Fig. 1: A diagrammatic explanation of the sequential stacking approach for link prediction in temporal networks.

We use q consecutive temporal layers in a stacked feature vector to predict links in the target layer; we call q the “flow variable” (blue: q = 2 in the diagram; q = 3 throughout all of our experiments here). We train the prediction (see Methods) using u layers before the target (u > q); we call u the “search variable” (green: u = 4 in the diagram; u = 6 throughout all of our experiments here). Features are generated for sampled dyads (node pairs) in each layer and stacked across q consecutive layers, with edge presence/absence labels from the following layer (green for training and red for testing). We then use standard supervised learning algorithms to train and generate link predictions in the target layer. As diagrammed here, no network information is used from the target layer, only the edge presence/absence labels (the “completely-unobserved setting''). When we consider the “partially-observed setting'', the sequentially stacked features include static topological features as calculated from the partially-observed target layer (throughout our partially-observed experiments here we 5-fold cross-validate the target layer, uniformly sampling 80% of node pairs to predict links on the remaining 20%).