0

I have read plenty of papers on link prediction on dynamic networks, but I am getting confusion on how to do it. I know link prediction on dynamic networks is different from static network.In dynamic network your given a series of snapshots of the graphs(G=(G1, G2....Gn)) and trying to predict link in time T+1

For example, I have a file with a bunch of edge lists (link http://networkrepository.com/ia-enron-email-dynamic.php) each edge list is associated with the time in which the edge was formed, here is the example on how I read the datasets

G=nx.read_edgelist('enron.txt',create_using=nx.Graph(), nodetype=int) but the final graph when you draw it is just one graph, how can I have those series of graphs? for example if I want to find the adjacency matrix for each snapshot, how to do it? Do i need to load the whole file as i did above to do predictions?

Any help including blogs with practical examples will be appreciated

TinaTz
  • 311
  • 3
  • 16
  • 1
    Hi Christina, I believe that each edge in the data (representing an email) also has a Unix time stamp. For example: 1004411463 is 3:11am on 10/30/2001 (see https://www.unixtimestamp.com/index.php). Python can make this conversion (see: https://stackoverflow.com/questions/3682748/converting-unix-timestamp-string-to-readable-date). You could then filter the data and make graphs for each month or some other time frame, and then proceed with link prediction. – Johannes Wachs Apr 15 '20 at 10:22
  • @JohannesWachs thank a lot, at least i got the way to go with your comment, can you tel me something, so for example if i calculate common neighbor index, should i do it for each year?how about for example if i am finding AUC should i do it for every year? this is where am stacking – TinaTz Apr 15 '20 at 12:22
  • 1
    Let's say we are talking about the year 2000. The common neighbors described how many shared email partners two people, A and B have. If they have many neighbors, we'd like to predict that in the future, they are more likely to exchange emails than other pairs with fewer common neighbors. You can test this by predicting who will email with who in 2001 with some classifier. Then you can check how well your predictions do in 2001 using AUC. Makes sense? – Johannes Wachs Apr 15 '20 at 13:28
  • 1
    @JohannesWachs, thank you very much for your useful answer, its real helpful, let me work on it – TinaTz Apr 15 '20 at 13:56

0 Answers0