pcy1302 / dmgi Goto Github PK

I downloaded the DBLP-V8 from Aminer website. It doesn't have paper abstracts field. Can you share your raw datasets that you have used?

Hi, I am excited to see this work which implemented Deep Graph Infomax on Multiplex Network!
And I want to test ACM, DBLP, Amazon Dataset based on your code, can you provide these data sets and code about preprocess about them?
I will be appreciated for your reply as soon as possible, thank you very much!

hi,sorry to interrupt you again~
we can see you get a so big improvement on amazon dataset !
i run your code on amazon dataset , and search parameters as your paper said "α; β; from 0:0001; 0:001; 0:01; 0:1".
but i am sad i can not get a comparable result as your paper. some details is below:
nmi: my best result is 0.2858 | 0.2954 for DMGI-att and DMGI. your paper is 0.412 and 0.425
macro | micro : my best result is 0.7426 | 0.7462| 0.7380| 0.7414 for DMGI-att and DMGI. your paper is
0.758 | 0.758 | 0.746 | 0.748. We can see my result is very close to you .
sim: my best result is 0.820 and 0.809 for DMGI-att and DMGI. , your paper is 0.825 and 0.816 . we can see my result is also very close to you .

so i am very confused about the performance about the metric nmi. do you meet the same problem or can you give me some suggestions?

If you have the raw datasets of ACM and IMDB, can please share it along with its preprocessing scripts?

HI, the web page of ACM & DBLP cannot be opened.Can you update it? Thank u

Sorry to interrupy you again, your link is error,Can you share DBLP and amazon using google drive again?

Hello, your work is impressive and has benefited me greatly. I'm a novice in deep learning, and I have some questions about your code.
I noticed that you've set a variable called "patience" in your code, and training stops when the loss increases for 20 consecutive times(parser.add_argument('--patience', type=int, default=20)).

if loss < best:
    best = loss
    cnt_wait = 0
    torch.save(model.state_dict(), 'saved_model/best_{}_{}_{}.pkl'.format(self.args.dataset, self.args.embedder, self.args.metapaths))
else:
    cnt_wait += 1

if cnt_wait == self.args.patience:
    break

Is this strategy reasonable? I haven't come across this in your papers.

I have an error when running the example. If I don't have the saved model, how could I make the program run?

Hello
I am new to Python. According to my project, I invoked ready-made functions and I want to use them. The nature of functions is about finding intra-layer and inter-layer edges in a multiplex graph. But I am not able to call and draw conclusions from these functions. If possible, please guide me. Thanks

The desired code is shared.

class LayerNetwork():

def __init__(self, graphs):

    self.graphs = graphs
    self.total_layers = len(graphs)

    self.get_nodes()
    self.get_edges_within_layers()
    self.get_edges_between_layers()

def get_nodes(self):
    self.nodes = []
    for z, g in enumerate(self.graphs):
        self.nodes.extend([(node, z) for node in g.nodes()])

def get_edges_within_layers(self):
    self.edges_within_layers = []
    for z, g in enumerate(self.graphs):
        self.edges_within_layers.extend([((source, z), (target, z)) for source, target in g.edges()])

def get_edges_between_layers(self):
    self.edges_between_layers = []
    for z1, g in enumerate(self.graphs[:-1]):
        z2 = z1 + 1
        h = self.graphs[z2]
        shared_nodes = set(g.nodes()) & set(h.nodes())
        self.edges_between_layers.extend([((node, z1), (node, z2)) for node in shared_nodes])

Hi, thanks for you great work here!
I read the paper Deep Graph Infomax, DGI introduce a discriminator $D$ to maximize the mutual information. (Eq(1) in the paper). And it say this term should be maximized.
But in your final loss function, there is no minus sign in front of it.

I haven't understood DGI much, so I wonder if I got it wrong or there is a written mistake?

Regards,

Dear authors :
I use the "IMDB" dataset and run the program，I didn't change the parameters, but I got the different results from your paper. So，could you help me to explain this ? The following is the result.

DMGI:
[Classification] Macro-F1: 0.6509 (0.0101) | Micro-F1: 0.6519 (0.0076)
[Clustering] NMI: 0.2015
[Similarity] [5,10,20,50,100] : [0.6115,0.6022,0.5944,0.581,0.5674]

DMGI(attention):
[Classification] Macro-F1: 0.6163 (0.0114) | Micro-F1: 0.6184 (0.0083)
[Clustering] NMI: 0.1992
[Similarity] [5,10,20,50,100] : [0.5868,0.5754,0.5646,0.555,0.5449]