Towards Dynamic Visualization for Understanding Evolution of Digital ...

Towards Dynamic Visualization for Understanding Evolution of Digital Communication Networks Matthias Trier1 1

TU Berlin, Franklinstrasse 28/29, 10587 Berlin, Germany

[email protected]

Abstract. The capabilities offered by digital communication are leading to the evolution of new network structures that are grounded in communication patterns. As these structures are significant for organizations, much research has been devoted to understanding network dynamics in ongoing processes of electronic communication. A valuable method for this objective is Social Network Analysis. However, its current focus on quantifying and interpreting aggregated static relationship structures suffers from some limitations for the domain of analyzing online communication with high volatility and massive exchange of timed messages. To overcome these limitations, this paper presents a method for event-based dynamic network visualization and analysis together with its exploratory social network intelligence software Commetrix. Based on longitudinal data of corporate e-mail communication, the paper demonstrates how exploration of animated graphs combined with measuring temporal network changes identifies measurement artifacts of static network analysis, describes community formation processes and network lifecycles, bridges actor level with network level analysis by analyzing structural impact of actor activities, and measures how network structures react to external events. The methods and findings improve our understanding of dynamic phenomena in online communication and motivate novel metrics that complement Social Network Analysis.

1 Introduction

Electronic media are becoming one of the main means for interaction in the workplace (Fallows, 2002). In addition to changing personal social behavior (e.g. Kraut et al., 2002) these means of computer-mediated communication affect organizational structures. For example, e-mail has

Prefinal Draft Version of: -1Trier, M. (2008): Towards Dynamic Visualization for Understanding Evolution of Digital Communication Networks. Accepted for publication in Information Systems Research, to appear in 2008.

been shown to complement formal work networks and provide more diverse, participative and less formally aligned relations (Bikson and Eveland, 1990). In effect the capabilities offered by digital communication networks are leading to the evolution of new network structures that are grounded in communication patterns. Examples of such structures are evident, for example, in the growth of online communities. They are defined as groups of people interacting in a virtual environment with a purpose, supported by technology, and guided by norms and policies (Preece, 2000). Such communities are of considerable significance for the corporation as organizational network structures are knowledge intensive and can constantly adapt their connection patterns (Monge and Contractor, 2003, p.325). Contrary to conventional wisdom, in such virtual networks relationships and attachments are developed and maintained (e.g. Cho et al., 2005). In a shared organizational context, the reduced social overhead of contacting unacquainted people even allows information flows between people that have never met face-to-face (Garton et al., 1997). Despite this virtual means of communicating and the large size of the participating groups, Berge and Collins (2000) found that most actors still have the perception of community. The formation of social network structures via interaction of people over time (Krackhardt, 1991) renders communication structures and online communities an object of systematic research with Social Network Analysis (SNA; e.g. Wellman et al., 1996; Garton et al., 1997). Its explicit focus on quantitatively analyzing interdependent patterns of social relationships differentiates SNA from traditional statistics and data analysis (Wasserman and Faust, 1994, p.3). The analytical approach uses network graph visualization extensively to represent, describe, and analyze communication matrices of interrelated actors. However, in the context of describing and explaining evolving relationships within online


communication networks, SNA has the important methodological limitation, that “almost all SNA research is static and cross-sectional rather than dynamic” (Monge and Contractor, 2003, p.325). This denies the dynamic nature of social relationships (Emirbayer, 1997) and inherent formation processes cannot be analyzed. In fact, the sampling method of SNA usually aggregates the wealth of longitudinal communication data into a single cumulative social network structure. The resulting analysis can be misleading when temporal and structural change is an inherent network property; as with online communication networks with their complex processes of community formation based on massive timed message events. Further SNA researchers frequently generate lists of central actors without knowing how important persons came into a position or if their status is already declining. Another important drawback is the predominance of static network images for visual representation and interpretation of structural properties. Such graphs can not represent network change (Moody et al., 2005, p. 1207). To improve existing research methods and to create new insights about the dynamic properties of online social networks, this paper presents an approach that disaggregates relationships into their constituting events and suggests event-based dynamic network analysis. The introduced method has also been implemented in the associated exploratory social network intelligence software Commetrix (cf. Trier, 2004; Trier, 2005). Based on the notion that visualization of information is the appropriate way to amplify cognition in complex domains (Card et al., 1999), and that SNA can be augmented by improving current static visualizations (also cf. Moody et al., 2005), the approach is to utilize current advances in information visualization to extend perceptional and analytical inferences about large amounts of dynamic network data. The individual streaming events are retained together with their time stamps for a more accurate dynamic visualization and measurement. The software implementation and especially its


visualization are regarded as an important cornerstone that enables exploratory observation of dynamic network evolution. The proposed event-based approach is a promising foundation for complementing existing SNA methods. Examples for its extensions include the analysis of group formation and stabilization over time, of actor paths to central positions, or of process oriented activity patterns with a structural impact on the network. Explicit recognition of relational events is further able to capture the growth of relationships and the network’s reaction to external events. Generally, the method provides multiple integrated levels of analysis by linking actor attributes (e.g. types), actors’ activity patterns, and the resulting impact on general network structures. The broad research objective of this paper is to illustrate the advantages offered by the proposed event-based approach to dynamic network analysis in improving understanding of evolving processes of online communication networks. Specifically, it addresses the following research questions: 1) How can longitudinal network analysis overcome limitations and measurement artifacts of summative pictures provided by static SNA? How volatile is the formation of an online communication network and its actors’ positions? 2) What processes of general network and subgroup formation can be observed and described with event-based visualization and analysis? 3) How can event-based dynamic network analysis evaluate actor activity, i.e. the structural impact of actors who actively broker and integrate separate parts of the corporate network? Which organizational positions have such actors? 4) What is the impact of external events on the network structure and its levels of change? The paper begins with a brief introduction to Social Network Analysis followed by a discussion


of the main shortcomings of its aggregated data and visualization model. Related research is then summarized to subsequently present the method of event-based dynamic network analysis and the associated software Commetrix for visualizing and analyzing the dynamics of evolving online communication networks. The suggested approach is applied to study the dynamics of the corporate e-mail communication network of Enron Corporation.

