j ourna l homepage: www.e lsev ie r .com/ locate /dss
Bicentric diagrams: Design and applications of a graph-based relationalset visualization technique
Hyunwoo Park a,⁎,1, Rahul C. Basole b,1
a Tennenbaum Institute, Georgia Institute of Technology, 85 Fifth Street NW, Atlanta, GA 30308, USAb School of Interactive Computing & Tennenbaum Institute, Georgia Institute of Technology, 85 Fifth Street NW, Atlanta, GA 30308, USA
Article history:Received 20 February 2015Received in revised form 22 December 2015Accepted 2 February 2016Available online 12 February 2016
In an era where data on social, economic, and physical networks are proliferating at a rapid pace, the ability tounderstand the underlying complex structural connections, discover prominent entities, and identify clustersis becoming increasingly important. It is also well-established that interactive visualizations can amplifyhuman cognition and augment decision making. Motivated by a practical need articulated by corporate decisionmakers and limitations of existing visual representations, this research presents our journey in designing andimplementing bicentric diagrams, a novel graph-based set visualization technique. A bicentric diagram enablessimultaneous identification of sets, set relationships, and set member reach in integrated egonetworks of twofocal entities. Our technique builds on the well-established sociological theory of tie strength to visually groupand position nodes. We illustrate the broad applicability of bicentric diagrams with examples from four diversesample domains: university collaboration, technology co-occurrence, health app purchases, and interfirmalliance networks. We assess the value of our technique using an expert-based evaluation approach. Thepaper concludes with implications and a discussion of opportunities for implementation in real-world decisionsupport settings.
Keywords:Bicentric diagramSet visualizationNetwork visualizationStrength of ties
1. Introduction
Humans are visual thinkers [1]. We use visualizations to solve prob-lems, explore opportunities, communicate ideas, recognize patterns,and understand complexity [2]. The foundational challenge in informa-tion visualization research is to transform andmap raw data into appro-priate symbolic and spatial visual representations and couple it witheffective and intuitive interaction techniques [3]. It also requires adeep understanding on the user and the task/decision context inwhich the visualization will be used. Visualization is thus both an artand a science. When done correctly, however, visualizations cansignificantly amplify human cognition [4].
In an era where data on social, economic, and physical networks areproduced at a rapid pace, the ability to visually understand underlyingcomplex structural connections, discover prominent entities, andidentify clusters is becoming increasingly important [5,6]. It is thus notsurprising to see the research community call for a greater integrationof visual network analytics into decision support systems (DSS) [7].The study of visual decision support is not entirely new. Prior workhas incorporated visualization into DSS and evaluated its efficacy in
variousfields such as personal finance [8], healthcare [9,10], and nation-al power system [11]. More recently, studies have shown the particularvalue of visual analytics for complex business ecosystem decisionmaking [12,13].
Despite these notable attempts, several important challengesremain. Most network analysis studies still use relatively commonvisualization layout algorithms, such as circular or force-directed, torepresent complex networks [14]. These layouts, however, often resultin cluttered representations that are hard to analyze and interpret.The cluttered representation is particularly problematic for largenetworks because the number of edges increases much faster than thenumber of nodes. The DSS and information systems literature hasrecognized and confirmed the importance of representation inmanage-rial problem solving and decision making [15,16]. Existing studies,however, predominantly compare existing classic representationssuch as table, list, and line chart [17,18]. Novel visual representationsfor specific types of decision-making problems are rarely developed,despite their importance to problem solving. Motivated by this litera-ture gap and the practical need of corporate executives and investorsto map global innovation networks of competing firms, and our discov-ery of the significant shortcomings of existing visualization techniquesto help address questions related to sets (i.e., firm clusters), membersof sets, relationships between sets, and reach (i.e., distance from focalnode or relationship tier), we designed a visualization layout calledthe bicentric diagram.
65H. Park, R.C. Basole / Decision Support Systems 84 (2016) 64–77
Overlaying two concentric layouts in a clever way, a bicentricdiagram fuses ideas from graph and set visualization to depict sets,relationship between sets, and the reach of setmembers in the integrat-ed egonetworks of two focal entities [19]. The concentric layout isarguably the best layout for depicting an egonetwork [20]. It placesthe focal node at the center of a set of concentric circles. All neighboringnodes are placed around the circumference of one of the circles basedon how many steps they are away from the focal node at the center.Although the concentric layout works well when displaying a singlefocal node's egonetwork, it falls short when it comes to comparing ormerging two focal nodes' egonetworks. A great deal of ambiguity ariseswhen one attempts to portray two egonetworks in the same canvassuch as where to place a node that has direct connection to both focalnodes. The bicentric diagramwe propose in this paper provides specificvisual encodings for simultaneous display of two egonetworks whenthe interest lies in identifying shared and exclusive componentsbetween the two networks.
During our early design discourse [21], it became increasinglyevident that analysis of such set-related issues are not just pervasivein business network contexts, but also in sociology, economics,healthcare, politics, and engineering. In social networks, for instance,a researcher may be interested in identifying the overlap and exclusionof friends and acquaintances; in the healthcare context [22], patientsmay be interested in understanding the co-occurrence structure ofdrug side effects.
The contribution of our study is multifold. To the best of ourknowledge, there are no existing techniques that enable simultaneousidentification of sets, set relationships, and set member reach inintegrated egonetworks of two focal entities. By developing a uniquegraph-based set visualization technique with broad application appeal,we contribute to information visualization and human–computer inter-action in decision making. Moreover, building on the well-establishedtheory of tie strength [23], we classify shared sets in the bicentric diagraminto four types: (i) strong/strong, (ii) weak/weak, (iii) strong/weak orweak/strong, and (iv) exclusive-strong or exclusive-weak. Thisclassification enables us to formally characterize the different typesof questions one may address with a bicentric diagram. Our imple-mentation of bicentric diagrams also contributes to the decisionmaking and DSS literature. Considering decision making as an iterativemulti-step process [24,25], our study focuses on improving the perspec-tive development and synthesis step. Moreover, many comtemporaryrelational and network decisions involve identification of specific keynodes (e.g. structural holes) and exclusive assets in the overall network[26]. Our novel representation immediately supports such tasks. TheDSS community also calls for further work on web-based DSS [25].Our system is deployed on the web and accessible for a wide range ofusers, enabling greater democratization of decision making. By consid-ering an appropriate layout and visual encodings, we also contributeto the aesthetic aspect of DSS design [24].
We illustrate the broad applicability of bicentric diagrams with ex-amples from four diverse sample domains: university collaboration,technology co-occurrence, health app purchases, and global strategicalliances network. Lastly, we assess an interactive prototype version ofthe bicentric diagram using a value-driven expert evaluation approachand discuss opportunities for implementation in real-world settings.Overall, respondents give above-average scores to all surveyed aspectsof our system. In particular, we find that survey respondents withhigher visualization literacy levels tend to give higher scores to ourimplementation of bicentric diagrams in performance areas of time re-duction, insight generation, and essence extraction. Even respondentswith relatively lower visualization expertise give satisfactory scores inmost aspects, indicating that experience and expertise are not a strictprerequisite to appreciate the value of our system.
The remainder of the paper is organized as follows. Section 2provides an overview of the key related literature from informationvisualization, business analytics, and network analysis. Section 3
explains the detailed machineries of the bicentric diagram.Section 4 showcases various domains where the bicentric diagramcan be a useful tool. Section 5 presents our user study that confirmsvalue of the bicentric diagrams in four key performance areas.Section 6 concludes the paper.
2. Related work
In the DSS literature, the decision-making process has beenmodeledas an iterative process of problem recognition, perspective develop-ment, perspective synthesis, actions, and results [24]. The role of DSSis to facilitate human decision-making parts by providing structuredviews about information pertaining to the decision-making problemspace [25]. As more network-centric data emerges, many decision-making contexts are now also network related. Given that the choiceof representation of data influences decision-making outcomes [15,16],DSS must take this shift in decision-making contexts into account andnovel representations are potentially needed.
Data visualization is a well-established method to support decisionmaking in a wide variety of domains. For instance, a visualization ap-proach was adopted to display geographic distribution and marketpower structure of the U.S. electric power system [11] and a visualiza-tion model of adjacency data illustrated college selection decisions[27]. Recent studies show how visualization affects and supports userdecision making in the context of financial services [8] and healthcaredelivery [9], respectively. Thanks to technological advancements inpersonal computers, the role of interactivity in visualizations has beenhighlighted in the literature. An interactive version of self-organizingmap visualizations is shown to reduce the cognitive loadduring Internetbrowsing tasks [28] and a suite of interactive visualizations was devel-oped and used to explore auction databases [29]. Departing from thedecision-making context, a visualization approach is also shown to beuseful in the strategic planning process as well [30].
