Difference between Loading Graph from CSV and Creating Custom Dataset.

[mprithiv]

Hi,

I am trying to create an unsupervised graph embedding using heterogenous graph. I am using my own dataset that I have in a CSV format. I tried to use "Load Graph from CSV" but then when trying to create a GNN model as per the example, I am stuck with not having train_mask. Does this mean, I need to create a dataset first and then load the csv files? What is the difference between loading graph from CSV and Creating a custom dataset?

Thanks

[rusty1s]

If you do not have a dedicated train mask, you can create one after graph creation as well, e.g., via the torch_geometric.transforms.RandomNodeSplit transform.

The torch_geometric.data.Dataset comes in handy when working with multiple graphs, and it provides support for downloading, processing and saving data objects. In your case, there isn't a real difference.

[mprithiv]

Does the RandomNodeSplit work for HeteroData object or train mask has to be created for individual nodes? Is there a way to split on a HeteroData object.

With the data object looking like this,

HeteroData( buyerId={ num_nodes=7813, x=[7813, 1152] }, sellerId={ num_nodes=884, x=[884, 387] }, (buyerId, txn_amount, sellerId)={ edge_index=[2, 335959] }, (buyerId, txn_date, sellerId)={ edge_indes=[2, 335959] } )

performing a random RandomNodeSplit on the whole HeteroData fails. The task is, creating an embedding for the buyer and seller with txn_amount and date as edges, this embedding will be used for many downstream applications. I am intended to try the GAT model under Creating Heterogeneous graphs

[rusty1s]

I see that this might be iriritating. The RandomNodeSplit will only split on node types that contain a certain key (see the key attribute of RandomNodeSplit. You may try to set it to None.

[mprithiv]

I tried setting the key to None. Below represents the data object

HeteroData( buyer={ num_nodes=7813, x=[7813, 1152] }, seller={ num_nodes=884, x=[884, 387] }, (buyer, txn_amount, seller)={ edge_index=[2, 335959] }, (buyer, txn_date, seller)={ edge_indes=[2, 335959] } )

There is no common feature column between buyer and seller. The mapping has buyerId, sellerId, tnx_amount and txn_date. Splitting with key=none

nodesplit = RandomNodeSplit(split='random', key=None) nodesplit(data)

results in
HeteroData( buyer={ num_nodes=7813, x=[7813, 1152] }, seller={ num_nodes=884, x=[884, 387] }, (buyer, txn_amount, seller)={}, (buyer, txn_amount, seller)={ edge_index=[2, 335959] }, (buyer, txn_date, seller)={ edge_indes=[2, 335959] }, (seller, rev_txn_amount, buyer)={ edge_index=[2, 335959] } )

I don't see any split happening. Whereas, for RandomLinkSplit

edge_split = RandomLinkSplit( num_val=0.05, num_test=0.1, neg_sampling_ratio=0.0, edge_types=[('buyer', 'txn_amount', 'seller')], rev_edge_types=[('seller', 'rev_txn_amount', 'buyer')], )

results in
Train Data

HeteroData( user={ num_nodes=7813, x=[7813, 1152] }, seller={ num_nodes=884, x=[884, 387] }, (user, txn_amount, seller)={}, (user, txn_amount, seller)={ edge_index=[2, 285567], edge_label=[285567], edge_label_index=[2, 285567] }, (user, txn_date, seller)={ edge_indes=[2, 335959] }, (seller, rev_txn_amount, user)={ edge_index=[2, 285567] } )

Test Data

HeteroData( buyer={ num_nodes=7813, x=[7813, 1152] }, seller={ num_nodes=884, x=[884, 387] }, (buyer, txn_amount, seller)={}, (buyer, txn_amount, seller)={ edge_index=[2, 302364], edge_label=[33595], edge_label_index=[2, 33595] }, (buyer, txn_date, seller)={ edge_indes=[2, 335959] }, (seller, rev_txn_amount, buyer)={ edge_index=[2, 302364] } )

Validation Data

HeteroData( buyer={ num_nodes=7813, x=[7813, 1152] }, seller={ num_nodes=884, x=[884, 387] }, (buyer, txn_amount, seller)={}, (buyer, txn_amount, seller)={ edge_index=[2, 285567], edge_label=[16797], edge_label_index=[2, 16797] }, (buyer, txn_date, seller)={ edge_indes=[2, 335959] }, (seller, rev_txn_amount, buyer)={ edge_index=[2, 285567] } )

Also,

• I see reverse-edge being created for only one edge types after performing transform = T.ToUndirected()(data_copy). Not sure if this is normal.
• When I do a node split/edge split on a heterogeneous graph with multiple node and edge types like aforesaid, how could I make sure the split happen across all the node/edge types?

[rusty1s]

I think you need to use split="train_rest" in order to allow for splits independent of any ground-truth labels.