2 SNA concepts and their shortcomings for dynamic analysis

The methodological body of Social Network Analysis (SNA) is frequently applied to observe and analyze online social networks (e.g. Garton et al., 1997; Cho et al., 2005). SNA typically builds a network of actors as nodes and their mutual relationships as ties. An overview of typical measures of SNA is provided in Table 1. These measures include composition variables, i.e. the number and properties of actors, or structural variables, i.e. the properties of relationships. In an online communication context, a relationship can be derived by counting exchanged messages. Relationship strength differs across communication media. For example, compared to e-mail, instant messages have much higher frequencies of interaction. However, in relative terms, strong and weak relationships can be identified for a defined technology of electronic communication. Actors who maintain strong ties are more likely to share the resources they have (Wellman and Wortley, 1990). Another basic property is network size (cf. Table 1). Larger social networks tend to have more heterogeneity in their social characteristics and more complexity in their structure (Wellman and Potter, 1997). Large heterogeneous networks (such as those often found online) are more likely to exhibit weak ties to different social circles which are beneficial for obtaining more diverse information (Granovetter, 1973; Garton et al., 1997). A further important property often studied


in network analysis is the centrality of selected actors (cf. Table 1). It has been identified as an indicator of satisfaction or importance of actors within a network (e.g. Brass, 1984). Although these measures and roles provide elaborated methods to analyze networks, they concentrate on structural issues. The snapshot of the final network does not describe, how central actors achieved their final positions or if the network or its clusters experience stability or decay. Tab. 1. Overview of structural SNA measures and network roles (for formalized definitions cf. Wasserman and Faust, 1994).

Network Size Relationship Strength, Tie Strength Degree (Activity vs. Prominence) Diameter

Density

ClusteringCoefficient Centrality Betweenness

Centrality Closeness Centrality Degree Reciprocity Broker role (Gatekeeper) Hub role Isolate role Transmitter, Receiver, Carrier role Pulsetaker role

Number of nodes in a network, e.g. participating actors. The strength of the relationship between two actors. It can indicate the frequency of interactions (daily, monthly), count actual interactions, or measure intensity of relationships. In the communication context, relationship strength is increased via timed events in the form of initiated and received messages. The number of adjacent contacts a node has, e.g. e-mail communication partners. If the direction of the events is contained in the dataset, activity (out-degree) measures the relationship forming events initiated by the observed actor, e.g. establishing the contact, referring to another authors work, sending messages etc. Prominence (in-degree) measures the events initiated by actors adjacent to the observed node. Longest shortest path (distance in terms of steps) between two nodes in the network, e.g. the longest process (in terms of steps) of forwarding a mail in a network from one side of the network to the other. The larger the diameter, the less likely is the arrival of information on the other end of the network. Connectedness of the network’s nodes. Proportion of pair wise connections realized between n nodes of a network divided by the number of theoretically possible relationships between those n nodes. Communication networks usually have a low density (sparse network) as not all actors are connected to all others. Measure of sub-group formation and of the density of an ego-network. The proportion of links between the direct contacts of an observed ego-node divided by the theoretically possible links between its direct contacts. In a communication networks, this shows if contacts of an actor tend to share information directly (transitivity). Measure of communication control. Number of shortest paths between pairs of nodes, which run through the observed node. In an e-mail network this could be the person who forwards important messages and thus is important for the information transfer between pairs of actors. This can be an important network position but is also critical for information transfer in a communication setting. Distance of a node to all other nodes in the network measured with average shortest path length. In a digital network this measure indicates how fast or efficient an actor can access the network and how likely it is, that information reaches him. A simple centrality measure, counting the relative share of contacts of a node. Symmetry of relationships. If there is a relationship from node A to node B and vice versa, then this relationship is called reciprocal. In online communication settings, it can also be a weighting of the links from A and B versus the links from B to A. Network position, which is located on an exclusive path between two cliques or subcomponents. If removed, adjacent subcomponents get disconnected. Brokers thus control the flow between sections of the network. They tend to have a high betweenness. A hub is a central actor (i.e. with a high degree). Many messages pass this position. An isolate has a degree of zero and has thus no relationships to others in the network. Transmitters have an in-degree of 0 and an out-degree above 0. They have only sent messages to the network but did not receive any. Receivers have an out-degree of 0 and an in-degree of above 0. Carriers have in-degrees and out-degrees above 0 (normal case) and thus received and transmitted information between other nodes. A pulse taker has a small degree but connects to nodes with a high degree (e.g. hubs). The quotient between indirect links and direct links is high. This can be an efficient position as most information is likely to arrive without the need to maintain many contacts.


The structure of recent changes remains invisible and unexplored as does the shift of central positions between nodes. Centrality measures alone do not convey, if a central position is beneficial for the network evolution or if it is a critical weak point. The actual activity of actors and their impact on the lifecycle of the community cannot be observed. Such gaps in recognizing dynamic processes have been long criticized by researchers: "Models of structure are not sufficient unto themselves. Eventually one must be able to show how concrete social processes and individual manipulations shape and are shaped by structure” (White et al. 1976, p. 773; also cf. Emirbayer, 1997). According to Doreian and Stokman (1996) studying network processes therefore requires the use of time, i.e. temporally ordered information in addition to descriptions of network structures as summarized information. Empirical analysis of social network change started with the collection of small numbers of separate waves of relationship data with a primary focus on aggregated interim states of a network (e.g. Hammer, 1980; Freeman, 1984; a comprehensive overview is given in Doreian and Stokman, 1996, p.6). These methods are limited to comparative studies of general differences between these states on the aggregated network level. The actual sequence of activities is lost and changes in the relationship pattern can average out between two points of observation. Hence, such comparative analysis may be employed in domains with little temporal change (e.g. kinship networks) but seems inappropriate for studying fast paced online communication. An approach that takes some repeatedly collected waves of relationship data as input and estimates the existence of certain dynamic effects in a network is the stochastic actor-driven model (e.g. Snijders, 2001). It is based on simulating Markov chains of networks between consecutive observations and assumes that actors analyze their current embeddedness in a network structure to change their links according to a pre-defined value function. It contains


factors expressing theoretic network effects (e.g. maximization of reciprocity or similarity among actors). Each factor has a parameter that can be estimated based on the waves of empirical data. This approach is close to another approach that employs probabilistic ties and uses a multi-agent based simulation model to predict network behavior (Carley, 2003). Such studies typically computed general variables at the network level using only a few waves of aggregated data (also cf. Moody et al., 2005), and did not relate structural change directly to time units. Thus the notion of pace or fluctuation of the network is not addressed. In terms of insightful visual representation, the studies mainly rely on line graphs with one or more variables (e.g. transitivity, reciprocity, density, and centrality) over a time-axis. Despite the key role of imagery in network research (Freeman, 2000), the above approaches do not exploit dynamic visualization to leverage the analysis. Other approaches in the field of visualization of dynamic networks do so; these are briefly discussed next.

