Change detection in dynamic attributed networks

Abstract

A network provides powerful means of representing complex relationships between entities by abstracting entities as vertices, and relationships as edges connecting vertices in a graph. Beyond the presence or absence of relationships, a network may contain additional information that can be attributed to the entities and their relationships. Attaching these additional attribute data to the corresponding vertices and edges yields an attributed graph. Moreover, in the majority of real‐world applications, such as online social networks, financial networks and transactional networks, relationships between entities evolve over time. Change detection in dynamic attributed networks is an important problem in many areas, such as fraud detection, cyber intrusion detection, and health care monitoring. It is a challenging problem because it involves a time sequence of attributed graphs, each of which is usually very large and can contain many attributes attached to the vertices and edges, resulting in a complex, high‐dimensional mathematical object. In this survey we provide an overview of some of the existing change detection methods that utilize attribute information. We categorize these methods based on the levels of structure in the graph that are exploited to detect changes. These levels are vertices, edges, subgraphs, communities, and the overall graph. We focus attention on the strengths and weaknesses of these methods, including their performance and scalability. Furthermore, we discuss some publicly available dynamic network datasets and give a brief overview of models to generate dynamic attributed networks. Finally, we discuss the limitations of existing approaches identifying key areas for future research. This article is categorized under: Fundamental Concepts of Data and Knowledge > Data Concepts Algorithmic Development > Spatial and Temporal Data Mining Technologies > Machine Learning Application Areas > Business and Industry

Combining Information from Multiple Views for Vertex-Based Change Detection in Dynamic Networks: A Comparative Study

“…15 The four tensor slices with edge distribution F 1 , during group-change scenario for complex situation Fig. 16 The four tensor slices with edge distribution F 1 , during hetero-to-homo for complex situation Fig. 17 The four tensor slices with edge distribution F 1 , during fragment for complex situation the baseline methods.…”

“…One of the important tasks in dynamic network analysis is change detection, where the objective is to continuously monitor a network for changes in the behavior of entities and their relationships. Since this problem has many vital applications, a lot of solutions have been proposed in the literature [2,16,20,38,51].…”

“…Although there are hardly any approaches in literature that focus on entitybased changes in the multi-view network, we find several interesting tensor-based approaches that detect changes in the overall behavior of the network. A majority of these change detection methods apply tensor decomposition to obtain a low-rank approximation of the tensor at each time instant, and then track a suitable graph feature to detect changes [16]. For example, in [41], the reconstruction error of the decomposed tensor is tracked to detect change points in the network.…”

A Simplified Algorithm for Identifying Abnormal Changes in Dynamic Networks

“…Some articles mix both of those proposals in the same model ( [17]). As presented in [14]: a (discrete) dynamic weighted network can be mathematically represented as a time sequence of weighted graphs, (Gt = (Vt, Et, ωt)) t∈N , where Vt and Et are the set of vertices and the set of edges, respectively, and ωt : Et → R + is a weight function at time t. If the set of vertices remains unchanged over time; that is, Vt = V for all t, then the weighted graph at any time t will be simply Gt = (V, Et, ωt). In this case, a dynamic weighted network can be encoded as a weighted matrix, W (t) = (Wi,j(t)), where Wi,j(t) is the weight of the edge between vertices i and j at time t. Now, by applying the method described in Subsection 2.2 on each weighted graph Gt, we end up with a times series of persistence diagrams (∆t)t. To induce a scalar time series from (∆t)t, we consider a fixed persistence diagram ∆, then the 2-Wasserstein distances, W2(∆, ∆t), between the persistence diagrams ∆t and ∆, give rise to a (scalar) time series.…”

“…A network is an efficient representation of a given set of entities and the existing relationships between them. Actually a network is a graph which entities are the vertices and the relations between them are presented as edges (see for instance [14] and [17]). Usually, these relations are quantified by (mostly positive) real numbers and in this case we say that the graph (the network) is weighted.…”