Pending the nature of the data and use context,many different formsof visual representation and interactions exist [31,2]. Multivariatedatasets are often visualized using parallel coordinates, starplots, orglyph-based techniques [32]. Hierarchical data are often depictedusing a variety of space-filling techniques, such as treemaps, sunbursts,and circle packing [33,34].
One context in which data visualizations are becoming increasinglypervasive is in the analysis of networks. Virtually any aspect of oureconomic, technical, and social contexts can be described usingnetworks [35]. Node-link diagrams and adjacency matrices are thepreferred visualization method for network and graph data [14]. Suchvisualizations of network evolution can complement and even enhancethe associated statistical analysis [36].
Network visualizations have been used to help identify global supplynetwork risks [20], examine criminal networks [37], and understand theevolution of digital communication networks [38].
Following the recognition that many contexts of decision makinginvolve analysis of networks, recent DSS studies have paid attention tothe network perspective [39,20]. However, it is the inherent complexityof network data that often hinders the conversion of data into decision-making insights; this is amplified when the scope and size of theunderlying data are large and wide [40,13]. Visualizations can helpusers overcome this issue by providing a graphical representation ofthe underlying network data structure [41]. When coupled with inter-action techniques, network visualizations can facilitate the knowledgeconstruction and sharing process and provide confirmatory and newinsights [42,43]. Visual analytic systems that afford interactive visualexploration and analysis are thus invaluable for supporting decision-making processes.
Real-world networks are often characterized by sets, groups, andclusters. Visualizing sets and set-typed data has also been a topic ofsubstantive interest in the information visualization community [44].Sets are common data structures in information visualization and are
66 H. Park, R.C. Basole / Decision Support Systems 84 (2016) 64–77
characterized as a group of elements sharing a given property, inherent-ly present in the data, or generated by an algorithm. Terms such as set,group, and cluster are often used interchangeably.
Set visualizations are motivated by tasks associated with elements(e.g. finding elements belonging to a set, finding sets containing aspecific element), sets and set relations (e.g. finding total number ofsets in the data, analyzing inclusion/exclusion/intersection relations,finding set similarities), and element attributes (e.g. finding theattribute value of a specific element, finding the distribution of anattribute in a certain set/subset).
To address these tasks, many different techniques have beendeveloped including Venn and Euler diagrams, variants of Eulerdiagrams (e.g. fan diagrams and region-, line-, and glyph-basedoverlays), node link representations, matrix representations, aswell as aggregation-based techniques. An excellent state-of-the-artoverview to set visualizations by [45] covers different techniquessampled in Fig. 1.
Set visualizations commonly consist of spatially arranged clusters ofset members [46]. In non-fragmented datasets, these clusters are oftenoverlayed with a layer of convex hulls [47]; in contexts where datasetsare fragmented, concave boundaries are often used [48]. In contextswhere the connectedness between set members and sets matter,relations are generallymodeled as edges of a bipartite graph. Prominentexamples include parallel lists [49], anchored maps [50], Circos [51],and RadialSets [45]. Anchored maps use a circular layout to visualizebipartite graphs by placing set nodes around a circle and elementnodes depending on their set membership. Elements belonging exclu-sively to a set are placed as a cluster of nodes outside the circle corre-sponding to the respective set node. Nodes that are shared between
Fig. 1. Relevant set visualizations are shown in close-up views. From left-to-right: (1) A multdifferent sets. Position and color-encoding shows the overlaps; (2) untangled Euler diagramselements; (3) bubble sets [48] are a visualization technique for dynamically creating bubble™(4) Kelp diagrams [73] depict set relations over points, where elements have predefined posiaccount; (5) PivotPaths [74] exposes facets and relations as interactive visual paths. Facets andrelationships; (6) RadialSets [45] uses a scalable frequency- and graph-based representation to
multiple sets are placed inside the circle relative to their set nodes.Circos also uses a circular layout for set nodes and uses the thicknessof stripes connecting the nodes to encode the number of elementsfalling into each category. A similar set encoding approach is used byRadialSets. In this approach, however, links originate from the samelocation, emphasizing that elements in a particular set overlap canalso belong to other sets.
3. Bicentric diagram
3.1. Design goals
Our main objective when designing the bicentric diagram was tosupport simple readability tasks in egonetworks of two focal entitiesnot available through other set visualization techniques. Anegonetwork is a graph of nodes having direct or indirect connectionsto the focal node. By definition, the concentric layout is well suitedwhen visualizing an egonetwork. However, it does not allow compari-son between two egonetworks by itself. Thus, we focused on five designgoals:
• Allow users to efficiently identify shared direct connections (G1).• Allow users to efficiently identify the set membership of each dataelement (G2).
• Allow users to efficiently identify shared indirect connections (G3).• Allow users to efficiently identify the distribution of data elementswithin a set and by level of reach (G4).
• Provide users an intuitive visual metaphor to identify sets and dataelements (G5).
i-set Venn diagram [71] shows all possible logical relations between a finite collection of[72] provide a simplified, compact representation and improve readability of sets and set
borders around sets of visual objects, while avoiding inclusion of items not in the set;tions. The graphs are designed to take aesthetic quality, efficiency, and effectiveness intoitems are spatially arranged and connected by subtle curves indicating facet–item–facetenable quick identification and analysis of overlapping sets.
67H. Park, R.C. Basole / Decision Support Systems 84 (2016) 64–77
A direct connection refers to nodes that are directly connected(i.e. neighbors) to one of the focal nodes. An indirect connectionrefers to nodes that are not directly connected to the focal nodesbut are connected to neighbors of the focal nodes. In graph-theoreticterms, these connection types are captured by the concept of reach.It is also referred to as degree of separation, popularized by thesmall world experiment by [52]. Direct and indirect connectionsthus are first and second level of reach away from one of the focalnodes, respectively.
3.2. Design motivation
The idea of the bicentric diagram was stimulated by discussions wehad with corporate decision makers who were interested in visuallycomparing their global innovation networks to their competitors [21].We considered many different visualization approaches. While force-directed graph layout techniques allowed an understanding of theoverall network structure and pockets of collaboration [53], it failed toclearly identify shared direct and indirect partners. Force-directedlayouts also did not allow determining what network tier partnersbelonged to, limiting the understanding how partners were distributed.Similarly, while existing set visualization approaches were able todepict groups, clusters, and relationships, they generally fell short inconsidering relationships between members and sets as well as differ-entiating the results by reach from the focal entities. Concentricnetwork layouts were the most effective in depicting a single firm'segonetwork but did not allow a simultaneous examination of twofocal firms of interest [20].
Fig. 2 showcases the evolution of our design considerations. Giventhe intuitive understanding of concentric network visualizations,we recognized that cleverly overlaying two concentric layouts (seeFig. 3a) to create a bicentric diagram (Fig. 3b) would provide possiblevalue. We presented our idea to these decision makers in a series ofbrainstorming sessions, involving multiple paper sketches and whiteboard drawings, and found that it addressed exactly those questionsinvolving sets, relationships, and tiers.
3.3. Layout
The bicentric diagram layout builds on the concentric networkvisualization and uses ideas from set visualization [45]—specificallynode-link and overlay techniques—to provide an effective representa-tion of two focal nodes as well as their shared direct and indirectnodes while organized by tier. A conceptual representation of this lay-out is shown in Fig. 3c. The bicentric layout provides a graph-based setvisualization of two focal nodes (A and B) positioned at a distance dapart. Two concentric circles with radii d/2 and d are drawn aroundeach focal node, where other non-focal nodes are placed. Each circlerepresents a tier from the focal node. The inner circle contains the
Fig. 2. Our design consideration evolved ov
focal node's 1st-tier nodes (1-step away neighbors). Similarly, theouter circle contains 2nd-tier nodes (2-steps away). Thus, a node thatis connected/related with a 1st-tier node but not with a focal nodewould be 2 steps away from the focal node in the outer circle.
The shaded arcs and intersecting points of the concentric circlesrepresent areas for positioning nodes corresponding to the variouscombinations of set membership. Nodes directly connected to bothfocal nodes (i.e. are 1 step away) are placed in a cluster at the centerof the focal dyad (area 1). Nodes that have direct or indirect connectionto only one of the focal nodes are positioned on the arcs (i.e., semicircle)on the side of that focal node. Those connected directly to only focalnode A (or B) are placed on the inner arcs—area A1 (or area B1), whileconnections to only those 1st-tier nodes (i.e., indirectly connectedto the focal nodes) are placed on the outer arcs—areas A2 and B2.Nodes one step apart from Node A and two steps away from Node Bare placed around area 2a and 2b. We differentiate nodes placed atthe top (area 2a) and the bottom (area 2b) by whether they belong tothe main component.