3 Related approaches in Visualizing Dynamic Social Networks

Since the beginning of graph theoretic analysis, there is a slow but continuous evolution of technical approaches to social network visualizations culminating in the creation of advanced tools to measure and visualize networks. Until recently, these visualizations simply compared graphs of the cumulative networks states at different times. A related strand of research, not directly focused on the quantitative analysis of relationship structures, developed rich and animated representations of online social spaces of electronic communication. Further, software libraries for dynamic graph drawing have been recently introduced. Finally, approaches that explicitly discuss and target dynamic network visualization and analysis of continuous (streams of) data with high sampling rates have begun to emerge. These related concepts are now


elaborated in more detail and discussed in relation to the presented Commetrix approach of visualizing event-based social network data. The static social network graph was first introduced by Moreno in 1934. This “sociogram” contained actors as nodes and their relationships as links between the nodes. Since its invention, changes occurred only in the technical methods to produce the graphs. Until today, powerful software tools for semi-automated analysis and visualization of large network structures have developed (for a comprehensive overview see Freeman, 2000). Examples for current analytical software packages are Ucinet (Borgatti, Everett & Freeman, 1992) or Pajek (Batagelj and Mrvar, 1998). They usually import formatted data files and provide sophisticated statistical analysis. They further can generate structural network graphs, which can then be exported as images or 3D models. Although Pajek recently introduced means to define in which time periods nodes or links were present in order to compute partial networks, such tools are based on data about aggregated structures and do not automatically capture, evaluate or animate dynamic data and events from communication sources. An alternative family of approaches comes from visualizations of online social spaces of electronic communication. They suggest various intuitive metaphors to represent online social activity, e.g. graphical tree-like hierarchies of postings (e.g. Smith and Fiore, 2001), a garden with flower petals, or a tree with leaves to convey the ‘health’ of the electronic group (e.g. Girgensohn et al., 2003). This has also resulted in the formulation of the concept of Social Translucency, as “an approach to designing digital systems that emphasizes making social information visible within the system” (Erickson and Kellogg, 2000). This family of approaches was the first to employ motion for insightful and ‘living’ virtual representations of changes in the conversation. However, compared to event-based network analysis, those concepts were


developed to aid the user in visually navigating online spaces. They do neither provide for a quantitative network analysis of the displayed dynamic structures nor do they explicitly focus on relationships. A further related development is the advancement of general graph drawing packages. One example is Graphviz of AT&T Labs Research (Ellson et al., 2004). As an open source graph visualization package, it is a collection of software for viewing and manipulating abstract graphs in the software engineering, networking, databases, knowledge representation, and bioinformatics. All early algorithms of Graphviz concentrated on static layouts, until Dynagraph was introduced in 2004 which includes algorithms, that “maintain a model graph with layout information, and accept a sequence of insert, modify or delete subgraph requests, with the subgraphs specifying the nodes and edges involved” (Ellson et al., 2004, p.14). The focus, though, is on interactive editors for general graph drawing with applicable technical layout concepts and software libraries to dynamically update a graph view. The libraries include no network analytical approach or perspective and are not focused on social networks. There are three contemporary approaches that, similar to the method presented in this paper, work on the actual integration of Social Network Analysis and changing graphs. Perer and Shneiderman (2006) introduced an approach that includes some functions to trace changes in network data by hiding links outside a selected moveable time window. Nodes maintain a fixed position based on the final network configuration. This mode has been termed flipbook by Moody et al. (2005, p.1234) as it is a static technique that reveals how a network structure unfolds over time based on interactions. However, the lack of dynamic repositioning of nodes yields interim networks with uninformative layouts. For example, nodes with an early but weak relationship would eventually be placed far apart, but early in the sequence would better be


positioned near each other and then move apart to slowly give room for later but stronger relationships between them. It is hence less suitable for recognizing cluster formation or sudden changes in actor’s network positions. Beyond this flipbook technique, the two more advanced approaches of dynamic network visualization by Gloor et al. (2004) and Moody et al. (2005) try to represent structural change as motion in a social network graph. Both segment longitudinal data into subsequent time windows and render their individual network graphs, which are then visualized as an animated sequence. To provide visual consistency for the changing node locations, positional transitions are computed between subsequent visualization frames. However, the suggested techniques based on transitions between time frames produce much unnecessary node movement that result in many crossings or long edges in the dynamic layout. This is likely to decrease readability for datasets larger than 50 to 100 nodes due to much simultaneous motion. A further obstacle to dynamic network research is that these software tools provide extensions to visualize network data but lack a direct integration with functionality to compute SNA metrics for selected network sections. The user interface still exhibits much potential for improving exploratory analysis and in-depth quantitative insights of the visualized networks or for manipulations of the dataset (e.g. filtering out a subset). On the other hand, the approach of Perer and Shneiderman (2006) is focused on easy exploration but does not fully exploit the opportunities for dynamic visualization. All employed animation algorithms also have potential for improvement and enrichment to better convey changing properties of actors and relationships over time. In summary, conventional SNA methods have developed comparative analysis and stochastic parameter estimations but are lacking in advanced visualization capabilities for observation and


verification. Only a few recent approaches have started to develop visual means for observing change in social networks, but they do not study the impact of activities or external events on the final network structure. Extant visualization techniques still suffer from some limitations and are not comprehensively connected to exploratory network measurement. Without such integration, novel measures that better capture network dynamics remain unattainable.

4 A Methodology for Dynamic Visualization and Measurement Approaches

Commetrix is a java-based tool constructed for event-based dynamic network analysis and attempts to address the limitations of current approaches. The development of this tool started at about the same time as the above related approaches (cf. Trier 2004, 2005) and has yielded a comprehensive set of software-based methods for exploratory static and dynamic visualization with integrated analysis of social network measures. The underlying framework for event-based dynamic network analysis consists of a data model that contains information about the network including the timing of network events. Integrated with that is a sophisticated visualization technique based on a 2D/3D spring embedder (cf. Fruchterman and Reingold, 1991) that allows for adding and deleting network elements to a graph representation. Finally, a special method for smooth graph transitions has been developed. First, the fast paced communication data needs to be sampled and stored in a data model for systematic analysis. Conventional SNA datasets are based on a graph G = (N, L) which consists of a finite set of nodes N and a finite set of lines L that are constituted by pairs (ni,nj) of nodes (Wasserman and Faust, 1994, p.122). If nodes represent actors and edges represent relationships, such a graph is also referred to as a sociogram. The respective matrix which stores relationships between each pair of actors is called a sociomatrix. The event-based approach now implies


several changes for storing the network data. Relating to Doreian’s and Stokman’s (1997, p.3) definition of a network process as a “series of events that create, sustain, and dissolve social structures”, relationships are not directly considered but their constituting timed events are captured. In communication network analysis such relational events are created by exchanging messages with others. From these events, relationships can be aggregated. In the most basic sample procedure every message event will increment the relationship’s strength by a value of 1. The simple case of dichotomous relationships (absent vs. present ties) can be covered by only modeling a single timed event that creates the relationship at a specific time. In studies of online communication, replies and carbon copy e-mails can be stored as relational events or can be intentionally ignored in the sampling process.

