We consider anomaly/change point detection given a time series of graphs with categorical attributes on the edges. Various attributed graph invariants are considered, and their power for detection as a function of a linear fusion parameter is presented.Index Terms-Anomaly Detection, Attributed Random Graphs, Fusion, Random Dot Product Graphs
TIME SERIES OF ATTRIBUTED GRAPHSGiven a time series of attributed graphs G(t) = (V, ϕ(·, t)), t = 1, 2, · · · , where the vertex set V = [n] = {1, · · · , n} is fixed throughout and edge attribution functions ϕ(·, t) : V
2→ {0, · · · , K} are time-dependent, we wish to detect anomalies and/or change points. Let us consider vertices to represent "actors," and an edge between vertex u and vertex v at time t (uv ∈ E(t), where the edge set E(t) is given by E(t) = {uv ∈ V 2 : ϕ(uv, t) > 0}) represents the existence of a communications event between actors u and v during the time period (t − 1, t]. Categorical edge attributes ϕ(uv, t), when non-zero, represent some mode of the communication event between actors u and v during (t − 1, t]; for instance, a topic label derived from the content of the communication. We will not consider directed edges or hypergraphs (hyper-edges consisting of more than two vertices) or multi-graphs (more than one edge between any two vertices at any time t) or self-loops (an edge from a vertex to itself) or weighted edges, although all of these generalizations of simple attributed graphs may be relevant for specific applications.The specific anomaly we will consider is the "chatter" alternative -a small (unspecified) subset of vertices with altered communication behavior during some time period in an otherwise stationary setting, as depicted in Figure 1. This figure notionally depicts the entire vertex set V behaving in some null state for t = 1, · · · , t * − 1; then, at time t in their null state throughout.) Our statistical inference task is to determine whether or not there has emerged a chatter group at some time t = t * . Fig. 1. Notional depiction of a time series of graphs in which the entire vertex set V behaves in some null state for t = 1, · · · , t * − 1 and then, at time t * , a subset of vertices V A exhibits a change in behavior.The latent process model for time series of attributed graphs presented in [1] provides for precisely this temporal structure. Each vertex is governed by a continuous time, finite state stochastic process {X v (t)} v∈V , with the state-space given by {0, 1, · · · , K}. The probability of edge uv at time t is determined by the inner product of the sub-probability vectors specified by p w,k (t) = t t−1The attribute ϕ(uv, t) for edge uv at time t, given that there is indeed an edge, is given byFor the scenario depicted in Figure 1, the vertex processes {X v (t)} v∈V A are stationary until time t * − 1 and then undergo a change point, while the processes {X v (t)} v∈V \V A remain stationary throughout all time.Our latent process model produces a dependent time series of attributed graphs G(t), each of which i...