The main component refers to the largest connected component ofan undirected graph and the largest weakly connected component of adirected graph, where all nodes within that component can reacheach other through some paths [54]. The notion of connected compo-nents has been known in graph theory from early days [55]. For exam-ple, suppose that we have a graph consisting of ten nodes. Five, three,and two nodes are connected to each other, respectively. In graphtheory, each group is called a connected component. Among threegroups, the group that contains the most nodes is themain component.Distinguishing the main component is of great interest in networkanalysis research [56,57]. In our system, visually differentiating themain component from other components allows us to further disentan-gle the sub-tier activities and reduce the number of edge crossing andlong edges in the visualization. This differentiation is also in accordancewith the visual design principle because it minimizes the number ofvertically crossing edges which clutters the overall visualization.Nodes two steps apart from both Node A and B are placed at areas 4aand 4b. The same top (area 4a) and bottom (area 4b) differentiationby membership to the main cluster applies.
Once nodes are placed, we adopt various visual encodings for nodesand edges to enhance user understanding of the data. An edge betweentwo nodes represents a relationship/connection between those twonodes. For instance, in product networks, one product i is connectedto another one j if it was recommended or co-purchased. The thicknessof an edge corresponds to the number of times the nodes were connect-ed/related in the dataset. The node diameter is sized proportionally tothe square root of the number of node occurrences. The larger thenode, the more common the node is in the dataset. The node colorcorresponds to the category the node belongs to. The edge color isgray on default and switches into corresponding color of the hoveredover node. Within each arc, nodes are grouped by category (counter-
er a series of brainstorming sessions.
A B
(a)
A B
(b)
(c)
A B
(d)
1
2a
2b
3a
3b
4a
4b
A1A2 B1 B2
N/A Weak Strong
B
A
Str
ong
Wea
kN
/A
I
II
II
III
IV
IV
Fig. 3. (a, b) The bicentric layout builds on the concentric network visualization to provide an efficient representation of two focal nodes as well as their shared direct and indirectnodes while organized by tier. (c) This schematic provides an overview of node placements in the bicentric layout. (d) Classification of relationship types identifies four combinationsof tie strength.
68 H. Park, R.C. Basole / Decision Support Systems 84 (2016) 64–77
)clockwise for (left or) right focal nodes, respectively. To further reducevisual clutter and improve readability and aesthetics, we apply a“no overlap” and node jittering rule. Such a rule adds randomperturbation to the position of all nodes until every node is notoverlapped with another.
3.4. Decision-making tasks using bicentric diagrams
The bicentric diagram is particularly suited to address questionsrelated to sets of strong, weak, and exclusive ties of two focal entities.Strong ties refer to direct connections of a focal entity (i.e. friends);weak ties refer to indirectly connected entities (i.e. friends-of-friendsor acquaintances) [23]. Consider the 3 ×3 tie type matrix shown inFig. 3d. Direct connections shared by two focal entities refer tostrong/strong ties (Cell I). These represent the common ties of twofocal entities. In the case of friendship networks, this would be ashared friend. In product recommendation networks, this wouldrepresent a product that was co-purchased by both. At the oppositeend are indirect connections to both focal entities. We refer tothese as weak/weak ties (Cell III). Cells II represent a direct connec-tion to one focal entity, and an indirect connection to the other(and vice versa). We refer to these as strong/weak and weak/strongties, respectively. Cells IV represent those direct and indirect tiesexclusive to each focal node, respectively. We refer to these asstrong-exclusive and weak-exclusive ties.
The intersecting and exclusive node placement areas in our bicentricdiagram thus correspond to one of these tie types. The visual separationenables rapid identification of these tie types and examination of a widerange of questions for making decisions in the network context. Forinstance, a corporate strategy manager may search the next potentialalliance partner that has been exclusively collaborating with thecompetitor firm. Another more everyday-level example is a consumerconsidering to buy another video game. One has a few favoritegames and wants to discover what other people purchase togetherwith the focal favorite games. The bicentric diagrams help quicklyidentify and generate candidate choice set for this type of decision-making problems.
3.5. UI implementation
The user interface (UI) for our prototype consists of twomain areas:the selection pane and the visualization pane (see Fig. 4). The selectionpane contains multiple drop-down lists that contain the data contextlist (Panel A) and the node lists for the chosen context (Panel B).We in-troduce the available contexts inmore detail in the applications section.The number next to each list entry indicates its occurrence frequency inthe dataset (Panel B). For this example, “Videoconferencing Software”was mentioned 35 times. Since the number of nodes can be over the
thousands, it is an important design consideration to make surethat the lists are scrollable and searchable (Panel B). Once a user hasselected two focal nodes of interest, the “Compare” button updatesthe visualization screen with the corresponding bicentric diagram(Panel G). The UI provides several interactive tools to manipulate thevisualization. Users can collapse and expand the visualization windowas well as zoom-and-pan the bicentric diagram (Panel C). To maximizereadability, users can show/hide edges, layout gridcircles, node labels,and the color category legend (Panel D). Users can also locate specificnodes in the visualization using a quick search box function (Panel F).Hovering over a node provides detailed information in a label(Panel E). It also highlights its directly connected edges and nodes(while fading others). Clicking on a node will fix the highlights forconvenient path tracking. Clicking on the background cancels thecurrent selection. Keyboard shortcuts for these UI functions are alsoprovided. Lastly, we setup five tutorial pages to assist first-timebicentric diagram users.
The interactive front-end UI is developed and implemented usingD3.js [58] and Bootstrap. The back-end is written in Python andNetworkX [59]. The back-end Python script returns initial nodepositions based on node attributes and network metrics (e.g., connectedcomponents) upon request from the front-end. The front-end imple-ments interactivity and adjusts node positions to minimize overlaps.The application was hosted and made available to users on the Herokuplatform at http://bicentric.herokuapp.com. For purposes of this study,we did not develop a public data import tool, mainly due to concerns ofdata security, authentication, and intellectual property rights. However,we envision to implement one for a future release.
3.6. Beta Testing
Prior to releasing our interactive UI to participants for evaluation,weconducted a rigorous iterative beta testwith four experienced visualiza-tion users. The aimwas to get in-depth feedback on the overall usabilityand usefulness of the tool, the steps of our tutorial, and the specificwording of our evaluation survey. The feedbackwas generally favorablein both visual representation and interactivity of our implementation.Specific suggestions included elaboration of the tutorial with concreteexamples, resolution of ambiguous domain-specific terms (e.g. “tier”versus “reach”) and graph-theoretic concepts (e.g. “what is a maincomponent?”), and learning tasks to get familiar with the UI ratherthan a free exploration. We revised the UI, tutorial, and instructionsaccordingly.
4. Applications
The bicentric diagram layout can be used in numerous differentcontexts. We focus our discussion to four diverse specific application
Fig. 4. A screenshot of the system UI is presented. The current context is set as “Technology Co-Occurrence” (Panel A). Two technologies (videoconferencing software and operationssupport systems) in the technology co-occurrence dataset are selected (Panel B). Various controls help users zoom and pan the diagram (Panel C). The bicentric diagram shows thetechnologies that co-occur with these two technologies; nodes are positioned at the appropriate set location (Panel G). The color legend (Panel D) and help guidelines are turned on(Panel F) to aid users with reading the diagram. The user hovers over a second-tier technology (i.e. virtual private networks). Its connections are highlighted and the label showsdetails (Panel E).
69H. Park, R.C. Basole / Decision Support Systems 84 (2016) 64–77
areas: university collaborations, technology co-mentions, health apppurchases, and interfirm alliance network.
We chose thefirst three use cases to demonstrate theuniversal valueand applicability of bicentric diagrams beyond the business setting thatmotivated this paper. We conclude with an example explicitly in thebusiness domain.
Each use case represents a different network data context. Theuniversity collaborations network, for instance, captures an institutionalsocial network aggregating individual collaborations. Other potentialexamples that could have been used include Facebook friendshipor Twitter follower networks. The technology keyword dataset is aco-occurrence network context. Other examples include the tagco-occurrence network in StackOverflow questions, daily deals inGroupon [60], or movie recommendations in Netflix. The healthapp network represents an example of a commerce network. Other
Table 1Possible application areas of bicentric diagrams.
Domain Example
Business and economics Supply networkAlliancesVenture capitalPatent citationGlobal trade
Commerce Product comparisonReviewsMusic
Healthcare and life sciences Process mapsSystems biology
Social networks Friendship networkFollower networkPolitical network
Manufacturing and Product designEngineering and computer science Software design
examples of such a network may be related products in Amazon orrestaurant reviews in Yelp.
Lastly, the interfirm alliance network is also a type of organizationalcollaboration network similar to the university collaborations network.However, it is an important example for the bicentric diagram as itlinks back to our original motivation of designing this diagram: theintelligence needs of corporate decision makers.
We must stress that the application areas introduced in this sectionshould merely be interpreted as possibilities rather than limitingconstraints. Bicentric diagrams can prove valuable to any contexts inwhich comparing two entities' relationship structure matters. Table 1highlights some of the possible application areas. We identify sixpotential application domains. Beside business decision making andsocial network application domains, product and software design arenotable domains that can benefit from applying the bicentric diagram.
