Fast PyG Data Object for Dynamic Node Label / Features

[shenyangHuang]

implement fast and efficient data object in PyG for dynamic node label / features, a few key characteristics as below. Node label and node feature can be treated interchangeably here as they are both vectors.

• should support tuple like representation (time, node id, node label vector).
• [!] not all nodes have labels for all timestamps
• [!] node label timestamps are often more coarse grain than edge timestamps and there might not be edges that share timestamp of the node labels
• should update or remove find_next_labels_batch function in dataset.py
• node labels should be broken down into days (or a fixed time intervals depending on the application)
• if no node label for a given day, return None when querying the dataset object

[shenyangHuang]

As an improved approach, try the following:

• construct a dict of node labels where key is the unix_timestamp of the day, item = (node_id, node_label_vec)
• ensure the timestamps of the node labels are assigned correctly, the day labels are shift by 1 as the edges from the day are used to predict the future (double check)
• find_next_labels_batch function in dataset.py should now be improved to query the nearest timestamp in the batch or signaled by the edgelist