Analytical Models for Motifs in Temporal Networks: Discovering Trends and Anomalies
Dynamic evolving networks capture temporal relations in domains such as social networks, communication networks, and financial transaction networks. In such networks, temporal motifs, which are repeated sequences of time-stamped edges/transactions, offer valuable information about the networks' evolution and function. However, currently no analytical models for temporal graphs exist and there are no models that would allow for scalable modeling of temporal motif frequencies over time. Here, we develop the Temporal Activity State Block Model (TASBM), to model temporal motifs in temporal graphs. We develop efficient model fitting methods and derive closed-form expressions for the expected motif frequencies and their variances in a given temporal network, thus enabling the discovery of statistically significant temporal motifs. Our TASMB framework can accurately track the changes in the expected motif frequencies over time, and also scales well to networks with tens of millions of edges/transactions as it does not require time-consuming generation of many random temporal networks and then computing motif counts for each one of them. We show that TASBM is able to model changes in temporal activity over time in a network of financial transactions, a phone call, and an email network. Additionally, we show that deviations from the expected motif counts calculated by our analytical framework correspond to anomalies in the financial transactions and phone call networks.