Content uploaded by Narges Mahyar
Author content
All content in this area was uploaded by Narges Mahyar on Feb 10, 2015
Content may be subject to copyright.
Supporting Communication and Coordination in Collaborative
Sensemaking
Narges Mahyar and Melanie Tory
Abstract—When people work together to analyze a data set, they need to organize their findings, hypotheses, and evidence, share
that information with their collaborators, and coordinate activities amongst team members. Sharing externalizations (recorded infor-
mation such as notes) could increase awareness and assist with team communication and coordination. However, we currently know
little about how to provide tool support for this sort of sharing. We explore how linked common work (LCW) can be employed within
a ‘collaborative thinking space’, to facilitate synchronous collaborative sensemaking activities in Visual Analytics (VA). Collaborative
thinking spaces provide an environment for analysts to record, organize, share and connect externalizations. Our tool, CLIP, extends
earlier thinking spaces by integrating LCW features that reveal relationships between collaborators’ findings. We conducted a user
study comparing CLIP to a baseline version without LCW. Results demonstrated that LCW significantly improved analytic outcomes
at a collaborative intelligence task. Groups using CLIP were also able to more effectively coordinate their work, and held more dis-
cussion of their findings and hypotheses. LCW enabled them to maintain awareness of each other’s activities and findings and link
those findings to their own work, preventing disruptive oral awareness notifications.
Index Terms—Sensemaking; Collaboration; Externalization; Linked common work; Collaborative thinking space.
1 INTRODUCTION
Supporting collaborative sensemaking has been identified as an im-
portant challenge in collaborative visualization [20]. Sensemaking in
collaborative VA is a very time consuming and demanding process, re-
quiring the analysts to iteratively exchange and discuss results to form
and evaluate hypotheses, derive conclusions, and publish findings.
Team members also need to maintain awareness of each other’s work,
including both activities that people are working on and results and
evidence that they have found. Tools that provide externalization sup-
port (i.e., ability to record insights, questions, and findings, e.g., as text
notes) can help teams to organize and share their results [6,18, 22,41],
and those that provide awareness channels should enhance collabo-
ration, communication and coordination [12]. However, to date, we
have a very limited understanding of how to provide externalization
and awareness support for collocated collaborative teams. How should
such tool support look and behave within VA tools?
We investigate the use of Linked Common Work (LCW) to facil-
itate synchronous collaborative sensemaking. With LCW, common
work elements such as similar findings are automatically discovered,
linked, and visually shared among the group. We built this technique
within a ‘collaborative thinking space’ that enables analysts to record,
organize and schematize their externalizations. Linked common work
reveals similarities in people’s externalizations, enabling analysts to
acquire awareness of each other’s findings, hypotheses, and evidence.
Moreover, each individual analyst can review and merge others’ work
from within his/her workspace. Our results demonstrate that applying
LCW to externalizations, and providing the ability to integrate collab-
orators’ findings together within one view, noticeably improve team
awareness, coordination, communication, and analytic outcomes.
Our work focuses on supporting teams of investigative analysts,
for example in the domain of intelligence analysis. Intelligence ana-
lysts need to sift through large document collections, determine which
pieces of data are relevant, and gradually build up an explanation sup-
ported by evidence. Field studies have revealed that professional an-
alysts need to share sources and data, view each other’s work, and
combine findings together in order to build common ground, resolve
•Narges Mahyar and Melanie Tory are with the University of Victoria.
E-mail: {nmahyar, mtory}@uvic.ca
Manuscript received 31 March 2013; accepted 1 August 2013; posted online
13 October 2013; mailed on 4 October 2013.
For information on obtaining reprints of this article, please send
e-mail to: tvcg@computer.org.
conflicts, and validate each other’s findings and hypotheses [8, 25].
The sensemaking process of intelligence analysts has been stud-
ied in some depth, and has been described as involving two iterative
loops: the information foraging loop and the sensemaking loop [34].
The information foraging loop involves searching for relevant data and
reading, filtering, and extracting information, whereas the sensemak-
ing loop involves iteratively developing a mental model, forming and
evaluating hypotheses, and publishing the results. We focus primarily
on supporting later stages of the sensemaking process (i.e., the sense-
making loop), when teams are more likely to work together in a syn-
chronous, collocated fashion [25]. This synthesis phase is reported to
be the most difficult and time-consuming phase of analysis [25].