Data Model has

Actor

has

Visualization

Property (Type,…)

Network has

Event

has

Property

(Relationship)

(Time, Content, Type,…)

Fig. 1. The data model stores actors with properties like name, function, type, etc. and events with properties like time, content, or type. Relationships are time oriented aggregations of events. The visualization represents actors as nodes and relationships as arcs and utilizes different visual variables (size, color, saturation, etc.) to encode the properties.

The data model underlying the approach consists of actors, actor properties, events, and event properties (also cf. Figure 1). For each event several properties are captured. For example, the time stamp of each message event is recorded as a message property. Hence, the sequence of messages and the change in relationship structure or strength is represented as a series of relational events in the data model. Examples of further event properties are keywords, contents, coded communication types (e.g. socialization vs. task organization), or evaluations, that can


then be used for content-oriented analysis or similarity based grouping. In addition to these important changes in capturing relationships, the actual actors are modeled together with their actor properties. The latter can include names, organizations, evaluations, organizational ranks, types, or locations. The visualization represents the data model graphically. As in the conventional sociogram (social network graph), actors are represented by nodes and edges represent the relationships as flexible aggregations of message events. The sociogram extended with additional means for information visualization and the capability to adapt to longitudinal network change yields a dynamic graph termed ‘communigraph’. Utilizing Bertin’s (1967) concept of visual variables to encode information, properties can be visualized by label, node size, node color (brightness, transparency), or a number of rings around the node. Relationship properties are graphically represented using colors, thickness, length, and labels. In the domain of dynamic analysis, the representation of change in the graph is a fundamental part of the visualization. It requires algorithms for handling transitions between incremental network states in order to represent structural changes with organic movement. This major aspect of dynamic visualization can be termed transition problem. Due to its role in differentiating among alternative approaches to dynamic network visualization, this aspect is now discussed in more detail. As already introduced, the related work of Moody et al. (2005) and Gloor et al. (2004) is based on a sliding time window that is moved through the overall sample period. For each of these time windows (frames) a network layout is computed. If structural change occurs, two subsequent network layouts differ in their node’s position. To create a consistent transition, the authors render interim frames. The visualization is then “gradually adjusting node coordinates and


adding or deleting nodes and arcs“(Moody et al., 2005) or as (Gloor et al., 2004) describe it: “the animation of the changing layout is interpolated between […] keyframes”. Both approaches thus calculate network graph layouts at different states (e.g. per day) and then linearly move nodes from their position in the network of the first time window to their position in the subsequent time window. Careful examination of such layouts shows that their rendering of transition frames disturbs the impression of organic evolution of network structures. Nodes cross other nodes, swap their position without need, or move at unintuitive changing speeds or in quickly changing directions across the screen. The inconsistent motion is caused by two conflicting relocation strategies. Node movement is alternately governed by the network layout algorithm of timeframe 1 and then by the positional transition algorithm that moves nodes to their new optimal network position in timeframe 2. Being trained to evaluate stable parts by their inertia, the observer is distracted from observing how new nodes find their position while large ‘established’ centers also shift positions and all adjacent nodes in their clusters with them. The result is a suboptimal impression of transitions between separate layouts instead of observing network behavior with its events and their impact on the remaining structure. To create smoother transitions across time frames, the visualization implemented in the Commetrix tool avoids linear transitions between rendered keyframes. Rather, new nodes are added directly to the visual representation at the time, when the resulting event actually occurs. The technique literally ‘throws’ additional communication elements into the network layout at the according time to let them find their natural place. As the supplemental videos (available at www.commetrix.de/enron) show, this results in a very organic view on network evolution. The novel technique necessitated the development of a dynamic version of the spring embedder


layout algorithm (Fruchterman and Reingold, 1991). It can accommodate new nodes into an existing network layout. A major reduction in unnecessary node movement has been achieved by relating node inertia to their number of contacts (degree). As a result, larger structures become more inert and less connected nodes quickly move towards them. This keeps established parts as stable as they should appear, while drawing the user’s full attention to moving areas where the actual change happens. The movement in the evolving graph of online communication thus directly represents structural changes and in effect, the social network looks like a real living system of interactive elements in a network relation. In analogy, relationships and nodes older than the observed time window can be dynamically taken out of the layout procedure. This yields visualizations that directly show the recent changes in the network’s evolution.

5 Analysis and Discussion

The research questions posed in the introduction are now addressed by illustrating the capabilities of this approach with a sample of corporate e-mail data of Enron. The data were originally published by the Federal Energy Regulatory Commission in May 2002 as a consequence of the investigations into the fraud and bankruptcy scandals of Enron in December 2001. The original dataset covered 619446 messages (around 92% of monitored e-mails) in 3500 personal e-mail folders over a period of three and a half years. This sample has been refined by Gervasio of SRI International for the CALO Project (Cognitive Assistant that Learns and Organizes) and subsequently by Shetty und Adibi from the University of Southern California's Information Sciences Institute, resulting in a corrected network of 517431 mails of 151 actors (cf. Shetty and Adibi, 2004; the authors also provide a link to the data source). The managers, traders and employees were working at different physical locations. In the study presented here,


of this set all those 19811 messages have been considered that originated and terminated within this set of 151 actors. The sample duration is 38 months, i.e. from May 5th, 1999 to June 21st, 2002. Discussed topics include regulations, internal projects, company image, political relationships, operations, logistics of arrangements, reports of business trips, and information about partnerships. The data also includes information exchange of a more personal nature in the professional context. The sample is very suitable to analyzing dynamic network evolution, as the e-mail contents are known and it consists of strong relationships of timed electronic communication, required to demonstrate network dynamics. The years 1999 and 2000 represent everyday operations of the sampled population whereas the years 2001 and 2002 reflect several external events in the context of Enron’s bankruptcy scandal, whose impact on the network dynamics is studied. Isolating Volatility in Communication Patterns and Positions The first research question concerned the artifacts created by conventional summative SNA. This method would only use the final static picture as shown in Figure 2d. This cumulated network contains one large component of 150 actors (1 isolate node has been removed in the graph). During the sample period, 1526 relationships can be observed with the average relationship strength of 26 exchanged messages. For better reference to particular sections of the network layouts, several borders have been manually added based on visual inspection. The final structure shows that the e-mail network, although completely connected, forms larger subgroups, which are in the cumulative graph of Figure 2d connected to a very dense center (named section 1) via a larger number of links. The more peripheral sections are smaller and have a stronger connection within than between sections and thus appear more separated and peripheral. Nodes have on average 20 contacts. The most central node is node 87 who is connected to 50% of all


