WalkingTime: Dynamic Graph Embedding Using Temporal-Topological Flows David Bayani dcbayani@alumni.cmu.edu 0 1 Galway, Ireland. International Conference on Knowledge School of Computer Science, Carnegie Mellon University , 5000 Forbes Ave. , Pittsburgh, PA , USA 2020 19 20

Node embedding - the process of generating final embeddings while simultaneously sparyears, with many contemporary developments cuss its on-going evaluation across a series

 -

walk framework [2], allowing us to leverage a Skip-gram inspired back end to produce the ing the need to discretize or align time-related attributes in the network. In our lightning talk, we plan to overview WalkingTime, disof tasks, and detail the potential value of both our method and the novel perspective underlying it.

Acknowledgments We would like to thank Reihaneh Rabbany for her thoughts, guidance, and support while

 Eficient estimation of word representations in vector space, arXiv:1301.3781 (2013). arXiv preprint

able feature learning for networks, in:

 Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016.

a relatively low dimensional vector to summarize a vertice’s many roles in a network - has received increased attention in recent building of of random-walks and NLP-inspired embedding methods, specifically Skip-grams [1]. Of particular focus within the last five years has been the development of techniques more suitable for dynamic networks, aiming to utilize the rich temporal structure present to better inform the embeddings produced.

Existing dynamic node embeddings, however, consider the problem as limited to the evoludiscrete states. Based on a fundamentally different handling of time, we propose a novel embedding algorithm, WalkingTime. While prior works considered time as a ordered collection of separate networks , WalkingTime allows for the local consideration of continuously occurring phenomena; while others consider global graph snap-shots to be firstorder citizens , we hold flows comprised of temporally and topologically local interactions as our primitives. Our temporal-topological lfows eloquently extend node2vec’s randomtion of a topology over a sequence of global, conducting this work.