 1 also really complicates implementation in TensorFlow.

For the sake of my sanity and others' understanding, I made an easier to understand reference implementation in PyTorch if you're interested: https://github.com/williamleif/graphsage-simple

I did not actually write the aggregate(...) function in this TensorFlow version, but I'll try to explain a bit, though it is really confusing (thus why I made the easier-to-read pytorch version):

The reason for the two loops is that to get the final "output" representation of a node, we need representations of that node at each of the preceding layers (in the notation of Algorithm 1, to get the iteration k=K representation of a node, we need its representations for iterations 1...K_1). So basically the first iteration of the inner-loop (when layer=0 and hop=0) corresponds to getting the iteration k=1 representations for the set of nodes that we want to have embeddings for at the final "layer"/iteration k=K. And the inner-loop when layer=0, hop=1 computers the iteration k=1 embeddings for nodes that are needed at iteration k=2, and the layer=0, hop=2 inner-loop computes the embeddings for the nodes that are only used at iteration k=1. This is horrendously confusing imo haha T_T, but it works and avoids unnecessary computations compared to more naive TensorFlow implementations.

from graphsage.

Yes that complication is for efficiency reasons...

from graphsage.