We are exploring the design of visual thinking spaces that support
the sensemaking loop in collaborative VA. A collaborative thinking
space should enable analysts to record and organize findings, evidence,
and hypotheses; moreover, it should facilitate the process of sharing
findings amongst collaborators, to minimize redundant work and help
investigators identify relationships and build a shared understanding.
In this paper, we examine the value of employing LCW to relate and
integrate team members’ visual thinking spaces. The notion of LCW
closely resembles collaborative brushing and linking [21] in which
certain actions of each investigator are visible to collaborators through
their own views. However, collaborative brushing and linking was
only applied to search queries and retrieved documents and did not
cover externalizations. It also focused on supporting only information
foraging activities. In contrast, our work facilitates later stages of the
collaborative sensemaking process (i.e., the sensemaking loop), by ap-
plying the linking concept to people’s externalizations (i.e., recorded
findings and notes). We anticipate that enabling analysts to see how
their findings relate to each other should make it easier to maintain
awareness of each others’ work, build common ground, and solve an-
alytic problems. We address the following research questions (RQs):
•RQ1: Does linking collaborators’ externalizations lead to better
analytic outcomes?
•RQ2: Does linking collaborators’ externalizations improve com-
munication?
•RQ3: Does linking collaborators’ externalizations help collabo-
rators to coordinate their work more effectively?
•RQ4: Does linking collaborators’ externalizations increase col-
laborators’ awareness of each others’ findings and activities?
To answer these questions, we designed and implemented CLIP, a vi-
sual thinking space to support collaborative sensemaking. CLIP allows
analysts to record their findings in the form of a node-link graph and
timeline, add evidence to facilitate evidence marshaling, and add free
form text to record hypotheses, questions, to-do-lists, etc. Most im-
portantly, CLIP incorporates LCW to relate and integrate the findings
of different collaborators. We assessed the value of LCW by compar-
ing CLIP to a baseline tool (BT) without the LCW features. Results
of our user study demonstrated that LCW led to more effective group
coordination and communication as well as better analytic outcomes.
2 RE LATE D WORK
The design of CLIP draws upon prior research on sensemaking and
collaboration support. For individual work, many tools have been de-
veloped to support both phases of sensemaking (e.g., [4,19,42]). How-
ever, in the context of collaborative sensemaking much less has been
done. Most of the existing tools either focus on the information forag-
ing loop [21] or asynchronous collaboration [7, 41].
In the remainder of this section, we summarize the existing guide-
lines on how to support collaborative sensemaking activities, and how
those relate to CLIP’s design and our experimental study. To gather
these guidelines, we reviewed relevant field and observational studies
to extract features that focused on the sensemaking loop.
2.1 Externalization, Schematizing, and History
Mahyar et al. [27] demonstrated the critical importance of external-
ization during collaborative analysis, and the lack of support for this
process in many current visualization tools. Correspondingly, Kang
and Stasko [25] suggested that supporting history of previous discov-
eries and sanity checking could save time during report-writing. Their
field observations showed that analysts spent substantial time return-
ing to original sources to find the supporting references and rationale
behind their statements. Vogt et al. [38] and Pirolli and Card [34]
similarly pointed out the need to record findings, hypotheses and evi-
dence. Several studies in the intelligence analysis domain signified the
importance of schematizing results [8,22, 24]; in other words, organiz-
ing results and other externalizations into a structured format. Various
structured formats can be useful, including timelines, spreadsheets,
lists, and networks [8, 22]. For instance Zhang [43] discussed the na-
ture of external representations in cognition and mentioned diagrams,
graphs, and pictures as a few typical types of external representations.
For meetings or tasks that require flexibility, such as brainstorming and
collaborative design, freeform graphical input could be a better option
to support flexibility [23, 32]. Other structures include casual loop di-
agrams, mind maps, diagrams, graphs, and pictures [23, 31, 32]. We
expect schematizing to be even more critical for collaborative work,
since the structure may also help with communication. CLIP therefore
includes node-link graph and timeline schemas for representing find-
ings. We anticipate that integrating collaborators’ schematic views
will help them to build common ground and relate findings.
