Bodong Chen

Crisscross Landscapes

Notes: Carley. (2003). Dynamic network analysis



Citekey: @Carley2003-zk

Carley, K. M. (2003). Dynamic network analysis. In R. Breiger, K. Carley, & P. Pattison (Eds.), Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers (pp. 133–145). Washington, DC: National Academies Press. Retrieved from






Clearly social network analysis can be applied to the study of covert networks (Sparrow, 1991). Many are stepping forward suggesting that to understand these networks we just need to “connect the dots” and then isolate the “key actors who are often defined in terms of their “centrality” in the network. To an extent, this is right. However, it belies the difficulty of “connecting the dots” in terms of mining vast quantities of information, pattern matching on agent characteristics for people who go under multiple aliases, and still ending up with information the may be intentionally misleading, inaccurate, out-of-date, and incomplete. Further, this belies the difficulty in “knowing” who is the most central when you have at best only a sample of the network. Finally, and critically, this approach does not contend with the most pressing problem – the underlying network is dynamic. (p. 3)

Limitations to Traditional SNA (p. 3)

Traditionally, social network analysis (SNA) has focused on small, bounded networks, with 2-3 types of links (such as friendship and advice) among one type of node (such as people), at one point in time, with close to perfect information. To be sure there are a few studies that have considered extremely large networks, or two types of nodes (people and events), or unbounded networks (such as inter-organizational response teams); however, these are the exception not the norm. (p. 3)

Dynamic Network Analysis (p. 4)

Recently there have been a number of advances that extend SNA to the realm of dynamic analysis and multi-color networks. There are three key advances: 1) the meta-matrix, 2) treating ties as probabilistic, and 3) combining social networks with cognitive science and multi-agent systems. (p. 4)

Meta-Matrix: Carley (2002) combined knowledge management, operations research and social networks techniques together to create the notion of the meta-matrix – a multi-color, multiplex representation of the entities and the connections among them. (p. 4)

For our purpose, the entities of interest are people, knowledge/resources, events/tasks and organizations – see table 1. This defines a set of 10 inter-linked networks such that changes in one network cascade into changes in the others; relationships in one network imply relationships in another. (p. 4)

on the basis of this meta-matrix new metrics can be developed that better capture the overall importance of an individual, task, or resource in the group. An example of such a metric is cognitive load – the effort an individual has to employ to hold his role in the terrorist group and it takes in to account, who he interacts with, which events he has been at, which organizations he is a member of, the coordination costs of working with others in the same organization or at the same event or in learning from an earlier event or training for an upcoming event. (p. 4)

A key difficulty from a growth of science perspective, is that as we move from SNA to DNA the number, type, complexity, and value of measures changes. A core issue for DNA is what are the appropriate metrics for describing and contrasting dynamic networks. (p. 4)

For example, cognitive load, which measures the cognitive effort and individual has to do at one point in time has been shown to be a valuable predictor of emergent leadership (Carley and Ren, 2001). Cognitive load is a complex measure that takes into account the number of others, resources, tasks the agent needs to manage and the communication needed to engage in such activity. (p. 4)

The point is that these are a candidate set that have value and that as a field we need to develop a small set of metrics that can be applied to networks, regardless of size, to characterize the dynamics. (p. 5)

Table 1. Meta-Matrix (p. 5)

Probabilistic Ties: The ties in the meta-matrix are probabilistic. Various factors affect the probability, including the observer’s certainty in the tie and the likelihood that the tie is manifest at that time. Bayesian updating techniques (Dombroski and Carley, 2002), cognitive inferencing techniques, and models of social and cognitive change processes (Carley, 2002; Carley, Lee and Krackhardt, 2001) can be used to estimate the probability and how it changes over time. (p. 5)

Multi-Agent Network Models: A major problem with traditional SNA is that the people in the networks are not treated as active adaptive agents capable of taking action, learning, and altering their networks. There are several basic, well known, social and cognitive processes that influence who is likely to interact with whom: relative similarity, relative expertise, and coworker. Carley uses multi-agent technology in which the agents use these mechanisms, learn, take part in events, do tasks to model organizational and social change. The dynamic social network emerges from these actions. (p. 5)

Dynamic Network Theory (p. 6)

To move beyond representation and method, we need to ask, “How do networks change?” What are the basic processes? From the meta-matrix perspective, the processes are easy – things that lead to the adding and dropping of nodes and/or relations – see table 2. (p. 6)

Similarly, there are a set of processes that lead to the addition and removal of relations. Basic processes are cognitive, social and political in nature. Cognitive processes have to do with learning and forgetting, the changes that occur in ties due to changes in what individuals know. Social changes occur when one agent or organization dictates a change in ties, such as when a manager re-assigns individuals to tasks. Finally, political changes are due to legislation that effect organizations and the over-arching goals. To illustrate what is meant, a limited number of such processes are described in Table 3. (p. 6)

DyNet (p. 7)