Nodes Edges
Supplier, customers Exchange of products and moneyFirms Formed allianceInvestors Co-invested experiencePatents Citation directionCountries Flow of import and exportProducts SimilarityUsers Co-review behaviorSongs, albums, artists Co-purchase or similarityActivities, medications Flow of resourceMolecules Chemical reactionPeople FriendshipTwitter IDs, Facebook Follower-following relationshipPoliticians Co-sponsor of a billComponents Mechanical/electrical linkObjects, classes Inheritance
70 H. Park, R.C. Basole / Decision Support Systems 84 (2016) 64–77
Even these highly technical design areas are increasingly adoptingvarious novel visualization approaches to enhance alignment amongstakeholders in different levels [61].
4.1. University collaborations
4.1.1. Context and dataInstitutional collaboration between universities and corporations is
a common phenomenon in scholarly research. Those institutionalcollaborations are macro patterns aggregated from the underlyingfiner-grained individual researchers' collaboration network. Locatingthe right set of collaborators is a great career concern for scholars. If atarget institution is too far from reach, knowing a close collaborator ofthe target may help connect to the target eventually. Identifying poten-tial collaborators and a path to reach them is equally important at theorganization level. Such organizational search problem in the networkcontext has been characterized through the sociological lens [23] andstudied extensively in organization science [62].
We source our data from the ACM Digital Library that contains a listof co-authoring institutions as well as their profiles. After collecting theprofiles of top 1000 institutions by publication count, we construct anetwork based on the co-authorship among them as of September 15,2014. Thus, a node represents an institution and an edge between twonodes indicates that they co-authored a publication. Nodes are sizedby the institution's total publications count and colored by geographicregion. All nodes are categorized into five regions—Europe, NorthAmerica, Asia, Australia, and Others. Europe has the greatest numberof institutions in the dataset while North American institutions aremuch larger in terms of per-institution publications. Edge thicknesscorresponds to the total number of co-authored publications betweenthe two institutions.
The original dataset tracks a full list of collaborators for each institu-tion. Even a pair of institutions that co-author only a single publication isthus captured as an institutional collaboration. We therefore applied amix of relative and absolute thresholds to filter out collaboration rela-tionships with minimal significance. We set the relative threshold to20% of themaximumedgeweight for each node. For instance, if an insti-tution published 1000 articles with its most frequent collaborator, alledges starting from this institution having less than 200 co-authoredpublications would be dropped. The absolute threshold was set to five
SeoulNational
University
UniversityUrbana-Ch
UniversitySouthernCaliforni
Europe
North America
Asia
Australia
Others
Fig. 5. A screenshot of institutional collaboration network between Seoul National University (and color legend are added externally using separate annotation software after capturing the s
publications so as to drop any collaborations resulting in less than fiveco-authored articles.
4.1.2. ExampleWe incorporate the aforementioned dataset into our interactive
bicentric diagram generator. As an illustrative example, considerthe bicentric diagram of the two focal universities—Seoul NationalUniversity (SNU) and Carnegie Mellon University (CMU)—shown inFig. 5. SNU and CMU directly share (Type I ties) only two collaborators,University of Southern California (USC) and University of Illinois atUrbana–Champaign (UIUC) (see Fig. 6 for the magnified version). Onestriking observation of the bicentric diagram is SNU's virtually exclusivecollaboration (Type IV ties) with institutions from Asia (depictedby nodes in green). Another immediate observation is that CMU's directnetwork consists of institutions primarily from North America; its two-step network, however, is substantially larger than that of SNU andregionally diverse. CMU's Type II ties form a large interconnected clusterwith Tsinghua University, resulting to be the only Asian connection.Following [23], the institutions in this cluster have weak (or indirect)ties with SNU but strong (or direct) ties with CMU. In addition to UIUCand USC, these institutions thus play a potential bridging role betweenCMU and SNU because from SNU's perspective it would be easier toestablish connection with its weak ties than with the Israeli universitiesshown in the zoomed-in version in Fig. 6. Another confirmatoryevidence that deserves mentioning is the connection intensity betweenCMU and the University of Pittsburgh. The collaboration is thicker thanany other direct connections that CMU has suggesting that the twouniversities would probably collaborate more frequently given thegeographic proximity.
4.2. Technology co-mentions in press releases
4.2.1. Context and dataMarket researchers and analysts are frequently interested in under-
standing how different technologies relate to each other. Changes andtrends in technological landscape influence strategic planning andpositioning for companies from start-ups to large corporations alike,and thus they are essential to the decision-making process for corporatetechnology management. Although press releases and news articlescontain a wealth of information on emerging technologies, search
CarnegieMellon
University
of Illinoisampaign
of
a
Tsinghua University
green focal node) and Carnegie Mellon University (orange focal node) is presented. Labelscreenshot.
Fig. 6. Zoomed-in snapshot of the collaboration network between Seoul National University and Carnegie Mellon University. User has selected CMU as the focus. Label indicates detailedinformation; edge thickness reveals the extent of institutional collaboration.
71H. Park, R.C. Basole / Decision Support Systems 84 (2016) 64–77
results can be overwhelming and their linear presentation format doesnot always provide necessary systemic insights. We use the data fromNorthern Light SinglePoint (http://www.northernlight.com) portal tobuild a co-occurrence network of technologies mentioned in the pressover the past year from September 10, 2014. We identify 834 uniquetechnologies (e.g., “Smartphones,” “Cloud Computing and Storage,”“Predictive Analytics”) and classify them into 10 categories adaptedfrom [12]. For each technology, we collect co-occurrence data withother technologies. Technologies thus represent nodes in the network.An edge between two nodes indicates that they co-occurred in adocument. Nodes are sized by the frequency of occurrences andcolored by category. Edge thickness corresponds to the number ofco-occurrences between two technologies.
Note that we model the technology co-mention network as adirected graph in a similar way for the institutional collaboration dataset used in the previous section. For instance, suppose we have afrequent/popular technology (e.g., “Smartphones”) and a relativelyinfrequent one (e.g., “Smart Fabric”). It may be the case that wheneverSmart Fabric is mentioned, it is also mentioned with Smartphones.However, the reverse statement may not necessarily be true. The maintechnologies that are co-mentioned with Smartphones also are likelyto be broad and popular technologies such as “Android OperatingSystems” or “iOS.” In sum, Smart Fabric can mainly co-occur withSmartphones, but from Smartphones' perspective co-occurrences withSmart Fabric can be ignored.
4.2.2. ExampleConsider a start-up company working on tangible media and
information visualization. Leveraging its own expertise, it considers toexpand into an adjacent technology space that can potentially provideadditional traction to the company. The decision thus relates to whattechnology path it should possibly explore. Fig. 7 shows the zoomed-
in and labels-on bicentric diagram of two technologies—wearables anddata visualization. We first notice that these two technologies share arange of different technology categories (Type I ties), includingplatforms (pink), mobile (olive green), and networking and communi-cations (orange). Interestingly, Type II ties of wearable technologiesare far less (Bluetooth, HomeAutomation, GPS, and PC)while the appli-cation category (blue) dominates Type II ties for Data Visualization. Byselecting the two focal technologies, then seeingwhich ones are directlyand indirectly related, the decision maker gains an overview of theoverall related technology space and can then make more informeddecision which technology path to pursue.
4.3. Health app purchases on Amazon.com
4.3.1. Context and dataProduct recommendation is a pervasive feature on e-commerce
sites. A multitude of individual consumers make purchase decisionsrelying on such recommendation data. However, most sites onlyprovide long scrollable lists that are hard to navigate. List formatdoes not afford exploration beyond the immediate related appsnot to mention comparison. A visual comparison of related orco-purchased products beyond the immediately related apps willbe helpful for consumers to generate candidate apps for potentialpurchase. To demonstrate the effectiveness of bicentric diagrams inthis type of consumer decision-making context, we collect data ofall Android health apps from Amazon.com. For each app, Amazonprovides a list of (up to 100) related apps. Related apps are deter-mined based on consumers' co-purchasing behavior. If consumersfrequently purchased both apps A and B, for instance, we considerthese two apps related to each other. Each health app is categorizedby Amazon into one of six categories (Diet, Medical, Exercise,Meditation, Pregnancy, and Sleep). For illustrative purposes, we
Fig. 7. Technology co-occurrence in press releases related to wearable technology (turquoise focal node) and data visualization (red focal node) is shown.
72 H. Park, R.C. Basole / Decision Support Systems 84 (2016) 64–77
filtered out those apps with less than five reviews to focus on theones that have some established user base. This resulted in a finalset of 711 health apps.
