Decision Support Systems 49 (2010) 151–161
Contents lists available at ScienceDirect
Decision Support Systems j o u r n a l h o m e p a g e : w w w. e l s e v i e r. c o m / l o c a t e / d s s
Visualizing social network concepts Bin Zhu a,⁎, Stephanie Watts a, Hsinchun Chen b a b
IS Department, School of Management, Boston University, 595 Commonwealth Avenue, Boston, MA 02215, United States MIS Department, University of Arizona, Tucson, AZ 85721, United States
a r t i c l e
i n f o
Article history: Received 23 January 2008 Received in revised form 12 November 2009 Accepted 2 February 2010 Available online 10 February 2010 Keywords: Network visualization Social network analysis Information categorization Information analysis
a b s t r a c t Social network concepts are invaluable for understanding the social network phenomena, but they are difficult to comprehend without computerized visualization. However, most existing network visualization techniques provide limited support for the comprehension of network concepts. This research proposes an approach called concept visualization to facilitate the understanding of social network concepts. The paper describes an implementation of the approach. Results from a controlled laboratory experiment indicate that, compared with the benchmark system, the NetVizer system facilitated better understanding of the concepts of betweenness centrality, gatekeepers of subgroups, and structural similarity. It also supported a faster comprehension of subgroup identification. © 2010 Published by Elsevier B.V.
1. Introduction Social networks link organizations, groups and individuals throughout the world. Analysis of these networks has become an important capability in many organizations. For example, firms are using social network analysis to make hiring and transfer decisions, to optimize the flow of information among their employees, and to get the most out of talent and ideas that are embedded in the social networks of their staff [7]. Decision-making in fields such as expert assessment [25], criminal investigation [6], and community understanding [36] relies increasingly on the analysis of network information. Practitioners in many fields are demanding support for effective network analysis, and various computerized network visualization tools have been developed in response. This paper presents an approach that provides strong visualization support for social network analysis. Comprehension of the social network data is unique in that valuable information is derived from the relationships between data points, in addition to the data points themselves. Since the 1950s social network researchers have identified concepts for characterizing these networks, such as network centrality [12,33,37], betweenness [1,11,30], and structural similarity [23]. These network concepts enable an understanding of network information that would not be possible otherwise [40]. And because of the very large data sets underlying social networks, decision-makers use computerized visualizations to interpret structural components of a network such as centrality, betweenness, and structural similarity. Hence the visual interface of a system designed to support network analysis plays a crucial role in the decision-making process. ⁎ Corresponding author. E-mail addresses:
[email protected] (B. Zhu),
[email protected] (S. Watts),
[email protected] (H. Chen). 0167-9236/$ – see front matter © 2010 Published by Elsevier B.V. doi:10.1016/j.dss.2010.02.001
Most network visualization research has presented social network data using the node–link format to describe social concepts. This is based on the assumption that a clearly drawn network representation will automatically deliver the topological features described by social network concepts. This assumption is not always valid, however. Visualizations based on the node–link format do not generally present the data in terms of standard network concepts, therefore, comprehending these concepts from them can be challenging. For the most part, these visualizations only imply the network concepts that are embedded in their underlying data. To interpret them, users must integrate the various visual cues presented in the visualization and then infer the network concepts of interest. And these integration and inference processes may be limited by the working memory capacity (WMC) of the user [26]. For example, betweenness centrality is a concept that measures how powerful a network actor is in the entire network. It is calculated based on how many other actors a network actor is connected to, when these other actors are not connected to each other. Comprehending this concept from the existing node and link network visualizations is not an easy job. Fig. 1a and b below present two visualizations derived using the standard node–link format. They are included here to illustrate how the nature of the underlying data affects the appearance of the visualization when this algorithm is used. In both of these illustrations of this algorithm, network actor C has a larger value of betweenness centrality than network actor D has. However, this is more clearly apparent from the visualization in Fig. 1a than it is from the one in Fig. 1b. Using Fig. 1b on the right, a user must determine if the network actors that are connected to actor D are also connected to each other directly or indirectly, then do this for actor C also, and compare the results. Clearly this becomes a cognitively overwhelming task for large underlying data sets. The important point here is that the layout algorithm is the same for both of these figures, but the
152
B. Zhu et al. / Decision Support Systems 49 (2010) 151–161
Fig. 1. a. Situation A for the comparison of betweenness centrality. b. Situation B for the comparison of betweenness centrality.
nature of the underlying data can make it more or less difficult for the user to identify network concepts from them. With this standard algorithm, the task increases in complexity as the number of actors connected to the focal actor increases, at some point exceeding the cognitive resources of the user. More recent visualization research [15,16,29] seeks to address this problem by providing additional views (e.g., a textual list of network actors ranked by their values of betweenness centrality) of the results of social network analysis adjacent to the node–link representation. Users thus need to move between multiple views of the same network data to synthesize information in order to comprehend network concepts and their implications. This is still difficult. The study in [43] demonstrated that it is more effective and efficient to present information explicitly on one screen than to require users to integrate information from several screens. Therefore, it is important for users to be able to comprehend social network concepts directly from one single network node–link visualization. Otherwise, important patterns embedded in large network-based data sources will not be apparent to analysts. To address this problem, we propose an approach called concept visualization to facilitate the visual comprehension of several social network concepts directly from one single node–link network visualization. This approach pre-processes the underlying data such that ultimate visualizations of it reflect network concepts. The rest of this paper explains this approach in detail and presents an implementation of it using criminal network data from the Tucson, Arizona, Police Department. We then describe a laboratory experiment that was performed to demonstrate the effectiveness of this implementation and, more generally, the concept visualization approach. In the experiment we compared the concept visualization approach with a conventional network visualization method. Results indicate that, compared with its conventional counterpart, the concept visualization approach facilitates better comprehension of the network concepts of betweenness centrality, group gatekeepers, and structural similarity between network actors. This approach also enables faster identification of subgroups. The paper is organized as follows. Section 2 develops research questions by reviewing and synthesizing previous visualization studies. In Section 3 we elaborate on the proposed concept visualization approach. Section 4 presents a network visualization utilizing the proposed approach, and Section 5 describes the empirical study that was used to compare this visualization with a conventional network visualization, using real-life tasks. We conclude in Section 6 with a discussion of the implications of this work. 2. Related work and research development 2.1. Social network concepts Social network analysis (SNA) is used to investigate networks in which the nodes are actors or groups of actors that move dynamically amongst each other, such that some are closer to others. This creates
emergent patterns that reflect the underlying social dynamics. Previous SNA studies have invented various social network concepts to depict the structural patterns of a network, but it is beyond the scope of this paper to discuss them all. Interested readers are referred to [40] for a detailed description of all network concepts. In this paper we describe several important network concepts that are central to the network theory, are used by network analysts in their work, and comprise the assigned tasks of our laboratory experiment. Network centrality is a concept that describes the most important nodes or actors in a network — important actors have high centrality. Two widely investigated types of network centrality are degree centrality and betweenness centrality. Degree centrality measures the popularity of an actor based on the number of other actors that the actor is linked to [40]. An actor with high degree centrality is connected to many other actors in the network, serving as a hub in the network. Since the most effective way to communicate with a group is to connect directly to the hub of that group, this hub actor is very influential in this network. Organizations find it useful to know which of their members have high degree centrality, since these members are likely to be able to successfully diffuse information throughout their networks. It is often not desirable to transfer members with high degree centrality out of their networks, since this is likely to disrupt knowledge movement. Betweenness centrality goes beyond degree centrality to explain how powerful an actor is in a network [30,38]. It