2.2 Communication and Coordination
Various studies have found that the closeness of groups’ collaboration
styles directly affected outcomes [5,22,25,38]. Isenberg et al. [22] and
Vogt et al. [38] found that teams who collaborated more closely were
more successful, and Bradel et al. [5] reported that members in more
successful teams understood each other’s work better. Groups should
not always work in a closely coupled fashion, however. Kang and
Stasko [25] found that in a long term project, collaboration was loose
during information collection but tight when synthesizing findings and
writing a report. A good collaborative system, then, should encour-
age groups towards closer collaboration styles when there are relevant
findings to be connected, but allow loose collaboration at other times.
One way to encourage closer collaboration is through awareness
mechanisms that provide information to each investigator about their
collaborators’ activities and findings. Paul and Reddy [33] suggested
providing support for both action awareness and activity awareness,
showing actions that led to a particular activity. Various techniques
have been developed to help collaborators maintain awareness. For
instance, Hugin [26] provides awareness support in a mixed-presence
setting through a layer-based GUI design. CoSpaces [28] places each
person’s information in a separate view. Users must then compare and
reconcile different views, a potentially cumbersome process. People
may also miss relevant changes since they are often hidden from view.
Nonetheless, Mahyar et al. [28] reported that separate views were use-
ful for exploring other people’s work in a non-disruptive way.
In contrast, integrating everyone’s information into one view could
cause disruption to individual work as the view constantly updates.
Brennan et al. [6] implemented a visualization of externalizations and
explored ways to merge collaborators’ content. Similar to CLIP, Bren-
nan et al.’s approach provides common ground to support collabora-
tion. While CLIP does not address confidence ratings as explicitly as
their system (though they could be added to notes), the simpler visual
design of CLIP should make it easier to communicate: consistency in
visual encoding should make it easy to understand other users’ per-
spectives and LCW should make it easy to find commonalties. Sim-
ilarly, CoMotion [9] enabled analysts to share data views and notes.
However, neither project evaluated whether the shared view was help-
ful in practice. Another related system is Cambiera [21], which pro-
vided awareness cues about related searches conducted by a collabo-
rator during information foraging. Isenberg et al. [22] reported that
these cues, which they termed collaborative brushing and linking, en-
couraged closer collaboration. We emphasize that Cambiera did not
consider how linking could be applied to externalizations. Our work
extends the linking idea to externalizations and support for the sense-
making loop. Note that we use the term ‘linked common work’ instead
of ‘collaborative brushing and linking’ to generalize the linking idea
to cases where there is no active ‘brushing’ action to initiate the link.
Which is more important, having everyone’s work visible in a single
shared space, or reducing disruption to individual work by keeping no-
tifications and visual changes to a minimum? Our study addresses this
question by comparing CLIP, which allows everything to be integrated
within one view, to a baseline tool with a separate view approach.
It is also important to support task coordination. Mark and
Kobsa [29] identify coordination support as a challenge for designers
and emphasize the need to choose appropriate visual representations
to help teams during VA tasks.
2.3 Shared and Individual Workspaces
The importance of providing both individual and group workspace is
well known [15,21,22]. For example, Wallace et al. [39] demonstrated
that groups had better outcomes when they were able to share their
results together. The need for shared and individual workspaces ap-
plies equally well to externalization spaces. Even though collocated
teams often assign one note taker [27, 36, 38], individuals still occa-
sionally need to record private notes [27]. This suggests that it is
important to provide both shared and individual notetaking spaces.
Mahyar et al. [27] revealed that taking notes on paper reduced the
note taker’s awareness of other activities, suggesting that notetaking
tools should be integrated with the investigative application. How-
ever, a separated digital view can suffer from the same problem. For
instance, in Bradel et al.’s [5] study, only one user could take notes
in the shared space; others had to work on separate views and this
reduced awareness. Kang and Stasko [25] recommended “promot-
ing individual workspaces as well as...the ability to share sources and
data, view and comment on others’ work, and merge individual work
together.” Brennan et al. [6] similarly suggested providing individual
perspectives on a shared space. McGrath et al. [30] introduced Branch-
Explore-Merge, a structure for transitioning between individual and
group workspaces. We note that their concept was applied to direct
views of data, not within a collaborative thinking space. With CLIP,
we address the need for individual workspace plus awareness of oth-
ers’ work by providing each user with flexible control over how much
of the collaborators’ information is shown in their view.