We construct the co-purchase network of health apps. Healthapps (nodes) are connected when co-purchased (edge). Nodes aresized by the total number of reviews and colored by the health app
Fat Burning Yoga
Diet
Medical
Exercise
Meditation
Pregnancy
Sleep
Fig. 8. Bicentric diagram of fat-burning yoga and diab
category defined by Amazon. Edge thickness proxies howmany peopleco-purchased both apps by inversing the order of appearance in thelist of related apps. Suppose an app A has two related apps: B and C.B appears in the first page of the list of related apps, while C appearsin the last page in the list. We interpret B to be co-purchased morefrequently with A than C.
Diabetic Recipes
etic recipes apps with labels turned off is shown.
73H. Park, R.C. Basole / Decision Support Systems 84 (2016) 64–77
4.3.2. ExampleConsider the example of an individual who loves yoga but is a
diabetic. She currently owns apps related to each of these issues.Fig. 8 shows the bicentric diagram of her two apps: a fat-burningyoga app on the left and the diabetics recipes on the right. Mostapps exclusively associated (Type IV ties) with the yoga app areexercise related (as depicted by the green nodes). Most commonapps (Type I ties) are related to meditation (depicted by red nodes).Not surprisingly, diet apps are more frequently co-purchased with thediabetic recipes app (Type II tie). This bicentric diagram provides clearvisual demarcation among different sets of related apps that she mayfind interesting. If she wants to explore apps exclusively related to heryoga app not to her recipe app, she can easily direct her attentiontoward the left semicircle.
4.4. Global strategic alliances network
4.4.1. Context and dataWe conclude with an example from the business domain. Today,
firms rarely create value alone. They form strategic alliances to accesscomplementary resources while focusing on their core competencies.In this increasingly complex and interconnected business environment,corporate decisionmakers are interested in understanding the structureof complex strategic alliance networks of competing organizationalentities, enabling them to identify key shared direct and indirectpartners and determine which partners are exclusive to them. Usingthe SDC Platinum dataset and following [53], we construct a global stra-tegic alliances network of firms in the information and communicationstechnology (ICT) industry from 1990 to 2012. Firms are categorized intofive broad industry segments based on their primary standard industrialclassification (SIC) code: hardware components, hardware equipment,software, telecommunications, and media. Filtering out firms with lessthan five alliances, the final sample contains 1755 firms. Entities directlyconnected to the focal entity are considered first-tier partners. Those in-directly connected are second-tier partners. As hierarchy prevails inbusiness settings, the notion of a tiered network structure makes clearsense. Entities are color coded by industry segment they belong to.
Hewlett-Packard
Hardware Components
Hardware Equipment
Software
Telecommunications
Media
Fig. 9. Global interfirm alliances network of Hewlett–
Node size and edge thickness encodes weighted degree of a firm andthe number of alliances between a specific pair of firms, respectively.
4.4.2. ExampleConsider the bicentric diagram of Hewlett–Packard (HP) (left) and
Intel (right) shown in Fig. 9. Edges and help gridcircles are turned offfor comparison if having them on creates bias or helps interpretation.A decision maker may be interested in identifying a potential futuresoftware alliance partner based on their extended alliance networks.The visualization reveals stark difference between the two firms'egonetworks. HP is larger than Intel, suggesting HP has more alliancesthan Intel. Indeed, many of shared direct partners of the two firms arelarger than other nodes. However, HPdoes not have any exclusive directpartners that does not have connection with Intel. We can see that onlya few indirect partners of HP are exclusive. On the other hand, not onlydoes Intel have many direct and indirect exclusive partners, butthose exclusive partners are relatively equally distributed across allfive industry segments compared to the indirect exclusive partnersof HP. The large green node prominent on the right-hand side of Intelis Microsoft. It provides a refreshing insight that Intel has a directconnection with Microsoft, while Microsoft does not even have anindirect connection to HP.
5. Value-driven evaluation
There is an open, rich, and perhaps even heated discussion in thebroader HCI community on how to best evaluate novel visualizationtechniques [63,64]. Some researchers argue that only a controlledlaboratory experiment provides valid insights. Others advocate alongitudinal, “in the wild” approach. Unquestionably, each evaluationapproach has advantages and disadvantages. Comparing individualvisualization techniques to each other—for usefulness or performancebenchmarks—is increasingly advised against as it essentially comparesapples to oranges and reduces the potential of understanding the trueunique value and purpose of each technique [65,66,67]. Given ourstudy context, we follow Stasko's value-driven evaluation approach tounderstand bicentric diagrams [67]. This evaluation approach consists
Intel
Microsoft
Packard and Intel without gridcircles is shown.
74 H. Park, R.C. Basole / Decision Support Systems 84 (2016) 64–77
of an assessment of four key components: time, insight, essence, andconfidence. For each component, we develop a set of questions towhich respondents answer on a five-point Likert scale (1 is labeled as“Strongly Disagree” and 5 as “Strongly Agree”).
Time refers to the visualization's ability to minimize the total timeneeded to answer a wide variety of questions about the data. Aneffective visualization allows a person to identify many values from adata set and answer different questions about the data simply byviewing and interacting with the visualization. Effective visualizationsassist users to rapidly answer “low-level” questions such as findingnodes, retrieving values, and characterizing distributions [68]. We usefour statements to evaluate time:
• T1: The visualization reduced the time I needed to find nodes sharedby the focal nodes.
• T2: The visualization reduced the time I needed to find sets of nodesrelated to the focal nodes.
• T3: The visualization reduced the time I needed to find how nodeswere distributed by tier.
• T4: The visualization reduced the time I needed to find how nodeswere connected.
Insight refers to the visualization's ability to spur and discoverinsights and/or insightful questions about the data. An effective visuali-zation allows a person to learn about and make inferences from a dataset that would be much more difficult to achieve without it. Insightdoes not necessarily only equate to the unexpected or the “aha!”moment, but also to knowledge building and confirmation. Insightcan thus be considered a by-product of exploration facilitated by thevisualization [69].
• I1: The visualization enabled me to discover insights about the data.• I2: The visualization enabled me to ask insightful questions aboutthe data.
Essence refers to the visualization's ability to convey an overallessence as well as a take-away sense of the data. An effective visualiza-tion allows a person to gain a “broad, total sense of a potentially largedata set beyondwhat can be gained from learning about each individualdata case and its attributes.” [67].
• E1: The visualization conveyed an overall essence of the data.• E2: The visualization provided a take-away sense of the data.
Confidence refers to the visualization's ability to generate confidence,knowledge, and trust about the data, its domain, and context. An effec-tive visualization allows a person to learn and understand more thanjust the raw information; it promotes a broader understanding of andtrust about the data. It can help identify embedded problems in thedata set, such as missing, erroneous, or incomplete values, or highlightareas where additional data are needed.
• C1: The visualization helped me generate confidence about the data,the domain, and context.
• C2: The visualization helped me generate knowledge about the data,the domain, and context.
• C3: The visualization helpedme generate trust about the data, the do-main, and context.
5.1. Participants and procedure
As our goal was to evaluate the benefits of the visualization, and notthe specific content domain, we recruited experienced academics andpractitioners from the information visualization community. Wecontacted 28 experts by email and sent them links to the web-basedvisualization tool, tutorial, and post-use survey. 4 declined to
participate, 9 did not respond, and 15 completed the evaluation, for anet response rate of 54%. Participants hold positions such as professor(full, associate, assistant), postdoctoral researcher, PhD candidate,lecturer, user experience manager, and vice president of data analytics.Our sample consists of 11 participants from academia and 4 participantsfrom industry. We administered the survey over 10 days.
5.2. Results
Table 2 and Fig. 10 shows the result of our value-driven evaluation.Table 2 tabulates the 11 survey questions and the summary statisticsof responses for all questions along with self-assessment on expertisein information visualization. Each question is answered on a five-pointLikert scale. The greater the average of responses is, the more agreeablerespondents find the question. Fig. 10 provides a histogram summary ofresults corresponding to each evaluation statement; the thick verticalblack line in each histogram marks the mean score for each statement.
Overall, our visualization scored high in both reducing time toconduct common tasks (T1: 4.6, T2: 4.5, T3: 4.4, T4: 4.5) and providinginsight (I1: 4.5, I2: 4.3). On the other hand, our visualization scoredsomewhat lower in providing essence (E1: 4.0, E2: 3.9) and generatingsufficient confidence (C1: 3.7, C2: 3.9, C3: 3.7). All mean scores aresignificantly greater than 3 (pb0.01 from one-sample t-tests), so westatistically confirm above-average scores to all questions. The particu-larly favorable evaluation of our time criteria (T1–T4) correspondsdirectly to our design goals of creating a visualization that enables effi-cient identification of shared direct connections (G1), set memberships(G2), indirect connections (G3), and distribution (G4). This finding isencouraging as it provides support for our underlying hypothesis ofthe value of bicentric diagrams. Users commented that our tool wasparticularly helpful in “rapidly identifying shared nodes” of two focalentities and enabled a quick scan of the sets embedded in the structure.