The purpose of the DyNet project is to develop the equivalent of a flight simulator for reasoning about dynamic networked organizations. Through a unique blending of computer science, social networks and organization theory we are creating a new class of tools for managing organizational dynamics. The core tool is DyNet – a reasoning support tool for reasoning under varying levels of uncertainty about dynamic networked and cellular organizations, their vulnerabilities, and their ability to reconstitute themselves. Using DyNet the analyst would be able to see how the networked organization was likely to evolve if left alone, how its performance could be affected by various information warfare and isolation strategies, and how robust these strategies are in the face of varying levels of information assurance. (p. 7)

The DyNet tool is a step toward understanding how networks will evolve, change, adapt and how they can be destabilized. (p. 8)

DyNet, which is a computer model of dynamic networks, can also be thought of as the embodiment of a theory of dynamic networks. The focus of this theory is on the cognitive, and to a lesser extent, social processes by which the networks in the meta-matrix evolve. The basic cognitive forces for change in DyNet are learning, forgetting, goal-setting, and motivation for interaction. The basic social forces for change are recruitment, isolation, and to a limited extent the initiation of rumors and training. (p. 8)

The basic motivations for interaction are relative similarity, relative expertise or some combination of the two. Relative similarity is based on the fundamental finding of homophilly, the tendency of interacting partners to be similar. Arguments surrounding this fundamental process include the need for communicative ease, comfort, access, and training. Relative expertise is based on the fundamental finding that when in doubt people will turn they view as experts for information. (p. 8)

Among the attrition strategies are removal of the most “central” individual, removal of the individual with the highest cognitive load, and removal of individual’s at random. (p. 8)

Agents can be distinguished based on fixed characteristics such as race, family and gender, and on knowledge (or training). (p. 9)

Finally, the basic networks can be extracted continually in order to see the system evolve. Among the networks that can be extracted are the knowledge network, the overall social network, the emotive or “friendship” networks, and the acquisition or “advice” network. The network evolutionary strategies include learning (during interaction), forgetting, personnel attrition, misinformation, and changing task demands. (p. 9)

Results (p. 9)

The Structure of the Network Matters (p. 10)

The first finding, and it is quite robust, is that the structure of the network matters. That is, random networks in which the relations are distributed in an independent and identical fashion, hierarchies, and cellular networks all evolve quite differently, require different strategies to destabilize, have different abilities to diffuse information, and exhibit different performance for the same task. (p. 10)

A second key finding is that networks are generally able to heal themselves. That is isolation of a node that links disparate groups together typically does not leave those groups disconnected. Rather the basic social and cognitive processes outlined lead individuals to seek alternative contact points to interact with. (p. 10)

Full Information is Not Necessary (p. 11)

Notice, that in traditional SNA, typically we have close to full information. For covert networks we do not. Information may be missing because we don’t know some of the nodes – the people involved, or because we don’t know some of the relations. (p. 11)

Clearly having close to perfect or perfect knowledge means that more effective isolation strategies are found. Note, however, that any isolation is better than none, assuming our goal is to degrade the performance and that we don’t need perfect information to be quite effective. (p. 12)

Essentially, when we don’t really know the underlying social and knowledge network we may overestimate the primacy of a person, who although not the key in terms of degree centrality, is more central in terms of cognitive load. Thus, in effect, less knowledge makes both the centrality and the cognitive load strategies more similar resulting in on average lower performance due to the fact that cellular networks are more devastated by the extraction of such emergent leaders, at least in the short run. Further, reduced information about relations makes all isolation strategies more mixed thus inhibiting the ability of the opponent to engage in meta-learning. (p. 12)

Summary (p. 12)

Thinking about networks from a dynamic perspective is absolutely essential to understanding the modern world. An approach toward dynamic networks has been outlined. There are several distinctive hallmarks to this approach. First, in contrast to other multi-agent work, the agents we describe are in actual social networks. Here, the networks and the agents co-evolve. Secondly, the web of affiliations connects not just agents, but agents and other entities such as knowledge, tasks and organizations. (p. 12)

In contrast to traditional SNA, DNA considers the role of the agent in terms of processes and not just position. That is, the agents can do things – communicate, store information, learn. Further, the networks are dynamic and changing even as the agents change. The links are probabilistic, the networks multi-colored and multi-plex to the extent that the set of networks combine in to one complex system where changes in one sub-network inform and constrain changes in the others, often leading to error cascades. Finally, DNA explores the sensitivity of the measures and the impacts to error. (p. 13)

However, since all human action is cognitively mediated – it is unlikely that such mechanisms will not be derivable, at a basic level from what the physical and physiological constraints, what the agent knows, the basic learning and information processing mechanisms, and the way in which groups, organizations and institutions store such information. To create a truly dynamic network theory we need to create the equivalent of a quantum dynamics for the socio-cognitive world, where the fundamental entities, the people, unlike atoms, have the ability to learn. (p. 13)