is determined on the basis of how many other actors an actor is connected to, when these other actors are not connected to each other. This is important, since being the hub of a set of interconnected actors is very different from being the hub of a set of isolated actors. In this latter case, actors that are linked to the hub person must communicate through him or her in order to communicate with each other. Removal of the hub actor in a network of isolated actors leaves the other actors without a communication channel to each other. Betweenness centrality is higher for hub actors who are linked to isolated members than for hub actors who are linked to inter-networked members. Organizations use betweenness centrality when they need to transfer a member out of a network, to determine which member would be the least deleterious to remove from that network. Because of its capacity for disrupting networks, betweenness centrality is a very important network concept. And for dense networks, it is extremely painstaking, if not impossible, to determine it from a visualization based on the conventional node–link approach. Organizations usually need to know both the degree of impact a network actor has and whom this actor has impacted. Therefore, simply providing the value of the betweenness centrality of each network actor is not sufficient. Only when the concept of betweenness centrality is presented directly on a node–link visualization can the network information support this task. Another important network analysis concept is structural similarity. Structural similarity is a characteristic of two actors when they occupy similar positions in the network [23]. For example, two software developers would be structurally equivalent if they both developed a software for the same platform, in a network characterized by multiple platforms. This is an important information since when two actors in a network are structurally equivalent, one could potentially be replaced with the other. Businesses use structural similarity information to determine the different types of roles that multiple actors in a network engage in. This information is useful for assessing whether redundancy exists, for business process redesign, and even for vacation scheduling. One factor that differentiates social networks from other networks is that social actors form subgroups. Subgroup identification is an important network analysis task, explaining which actors belong to which subgroups within the network. A subgroup is defined as a densely linked group of actors. Actors in subgroups tend to have more mutual and frequent ties with those within their subgroups than with
B. Zhu et al. / Decision Support Systems 49 (2010) 151–161
those outside their subgroups. Identification of cohesive subgroups is important because it enables analysts to determine where communication travels most quickly, where norms are likely to be closely shared, where coalitions are most likely to form, etc. [3,8,10,13]. Gatekeepers are those actors who are critically positioned between subgroups of a network. They regulate communication and interaction between the various subgroups that they connect. While it is true that gatekeepers usually are network nodes with high betweenness centrality, knowing only which nodes have high betweenness centrality is not enough. Organizations also need to know how the nodes with high betweenness centrality affect the overall network structure. Therefore, this research also seeks to aid the identification of gatekeepers for a given group. In this paper, gatekeepers are associated with the specific subgroups to which they are connected. Organizations utilize the gatekeeper concept as a means for capitalizing on communication channels that bridge multiple groups, or alternatively, to serve communication ties between two or more groups. In summary, social network researchers invent network concepts to describe the topological features of social networks. These concepts are very helpful for any number of potential interventions into ongoing networks, and organizations are increasingly taking advantage of this information [7]. However, the existing visualization support is limited, since, while an entire network may be depicted in a node–link format, values for the concepts described above are generally presented numerically in a tabular format. This makes it very difficult for analysts to discover overall patterns and relationships among the concepts. For example, each network actor will typically be assigned with a numerical value of betweenness centrality, listed in tabular format. This format makes it easy to compare the levels of betweenness centrality across actors, but does not provide information about how those network actors with high betweenness centrality affect the rest of the network. This is an important information, since a network actor can have high betweenness centrality for various reasons. For example, an actor may have high betweenness centrality because this person connects other actors to an important actor with high betweenness centrality. Or, it could be that this network actor provides the only connection between two subgroups. Such distinctions become apparent when actors with high betweenness centrality are highlighted on node–link network visualizations in the context of the rest of the network. In these scenarios, users are likely to find it more valuable to be able to identify important network actors directly from one single visual representation of network data, rather than from a combination of tables of numerical values and a node–link network representation. This is also true for combinations of node–link representations and statistical diagrams displaying the values of network concepts, because these also require users to spend cognitive resources to integrate information from different views. 2.2. Network visualization Most network visualization researchers to date have focused on solving issues concerning the clarity and scalability of automatic network visualization methods [17]. According to [17], previous studies tackle these two issues from three aspects, the first being about the layout algorithm, the second focusing on the navigation and interaction with the network graph displayed, and the third working to reduce visual complexity. Development of a network visualization system usually combines technologies from these three aspects. One example is the system described in [35]. This system visualized a telephone network by integrating a node–link representation with the navigation technology of zoom and focus + context, and the visual complexity reduction technology of hierarchical clustering. Various network visualization systems have been developed in the domains of document visualization [34], citation analysis [4], and internet
153
visualization [27,32]. A complete review of all the visualization systems developed is beyond the scope of this study. Interested users are referred to [17] and [39] for a more detailed review. This study focuses on the first aspect of network visualization, the layout method, since this is the first step in visualizing a network. Therefore we will not address navigation, interaction, and visual complexity reduction technologies. In addition, this paper concentrates only on the visualization of un-directional networks. For a review of the layout methods for tree, a special format of network structure, and for directional networks, please see [17]. The most popular layout for visualizing un-directional networks is the force-directed layout algorithm [4,9]. The spanning tree [19] is another popular layout method that extracts and presents critical paths in a network. Both types use the node–link representation to directly present connectedness among network actors. The key issue that most existing layout methods seek to address is how to create easily decipherable ways to lay out network nodes (or “actors” as nodes are referred to in social network analysis). This is not trivial for dense networks, since the presentation of many nodes simultaneously can appear as clumps and, at a minimum, analysts need to be able to see how nodes are related to one another. However, most existing layout methods do not provide the additional support necessary for comprehending more complex network concepts such as betweenness and structural similarity. As discussed above, for a user to comprehend these concepts from a standard node–link network visualization, he or she must mentally map the visualization to the relevant network concept. The complexity of this mental mapping process varies with the density of the network and the particular network concept, but for most network concepts this mental mapping process is quite complex. More recent visualization research has sought to better support the exploration of network data by adding additional views or more advanced user-interface-interaction techniques to a conventional force-directed node–link network visualization. For instance, the Social Action system [29] enables users to rank and filter network nodes based on the values of network concepts. The MatrixExplorer system [16] complements the node–link view of a social network with a matrix view of the same network. The system described in [15] augments the conventional force-directed algorithm with continuous visual abstraction and the fisheye technique. However, those visualization systems are not developed specifically for the comprehension of network concepts. Users still need to synthesize several views of a network in order to associate the values of network concepts with the network's overall structure. On the other hand, as discussed above, it is crucial for both researchers and practitioners to be able to comprehend various network concepts directly from one single node–link network visualization rather than from several views of the same social network. Therefore, we believe that an effective way to support the comprehension of network concepts is to have a node–link layout algorithm that presents important network concepts explicitly to its users. By doing so, users can perceive important network concepts directly from one node–link visualization without having to synthesize information from multiple views or visual cues, since this synthesizing process demands additional cognitive resources from the user. The study reported in [43] demonstrated that visualizing information explicitly in one window is more effective and efficient than using several windows.
