deep-polyu / aane_python Goto Github PK

I try to run the Runme.py, in which the BlogCatalog set is trained. But when I used the embedding for node classification, the performance was terrible. The micro f1 was around 0.2, Why?

Hi,
Thanks for the code.
I tried your code on standard cora dataset. It gives me very poor results.
I could not understand the possible reasons. Could you provide some input.

Thanks a lot for sharing the code. I have been trying to implement your paper from scratch to understand ADMM. Your code is very helpful in understanding the paper and getting insights from your solution.

To understand your solution better, I try to evaluate the objective function as follows:

1. Minimize the objective function using scipy minimize.
2. Compare the objective function from H for your implementation for each iteration.

### A: attribute matrix
### W: weight matrix
S = A.transpose() * sparse.diags(np.ravel(np.power(A.power(2).sum(1), -0.5)))
S = S.toarray()
def objective(H, S):
    n = S.shape[0]
    len_H = H.shape[0]
    wid_H = H.shape[1]
    
    term1 = 0
    for i in range(n):
        for j in range(n):
            tmp = (S[i][j] - np.matmul(H[i], H[j].T))
            term1 += tmp * tmp
    
    term2 = 0
    for i in range(len_H):
        for j in range(wid_H):
            term2 += np.linalg.norm(H[i] - H[j]) * W[i,j]
    
    term2 = term2 * lambd
    
    sol = term1 + term2
    return sol

blog_catalog = sio.loadmat('BlogCatalog.mat', struct_as_record=True)
### For fast iteration, only consider 20x20 size of W and A
W = blog_catalog['Network']
W = W[:20,:20]
A = blog_catalog['Attributes']
A = A[:20,:20]
d = 3 #dimension of H (embedding representation)
e = 0.001
lambd = 0.05
rho = 5

max_iter = np.arange(50)
objectives_author = []
for i in max_iter:
    H_author = AANE(W, A, d, lambd, rho, i,'Att').function()
    objectives_author.append(objective(H_author, S))

However, I see that the objective function increases with the number of iterations as shown in the figure below:

Can you please give me some insights on this problem?

can we have yelp data set? thank you for it.

@Tomposon

Given the whole N nodes, you do five-fold cross-validation, i.e., four-fold nodes for training, denoted as the training group T1. We denote the test nodes as T2.

When "Training Percentage # nodes for embedding = 100%", all the four-fold nodes (T1) are used. We build a graph by including nodes in T1 and T2, and the edges between them. We embedding the constructed graph to learn H. And use embedding representations of T1 to train the classifier.

When "Training Percentage # nodes for embedding = 50%", 50% of the nodes in T1 are used to construct the new graph along with T2.

Thanks.

This python implementation is fast on mac os, but slow on Linux. It is because scipy.sparse.linalg.svds is used. Will fix it in the future.

emb_matrix = AANE_fun(data_net_matrix, data_attr_matrix, d = 128 )

when I use AANE to get the data_attr_net_matrix embedding matrix, is wrong.