GraphMixer
GraphMixer copied to clipboard
CongWeilin
``` Epoch 1 train mode with negative samples 1 ...: 1%|▏ | 1/184 [00:00
Hello, Is it possible to make your code compatible with datasets as are used in https://github.com/fpour/DGB/tree/main It will make benchmarking easy for further investigations.
Hello, Is it possible to create a Python-only version of GraphMixer by doing away with the C++ sampler? I want to run this code in the cluster, and I'm encountering...
作者您好！感谢您的出色工作~ 我发现在为root_nodes进行邻居节点采样时，dst节点的采样可能存在问题，以下是问题的概述与修复方式： 在每次采样时，root_nodes会以batch_size * (2+sample_num)的形式送入采样器，后者会对每个root_node返回其历史邻居。但是，看上去，root_nodes的前batch_size个节点一定是该batch中的src，第batch_size到2*batch_size一定是该batch中的dst。 因此，当采样器遍历前batch_size个root_node（src）时，采样器中的指针会根据时间变化，这导致在采样dst节点时，倒数neighbor个历史邻居很可能不满足采样要求，导致无法采集到任何邻居。 我想询问是否应该在dst采样前执行一次`sample.reset()`？ 示例代码如下，位于construct_graph.py中： ``` def get_mini_batch(sampler, root_nodes, ts, num_hops, extra_neg_samples): # neg_samples is not used """ Call function fetch_subgraph() Return: Subgraph of each node. """...
Hi authors, Thanks for your excellent works. However i have met some troubles in reproducing the results reported in paper. I found that there are two points may cause data...
Hi, I tried to compile the package according to the README on windows system, with python verison 3.11 or 3.9 Now there are many errors:  The main errors are:...
Hi, I’m doing experiments on financial data using GraphMixer and other temporal graph networks, and I’ve found that GraphMixer consistently outperforms the others by a significant margin. That said, I...