2.3. Working memory capacity vs. information visualization The process by which an individual comprehends the features of a visual representation in terms of its underlying meaning takes place in the working memory of that individual [41]. Attempts to acquire highly complex information can overload working memory capacity and lead to decreased performance [18]. However, appropriately
154
B. Zhu et al. / Decision Support Systems 49 (2010) 151–161
designed information visualization can extend working memory and hence cognition [41]. Prior visualization research has identified two cognitive processes that people undertake when using visual displays to reason and solve problems [21,28]. Perceptual processes entail mapping visual cues onto their respective semantic meanings to generate hypotheses [31]. Conceptual processes generate and refine hypotheses developed from the information acquired during the perceptual process. New information is inferred and derived during conceptual processing. Because conceptual processing occurs within the working memory, it demands cognitive resources in the form of WMC. On the other hand, perceptual processing takes place when people make meaningful inferences from visualized patterns without undertaking conceptual thinking. So, because perceptual processing can occur without conceptual processing, and conceptual processing demands WMC, perceptual processing requires fewer cognitive resources. For instance, a visual inspection of a network visualization to infer whether two network actors are connected is a perceptual process. By contrast, taking a count of the links of two actors in order to determine which one has higher degree centrality requires conceptual processing and hence greater WMC. One important way that visualizations can reduce demand for conceptual processing is by pre-processing the data in such a way that conceptual tasks are processed perceptually instead [42]. Therefore, good network visualizations support the complex network concept comprehension in part by converting conceptual processing into perceptual processing. They filter out excess demand for conceptual processing by presenting abstract concepts in concrete ways that visually reflect the nature of the underlying concepts. For example, the above-mentioned conceptual processing necessary to determine degree centrality by counting links could be converted into perceptual processing by presenting actors with high degree centrality as physically larger than ones with low degree centrality. 2.4. Research question development Visualization researchers are beginning to acknowledge the importance of supporting visual comprehension of network concepts and have developed several network visualization systems as a result. The CiteSpace system [4] creates links across citation networks to highlight actors with large values of betweenness centrality. These then indicate transitions across knowledge domains over time. Brandes et al. [2] propose a layout algorithm that uses location to display the centrality of network actors, helping users understand how the connections of an actor affect his centrality. This understanding is important because the centrality of an actor represents his structural status in the network [2]. At the same time, most existing network visualizations are designed for the general exploration of the social network data. However, effective network analysis invokes the concepts discussed above for understanding which actors play a central role in the network, which ones act as gatekeepers, which are structurally similar to others, etc. Network visualizations that do not support the comprehension of these concepts are generally less helpful for network analysis tasks than those that do provide this support. In particular, most existing visualization systems do not support the comprehension of these concepts. Instead, understanding these network concepts requires users to synthesize visual features [9,14] or information from multiple views [15,16,29] into a mental representation of the desired concept's meaning. This cognitive integration process requires sometimes a considerable conceptual processing, and hence may be limited by the user's working memory capacity [26]. Therefore the research question this paper seeks to address is how to design a network visualization that takes advantage of human perceptual processing capacity to facilitate visual comprehension of social network concepts.
3. Concept visualization We thus propose an approach called concept visualization to make network concepts more readily discernable to users. Network concepts such as betweenness centrality, subgroups, and structural similarity can be calculated from social network data using social network analysis algorithms. The results of such “pre-analyses” calculations can then be presented in such a way as to visually depict these concepts. When this is done, particular network concepts are made highly explicit and much more comprehensible than if they were presented implicitly. It is the incorporation of the results of “preanalysis” of social network data into a node–link visualization that distinguishes the concept visualization approach from other approaches, in which the node–link representation is presented by itself or in conjunction with textual/graphic display of the results of calculating social network concepts. To validate the effectiveness and feasibility of concept visualization, we propose a network visualization algorithm that makes several network concepts explicit, including degree centrality, betweenness centrality, subgroups, and gatekeepers. The rest of this section provides a detailed description of the proposed algorithm. The proposed algorithm consists of a circular version of the Self Organizing Map (SOM). As an information categorization and visualization tool, SOM was first proposed by Kohonen, who based his depictions of neural networks on the associative neural properties of the brain [20]. The network consists of an input layer and an output layer. The number of the input nodes is equal to the number of attributes associated with the input. After all of the input is processed, the result is a spatial representation of the input data, organized into clusters. While various clustering algorithms could be used to partition networks, we selected the SOM for clustering because of the inherent spatial representation of the output, in which the spatial proximity between actors indicates their structural similarity. We then added additional visual cues to display degree centrality, betweenness centrality, subgroups, gatekeepers between subgroups, and structural similarity. These visual cues were designed to enhance the perceptual processing of the information they reflect. The proposed layout algorithm consists of three steps. 1. A circular SOM is applied to categorize actors and lay them out along a circle. The input to the SOM is a list of all network actors, with each actor being represented by a vector consisting of this focal actor and any other actors which this one is directly connected to. As displayed in Fig. 2, the location of an output node on an SOM is decided by its coordinates (r, θ). Table 1 provides a detailed description of the SOM algorithm. Once this has been completed, each network actor is assigned with coordinates in the format of (r, θ), clustered into subgroups. Each subgroup will have its own color and all actors within the subgroup will have the same r coordinate. 2. Next, the betweenness centrality of each actor is calculated, and the r coordinate of each actor is adjusted to reflect its betweenness centrality, as follows. Actors high in betweenness centrality are moved toward the center of the circle and those lower in betweenness centrality are moved away from the center. These moves do not change the actors' θ coordinates. 3. Lastly, the degree centrality of each actor is calculated and the size of each actor is changed in proportion to his or her degree centrality. Thus the proposed algorithm entails various pre-processing steps to visually present aspects of the data that would not be apparent using the traditional layout algorithm. For example, it includes steps to calculate betweenness centrality and degree centrality, to lay out actors based on their structural equivalence, and to draw the boundaries of subgroups to assist identification of the groups and their gatekeepers. These social network concepts would be minimally
B. Zhu et al. / Decision Support Systems 49 (2010) 151–161
155
Below we summarize what this expert said about the helpfulness of each of the network concepts that the NetVizer system seeks to deliver.