actors (74 contacts). Size: 87 actors in 167 relationships

Size: 115 actors in 488 relationships

Node 87: Dgr, BtwC%, DgrC%: 5, 0.11%, 6% Rank BtwC: 47th Rank Dgr: 21st

Node 87: Dgr, BtwC%, DgrC%: 15, 1.14%, 13% Rank BtwC: 44th Rank Dgr: 101st

Section 4

Section 3 Section 1

a) July 1st, 2000

b) February 26th, 2001

Size: 147 actors in 1204 relationships

Size: 150 actors in 1526 relationships (+ 1 isolate)

Node 87: Dgr, BtwC%, DgrC%: 16, 0%, 11% Rank BtwC: 123rd Rank Dgr: 63rd

Node 87: Dgr,BtwC%,DgrC%: 74, 9%, 50% Rank BtwC: 1st Rank Dgr: 1st

Section 2

Section 3

c) October 24th, 2001

Section 4

Section 1

Section 3

Section 2 Section 4

Section 1

d) June 21st, 2002

Fig. 2. The cumulative evolution of the most central actor’s position in the network. Color represents degree and size represents the betweenness of the node at the respective time. Observed node is red. Network size, degree, betweenness centrality, and degree centrality for the observed node 87 is given. All measures and visual output were computed using Commetrix. The borders between sections were manually added for better reference. The original animated graph is available as a movie at http://www.commetrix.de/enron.

These findings of static analysis can now be contrasted with insights gained from analyzing the network’s structural change. Figure 2a-c shows three snapshots of the animated graph of the complete evolution. The changing node size of node 87 now highlights that this identified central actor (node 87) clearly did not establish its position in a steady growth but rather suddenly towards the end of the overall sampling period. The network metrics of node 87 over time (listed in Figure 2) show a centrality ranking of rank 47 out of 87 active nodes in period 1 with only 5 contacts. Subsequently, node 87 remains equally unimportant in terms of centrality until in the last quarter almost all of its centrality has been achieved (note the difference in node size between Figure 2c and 2d). Prefinal Draft Version of: -18Trier, M. (2008): Towards Dynamic Visualization for Understanding Evolution of Digital Communication Networks. Accepted for publication in Information Systems Research, to appear in 2008.

A detailed analysis highlights the most interesting period: between February 2nd and February 5th, 2002, where the centrality has increased by the factor 2.8 in just three days within the overall period of 1137 days (an animation visualizing this abrupt change is available at www.commetrix.de/enron). Afterwards there are no further significant positional changes until the end. Analyzing the broader context, node 87 represents the assistant of the leader of the wholesale trading division. That leader became Enron’s last president (node 44) in August 2001. This promotion seems to be an external change affecting the position of the assistant node 87. It can be concluded that the most central node highlighted by summative SNA established its position not in a steady increase but in a very fast burst of activity. Dynamic analysis hence directs the focus to unusual patterns in the overall network evolution which would not have been discovered with static analysis. Once temporal effects such as sudden changes have been identified in the sample, the analyst can focus on studying the temporal evolution of the network to decide whether node 87 should still be considered structurally important. With the underlying event-based data model, the analysis can hence seamlessly shift from the network level to the actor level. Generally, dynamic analysis highlights that in digital communication networks central positions can be very volatile due to the ease with which new relationships are created. This is especially the case in a corporate network, where a macro-context has an impact on the initialization of new contacts. Other examples, e.g. sudden changes triggered by the newly appointed CEO around August 24th, 2001, support this notion. In online communication, the measure of centrality hence strongly depends on the timing, e.g. SNA would identify another node (84) in period three. For the studied domain, static measures are thus likely to yield misleading results. Taking this point one step further, developing a general dynamic measure of


burstiness and volatility (i.e. variance) can help to establish, when dynamic analysis is necessary in order to prevent measurement errors from inappropriate aggregated sampling. At the actor level, sudden changes can further be related to actor properties (e.g. membership duration) or used to identify a-typical changes as a signal for possible suspect behavior. On the other hand, dynamic network analysis can also help to remove unrepresentative anomalies from the network data, which otherwise would result in an incorrect representation of the final structure. Network and Subgroup Formation Next to observing single nodes and their positional changes in the network’s structure, dynamic analysis provides improved means to describe the development of the complete network and its separation of sections over time (research question 2). The following descriptive analysis of the formation process of the Enron e-mail network is based on a combination of exploratory visual inspection and time dependent SNA measures performed using Commetrix. The focus was on identifying typical patterns by which certain social network architectures and subgroups emerge. Figure 2a-d is again used as a visual reference. Starting with a small integrated network, the center is increasing its density and section 3 emerges with a burst in activity around node 90 on January 1st, 2000 (Figure 2b). On this date, the node achieved a betweenness of 19%. The new section is connected to the main center by only a few nodes. On August 18th, 2000, the spike of section 4 occurs via an exclusive link between nodes 67 and 106. Minor traces of the slow and broad formation of section 2 are also visible. In the third quarter of the formation process, section 4 builds many connections to the center and almost becomes integrated. During this process the initiating node 106 and the initial exclusive link completely lose importance. Section 2 continues its slow separation in the winter of 2000/2001. Section 3 establishes a broader connection via many connecting nodes to the


center. In this process, the initiator of section 3, node 90, loses its brokering position (betweenness declines to 1.5%) and a new actor emerges with node 9 (betweenness 8%). Actors 6 and 76 in the center grow in their degree (denoted by node size). They largely contact nodes within the center and thus increase the density of that area. In the final period, all peripheral sections develop more separation but remain connected with each other via the center. This first detailed descriptive account of dynamic network formation processes with emerging and decaying sub-structures highlights some general process patterns. Separate sections were initiated by a very central and active node (e.g. node 90 started section 3), by a central and exclusive link (e.g. like between nodes 67 and 106 for section 4), or by very slow separation of many nodes (section 2). Such descriptions of structural processes show the potential of dynamic analysis to support the induction of general theories about dynamic patterns or antecedents of community formation from empirical data. A further insight is that such cluster formation is not uniform. Section 4 was moving towards integration with the center and the continuous separation of section 3 resulted from node 9 taking over the declining central position of node 90. This suggests a concept of several overlapping lifecycles of network sections instead of assuming steady and homogeneous growth across the overall network. Processes of change are even more visible if older messages outside the observed time window decay and get eliminated from the visualization. This emphasizes the added activity within the current time window (e.g. one day) without distractions from past accumulated structures. In this visualization mode, clusters are only persistent if the participating nodes reactivate their links within the defined time window. Otherwise relationships dissolve again. This gives a visual impression of networking speed (frequency) and helps to understand, who contacts whom to actually establish the network. Based on such process oriented analysis of


