Request for Baseline Code or Data Processing Method Due to Divergent Training Data Format

[liane886]

Dear Feng,

I've noticed your repository's training data format uniquely diverges from the common ['u', 'i', 'ts', 'label', 'idx'] structure used in temporal graph benchmarks, particularly with distinct labels for src and dst nodes. Could you share insights or provide baseline code on adapting this format for standard methods? Your guidance would greatly enhance our ability to align our research with your innovative framework.

Thank you for considering my request. I look forward to your response.

Best regards,

Liane

[tulerfeng]

Thank you for your interest.

Our training data format includes the following attributes: ['u', 'i', 'label_u', 'label_name_u', 'label_i', 'label_name_i', 'ts', 'idx'].

To align our training data format with the standard structure ['u', 'i', 'ts', 'label', 'idx'], you simply need to remove 'label_name_u' and 'label_name_i'. Subsequently, you can use either 'label_u' or 'label_i' as the 'label', depending on your classification needs. 'Label_u' corresponds to the label for the source node, while 'label_i' corresponds to the label for the target node in a directed edge. Alternatively, if your aim is to conduct an edge classification task rather than node classification, the attribute 'label' becomes redundant.