Fig. 2. The output nodes of a circular SOM.
apparent on a visualization that had not used the concept visualization approach. 4. Development of the NetVizer system We developed a prototype visualization system that we call NetVizer by applying the network layout algorithm proposed above. The prototype utilizes the network data of street gang groups provided by the Arizona Artificial Intelligence Lab. As described in [6], the network was constructed from 272 Tucson Police Department incident summaries of about 164 crimes committed from 1985 through May 2002, using the concept space approach developed by [5]. In this network, two people are considered to have a connection with each other if both of their names appear in the same case record. A detailed description of the construction of this network is provided in [6]. The selection of the criminal analysis domain as the testbed for the proposed algorithm was based not only on the availability of this data, but also on the importance of network analysis concepts to the users of this database. This importance was evidenced during a phone interview with a domain expert/detective who has been using this network data in the Tucson Police Department for twenty years. According to this domain expert, a detective typically explores the network data when he or she returns from a crime scene on the street. The network data is used to identify whom to bring in for questioning.
Table 1 Description of SOM algorithm. 1. Initialization: assign each output node displayed in Fig. 2 a weight vector of N random numbers. N equals to the number of actors in the network. 2. Present each network actor in order as the input: represent each actor i by a vector of N features (vi0, vi1,… … vin), vij = 0 when there is no direct link between actor i and actor j. Otherwise vij = m, where m is the value of the link between actors i and j. Present actor i to the system. 3. Compute distances to all nodes: compute distance dj between the actor and each output node j. 4. Select winning node j* and update weights to node j* and neighbors: select winning node j* that has the shortest distance to the actor. Update weights for node j* and its neighbors to reduce their distances (between the actor and output nodes). 5. Repeat 2–4 several times 6. Assign each network actor to an output node: assign a network actor to an output node that has the shortest distance to the network actor 7. Form subgroups: submit unit input vectors of single actors to the trained network and assign the winning node the name of input actor. Neighboring nodes that contain the same name then form a subgroup.
• Degree centrality: “Knowing a person [has] many links to others is certainly valuable to detectives. But we also have to be careful here, sometimes a person's having so many links simply indicates that this person has been contacted by police so many times. He might be the least intelligent person in the group.” Therefore, additional information about the nature of the degree centrality of an individual is very helpful to these detectives. • Betweenness centrality: “It is valuable to know… who are the people that other people need to connect to in order to connect to the rest of the network.” • Subgroups identification: “If a person is involved in a criminal activity… and if I know this person belongs to a gang group, this certainly helps me narrow down my suspect list.” • Gatekeepers between subgroups: “If there is a victim in green group, and we know that blue group and green groups are rivals, we certainly want to bring in people who know both green and blue groups for questions.” • Structural equivalence: “If two guys have the same friends in the database, it is most likely that they are best friends on the street. They are not connected in the database simply because they have not been caught together in the same crime yet.” Therefore when one person becomes the suspect and cannot be found, someone structurally equivalent to him or her is the best candidate to question about the suspect. The NetVizer system was developed using JAVA on the Microsoft Windows platform. Fig. 3 displays the layout of the 290 members in this database using the proposed algorithm. On the NetVizer interface, each member is represented by one square. Subgroups were determined on the basis of their relational strength to other actors in the network, and each group has its own color. The size of each square reflects the number of connections this person has; the larger the size, the more other members this person is connected to. The location of each square indicates its betweenness centrality — those closer to the center have higher betweenness centrality. Members tend to connect to members closer to the center to connect with others in the same network. A user can click on an actor's square to check all the direct connections that actor has, since this highlights these links. She can also double-click on an actor to display the connections of all other members of the subgroup. One advantage of using the circular SOM as the clustering tool is that actors in the same group are likely to be connected to each other or be connected to the same group of other actors. As displayed in Fig. 3, members “Barcely, Pau” and “Hilary, Antono” are not only connected to each other, but also have similar connections with other members. The dashed lines in Fig. 3 indicate the boundaries between subgroups identified by using the SOM. The NetVizer interface also supports the identification of gatekeepers: members with connections to members in other subgroups. 5. System evaluation In order to evaluate the concept visualization approach, we designed a pilot study and tested it with five PhD students, based on the criminal analysis task domain. Results of this pilot were positive and suggested the efficacy of this approach. Next we enrolled 60 participants recruited from a subject pool at a large northeastern university. Each participant performed several network analysis tasks and was paid $10 for their time and given a chance to win $100. Participants were asked to extract as much information related to street crime as possible from a social network database. Subjects were randomly divided into two groups. Each group worked on the same series of network comprehension tasks, but with one of two types of
156
B. Zhu et al. / Decision Support Systems 49 (2010) 151–161
Fig. 3. The interface of the NetVizer system.
network representations: either the NetVizer or the benchmark visualization, described below. We selected students as participants because of the difficulty of finding a sufficient number of domain experts to obtain statistically significant results. In addition, comprehending network concepts from a network visualization can be accomplished by a student after minimal training, since the comprehension of network concepts from a visualization does not require extensive domain knowledge. For instance, a centrally located network actor can be in a network of software developers, biochemists, or film production workers. An analyst can identify the centrality of this actor from a visual representation without having expertise in any of these industries. He or she needs only to synthesize visual cues from the visualization and find the network actor with the highest centrality value, a process that does not require that domain knowledge be retrieved from long-term memory (although it may subsequently be applied during the decision-making process). The results of the study conducted in [44] also demonstrated the domain independence of a network concept comprehension task. 5.1. Selection of the benchmark system Fig. 4 below displays the visual representation that we developed by applying the layout algorithm of [14] over the same criminal network data that we implemented the NetVizer. The reason we chose a force-directed algorithm is because the force-directed method and its variants are the most commonly used algorithms for visualizing network information. And most existing social network analysis softwares such as NetMinner and Uncinet utilize a forcedirected algorithm for network visualization. Other existing network layout algorithms are different from the force-directed algorithms solely in the way they achieve their visual aesthetic. They are fundamentally similar because they use the “node and line” format to represent connectiveness among network actors. While this layout
algorithm does a good job of presenting all the network actors and their relationships, unlike NetVizer it does not describe important network concepts such as structural similarity and betweenness centrality. In order to provide users with a reasonable ability to determine these values from the benchmark system, this standard force-directed algorithm was modified slightly. This was necessary in order to make it more comparable with the NetVizer system. Specifically, the size of each actor was made to be proportional to the number of links this actor had. We refer to this as the benchmark system in the remainder of this paper. Another reason for the selection of the benchmark system is that we are interested in evaluating the layout algorithm proposed in this paper. The development of a network visualization system usually consists of two parts: a layout algorithm that projects network data into a visual representation; and a combination of user-interfaceinteraction methods to facilitate the exploration of the visual representation developed. Therefore, it is appropriate to compare the proposed visualization algorithm with a widely used layout algorithm when both visual representations have similar userinterface-interaction. For this reason we did not choose the systems described in [15,17,29], or any other existing network visualization systems, as our benchmark system. We believe that the additional views and more advanced user-interaction methods used by those systems to complement a node–link representation would add a layer of complexity to the evaluation and that we need to understand the basic layout issue first. 5.2. Task design The series of tasks that the participants performed during the evaluation experiment was designed based on the objectives of the NetVizer system and on information gleaned during our interview with the domain expert. We were limited to a time constraint of 1 h by
B. Zhu et al. / Decision Support Systems 49 (2010) 151–161
157
Fig. 4. The network visualization using the conventional approach.
