yule-buaa/dyglib

A Library for Dynamic Graph Learning (NeurIPS 2023)

dyglib's Issues

Questions about the application of the so-called Patching Technique

First of all, thank you for your outstanding work.

My question arose when I read this description in your paper

Instead of focusing on the interaction level, we divide the encoding sequence into multiple non-overlapping patches to break through the bottleneck of existing methods in capturing long-term temporal dependencies.

Why should we divide the sequence into num_patches * patch_size's? It doesn't seem to affect the long-term temporal dependencies that exist over an extended period of time within a sequence without doing this step?
Or
Have you experimentally proved that the results do improve with the help of this 'Patching Technique', i.e., that the staleness problem does occur during the training? I mean, Self-Attention Mechanism already allows processing whole sequence in parallel theoretically.

[Help] Why Node ID starts from 1 instead of 0?

Thanks for your efforts!
I am confused by the reindex operation in preprocess_data.py since the JOIDE datasets implemented in PyG start node index from 0 instead of 1 here.

Could you tell me why DyGLib (and DGB) take such reindex operations?

A question about dynamic node features

Thank you for the wonderful work!

I am trying to use your temporal GNN models for my research project. But the problem is that the node features are dynamic rather than static in my case -- the feature vector of each node may change over time. I have checked your code but it seems to me that DyGLib does not support the dynamic node feature in terms of data structure and model updating (Please correct me if I am wrong). Do you have any suggestions on how to adpat your code to support the dynamic node features? Is it possible to support the dynamic_graph_temporal_signal data structure in pytorch_geometric_temporal? Thanks!

Question of dynamic node classification & edge classification?

Hii Yu, thanks a lot for your nice work and codes sharing!! I have a question about the dynamic node classification, after reading the codes, I found in the DG_data/DATASETS_README.md that it says the title of label column is the edge_label , but indeed it's the dynamic node classification task of src_node (shown in train_node_classification.py file). So there's a minor misleading issue here. Moreover, do you have plan to support edge classification in DyGLib? Thanks in advance! ^0^

Inference code

How do I infer the pretrained model? Thank you for reading and supporting me.

TypeError:'float!object cannot be interpreted as an integer

If you meet the problem: TypeError:'float!object cannot be interpreted as an integer, you could try to change "0.0" to "0" in DyGFormer.py (Line 372 and 378)

time_interval_aware策略下计算采样概率为什么用的np.cumsum

您好！请问一下 此处 代码中在time_interval_aware策略下计算采样的概率，为什么是除以 np.cumsum(exp_node_neighbor_times)，这样算出来的概率分布有什么特别的嘛？我试着输出了一下 sampled_probabilities，发现都是第一个是1，后续逐渐减小，似乎并没有特别的规律。
谢谢！

Creating custom dynamic and temporal dataset for link prediction

I have dynamic (Nodes are created and deleted over a period of time) and temporal (node features are changing over a time) data. Then, how to create a dataset which are suitable for GNN ? Also, node features are numeric values. So, how to represent it in a .npy format ? Thanks.