Wewere also encouraged to see the results for insight (I1-I2). Whilethese did not directly correspond to one of our main goals, it is the aimof any visualization to provide insight. Each use context generated someinteresting “aha!” moments.
For instance, in the institutional collaboration context, one userdiscovered that her institution does not collaborate directly with anynon-U.S. institutions, while a counter top institution in her field does(e.g., “No wonder they recruit so well from China.”). However, shealso identified that both institutions collaborate with the same corpo-rate partners, highlighting their head-to-head competition for industryresearch grants. In the technology co-occurrence context, one user ex-plored what topics relate to iOS and Android. Confirming his intuition,shared topics included smartphones and tablet computing. However,counterintuitively, social networking was only co-mentioned with iOS,while next-generation hardware-related topics were co-mentionedwith Android, possibly suggesting diverse platform strategies. Topics re-lated to data and analytics were surprisingly only indirectly connectedto both operating systems.
Perhaps the more disappointing outcomes of our value-driven eval-uation was the poor performance on essence (E1–E2) and confidence(C1–C3). We speculate that this was partially due to the applicationcontexts we made available to the users.
We certainly agree that prior domain knowledge or a vested interestin the domainmatters.We also believe that if we had used the bicentricdiagram with business users and our innovation network dataexclusively, our evaluation for essence and confidence would havebeen much higher.
Additionally, the Min and Max columns in Table 2 advocate oursystem from a different perspective. The values in these columns aremeant to show extreme values in the survey responses. We have all5 s in the Max column, which means at least one respondent foundour tool strongly agreeable for each survey question. On the otherhand, Min values vary from 2 to 4. We rarely see a 2 in this column.The mean score to question 3 is fairly high at 4.4, while that to question
Table 2Evaluation questions and results (N=15).
No. Area Question Mean Std. Dev. Min Max
1 T The visualization reduced the time I needed to find nodes shared by the focal nodes. 4.60 0.51 4 52 T The visualization reduced the time I needed to find sets of nodes related to the focal nodes. 4.47 0.52 4 53 T The visualization reduced the time I needed to find how nodes were distributed by tier. 4.40 0.83 2 54 T The visualization reduced the time I needed to find how nodes were connected. 4.53 0.64 3 55 I The visualization enabled me to discover insights about the data. 4.53 0.64 3 56 I The visualization enabled me to ask insightful questions about the data. 4.33 0.62 3 57 E The visualization conveyed an overall essence of the data. 4.00 0.85 3 58 E The visualization provided a take-away sense of the data. 3.93 0.80 3 59 C The visualization helped me generate confidence about the data, its domain and context. 3.73 0.59 3 510 C The visualization helped me generate knowledge about the data, its domain and context. 3.87 0.83 3 511 C The visualization helped me generate trust about the data, its domain and context. 3.73 0.80 2 5N/A N/A Please rate your level of visualization expertise. 4.47 0.64 3 5
Note: Respondents were given a five-point Likert scale to answer each question (1 as “Strongly Disagree” and 5 as “Strongly Agree”). The greater the number is, the more agreeablerespondents find the question. The legend for the performance area shorthand notations is as follows. T: Time, I: Insight, E: Essence, C: Confidence.
75H. Park, R.C. Basole / Decision Support Systems 84 (2016) 64–77
11 is relatively low at 3.7. This finding suggests that our system mayneed to improve in terms of trust and confidence-related aspects.
Correlation analysis between participants' expertise in informationvisualization and four performance area average scores supports our in-terpretation above. Pairwise correlation coefficients between expertiseand average scores in time, insight, essence, confidence are 0.34, 0.22,0.17, -0.08, respectively. Thus, the greater the respondent's expertise,the higher the scores in time, insight, essence, and the lower three
Fig. 10.Value-driven evaluation results are shown in the formof histograms by four performanc
confidence scores. Particularly, time performance receives highlyfavorable scores from participants with high expertise.
The reason for choosing visualization “experts” for evaluating oursystem was to ensure that our system was not just easy-to-use butalso conforms to visualization principles. Experts will be much moreversed in understanding the nuances of visualization principles. Theywill also be able to comment on the relative strengths of the systemcompared to other visualization approaches. This does not mean that
e areas. The vertical black thick line in each histograms denotes the respectivemean value.
76 H. Park, R.C. Basole / Decision Support Systems 84 (2016) 64–77
our tool is inaccessible to people with less experience, in fact, our anal-ysis shows that users with lower expertise scores still rated the toolquite favorably. We acknowledge that the sample size is small, and weplan to conduct a broader user study with wide range of characteristicsin future work.
We also received informal feedback from our users after the evalua-tion phase regarding the overall ease of use, aesthetics, and applicabilityof our visualization. With regards to ease of use, for instance, onerespondent commented, “I can quickly figure out which of the twofocal nodes has more relationships by comparing the number ofnodes and their distributions. I can get to know how a certain nodehas closer relationships among two focal nodes. I can see how categoryaffects relationships and which nodes have more cross-categoryrelationships.” Users complimented the look-and-feel of our visualiza-tion (e.g., “Overall, beautiful tool!” and “The implementation feels veryprofessional.”). Lastly, users commented the generalizability of ourvisualization to other decision-making contexts (e.g., “I loved thevisualization, very clever way of showing the connections betweentwo things.” and “I can see this visual representation applied to myown dataset!”).
6. Conclusion
Visualizations amplify human cognition and augment decisionmaking. As data on social, economic, business, and physical networksare proliferating at an accelerating pace, the ability to engage withsuch data interactively is becoming ever critical to explore underlyingcomplex structural connections, discover prominent entities, andidentify clusters. Motivated by a practical need articulated by corporatedecision makers and limitations of existing visual representations,this research presented our journey in designing and implementingbicentric diagrams, a novel graph-based set visualization techniquethat can be used for perspective development in the loop ofdecision-making process [24].
Bicentric diagrams enable simultaneous identification of sets, setrelationships, and set member reach in integrated egonetworks oftwo focal entities. Thus, it helps quickly generate potential solutioncandidates for organizational decision-making problems locating struc-ture holes from the network perspective. Our technique builds onthe well-established sociological theory of tie strength to visuallygroup and position nodes. In contrast to traditional network layoutapproaches, our technique provides an aesthetically appealing way tovisualize the intersections of weak and strong ties between two entities.We demonstrate the broad applicability of bicentric diagrams withexamples from four diverse sample domains: university collabora-tion, technology co-occurrence, health app purchases, and interfirmalliance networks.
The results of our value-driven expert evaluation revealed severalinteresting results and offered suggestions for future research. First,bicentric diagrams enable decision makers to rapidly identify keyplayers and their roles and positions in egonetworks of two focalentities. Second, our technique also facilitates gaining new insightsand led to various “aha!” moments, which helps form perspectives fora decision-making problem. Third, users appreciated ease of use,aesthetics, and broad applicability of the bicentric layout to differentdecision-making domains.
Our expert user study, however, also revealed a few limitations.For instance, depending on the size of networks, visual clutteringcan be an issue. Fisheye focus + context techniques [70] orintelligent labeling could be used to overcome this. Individual usercompetencies may also impact performance. Examining how visualperception and literacy affects a user's understandingwill be anotherinteresting path of inquiry. Others commented that a more compre-hensive user study in one specific domain, such as e-commerce prod-uct recommendations, could provide additional insights. We areconsidering a large-scale mechanical turk study to explore this
further. Contexts such as restaurant co-review networks (e.g. Yelp)or start-up investment’ networks (e.g. Crunchbase) may be of inter-est. Each of these limitations presents exciting future research andtheir pursuit will help further refine the usefulness and utility ofbicentric diagrams. We hope that bicentric diagrams can serve as aguide for other researchers and practitioners to develop novel repre-sentations for specific decision-making problems and implementthem into a web-based DSS.
Appendix A. Supplementary data
Supplementary data to this article can be found online at http://dx.doi.org/10.1016/j.dss.2016.02.001.
References
[1] K. Aspelund, Designing: An Introduction, Fairchild Books, New York, NY, 2014.[2] J. Heer, B. Shneiderman, Interactive dynamics for visual analysis, Communications of
the ACM 55 (2) (2012) 45, http://dx.doi.org/10.1145/2133416.2146416.[3] S.K. Card, J.D. Mackinlay, B. Shneiderman, Readings in Information Visualization:
Using Vision to Think, Morgan Kaufmann, San Francisco, CA, 1999.[4] J.-d. Fekete, J.J.V. Wijk, J.T. Stasko, C. North, The value of information visualization,
Information Visualization 4950 (2008) 1–18, http://dx.doi.org/10.1007/978-3-540-70956-5_1.
[5] A.-L. Barabási, Linked: how everything is connected to everything else and what itmeans for business, Science, and Everyday Life, Plume (2003)arXiv:0806.1610,http://dx.doi.org/10.1146/annurev.soc.30.020404.104342.