the rules of the student subject pool. We selected realistic tasks, conducted by actual users of this criminal network data, to test the efficacy of these two visualizations for presenting network concepts. The two objective measures recorded were effectiveness and efficiency, two measures that have been widely used in previous studies to evaluate visualization tools [22,24]. We measured effectiveness by task accuracy and efficiency by time-to-task. Subjective measures collected in this study included perceived ease of use of the visualization, perceived usefulness, and perceived learnability. Table 2 summarizes the tasks used in the experiment. The design of these tasks was based both on the usefulness of the task to users of the network data used in this paper and also on the functionalities of the benchmark and NetVizer systems. While the benchmark system portrays a visually decipherable network that displays direct connections among network actors, the NetVizer system focuses on the presentation of network concepts. We are interested in the extent to which the NetVizer system can present the network concepts more effectively and efficiently than the benchmark system. We are also assessing how well the NetVizer system performs on the tasks that the
Table 2 The summary of experimental tasks. Network concepts
Tasks
Degree centrality
1. Count links of a given actor 2. Compare number of links of two actors 3. Find actors with the maximum number of links 4. Who knows the most number of friends of a given actor? 5. If a given actor belongs to a group? 6. Identify gatekeepers of a given group 7. Which one of the two actors should be removed to disconnect more people in the network?
Structural equivalence Group identification Interaction between groups Betweenness centrality
benchmark system is designed for. We discuss these tasks in detail below. 5.2.1. Degree centrality Three types of tasks were selected to evaluate the delivery of degree centrality: a count of the number of links an actor has; a comparison of two actors to determine which has more links to other criminals in the network; and identification of the actor with the maximum number of links. According to our domain expert, a good visualization in this domain needs to support the comparison of the number of links an actor has, as well as the ability to find the actor with the maximum number of links relative to others. 5.2.2. Structural similarity As indicated by the domain expert, two actors who are structurally equivalent to each other in the database are most likely to be best friends on the street. However, during our pilot study we realized that it is very difficult to explain the concept of structural similarity to students who do not have a background in network analysis. Therefore, we simplified the design of this task slightly to one of finding the actor in the network who is connected to the highest number of connections of a given suspect. 5.2.3. Subgroups and gatekeepers Knowing that a person belongs to a particular subgroup can help detectives narrow down the list of suspects. In addition, the ability to identify gatekeepers who connect the various subgroups is crucial for finding appropriate suspects. Therefore two of the experimental tasks involved the concepts of subgroup and gatekeeper. The first consisted of determining if a particular suspect was a member of a subgroup, and the second involved locating all the gatekeepers of a given subgroup.
158
B. Zhu et al. / Decision Support Systems 49 (2010) 151–161
5.2.4. Betweenness centrality Betweenness centrality is a valuable concept to criminal analysts, but knowing the absolute value of the betweenness centrality of an actor in the network is not. Detectives are more interested in the relative betweenness centrality of two or more network actors. We therefore asked participants to compare the betweenness centrality of two network actors in order to evaluate the comprehension of this network concept. 5.3. Experiment process and results Sixty participants were randomly divided into two groups with one group working with the benchmark system and the other working with the NetVizer system. The same experimental procedure and tasks were used for both groups. The experiment began with a short introduction, followed by a questionnaire to gather demographic information. After that, the participants were given standardized instructions on how to use the particular visual interface he or she had been given and encouraged to ask questions. Participants then went through a training session before starting to work on the experimental tasks. During the training session, the participants were provided with questions similar to the actual questions they were about to be asked, along with answers. A detailed written description of each of the network concepts was presented in the training session. Participants took as long as they needed to familiarize themselves with the network concepts and the visualization system they would use, but none of them took longer than 5 min to do this. Following the training session the actual experimental session was administered by one of two experimenters, and these experimenters were randomly assigned to participants and visualization types. During the experiment, participants were asked to finish the seven tasks in sequence and were not allowed to look ahead. This was important to control for the effects of prior tasks on subsequent ones. The experimenter recorded the time each participant took to complete each experimental task. At the end of the experiment, we asked each participant to complete a survey in order to collect the data on perceived ease of use, perceived usefulness, and perceived learnability. Participants were asked to think aloud and this verbal protocol was tape recorded. Tables 3, 4 and 5 summarize the results for effectiveness, efficiency, and the perceptual measures of ease of use, usefulness, and learnability. We used independent t-tests in SPSS to calculate the results. In these tables, time-to-task is measured in seconds, and significant results are italicized. Overall we found that, compared with the benchmark system, the NetVizer system facilitated better understanding of the concepts of betweenness centrality, gatekeepers of subgroups, and structural similarity. It also supported a faster comprehension of subgroup identification. The rest of this section presents a detailed explanation of these findings. 5.3.1. Degree centrality On the interface of both the NetVizer and benchmark systems, degree centrality was explicitly indicated by the size of each actor. Hence, participants could simply visually inspect the size of the actors'
nodes to accomplish the tasks of comparing actors according to their number of links (task 2) and finding the actor(s) with the maximum number of links (task 3). However, node size was not very helpful when it came to counting links (task 1). To do this, participants of both visualizations needed to locate and click on the actor of interest and count the links manually. Since the layout algorithm used by the benchmark system was designed to address the aesthetic issues of network visualization, the benchmark interface appeared to have less link clutter than the NetVizer interface. As a result, we found the benchmark system to be significantly more efficient than the NetVizer system in supporting the counting task because the NetVizer system is not designed to support this task. 5.3.2. Structural similarity To accomplish task 4, participants looked for the actor who was connected to the greatest number of actors that was connected to the particular actor specified in the question. To accomplish this, participants usually found and clicked on the assigned actor of interest and inspected the connections of this actor. They then clicked around, trying to find the actor who connected to most of the connections of the actor of interest. The effectiveness of this task was calculated as follows: Effectiveness ¼ number of mutual connections between the actor selected by the participant andthe actor of interest; divided by the number of connections of the actor of interest:
This score indicated the extent to which the participants' answers were able to identify the actor with the most connections to the assigned actor. To find the answer to this question, participants needed to compare the connections of the assigned actor with those of the various other actors they clicked on as they attempted to answer this question. By observing this technique and examining the verbal protocol, it appeared that participants in both groups used similar inference processes to accomplish this task, a process that involved considerable trial and error. However, because NetVizer grouped actors with similar connections, NetVizer users did not have to click on as many actors as the benchmark users did. On the other hand, benchmark users needed to memorize more network actors than the NetVizer users. Consequently NetVizer users had significantly higher effectiveness than the benchmark users for this task (q = 0.008, t = 2.750). We did not find a significant difference in time-to-task between the two groups of users. 5.3.3. Subgroups and gatekeepers During the instruction session, the participants were told that a subgroup in the network is a group of people who are connected to each other. As they undertook task 5, we observed that benchmark users determined if a network actor belonged to a subgroup based on a visual assessment of the links among network actors near the actor specified in the task question. Benchmark users appeared to be able to accomplish this task without much difficulty. On the other hand, NetVizer provides more visual cues than the benchmark regarding subgroups, since each subgroup has its own unique color and dashed
Table 3 Experiment results for effectiveness. Network concepts
Tasks
Benchmark
NetVizer
t value
q
Degree centrality
Count links of a given actor Compare number of links of two actors Find actors with the maximum number of links Who knows the most number of friends of a given actor? If a given actor belongs to a group? Identify gatekeepers of a given group Which one of the two actors should be removed to disconnect more people in the network?