individual activities and their structural impact, the identification of correct important (i.e. active) players in online communication networks with high volatility can be improved. The Structural Impact of Brokering Actions One application of such activity oriented analysis of network dynamics arises in the context of research question 3. The changes brought about within subsequent time windows of one day are studied to analyze brokering activities that span large distances in order to integrate separate parts of the corporate network. Each action is considered a brokering activity, which creates shortcuts in path length of more than one step. This excludes connections resulting from the natural tendency of indirect paths of length 2 to become direct paths of length 1 (triadic closure), i.e. bypassing one intermediary node. For example, if nodes A and D were connected via three steps (e.g. A-B-C-D) then A performed a brokering activity if he directly created a relationship to D (A-D). Figure 3 gives an example of how dynamic visualization shows such a brokering situation. On the observed day, the marked node impacts the overall network structure by connecting three otherwise disconnected segments of the network. The process results in shorter network paths and thus contributes to the formation of a more integrated network structure. Research question 3 further concerns the organizational ranks of actors with a high brokering activity level. To establish this relationship, available data about 95 organizational positions is utilized. In the example shown in Figure 3, the observed node has the rank president (represented by node color and label) and its new contacts are also above management level: one director and two vice presidents. For the following analysis, it has to be noted that the sample is not a random sample but focuses on people, which where related to the Enron case of fraud conspiracy and the resulting bankruptcy in late 2001 and is thus biased towards upper levels of the hierarchy. The year 2000 is taken as a subset. This year’s activity is well before the unusual final year of 2001


and should thus give a representative account of networking processes. All brokering activities of the network have been counted and coded by visual inspection of the animated graph.

a) February 27th, 2000

b)

c) February 28th, 2000

Fig. 3. A node with job position president connects three otherwise separate clusters (via two vice presidents and one director) on February 28th, 2000. Four separate frames of the according animation. Time filter shows only one past month of mail activity.

Together, 74 actions have been classified as brokering actions in the year 2000. They spread evenly across the year. For 36 of these actions the job position of the involved actors is known. Out of these brokering actions, 9 instances (25 percent) have involved only top management positions and further 9 (25 percent) only employees. The majority of 18 connecting actions (50 percent) have been cross-hierarchical. This quantitative pattern is supported by the visual impression: Managers connect with distant employees in a brokering action to join separate parts of the network and form the single integrated component shown in Figure 2d. This study of brokering activity demonstrates the multiple levels of analysis facilitated by the approach. Event-based analysis relates actor attributes (e.g. organizational rank) with actor activities and their impact on network formation. The technique of sliding time window visualization and analysis enables to hide past cumulative structures in order to analyze a network from an activity oriented view. Network structures are disaggregated into individual networking processes and each incremental network development can be observed and


measured. By that, the structural impact of actor attributes and activities on the overall formation of the social network structure is uncovered. The Impact of External Events Related to the analysis of activities, dynamic visualization of sliding time windows can be utilized to learn about the reactions of an electronic communication network to external events (research question 4). The network’s level of activity and change is measured by setting a sliding time window (e.g. one month) and by moving forward in time taking measurements of active nodes, active relationships between them, and the current average relationship strength. The active nodes and relationships of one time window can be interpreted as the incremental addition of network activity. Figure 4 summarizes the quantitative analysis of the animation. The number of actors slowly increases until July 2001. Then it stagnates at the level of about 130 simultaneously active actors. Despite this stable number of active users per month, this period is marked by an unprecedented increase in active relationships and in relationship strength. A constant number of actors are creating new relationships among themselves and intensifying them, resulting in an increasingly dense network. This temporal effect is accompanied by a sharp increase in message frequency (middle chart in Figure 4). Further information about the context of the period reveals that this pattern of change happens at the time when the Security and Exchange Commission started their investigations into the Enron fraud scandal on October 31st, 2001, and Enron filed for bankruptcy in December 2nd, 2001. The Enron e-mail network seems to react to a fundamental external event with a contraction indicated by a quick increase in interaction frequency, network activity, and network integration.


14

Average Relationship Strength Linear (Average Relationship Strength)

12 10 8 6 4 2

02 M ai 02

M rz

02

01

Ja n

01

N ov

Frequency (sent msg in 30 day window)

Se p

01 M ai 01 Ju l0 1

M rz

01

00

Ja n

00

N ov

Se p

M rz

Ja n

00

99

99

N ov

Se p

2500

00 M ai 00 Ju l0 0

0 3000

2000 1500 1000 500 0 S ep- 99 Nov - 99 J an- 00 Mar - 00 May - 00 J ul - 00 S ep- 00 Nov - 00 J an- 01 Mar - 01 May - 01 J ul - 01

S ep- 01 Nov - 01 J an- 02 Mar - 02 May - 02

May-02

Mar-02

Jan-02

Nov-01

Sep-01

Jul-01

May-01

Mar-01

Jan-01

Nov-00

Sep-00

Jul-00

May-00

Mar-00

Jan-00

Nov-99

Number of Active Authors Number of Active Relations

Sep-99

500 450 400 350 300 250 200 150 100 50 0

Fig. 4. Level of network change. Time window shows only the last month of network activity to indicate the change in active actors, active relationships and current average relationship strength across the time period. A clear peak in change and in relationship strength is visible around the time when the Enron scandal was published.

This finding is another example of the improved analytical understanding of processes in online communication networks resulting from the combination of events and SNA. The study can be a starting point for further academic investigations about typical reactions to external events. Next to reactions, another important issue is the anticipation of events by analyzing network behavior to identify indicators for a current general external impact. This also builds a connection to the first research question which found anomalies that even affected the final structure of the network. Anomalies are likely to be more influenced by external factors than by internal structures. Further, community formation processes (cf. second research question) can be related to the impact of external events. Generally, the research scope extends from studying change to studying change of change, e.g. large tendencies and their likely reasons, sudden activity, or fast restructuring. Prefinal Draft Version of: -25Trier, M. (2008): Towards Dynamic Visualization for Understanding Evolution of Digital Communication Networks. Accepted for publication in Information Systems Research, to appear in 2008.