[6] D.J. Watts, The new science of networks, Annual Review of Sociology 30 (1) (2004)243–270, http://dx.doi.org/10.1146/annurev.soc.30.020404.104342.
[7] C. Holsapple, A. Lee-Post, R. Pakath, A unified foundation for business analytics,Decision Support Systems 64 (2014) 130–141, http://dx.doi.org/10.1016/j.dss.2014.05.013.
[8] O. Ben-Assuli, Assessing the perception of information components in financialdecision support systems, Decision Support Systems 54 (1) (2012) 795–802,http://dx.doi.org/10.1016/j.dss.2012.09.007.
[9] B. Kamsu-Foguem, G. Tchuenté-Foguem, L. Allart, Y. Zennir, C. Vilhelm, H. Mehdaoui,D. Zitouni, H. Hubert, M. Lemdani, P. Ravaux, User-centered visual analysis using ahybrid reasoning architecture for intensive care units, Decision Support Systems54 (1) (2012) 496–509, http://dx.doi.org/10.1016/j.dss.2012.06.009.
[10] R.C. Basole, M.L. Braunstein, V. Kumar, H. Park, M. Kahng, D.H.P. Chau, A. Tamersoy,D.A. Hirsh, N. Serban, J. Bost, B. Lesnick, B.L. Schissel, M. Thompson, Understandingvariations in pediatric asthma care processes in the emergency department usingvisual analytics, Journal of the American Medical Informatics Association (2015),http://dx.doi.org/10.1093/jamia/ocu016 (ocu016).
[11] T.J. Overbye, J.D. Weber, K.J. Patten, Analysis and visualization of market powerin electric power systems, Decision Support Systems 30 (3) (2001) 229–241,http://dx.doi.org/10.1016/S0167-9236(00)00101-9.
[12] R.C. Basole, C.D. Seuss, W.B. Rouse, IT innovation adoption by enterprises: knowl-edge discovery through text analytics, Decision Support Systems 54 (2) (2013)1044–1054, http://dx.doi.org/10.1016/j.dss.2012.10.029.
[13] R.C. Basole, Visual business ecosystem intelligence: lessons from the field, IEEE Com-puter Graphics and Applications 34 (5) (2014) 26–34, http://dx.doi.org/10.1109/MCG.2014.104.
[14] I. Herman, G. Melancon, M. Marshall, Graph visualization and navigationin information visualization: a survey, IEEE Transactions on Visualizationand Computer Graphics 6 (1) (2000) 24–43, http://dx.doi.org/10.1109/2945.841119.
[15] C. Speier, The influence of information presentation formats on complex taskdecision-making performance, International Journal of Human-Computer Studies64 (11) (2006) 1115–1131, http://dx.doi.org/10.1016/j.ijhcs.2006.06.007.
[16] J. Gettinger, S.T. Koeszegi, M. Schoop, Shall we dance? - the effect of informationpresentations on negotiation processes and outcomes, Decision Support Systems53 (1) (2012) 161–174, http://dx.doi.org/10.1016/j.dss.2012.01.001.
[17] W. Hong, J.Y. Thong, K.Y. Tam, The effects of information format and shopping taskon consumers' online shopping behavior: a cognitive fit perspective, Journal ofManagement Information Systems 21 (3) (2004) 149–184.
[18] J. Gettinger, E. Kiesling, C. Stummer, R. Vetschera, A comparison of representationsfor discrete multi-criteria decision problems, Decision Support Systems 54 (2)(2013) 976–985, http://dx.doi.org/10.1016/j.dss.2012.10.023.
[19] H. Park, R.C. Basole, Bicentric diagrams: design of a graph-based relational setvisualization technique, Proceedings of the 33rd Annual ACM Conference ExtendedAbstracts on Human Factors in Computing Systems—CHI EA ‘15, ACM Press,New York, New York, USA 2015, pp. 1815–1820, http://dx.doi.org/10.1145/2702613.2732752.
[20] R.C. Basole, M.A. Bellamy, Visual analysis of supply network risks: insights from theelectronics industry, Decision Support Systems 67 (2014) 109–120, http://dx.doi.org/10.1016/j.dss.2014.08.008.
[21] R.C. Basole, M.G. Russell, J. Huhtamäki, K. Still, N. Rubens, K. Still, H. Park,Understanding business ecosystem dynamics: a data driven approach, ACMTransactions on Management Information Systems 6 (2) (2015) 1–32, http://dx.doi.org/10.1145/2724730.
77H. Park, R.C. Basole / Decision Support Systems 84 (2016) 64–77
[22] R.C. Basole, H. Park, V. Kumar, M.L. Braunstein, J. Bost, D.H. Chau, M. Kahng, Bicentricvisualization of pediatric asthma care process activities, Proceedings of IEEE VIS2014 Workshop of Electronic Health Records, IEEE, 2014.
[23] M.S. Granovetter, The strength of weak ties, American Journal of Sociology 78 (6)(1973) 1360–1380, http://dx.doi.org/10.2307/2776392.
[24] J.F. Courtney, Decision making and knowledge management in inquiring organiza-tions: toward a new decision-making paradigm for DSS, Decision Support Systems31 (1) (2001) 17–38, http://dx.doi.org/10.1016/S0167-9236(00)00117-2.
[25] J. Shim, M. Warkentin, J.F. Courtney, D.J. Power, R. Sharda, C. Carlsson, Past, present,and future of decision support technology, Decision Support Systems 33 (2) (2002)111–126, http://dx.doi.org/10.1016/S0167-9236(01)00139-7.
[26] G. Ahuja, Collaboration networks, structural holes, and innovation: a longitudinalstudy, Administrative Science Quarterly 45 (3) (2000) 425–455.
[27] E. Condon, B. Golden, S. Lele, S. Raghavan, E. Wasil, A visualization model based onadjacency data, Decision Support Systems 33 (4) (2002) 349–362, http://dx.doi.org/10.1016/S0167-9236(02)00003-9.
[28] C.C. Yang, H. Chen, K. Hong, Visualization of large category map for internetbrowsing, Decision Support Systems 35 (1) (2003) 89–102, http://dx.doi.org/10.1016/S0167-9236(02)00101-X.
[29] G. Shmueli, W. Jank, A. Aris, C. Plaisant, B. Shneiderman, Exploring auctiondatabases through interactive visualization, Decision Support Systems 42 (3)(2006) 1521–1538, http://dx.doi.org/10.1016/j.dss.2006.01.001.
[30] M.J. Eppler, K.W. Platts, Visual strategizing, Long Range Planning 42 (1) (2009)42–74, http://dx.doi.org/10.1016/j.lrp.2008.11.005.
[31] J. Heer, M. Bostock, V. Ogievetsky, A tour through the visualization zoo,Communications of the ACM 53 (6) (2010) 59, http://dx.doi.org/10.1145/1743546.1743567.
[32] P.C. Wong, R.D. Bergeron, 30 Years of multidimensional multivariate visualization,Scientific Visualization, Overviews, Methodologies, and Techniques, IEEE ComputerSociety 1994, pp. 3–33.
[33] T. Asahi, D. Turo, B. Shneiderman, Using treemaps to visualize the analytic hierarchyprocess, Information Systems Research 6 (4) (1995) 357–375.
[34] J. Stasko, R. Catrambone, M. Guzdial, K. McDonald, An evaluation of space-fillinginformation visualizations for depicting hierarchical structures, InternationalJournal of Human-Computer Studies 53 (5) (2000) 663–694, http://dx.doi.org/10.1006/ijhc.2000.0420.
[35] A. Sundararajan, F. Provost, G. Oestreicher-Singer, S. Aral, Research commentary-information in digital, economic, and social networks, Information Systems Research24 (4) (2013) 883–905.
[36] N. Venkatraman, C.-H. Lee, Preferential linkage and network evolution: a conceptualmodel and empirical test in the U.S. video game sector, Academy of ManagementJournal 47 (6) (2004) 876–892, http://dx.doi.org/10.2307/20159628.
[37] J. Xu, H. Chen, Criminal network analysis and visualization, Communications of theACM 48 (6) (2005) 100–107, http://dx.doi.org/10.1145/1064830.1064834.
[38] M. Trier, Research note-towards dynamic visualization for understanding evolutionof digital communication networks, Information Systems Research 19 (3) (2008)335–350.
[39] M. Keith, H. Demirkan, M. Goul, The influence of collaborative technologyknowledge on advice network structures, Decision Support Systems 50 (1) (2010)140–151, http://dx.doi.org/10.1016/j.dss.2010.07.010.
[40] L. Bizzi, A. Langley, Studying processes in and around networks, IndustrialMarketing Management 41 (2) (2012) 224–234, http://dx.doi.org/10.1016/j.indmarman.2012.01.007.
