Bug of Normalisation

[Linnore]

Multi-GNN/data_loading.py

Lines 108 to 127 in 252b025

	#Adding ports and time-deltas if applicable
	if args.ports:
	logging.info(f"Start: adding ports")
	tr_data.add_ports()
	val_data.add_ports()
	te_data.add_ports()
	logging.info(f"Done: adding ports")
	if args.tds:
	logging.info(f"Start: adding time-deltas")
	tr_data.add_time_deltas()
	val_data.add_time_deltas()
	te_data.add_time_deltas()
	logging.info(f"Done: adding time-deltas")
	#Normalize data
	tr_data.x = val_data.x = te_data.x = z_norm(tr_data.x)
	if not args.model == 'rgcn':
	tr_data.edge_attr, val_data.edge_attr, te_data.edge_attr = z_norm(tr_data.edge_attr), z_norm(val_data.edge_attr), z_norm(te_data.edge_attr)
	else:
	tr_data.edge_attr[:, :-1], val_data.edge_attr[:, :-1], te_data.edge_attr[:, :-1] = z_norm(tr_data.edge_attr[:, :-1]), z_norm(val_data.edge_attr[:, :-1]), z_norm(te_data.edge_attr[:, :-1])

The data normalisation in line 127 does not normalise the last edge attribute, which corresponds to "payment format" . This is a desired behaviour if no ports nor time deltas are added. However, when ports or time deltas are added, the last edge attribute would not be "payment format" anymore and required normalisation.

[AlessioBoetti]

Agree. I'd like to add to this that when using Reverse MP, inside the create_hetero_obj function the edge_attr relative to the ports are swapped for the reverse edge, but this doesn't take into account the use of timedeltas, which if used together with the ports add other two columns, so the latest two columns are no more relative to the ports but relative to the timedeltas.