Question on the feature format for generated samples of ENZYMES dataset

[lizaitang]

Dear Author, you paper really helps a lot, but I have a question that I want to pass the generated graph to some classifiers, but seems the node features generated by the GDSS is different from the original dataset in data scale. How can I solve it? Thanks

[harryjo97]

Hi lizaitang,

In our work, we used the degree of each node as the node features instead of the given node features of the original dataset. In order to use the node features of the original dataset, you can modify the code in

GDSS/utils/data_loader.py

Line 6 in 4d96334

def graphs_to_dataloader(config, graph_list):

and

GDSS/utils/graph_utils.py

Line 43 in 4d96334

def init_features(init, adjs=None, nfeat=10):

for loading the node features of the dataset.
After changing these, you could newly train the score models to generate both node features and the adjacency matrices.

[lizaitang]

Dear Author, thank you very much for your quick reply! How can we modify the code to load node feature of the dataset? Change the init to zeros or ones? Or we dirrectly change x_tensor = init_features(config.data.init, adjs_tensor, config.data.max_feat_num) to the feature of the dataset?
def init_features(init, adjs=None, nfeat=10):

if init=='zeros':
    feature = torch.zeros((adjs.size(0), adjs.size(1), nfeat), dtype=torch.float32, device=adjs.device)
elif init=='ones':
    feature = torch.ones((adjs.size(0), adjs.size(1), nfeat), dtype=torch.float32, device=adjs.device)
elif init=='deg':
    feature = adjs.sum(dim=-1).to(torch.long)
    num_classes = nfeat
    try:
        feature = F.one_hot(feature, num_classes=num_classes).to(torch.float32)
    except:
        print(feature.max().item())
        raise NotImplementedError(f'max_feat_num mismatch')
else:
    raise NotImplementedError(f'{init} not implemented')

flags = node_flags(adjs)

return mask_x(feature, flags)

[harryjo97]

You can change init_features in

GDSS/utils/graph_utils.py

Line 43 in 4d96334

def init_features(init, adjs=None, nfeat=10):

to take in graph_list as input and return the node features.
To be specific, each graph in the graph_list is a networkx Graph with node features.

Or you could directly modify x_tensor = init_features(config.data.init, adjs_tensor, config.data.max_feat_num) in

GDSS/utils/data_loader.py

Line 6 in 4d96334

def graphs_to_dataloader(config, graph_list):

to obtain the original node features from the networkx Graph.

Please refer to the networkx documentation for more details.

FYI, the attributed graphs of the ENZYMES dataset are loaded by this function:
https://github.com/harryjo97/GDSS/blob/4d96334fd0d07577f9891e9d5e81dae4d64a92fd/data/data_generators.py#LL131C13-L131C13

[lizaitang]

Dear Author,
Thank you so much for your quick reply, I have a minor question that I follow the format of graph_to_tensor to load the original node features, but for v, feature in g.nodes.data('feature') gives feature as none, could you please help to fix on it?

def feat_to_tensor(graph_list, max_node_num,max_feat_num):
    feat_list = []
    max_node_num = max_node_num

    for g in graph_list:
        assert isinstance(g, nx.Graph)

        node_feat_list = np.zeros([max_node_num,max_feat_num], dtype = float)
        i=0 
        for v, feature in g.nodes.data('feature'):
            
            node_feat_list[i]=feature
            
            i=i+1
        #print(node_feat_list)
       
        feat_list.append(node_feat_list)

    del graph_list

    feat_np = np.asarray(feat_list)
    del feat_list

    adjs_tensor = torch.tensor(feat_np, dtype=torch.float32)
    del feat_np

    return adjs_tensor 

[harryjo97]

In the graph loader code:

GDSS/data/data_generators.py

Line 131 in 4d96334

def graph_load_batch(min_num_nodes=20, max_num_nodes=1000, name='ENZYMES', node_attributes=True, graph_labels=True):

The node labels you are looking for are saved in g.nodes.data('label') (saved by Line 158 G.add_node(i + 1, label=data_node_label[i]))

You may want to try g.nodes.data('label') instead of g.nodes.data('feature').

[lizaitang]

Thanks for your reply, but if I want to generate graph with same format node features, shouldn't we use the feature instead of node labels?

[harryjo97]

I think the node features you want to use for the classifier are contained in the label.

[lizaitang]

Sorry to bother, but I try label, ```
[[2. 2. 2. ... 2. 2. 2.]
[2. 2. 2. ... 2. 2. 2.]
[2. 2. 2. ... 2. 2. 2.]

[harryjo97]

First of all, the label contains other values other than 2 (please see https://github.com/harryjo97/GDSS/blob/master/dataset/ENZYMES/ENZYMES_node_labels.txt)

Furthermore, if you want to use the node attributes in https://github.com/harryjo97/GDSS/blob/master/dataset/ENZYMES/ENZYMES_node_attributes.txt,
you may change the code in:

GDSS/data/data_generators.py

Line 266 in 4d96334

graphs = graph_load_batch(min_num_nodes=10, max_num_nodes=1000, name=dataset,

by setting the node_attributes=True which will load the node attributes file by
data_node_att = np.loadtxt(path + name + '_node_attributes.txt', delimiter=',')
in

GDSS/data/data_generators.py

Line 131 in 4d96334

def graph_load_batch(min_num_nodes=20, max_num_nodes=1000, name='ENZYMES', node_attributes=True, graph_labels=True):