Cyrus9721
commented
5 months ago Hi Jiale,
Thank you so much for your interest in this work. Here is the explanation:
In the original SGC paper (https://arxiv.org/abs/1902.07153), section 2.2, equation (8), the SGC can be reformulated as Y_{SGC} = softmax S^{K}XΘ. In this case, S^{K}X can be precomputed. Also, you can take a look at DGL's documentation at ( https://docs.dgl.ai/en/0.8.x/_modules/dgl/nn/pytorch/conv/sgconv.html#SGConv), the code chunk for '# compute (D^-1 A^k D)^k X'.
Best, Zizhang
On Thu, May 9, 2024 at 11:34 PM JialeDeng @.***> wrote:
Hello! The paper was quite well-written and this work inspired me a lot, but I have some concerns about the code. In particular, when computing the node influence, a SimplifiedGraphNeuralNetwork (which is on earth a LogisticRegression model) is trained instead of SGC model mentioned in the paper. The code is in https://github.com/Cyrus9721/Characterizing_graph_influence/blob/c0c87b6570b83f2d8a5ab4ed548d5cf9184cf9a1/calculate_node_influence.py#L108 [github.com] https://urldefense.com/v3/__https://github.com/Cyrus9721/Characterizing_graph_influence/blob/c0c87b6570b83f2d8a5ab4ed548d5cf9184cf9a1/calculate_node_influence.py*L108__;Iw!!DaRZpAeNFA!bl-fUti_j48y0d2Xiqrxddw9K15BCSulOmMXP4jN1DBzYZS-rNpLKCGxLvrC7SMQu4-CzvbEnTzFNeKbf9UH68XetoLoYQ$ It is quite confusing since the LogisticRegression model leverages no graph structure and computing influence score on them seems not an approximation to the SGC model. I would be very grateful if you could answer the questions. Thanks again for providing such inspiring work, and looking forward to your reply!
— Reply to this email directly, view it on GitHub [github.com] https://urldefense.com/v3/__https://github.com/Cyrus9721/Characterizing_graph_influence/issues/1__;!!DaRZpAeNFA!bl-fUti_j48y0d2Xiqrxddw9K15BCSulOmMXP4jN1DBzYZS-rNpLKCGxLvrC7SMQu4-CzvbEnTzFNeKbf9UH68Vit-1WCw$, or unsubscribe [github.com] https://urldefense.com/v3/__https://github.com/notifications/unsubscribe-auth/AVSZDPCCNUUDIL2CZFRC4STZBQ55ZAVCNFSM6AAAAABHP4LU42VHI2DSMVQWIX3LMV43ASLTON2WKOZSGI4DQOBUHEYDEOI__;!!DaRZpAeNFA!bl-fUti_j48y0d2Xiqrxddw9K15BCSulOmMXP4jN1DBzYZS-rNpLKCGxLvrC7SMQu4-CzvbEnTzFNeKbf9UH68XsTg0MRg$ . You are receiving this because you are subscribed to this thread.Message ID: @.***>
JialeDeng
Hello! The paper was quite well-written and this work inspired me a lot, but I have some concerns about the code. In particular, when computing the node influence, a
SimplifiedGraphNeuralNetwork(which is on earth a LogisticRegression model) is trained instead of SGC model mentioned in the paper. The code is in https://github.com/Cyrus9721/Characterizing_graph_influence/blob/c0c87b6570b83f2d8a5ab4ed548d5cf9184cf9a1/calculate_node_influence.py#L108 It is quite confusing since the LogisticRegression model leverages no graph structure and computing influence score on them seems not an approximation to the SGC model. I would be very grateful if you could answer the questions. Thanks again for providing such inspiring work, and looking forward to your reply!