1.00 3.72 1.00 0.369 1.933 0.825 0.9000
0.767 3.7 1.00 0.500 1.933 1.467 1.733
− 0.997 − 1.070
0.325 0.292
2.750 0.000 4.311 8.398
0.008 1.000 0.008 0.00
Structural equivalence Group identification Interaction between groups Betweenness centrality
B. Zhu et al. / Decision Support Systems 49 (2010) 151–161
159
Table 4 Experiment results for time-to-task (measured in seconds). Network concepts
Tasks
Benchmark
NetVizer
t value
q
Degree centrality
Count links of a given actor Compare number of links of two actors Find actors with the maximum number of links Who knows the most number of friends of a given actor? If a given actor belongs to a group? Identify gatekeepers of a given group Which one of the two actors should be removed to disconnect more people in the network?
60.53 142.5 24.467 87.8 44.87 78.4 126.142
79 155.5 20.433 75.53 29.97 91.6 105.263
1.991 − 0.149 − 0.586 − 1.252 − 3.089 1.008 − 1.107
0.051 0.882 0.560 0.216 0.03 0.318 0.275
Structural equivalence Group identification Interaction between groups Betweenness centrality
lines indicating its boundaries. These additional visual cues enabled NetVizer users to accomplish subgroup identification with a brief glance. Results of the experiment indicate that both types of visualizations were comparable in effectiveness, but the NetVizer system enabled significantly more efficient subgroup identification. For task 6, participants had been instructed that the gatekeeper of a subgroup is the actor(s) who connects the subgroup to other subgroups outside it. Deciding whether or not an actor belongs to a subgroup was relatively easy for those using the benchmark interface, but identifying all the gatekeepers of this subgroup was not, particularly because subgroups do not have an apparent boundary on this visualization. We observed that participants using the benchmark worked to try to memorize subgroup members and then click on them to see who they were linked to and whether these members belonged to a different subgroup. Since this information is only implied by this interface, benchmark users appeared to have more difficulty accomplishing this task than NetVizer users. This is not surprising since the NetVizer interface indicates group boundaries explicitly. NetVizer participants were able to double-click on any member of a given subgroup to bring up all the links of these group members, making the identification of gatekeepers a visual perception task. NetVizer users had a statistically better performance in this task but did not accomplish the task faster than benchmark users. We observed that of the thirty subjects working with the benchmark system, twenty-three of them stopped after finding one gatekeeper, despite the fact that multiple gatekeepers existed for this group, perhaps because of the extensive effort that this task entailed using the benchmark interface. Yet only seven NetVizer subjects stopped after finding one gatekeeper for the same task. It appeared that it took NetVizer users a longer time to accomplish this task because they tended to identify multiple gatekeepers and took the time to record these multiple answers on the answer sheet. 5.3.4. Betweenness centrality Task 7 involved comparing the betweenness centrality of two prespecified actors to determine which one would have the greatest impact on the network if removed from it. We observed that this concept was the most difficult one for benchmark users to comprehend, especially when the betweenness centrality of a network actor was not consistent with the participant's visual perception. Referring back to Fig. 1a and b as examples, one can see that on both figures, network actor C has a larger value of betweenness centrality than network actor D. Benchmark users had less difficulty obtaining the right answer in a situation similar to
Table 5 Experiment results for user satisfaction.
Perceived ease of use Perceived usefulness Perceived learnability
Benchmark
NetVizer
t value
q
4.53 5.49 3.89
5.05 5.70 1.90
1.816 0.651 − 1.539
0.075 0.518 0.129
Fig. 1a, when the larger value of betweenness centrality corresponded to a larger number of links. We observed that when one participant tried to investigate if there were links among the connections of the actor specified in the task, he appeared to be overwhelmed by the cognitive load required and ultimately made the decision using a simple visual inspection. And on the benchmark visualization, visual depictions of network linkages can sometimes suggest the wrong answer. As a result, twenty-six of thirty participants using the benchmark system arrived at the correct answer in Fig. 1a situation, but for the situation resembling Fig. 1b, in which the visual inspection suggested the wrong answer, only one of these thirty arrived at the correct answer. On the other hand, NetVizer users had a much easier time accomplishing this task because the value of the betweenness centrality of an actor is indicated by its location on this interface. This comparison task, which required much cognitive processing using the benchmark interface, was a perceptual task on the NetVizer interface. All thirty NetVizer users obtained the right answer for the situation similar to Fig. 1a, and twenty-two of them arrived at right answer for the situation similar to Fig. 1b. Overall we found NetVizer to be significantly more effective for supporting the task of comparing betweenness centrality than the benchmark system. For the subjective measures, NetVizer users reported significantly higher levels of perceived ease of use than the benchmark system users, while there were no significant differences in perceived usefulness and perceived learnability between the two systems.