6 Conclusions and Outlook

The approach described in this paper has two types of implications: insights about the dynamics of an e-mail network and methodical insights about how event-based dynamic network analysis can help researchers and practitioners to learn more about social networks with massive timed events. Dynamic analysis of Enron’s corporate e-mail creates a more detailed picture of processes in online communication networks: Central actors are not constantly maintaining their position but quickly rise and fall in their centrality ranking. Centrality is thus very volatile and dependent on time, reflecting a temporal utilization of the network by individuals to carry out organizational tasks. Very short bursts in activity can affect the overall network structure significantly. Network sections (and with that possibly communities) emerge and decay and are not necessarily a persistent structural element. This suggests several overlapping lifecycles of different subnets in the overall network. Three different activity patterns have been found to initiate such sections, exclusive nodes, exclusive links, or slow separation of a dense subgroup. Actor and activity oriented dynamic analysis uncovers that integrated network structures are a result of brokering activities. The analysis of actor attributes showed that managers primarily connected with distant employees across hierarchies to form the final integrated network. External events induced reaction patterns marked by fast network contraction with a sharp increase of message frequency accompanied by increasing network density, and intensified relationships among actors. From a methodological point of view, these findings demonstrate the novel research perspectives resulting from event-based dynamic network analysis. Networks are now less a static phenomenon but can be perceived as a versatile structure in constant change and motion. The main underlying methodological difference in the approach described here is disaggregating


relationships into ordered series of timed events, and explicit recognition of a variety of event and actor attributes. The resulting dynamic visualization and analysis is computationally intensive and thus requires sophisticated software support. For that, the exploratory java-based tool Commetrix has been employed. Its close integration of SNA and visualization overcomes an important weakness of other current approaches to network dynamics. During the process of visual inspection, network metrics such as degree or betweenness can be computed and exported for the visualized partial structures and their changes. In effect, researchers can now trace and measure how final structure emerges from single activities at different but connected levels of analysis. This provides an opportunity to overcome SNA’s current limitation of interpreting network structures based on a single level of analysis (cf. Monge and Contractor, 2003) despite the strong interdependency between the actor and the network level (Doreian and Stokman, 1997, p.15). With such integration of network and actor level analysis samples with unusual development become an opportunity rather than a threat. If change is detected, researchers scale their perspective from general static network properties down to patterns of change of actors and their activities. In addition to this bridging capability, the relevance of developing dynamic network visualization and analysis is substantiated by the finding that the core metric of SNA, i.e. centrality, is highly dependent on time. This motivates research into novel methods that identify important people based on their networking activities and their structural impact. Dynamic visualization has been emphasized as the primary means to induce hypotheses and theory from observed network data. The concept of visually moving through subsequent sliding time windows and removing older message events renders current zones of change an explicit object of analysis. Such a visualization mode highlights the immediate impact of recent events as


demonstrated by Enron’s peak of networking in coincidence with its bankruptcy filing. With that, researchers could extend studies of other drastic impacts on networks (e.g. catastrophes) to derive more informed prediction models for networking behavior. For the practitioner, the presented approach allows improved detection of emerging organizational communities and their developing integration with other groups (e.g. after reorganization). Active people which might not be detected by static metrics can be identified, or changes and activity levels of network areas can be analyzed to measure network reactions on external stimuli (e.g. campaigns). Future research will need to augment the exploratory study discussed in this paper to arrive at a methodology for robust scientific insights into network dynamics. Currently, important objectives include the quantification and automation of the dynamic measure brokering activity. Another challenging field of research is the design of algorithms that automatically identify the formation of online communities as (emerging) borders between sections of the network to support current visual inspection and to advance the current descriptive account of network evolution. This can enable the recognition of typical temporal interaction patterns in large networks of online communication. If future algorithms can compare masses of incremental subsequent subnets in order to identify and measure patterns or temporal relationships among patterns, stability in network structures can be advanced from a general description to a quantified measure to compare subgroup dynamics within an overall network. A final direction of our current research recognizes that the message event properties of the presented method for event-based network analysis can also store contents. Such a combination of content analysis with dynamic analysis allows new ways of studying innovation diffusion over time in online communities and can advance SNA towards Social Network Intelligence.


