Creating Custom Datasets

[mprithiv]

Hello,

I am working on creating a graph embedding using a datasets as buyers.csv, sellers.csv and bought.csv files. The buyers will have a uniqueBuyerID and 4 other features. The sellers data will have sellerUniqueID and 3 other features. The bought.csv will have uniqueBuyerID, uniqueSellerID, amount, date. I am trying to create a heterogeneous graph and train an unsupervised embedding with RGCN. I am using one of the columns in buyer.csv as a label. I tried using the movie rating example and create a graph. What would be the numpy array for both the nodes would look like? How would I include the node and edge features?

Nodes: buyers, sellers
Edges: amount, date

While performing RandomNodeSplit and RandomLinkSplit, how to make sure the split contains information across all the nodes and edges? What are the results of applying a simple GCN over this graph rather than using RGCN (do not see any implementation of that in PyG) provided I have a heterogeneous bipartite graph.

[rusty1s]

That looks correct to me. Buyers and sellers would be a node type, in which each node/row has 3 or 4 features, i.e., buyer.x has shape [num_buyers, 4] and seller.x has shape [num_sellers, 3]. Since you want to predict a label of a buyer, you can use RandomNodeSplit to further split the data.

Since the feature dimensionality of buyers and sellers is different, you will have to use some heterogeneous GNN to process the data. As an alternative, you can try to map the features of buyers and sellers first into a unified embedding space before using a homogeneous GNN for further processing. It's hard to tell which approach works better in practice in your case.

[mprithiv]

Trying out right away.

[mprithiv]

I tried on both Heterogeneous and Homogenous data.

Before NodeSplit on Homogenous data

Data(node_type=[8697], edge_index=[2, 671918], edge_type=[671918])

After NodeSplit on Homogenous data

Data(node_type=[8697], edge_index=[2, 671918], edge_type=[671918], train_mask=[8697], val_mask=[8697], test_mask=[8697])

On Heterogeneous

Original Data

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_index=[2, 335959] } )

Transformed Data

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

Does this supposed to split 7813/884 (or 8697 if homogenous) into 60-20-20? Or is it because of key=None, utilizes all the data for train, test and validation?

[rusty1s]

Can you clarify what you mean? You can specify num_val and num_test to denote how many nodes you want to use for validation and testing, respectively. The remaining nodes are then used for training.

[mprithiv]

I have specified num_val and num_test as 0.2. With that, I see all the train_mask, test_mask, val_mask as same [7813] which is the num nodes for buyer node type.

[rusty1s]

Yes, this is a mask: a boolean vector with a shape equal to the number of nodes, in which only a certain fraction of values are set to True.