6. Discussion and conclusions The formal social network concepts developed by previous network analysis researchers [1,11,12,23,30,33,37] have become an important means for understanding the social network phenomena. Formal social network concepts are aggregations of underlying network data, which makes them difficult for users to interpret without explicit visual support [2]. This is particularly true in the case of large data sets. However, the existing network visualization techniques provide limited support to analysts working to understand these concepts. Without additional support, analysts must cognitively infer network concepts from their data sets or synthesize information obtained from multiple views of the same network, which is challenging and error prone. In order to provide greater levels of support, we have proposed here a concept visualization approach that explicitly presents the network concepts of degree centrality, betweenness centrality, subgroup identification, gatekeepers, and structural equivalence. This same concept visualization approach could be applied for the delivery of other network concepts, such as closeness and prestige [40]. Due to the plethora of potential social network concepts to visualize and the limitations of a computer screen for presenting images, designers cannot visualize every network concept in a single image, even using the concept visualization approach. For this reason, visualization developers
160
B. Zhu et al. / Decision Support Systems 49 (2010) 151–161
need to select the most appropriate network concepts for the relevant domain and create images that reflect these critical network concepts. In this paper we have described a network visualization system called NetVizer that we used to demonstrate the concept visualization approach. NetVizer was developed by applying this approach to criminal network data extracted from the criminal database of the Tucson Police Department. The availability of this real-world network data was an important reason for the selection of this domain, as was information gleaned during our interview with a domain expert who stressed the importance of certain network concepts for supporting criminal analysis. While social network analysis is new to many organizations, it is increasingly being adopted by companies ranging from law firms to drug companies and financial services organizations [31]. Other organizational domains that rely on social network analysis are research and development, marketing, and knowledge management. We could certainly have demonstrated the viability of the concept visualization approach using one of these other domains and are confident that the benefits of the approach over standard forcedirected layout algorithms would have been equally as apparent. This paper presents a controlled lab experiment to validate the performance of the NetVizer system developed. Using one of the variants of the force-directed network layout algorithm [14] as the benchmark, findings indicate that the NetVizer users performed better in tasks of betweenness centrality comparison, gatekeeper identification, and finding structurally similar network actors, when compared with their benchmark counterparts. We also found that NetVizer users performed faster in tasks of identifying subgroups. Therefore we conclude that, compared with a conventional network visualization approach that focuses on the aesthetic aspects of network visualization, the concept visualization approach proposed here facilitates better a comprehension of the network concepts it selects to deliver. In addition, the NetVizer system was perceived by the participants to be much easier to use than the benchmark system for the tasks they performed. Importantly, results indicate that the concept visualization approach could not replace conventional network visualization. We found the two types of visualizations to be complimentary to each other, as follows. Conventional force-directed algorithms focus on the aesthetic aspects of the visualization, reducing clutter among links and thus making the task of counting the links of a node easier compared to the NetVizer system. On the other hand, the NetVizer system was designed to support the comprehension of more complex network concepts such as betweenness centrality, structural similarity, and gatekeepers. The node–link metaphor is the most appropriate representation of network information when the task does not entail comprehending complex concepts derived from aggregated network data. But, when the comprehension of network concepts is an important goal of the visualization, the inherent limits of conventional visualization approaches become apparent. Conventional approaches are constrained by presenting the underlying data directly, without transforming it to reflect network analysis concepts of interest. Under the concept visualization approach, the layout algorithm selected becomes a critical component of the analysis toolkit. Developers need to determine which kinds of visual cues best reflect the relationships in the underlying data without distorting it. They need to better understand how visual components can be used to enhance this understanding, and also how they can hamper it. This paper demonstrates the efficacy of the concept visualization approach using a laboratory experiment. Evaluation of these results supports the usefulness of both the approach and the artifact for understanding social network analysis concepts. Future research includes applying the concept visualization approach to social network data sets from other domains and understanding the impact of the concept visualization approach at both the individual and organizational levels. Another interesting avenue of future research
would be to augment the proposed network layout algorithm with the means used by existing network visualization systems [15,16,29] to improve the effectiveness of a conventional node–link network representation. Additional studies should be conducted to understand how such an improved NetVizer system would affect the users' network analysis tasks. Overall, the approach and the resulting artifacts can enable social network analysts to capitalize on the large quantities of data they face. This work advances our understanding of how to best provide visualization support to social network analysts. Acknowledgement The authors would like to thank detective Tim Peterson from the Tucson Police Department for taking time to interview with us and providing valuable domain knowledge about the social network data used in this research. This research is sponsored by the Boston University Institute for Leading in a Dynamic Economy (BUILDE). References [1] J.M. Anthonisse, The Rush in a Graph, Mathematische Centrum (Amsterdam, 1971). [2] U. Brandes, J. Raab, D. Wagner, Exploratory network visualization: simultaneous display of actor status and connections, Journal of Social Structure 2 (4) (2001). [3] R.S. Burt, Network items and the general social survey, Social Networks 6 (1984) 293–340. [4] C. Chen, CiteSpace II: detecting and visualizing emerging trends and transient patterns in scientific literature, Journal of the American Society for Information Science and Technology 57 (3) (2005) 359–377. [5] H. Chen, K.J. Lynch, Automatic construction of networks of concepts characterizing document database, IEEE Transactions on Systems, Man and Cybernetics 22 (5) (1992) 885–902. [6] H. Chen, W. Chung, J. Xu, G. Wang, M. Chau, Y. Qin, Crime data mining: a general framework and some examples, IEEE Computer 37 (4) (2004) 50–56. [7] R. Cross, A. Parker, The hidden power of social networks: understanding how work really gets done in organizations, Harvard Business School Publishing, Boston, MA, 2004. [8] R. Collins, Theoretical Sociology, Harcourt Brace Jovanovich, New York, 1988. [9] P. Eades, A heuristic for graph drawing, Congressus Numerantium 42 (1984) 149–209. [10] B. Erickson, The Relational Basis of Attitudes, in: B. Wellman, S.D. Berkowitz (Eds.), Social Structures: Network Approach, Cambridge University Press, Cambridge, England, 1988, pp. 99–121. [11] L.C. Freeman, A set of measures of centrality based on betweenness, Sociometry 40 (1977) 35–41. [12] L.C. Freeman, Centrality in social networks: I. Conceptual clarification, Social Networks (1) (1979) 215–239. [13] N.E. Friedkin, Structural cohesion and equivalence explanations of social homogeneity, Sociological Methods and Research 12 (1984) 235–261. [14] T.M.J. Fruchterman, E.M. Reingold, Graph drawing by force-directed placement, Software Practice and Experience 21 (1991) 1129–1192. [15] F.V. Ham, J.J.V. Wijk, Interactive visualization of small world graphs, Proceedings of IEEE Symposium on Information Visualization 2004, Austin, Texas, USA, 2004, pp. 199–206. [16] N. Henry, J.D. Fekete, MatrixExplorer: a dual-representation system to explore social networks, IEEE Transactions on Visualization and Computer Graphics 12 (5) (2006) 677–684. [17] I. Herman, G. Melanscon, M.S. Marshall, Graph visualization in information visualization: a survey, IEEE Transactions on Visualization and Computer Graphics 6 (1) (2000) 24. [18] J.A. Jacko, K.G. Ward, Toward Establishing a Link between Psychomotor Task Complexity and Human Information Processing, 19th International Conference on Computers and Industrial Engineering, 31, 1996, pp. 533–536. [19] D. Jungnickel, Graphics, Networks and Algorithms, Springer Verlag, 1999. [20] T. Kohonen, Self-Organizing Maps, Springer-Verlag, Berlin, Heidelberg, 1995. [21] J.H. Larkin, H.A. Simon, Why a diagram is (sometimes) worth ten thousand words, Cognitive Science (1987) 64–100. [22] M.D. Lee, M.A. Butavicius, R.E. Reilly, Visualizations of binary data: a comparative evaluation, International Journal on Human–Computer Studies 59 (2003) 569–602. [23] F. Lorrain, H.C. White, Structural equivalence of individuals in social networks, Journal of Mathematical Sociology 1 (1971) 49–80. [24] R. Mazza, V. Dimitrova, CourseVis: a graphical student monitoring tool for supporting instructors in web-based distance courses, International Journal of Human Computer Studies 65 (2) (2007) 125–139. [25] D.W. McDonald, M.S. Ackerman, Expertise Recommender: a Flexible Recommendation System and Architecture, CSCW’ 2000, ACM Press, 2000. [26] G.A. Miller, The magic number seven plus or minus two: some limits on our capacity for processing information, Psychological Review (67) (1956) 191–257.