References

Anklam, P. (2002): Knowledge Management: The Collaboration Thread. Bulletin of the American Society for Information Science and Technology, 28(6)2002. Batagelj , V., Mrvar, A. (1998): Pajek - Program for Large Network Analysis. Connections 21(2)1998, p. 47-57. Berge, Z. L., Collins, M. P. (2000): Perceptions of e-moderators about their roles and functions in moderating electronic mailing lists. Distance Education: An International Journal, 21(2000)1, p. 81-100. Bertin, J. (1967): Semiology of Graphics. The University of Wisconsin Press, Madison. Bikson, T.K., Eveland, J.D. (1990): The interplay of workgroup structures and computer support. In Galagher, J., Kraut, R., and Egido, C. (Eds.) Intellectual teamwork Norwood, NJ: Erlbaum. p. 243-290. Borgatti, S., Everett, M.,Freeman, L.C. (1992): UCINET IV, Version 1.0, Columbia: Analytic Technologies. Brass, D.J. (1984): Being in the Right Place: A Structural Analysis of Individual Influence in an Organization. Administrative Science Quarterly, 29(1989)4, p. 518-39. Card, S.K., Mackinlay, J.D., Shneiderman, B. (1999): Information Visualization, Using Vision to Think. Morgan Kaufmann Publisher, San Francisco, 1999. Carley, K. M. (2003): Dynamic Network Analysis. In: Breiger, R., Carley, K. M., Pattison, P. (Eds.) Dynamic Social Network Modelling and Analysis: Workshop Summary and Papers (2003). Cho H.-K., Trier, M., Kim, E. (2005): The Use of Instant Messaging in Working Relationship Development: A Case Study. Journal of Computer-Mediated Communication, Volume 10, Issue 4, July 2005. Doreian, P., Stockman, F.N. (1996): The Dynamics and Evolution of Social Networks. In: Evolution of Social Networks, edited by P. Doreian and Frans N. Stokman. New York: Gordon & Breach, p. 1-17. Ellson,, J., Gansner, E.R., Koutsofios E., North, S.C., Woodhull, G.(2004): Graphviz and Dynagraph – Static and Dynamic Graph Drawing Tools. Graph drawing software. New York: Springer-Verlag, 2004. Emirbayer, M., (1997): Manifesto for a Relational Sociology. American Journal of Sociology 103(1997)2, p. 281317. Erickson T., Kellogg W.A. (2000): Social Translucence: An Approach to Designing Systems that Mesh with Social Processes. In Transactions on Computer-Human Interaction. 7(1)2000, ACM Press, New York, 2000. p. 5983. http://www.research.ibm.com/SocialComputing/Papers/st_TOCHI.htm. Accessed: 2007-01-20. Fallows, D. (2002): Email at work - few feel overwhelmed and most are pleased with the way email helps them do their jobs. PEW Internet and American Life Project. URL: http://207.21.232.103/pdfs/ PIP_Work_Email_Report.pdf. Accessed: 2007-01-28. Freeman, L.C. (1984): The impact of computer based communication on the social structure of an emerging scientific speciality. Social Networks, (6)1984, p. 201-221. Freeman, L.C. (2000): Visualizing Social Networks. Journal of Social Structure. 1(1)2000. http://www.cmu.edu/joss/content/articles/volume1/Freeman.html. Accessed: 2006-10-20. Fruchterman, T. M. J., Reingold, E. M. (1991) Graph Drawing by Force-Directed Placement. Software - Practice & Experience 21(11)1991, p. 1129-1164. Garton, L., Haythornthwaite, C., Wellman, B. (1997): Studying Online Social Networks. Journal of Computer Mediated Communication, 3(1997)1. Girgensohn, A., Lee, A., Zhang, J. (2003): Social Browsers for Visualizing Web Communities. Proceedings of the ACM WWW 2003, Budapest, Hungary, May 2003. Gloor, P., Laubacher R., Zhao, Y., Dynes, S. (2004): Temporal Visualization and Analysis of Social Networks, NAACSOS Conference, June 27 - 29, Pittsburgh PA, North American Association for Computational Social and Organizational Science, 2004. Granovetter, M.S. (1973): The Strength of Weak Ties. American Journal of Sociology, 78(1973)6, p. 1360-80. Krackhardt, D. (1991): The strength of strong ties: The importance of philos in organizations. In N. Nohira and R. Eccles (eds.): Organizations and networks: Theory and practice. Cambridge, MA: Cambridge University Press, p. 216-239.


Kraut, R., Kiesler, S., Boneva, B., Cummings, J. N., Helgeson, V., & Crawford, A. M. (2002): Internet paradox revisited. Journal of Social Issues, 58(1)2002, p. 49-74. Moody J., McFarland D., Bender-DeMoll S. (2005): Dynamic Network Visualization. American Journal of Sociology. AJS Volume 110 Number 4 (January 2005), p.1206–41 Moreno, J. L. (1934): Who Shall Survive? Nervous and Mental Disease Publishing Company, Washington, 1934. Perer, A., Shneiderman, B. (2006): Balancing Systematic and Flexible Exploration of Social Networks. IEEE Transactions on Visualization and Computer Graphics 12(5)2006. Preece, J. (2000): Online Communities: Designing Usability, Supporting Sociability. Chichester and New York: John Wiley & Sons. Shetty, J., Adibi, J. (2004): The Enron Dataset Database Schema and Brief Statistical Report. URL: http://www.isi.edu/~adibi/Enron/Enron_Dataset_Report.pdf .Accessed: 2006-06-01. Smith, M.A., Fiore, A.T. (2001): Visualization components for persistent conversations. Proceedings of the SIGCHI conference on Human factors in computing systems, 2001. Snijders, T.A.B. (2001): The Statistical Evaluation of Social Network Dynamics. In: Sociological Methodology Dynamics, edited by M. Sobel and M. Becker, Basil Blackwell, Boston and London, 2001, p. 361 -395. Trier, M. (2004): IT-Supported Monitoring and Analysis of Social Networks in Virtual Knowledge Communities. In: Proceedings Sunbelt XXIV Conference, Portoroz, Slovene 2004. Trier, M. (2005): IT-supported Visualization of Knowledge Community Structures., Proceedings of 38th IEEE Hawaii International Conference of Systems Sciences, Hawaii, USA, 2005. Wasserman, S., Faust, K. (1994): Social Network Analysis: Methods and Applications. Cambridge University Press: Cambridge, 1994. Wellman, B., Potter, S. (1997): The elements of personal community. In: Wellman, B. (ed.): Networks in the global village. Norwood, NJ, 1997. Wellman, B., Salaff, J., Dimitrova, D., Garton, L., Gulia, M., Haythornthwaite, C. (1996): Computer networks as social networks: Collaborative work, telework, and virtual community. Annual Review of Sociology, 22(1996), p. 213-238. Wellman, B., Wortley, S. (1990): Different Strokes from Different Folks: Community Ties and Social Support. American Journal of Sociology. 96(1990), p. 558-88. White, H.C., Boorman S.A., Brieger, R.L. (1976): Social Structure from Multiple Networks: I. Blockmodels of Roles and Positions. American Journal of Sociology, (81)1976, p. 730-80.


Towards Dynamic Visualization for Understanding Evolution of Digital ...

Towards Dynamic Visualization for Understanding Evolution of Digital ...

Suggest Documents

Understanding the dynamic evolution of cyclic ... - Personal.umich.edu

Towards understanding the dynamic behaviour of ... - CiteSeerX

Computation and Visualization for Understanding

Understanding digital game evolution through developmental ...

Learning in Digital Space: Dynamic Visualization of Structural ...

Visualization for Understanding of ... - Semantic Scholar

Towards visual analytics for teachers' dynamic ... - ACM Digital Library

Visualization of Dynamic Reference Graphs - Association for ...

Towards Dynamic Pricing for Digital Billboard Advertising Network in ...

Towards Understanding the Evolution of Wars in Virtual ... - CiteSeerX

Understanding the dynamic evolution of the relativistic electron slot ...

Towards Understanding the Evolution of the WWW Conference

Understanding the dynamic evolution of cyclic variability at the ...

Towards understanding the dynamic behaviour of floodplains as

Towards an understanding of dynamic demand on ground support

Chess Evolution Visualization

Interactive visualization of dynamic models

Dynamic Visualization of Battle Simulations

Understanding of Dynamic Voltage

Towards an Understanding of Liminal Imagery in the Digital ... - Core

Heapviz: Interactive Heap Visualization for Program Understanding ...

Residual Generation and Visualization for Understanding ... - CiteSeerX

A Visualization Method for Understanding Forensic Statements

Visualization for Increased Understanding and ... - Semantic Scholar