We release a benchmark dataset for the evaluation and development of temporal network embedding methods that leverage such fundamental graph time-series representations. The benchmark consists of 26 temporal networks from a variety of different application domains. We also release the graphs in each of the different time-series representations as separate files that can be used for benchmarking purposes. This includes the $\tau$-graph time-series that decomposes the edge stream into a time-series of graphs based on an application time-scale such as 1 day, as well as the $\epsilon$-graph time-series that decomposes the edge stream into a time-series of graphs such that each graph has a constant number of edges. For a direct comparison between the $\tau$-graph and $\epsilon$-graph time series approaches, we require an equal number of graphs for each, that is, $|\mathcal{G}_{\tau}| = |\mathcal{G}_{\epsilon}|$. For this, we first fix $\tau$ to be an application time-scale (such as 1 hour, 1 day, and so on) and use this $\tau$ to derive a $\tau$-graph time-series $\mathcal{G}_{\tau}$. We then set the number of edges per snapshot in the $\epsilon$-graph time-series as $\epsilon = \ceil{\tfrac{|E|}{|\mathcal{G}_{\tau}|}}$, which results in an equal number of snapshots as desired. For each of the 26 temporal networks, we release a variety of temporal granulaties. For instance, for email-EU-core network, we have a directory called email-EU-core, inside that directory consists of the raw temporal network (edge stream) called email-EU-core.edges, which is a simple comma-delimited edge list where the first and second column are node ids, whereas the third is the timestamp of the edge. Furthermore, there are two other directories inside email-EU-core, namely, 1day and 1week, which indicates the temporal granulatiy, and inside either of those directories are two more directories called tau and epsilon, which contain the edge list files for each of the different graph time-series representations.

{"references": ["Ryan Rossi, Nesreen Ahmed, \"On Graph Time-Series Representations for Temporal Networks\""]}