B. Zhu et al. / Decision Support Systems 49 (2010) 151–161 [27] T. Munzner, P. Burchard, Visualizing the structure of the world wide web in 3D hyperbolic space, Proceedings of VRML'95, San Diego, CA, December 1995, pp. 33–38. [28] N.H. Narayanan, M. Suwa, H. Motoda, Hypothesizing behaviors from device diagrams, in: J. Glasgow, N.H. Narayanan, B. Chandrasekaran (Eds.), Diagrammatic Reasoning: Cognitive and Computational Perspectives, AAAI Press, Menlo Park, CA, 1995, pp. 501–534. [29] A. Perer, B. Shneiderman, Balancing systematic and flexible exploration of social networks, IEEE Transactions on Visualization and Computer Graphics 12 (5) (2006) 693–700. [30] F.R. Pitts, The medieval river trade network of Russia revisited, Social Networks (1) (1979) 285–292. [31] S. Patton, Who knows whom and who knows what? CIO Magazine Online (June 15, 2005) http://www.cio.com/archive/061505/km.html. [32] P. Pirolli, S.K. Card, M.M.V.D. Wege, Visual information foraging in a focus + context visualization, Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, Seattle, Washington, United States, 2001, pp. 506–513. [33] C.H. Proctor, C.P. Loomis, Analysis of sociometric data, in: M. Jahoda, M. Deutsch, S.W. Cook (Eds.), Research Methods in Social Relations, Dryden Press, New York, 1951, pp. 561–586. [34] S. Risch, D.B. Rex, S.T. Downson, T.B. Walters, R.A. May, B.D. Moon, The STARTLIGHT information visualization system, Proceedings of IEEE Conference of Information Visualization, 1997, pp. 42–49. [35] D. Schaffer, Z. Zuo, S. Greenberg, L. Bartram, J. Dill, S. Dubs, M. Roseman, Navigating hierarchically clustered networks through Fisheye and Full-zoom methods, ACM Transactions on Human Computer Interaction 3 (2) (1996) 162–188. [36] W. Sack, Conversation map: a content-based usenet newsgroup browser, The Proceedings of the International Conference on Intelligent User Interfaces, New Orleans, LA, Association for Computing Machinery, January 2000. [37] M.E. Shaw, Group structure and the behavior of individuals in small groups, Journal of Psychology (28) (1954) 139–149. [38] A. Shimbel, Structural parameters of communication networks, Bulletin of Mathematical Biophysics (15) (1953) 501–507. [39] O. Turetken, R. Sharda, Visualization of web spaces: state of the art and future directions, Data Base 38 (3) (2007) 51–81. [40] S. Wasserman, K. Faust, Social network analysis: methods and applications, Cambridge University Press, Cambridge, United Kingdom, 1999. [41] C. Ware, Information Visualization Perception for Design, Morgan Kaufmann Publishers, Inc, San Francisco, CA, 2000. [42] J. Zhang, The nature of external representations in problem solving, Cognitive Science 21 (2) (1997) 179–217.
161
[43] B. Zhu, H. Chen, Using 3D interfaces to facilitate the spatial knowledge retrieval: a Geo-referenced knowledge repository system, Decision Support Systems 40 (2) (2005) 167–182. [44] B. Zhu and S. Watts, Visualization of Network Information: The Impact of Task and Working Memory Capacity Differences, Information Systems Research, forthcoming (2010). Dr. Bin Zhu received her PhD degree in Management Information Systems from the University of Arizona in 2002. She is an assistant professor in the Information Systems department at Boston University. Her current research interests include business intelligence, information analysis, social network, human–computer interaction, information visualization, computer-mediated communication, and knowledge management systems. She has been a lead author for papers that have appeared in Information Systems Research, Decision Support Systems, Journal of the American Society for Information Science and Technology, IEEE Transaction on Image Processing, and D-Lib Magazine. Her research also received an IBM faculty award in 2003. Dr. Stephanie Watts is an associate professor of Information Systems at the Boston University School of Management. She was previously on the faculty of the Weatherhead School at Case Western Reserve University. Her research focuses on the various roles that information technology plays in organizations, with a focus on mediated knowledge sharing and the role of cognition in it. She has published academic papers in such journals as Information Systems Research, Organization Science, Journal of Strategic Information Systems, Information and Management, and Journal of Computer-Mediated Communication. Dr. Hsinchun Chen is the McClelland professor of Management Information Systems and Andersen professor of MIS at the University of Arizona, where he is the director of the Artificial Intelligence Lab and the director of the Hoffman E-Commerce Lab. He received his PhD degree in Information Systems from the New York University in 1989. His articles have appeared in Communications of ACM, ACM Transactions on Information Systems, IEEE Computer, Journal of the American Society for Information Science and Technology, Decision Support Systems, and many other journals. Professor Chen has received grant awards from NSF, DARPA, NASA, NIH, NIJ, NLM, NCSA, HP, SAP, 3COM, and AT&T. He serves on the editorial board of Decision Support Systems, Journal of American Society for Information Science and Technology, and ACM Transactions on Information System.
