gemslab / regal Goto Github PK

Representation learning-based graph alignment based on implicit matrix factorization and structural embeddings

Home Page: https://gemslab.github.io/papers/heimann-2018-regal.pdf

License: MIT License

Python 100.00%

representation-learning network-embedding network-alignment matrix-factorization scalability graph-matching node-embedding

regal's People

Contributors

Stargazers

Watchers

Forkers

wsgan001 treenewbee0 catyans codeinging songfgh naihemeng akrstova social-network-alignment cxxszz umiss mixs222 johannesschuele mightycrane xingye1999 kahei-cjp dawn-515

regal's Issues

Get best alignment

I'm trying to add a code to alignments.py, to get the best alignment the alignment matrix doesn't work, but it doesn't work for me, I don't know if it finds the error and could you help me with it?

def get_best_alignment(alignment_matrix):
  n_nodes = alignment_matrix.shape[0]
  result = []
  if not sp.issparse(alignment_matrix):
      sorted_indices = np.argsort(alignment_matrix)
  for node_index in range(n_nodes):
    target_alignment = node_index #default: assume identity mapping, and the node should be aligned to itself      

    if sp.issparse(alignment_matrix):
          row, possible_alignments, possible_values = sp.find(alignment_matrix[node_index])
          node_sorted_indices = possible_alignments[possible_values.argsort()]
    else:
          node_sorted_indices = sorted_indices[node_index]

    result.append([target_alignment, node_sorted_indices[-1:][0]])

  return result

The idea is that of a result type [A -> B...].

Hi Mark, how to construct the joint adj A and the truth alignment adj P?

Hi Mark,

Sorry to bother you again. Now I try to use some datasets(arenas-meta.edges) get online as A1 for your REGAL program. As I learned from your paper, the input for this program is a jointed matrix in the form as A = [A1,0;0,A2]. 
The adjacency matrix A1 is the true adj of the network we focus on, while the other matrix A2 is just generated by A2 = P * A1 *P' (P is randomly generated).  
Now, I have constructed both the jointed adjacency matrix A and the true-alignment matrix P as input in your method. But my output result is really bad... I only got 0.02 for score top10. 
I've spent 2 days on contructing the correct-format input file for the program. The input file I constructed and the corresponding code generating it have been upload as appendix.
I'm afraid it would be a bother for you to read my code or find mistakes in it... So would you like to give some more detail about how to construct the joint adj A and the truth alignment file P which respectively respond to the file named "arenas_combined_edges.txt" and the file named "arenas_edges-mapping-permutation" in your program? It would be better if you already have some program deal with this problem and if you can share it with me...

I have been confused on this problem for a long time... I'm looking forward to your reply.

Best wishes,
Weifan

code.zip

I am confused with the dataset arenas.

Dear Authors,

I found from the anchor pairs file: "arenas_edges-mapping-permutation.txt" , see below records:

anchor = pickle.load(open('data/arenas_edges-mapping-permutation.txt','rb'))
anchor[0]
334
anchor[334]
1120
anchor[1120]
1103

Does it mean the pairs (0,334), (334,1120), (1120,1103)..... are all anchors ?

Also, I found the from the edge list file "arenas_combined_edges.txt", there are 2270 nodes totally (node ids range from 0 to 2269), and there are 1135 pairs of anchors, but all the anchor node ids are range from 0 to 1134, no one is in the range from 1135 to 2269.

Could you please help to make it clear?

Regards,
Bryan Chen

I

Questions on datasets

May I share ppi and arxiv data sets, please.thanks

Some question

get_estimated_alignment_matrix function is missing?

In your code alignments.py, I can't find the declaration of get_estimated_alignment_matrix in row 49, maybe it's missing.

Where can I find the complete datasets of different networks mentioned in the paper?

Specially the the Arenas email network with 1% noise, how can I get the same dataset with other percent of noise like 5%...?

How to imply this into multiple network (the number of layers >= 3)?

Question about fig2

B has a 1-hop neighbor A whose degree is 3, so why d^1_B is not [0,0,1,0], but in paper is [1,0,0,0] ?

Graph alignment problem with different node amounts in two graphs?

Hello, I'm curious about that is REGAL enable to deal with the graph alignment problem where there are different node amounts in two graphs. Thanks!

Question about figure 4 in the paper!

Glad to read your paper,it's a good idea. I tried to realize the regal as your paper said , but I can only reach the accuracy shown in figure 4 when using top-5, I'd like to konw whether you use top-5 accuracy in figure 4 as well. By the way, file, arenas_combined_edges.txt seems not to flip edges.