BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Learning evolving node and community representations on dynamic gr
 aphs - Simeon Spasov
DTSTART:20200310T131500Z
DTEND:20200310T141500Z
UID:TALK140476@talks.cam.ac.uk
CONTACT:Mateja Jamnik
DESCRIPTION:Representation learning of static and more recently dynamicall
 y changing graph-structured\ndata has gained noticeable attention\, especi
 ally in the unsupervised case. The majority\nof prior research work has fo
 cused extensively on modelling the evolutionary dynamics of\nindividual no
 des with applications\, such as link prediction and node classification. H
 owever\,\nthere is a lack of methods for unveiling more global patterns of
  dynamic graph evolution.\nThus\, in this work\, we present a generative m
 odel which is able to learn evolving node\nand community representations. 
 We can use these dynamic representations to investigate\nthe community tra
 nsition patterns for individual nodes\, or to study macro-trends\, such\na
 s the evolution of graph communities over time. The proposed co-evolutiona
 ry model\nis trained via variational inference by optimizing the ELBO of t
 he observed edges in the\ndynamic graph. We demonstrate competitive or sup
 erior performance of our method against\nstate-of-the-art baselines on dyn
 amic link prediction. We then analyze the evolution of\ninterpretable comm
 unities in the DBLP bibliography\, and also the community transition\nprob
 abilities of individual authors to showcase the capabilities of our co-evo
 lutionary model\nas a tool to extract dynamic graph insights.\n
LOCATION:SS03\, Computer Laboratory\, William Gates Building
END:VEVENT
END:VCALENDAR