[41] P. Fox, J. Hendler, Changing the equation on scientific data visualization, Science 331(6018) (2011) 705–708, http://dx.doi.org/10.1126/science.1197654.
[42] J. Thomas, K. Cook, A visual analytics agenda, IEEE Computer Graphics andApplications 26 (1) (2006) 10–13, http://dx.doi.org/10.1109/MCG.2006.5.
[43] W.A. Pike, J. Stasko, R. Chang, T.A. OConnell, The science of interaction, InformationVisualization 8 (4) (2009) 263–274, http://dx.doi.org/10.1057/ivs.2009.22.
[44] B. Alsallakh, L. Micallef, W. Aigner, H. Hauser, S. Miksch, P. Rodgers, Visualizing setsand set-typed data: state-of-the-art and future challenges, Eurographics Conferenceon Visualization (EuroVis), 2014.
[45] B. Alsallakh, W. Aigner, S. Miksch, H. Hauser, Radial sets: interactive visual analysisof large overlapping sets, IEEE Transactions on Visualization and Computer Graphics19 (2013) 2496–2505, http://dx.doi.org/10.1109/TVCG.2013.184.
[46] N. Watanabe, M.Washida, T. Igarashi, Bubble clusters: an interface for manipulatingspatial aggregation of graphical objects, ACM SymposiumonUser Interface Softwareand Technology 2007, pp. 173–182, http://dx.doi.org/10.1145/1294211.1294241.
[47] J. Heer, D. Boyd, Vizster: visualizing online social networks, IEEE Symposium onInformation Visualization (InfoVis) 2005, pp. 33–40, http://dx.doi.org/10.1109/INFVIS.2005.1532126.
[48] C. Collins, G. Penn, S. Carpendale, Bubble sets: revealing set relations with isocontoursover existing visualizations, IEEE Transactions on Visualization and Computer Graphics15 (2009) 1009–1016, http://dx.doi.org/10.1109/TVCG.2009.122.
[49] J. Stasko, C. Görg, Z. Liu, R. Spence, Jigsaw: supporting investigative analysis throughinteractive visualization, Information Visualization 7 (2) (2008) 118–132, http://dx.doi.org/10.1057/palgrave.ivs.9500180.
[50] K. Misue, Anchored maps: visualization techniques for drawing bipartite graphs,Human–Computer Interaction, Interaction Platforms and TechniquesSpringer,Berlin Heidelberg 2007, pp. 106–114.
[51] M. Krzywinski, J. Schein, I. Birol, J. Connors, R. Gascoyne, D. Horsman, S.J. Jones, M.A.Marra, Circos: an information aesthetic for comparative genomics, GenomeResearch 19 (9) (2009) 1639–1645, http://dx.doi.org/10.1101/gr.092759.109.
[52] S. Milgram, The small-world problem, Psychology Today.
[53] R.C. Basole, H. Park, B.C. Barnett, Coopetition and convergence in the ICT ecosystem,Telecommunications Policy 39 (7) (2015) 537–552, http://dx.doi.org/10.1016/j.telpol.2014.04.003.
[55] J. Hopcroft, R. Tarjan, Algorithm 447: efficient algorithms for graph manipulation,Communications of the ACM 16 (6) (1973) 372–378, http://dx.doi.org/10.1145/362248.362272.
[56] M.A. Schilling, C.C. Phelps, Interfirm collaboration networks: the impact of large-scalenetwork structure on firm innovation,Management Science 53 (7) (2007) 1113–1126.
[57] L. Rosenkopf, M.A. Schilling, Comparing alliance network structure across industries:observations and explanations, Strategic Entrepreneurship Journal 1 (3–4) (2007)191–209, http://dx.doi.org/10.1002/sej.33.
[58] M. Bostock, V. Ogievetsky, J. Heer, D3: data-driven documents, IEEE Transactions onVisualization and Computer Graphics 17 (12) (2011) 2301–2309, http://dx.doi.org/10.1109/TVCG.2011.185.
[59] A. Hagberg, P. Swart, D.S. Chult, Exploring Network Structure, Dynamics, andFunction Using NetworkX, Tech. rep. Los Alamos National Laboratory (LANL), 2008.
[60] H. Park, K. Lee, Dependence clustering, a method revealing community structurewith group dependence, Knowledge-Based Systems 60 (2014) 58–72, http://dx.doi.org/10.1016/j.knosys.2014.01.004.
[61] R. Basole, A. Qamar, H. Park, C. Paredis, L. McGinnis, Visual analytics for early-phasecomplex engineered system design support, IEEE Computer Graphics andApplications 35 (2) (2015) 41–51, http://dx.doi.org/10.1109/MCG.2015.3.
[62] G. Vasudeva, A. Zaheer, E. Hernandez, The embeddedness of networks: institutions,structural holes, and innovativeness in the fuel cell industry, Organization Science.
[63] S. Carpendale, Evaluating Information Visualizations, in: Information Visualization,Springer Berlin Heidelberg, Berlin, Heidelberg, 2008 19–45, http://dx.doi.org/10.1007/978-3-540-70956-5_2.
[64] H. Lam, E. Bertini, P. Isenberg, C. Plaisant, S. Carpendale, Empirical studies ininformation visualization: seven scenarios, IEEE Transactions on Visualizationand Computer Graphics 18 (9) (2012) 1520–1536, http://dx.doi.org/10.1109/TVCG.2011.279.
[65] M. Tory, User studies in visualization: a reflection onmethods, Handbook of HumanCentric Visualization, Springer 2014, pp. 411–426.
[66] B. Shneiderman, C. Plaisant, Strategies for evaluating information visualization tools:multi-dimensional in-depth long-term case studies, Proceedings of the 2006 AVIWorkshop on Beyond Time and Errors: Novel Evaluation Methods for InformationVisualization, ACM 2006, pp. 1–7.
[67] J.T. Stasko, Value-Driven Evaluation of Visualizations, in: Proceedings of theWorkshop: Beyond Time and Errors—Novel Evaluation Methods for Visualization(Paris, France, November 2014), 2014.
[68] R. Amar, J. Eagan, J. Stasko, Low-level components of analytic activity in informationvisualization, IEEE Symposium on Information Visualization (InfoVis) 2005,pp. 111–117, http://dx.doi.org/10.1109/INFVIS.2005.1532136.
[69] R. Chang, C. Ziemkiewicz, T.M. Green, W. Ribarsky, Defining insight for visualanalytics, IEEE Computer Graphics and Applications 29 (2009) 14–17, http://dx.doi.org/10.1109/MCG.2009.22.
[70] J. Lamping, R. Rao, P. Pirolli, A focus + context technique based on hyperbolicgeometry for visualizing large hierarchies, Proceedings of the SIGCHI Conferenceon Human Factors in Computing Systems - CHI ‘95, ACM Press, New York,New York, USA 1995, pp. 401–408, http://dx.doi.org/10.1145/223904.223956.
[71] L. Wilkinson, Exact and approximate area-proportional circular Venn and Eulerdiagrams, IEEE Transactions on Visualization and Computer Graphics 18 (2012)321–331, http://dx.doi.org/10.1109/TVCG.2011.56.
[72] N.H. Riche, T. Dwyer, Untangling Euler diagrams, IEEE Transactions on Visualizationand Computer Graphics 16 (2010) 1090–1099, http://dx.doi.org/10.1109/TVCG.2010.210.
[73] K. Dinkla, M.J. van Kreveld, B. Speckmann, M.a. Westenberg, Kelp diagrams: pointset membership visualization, Computer Graphics Forum 31 (2012) 875–884,http://dx.doi.org/10.1111/j.1467–8659.2012.03080.x.
[74] M. Dork, N. Henry Riche, G. Ramos, S. Dumais, PivotPaths: strolling through facetedinformation spaces, IEEE Transactions on Visualization and Computer Graphics 18(12) (2012) 2709–2718, http://dx.doi.org/10.1109/TVCG.2012.252.
Hyunwoo Park is a postdoctoral fellow of the Tennenbaum Institute at the GeorgiaInstitute of Technology. His research interests include technological innovation, platformstrategy, visual analytics, and computational strategy science. Park has a PhD in industrialengineering from the Georgia Institute of Technology and a master's in informationmanagement and systems from the University of California, Berkeley. Contact himat, [email protected].
Rahul C. Basole is an associate professor in the School of Interactive Computing, theassociate director for Enterprise Transformation in the Tennenbaum Institute/IPaT, andan affiliated faculty member in the GVU Center at the Georgia Institute of Technology.He is also a visiting scholar in HSTAR at Stanford University and a Fellow of the BattenInstitute at the Darden School of Business. He is also the editor-in-chief of the Journal ofEnterprise Transformation.His research and teaching focuses on computational enterprisescience, information visualization, and strategic decision support. He holds a PhD inindustrial and systems engineering from the Georgia Institute of Technology. Contacthim at, [email protected].