2.4 Studies of Collaborative Thinking Spaces
Most similar to our work are studies of collaborative visual analytics
work involving thinking spaces, particularly studies that explore how
to integrate or provide access to different users’ views. What level of
integration is appropriate for different kinds of shared information?
Chen et al. [7] built a tool that enabled asynchronous collaborators
to record and share insights, and demonstrated that people could learn
from others’ past insights. Similarly, Willett et al. [41] demonstrated
that asynchronous collaborators benefited from the ability to classify
their text comments using tags and from the ability to link comments
and views; however, in their system, users had to create links manually.
Neither of these studies examined the value of automatically linking
common work, nor did they look at synchronous collaboration.
For synchronous collaboration, most evidence suggests that highly
integrated views should be a good approach. Balakrishnan et al. found
that a shared view of a network diagram supported better performance
at a collaborative intelligence task than separated views [2]. We note,
however, that their network diagram was a simple social network
rather than a dynamic thinking space for recording and organizing ex-
ternalizations. We are motivated by their later study [3], where they
suggested that it might help to additionally visualize partners’ activi-
ties and externalizations. Bradel et al. [5] used Jigsaw’s tablet view to
allow analysts to record their findings in the form of notes, graphs, and
timelines. They found that Jigsaw’s tablet view was inadequate for col-
located collaboration, because participants wanted to use it as a shared
note space, but it accepted only one person’s input at a time. Their
study did not examine separate linked views of the thinking space, but
they suggested that it may be a good idea.
3 LINKED COMMON WORK (LCW)
The LCW technique employed in CLIP reveals similarities between
collaborators’ findings by discovering, linking, and visually represent-
ing the common work. This approach is based on research in social
interaction such as Clark and Brennan’s [10], that showed the impor-
tance of a shared understanding for effective collaboration. Subtle vi-
sual cues enable analysts to gain awareness of each other’s findings,
hypotheses, and evidence with minimal disruption [16, 21]. This is
what we called partial merging. Then, if an analyst wants to more
closely monitor others’ work the full merging option is available to in-
tegrate others’ findings directly into his/her workspace. Partial and
full merging are described in detail below. When views are fully
merged, the layout computation only updates positions of common
nodes (those in common between the user’s graph and their collab-
orator’s). Common nodes are overlaid and their edge shapes are re-
computed as necessary. All remaining nodes (i.e. uncommon nodes)
are placed where they originally were, and the user can move them if
necessary. We avoid making automatic changes in the layout of the
local graph in order to preserve the user’s mental map of nodes and
relationships.
As a proof of concept, we used Jigsaw to extract evidence items and
lists of entities (People, Locations, Organizations, Chemicals, Events,
etc.) from the document corpus to ensure perfect matching between
participants’ externalizations. There are many ways to improve both
the visual representation and the merging algorithm used for LCW. For
example, there are algorithms to merge entities that are named differ-
ently but semantically related (e.g., lexicon chains to find synonyms
and related words). However, the main goal at this stage of our work
was to demonstrate value of LCW. In future work we discuss how to
improve and extend the LCW technique.
4 SYSTEM DESIGN
CLIP, our Java-based prototype, was designed to facilitate collabora-
tive analysis by providing a space for teams to record and share hy-
potheses, conjectures, and evidence. CLIP’s design (Figure 1) takes
into account the design guidelines outlined earlier. For example,
CLIP enables analysts to record externalizations in structured formats
(graph, timeline, and notes), addressing the need for externalization
and schematizing. CLIP also enables analysts to work individually but
merge their findings, supporting the principle of shared and individual
workspaces. CLIP facilitates validation of results by enabling people
to add evidence to each finding and by visually representing evidence
both around the nodes as well as in the evidence cloud. Most impor-
tantly, we implemented LCW to enhance awareness by revealing rela-
tionships between collaborators’ externalizations. Collectively, these
features help analysts to see who did what and follow the trail of how
other analysts came to a particular conclusion. To facilitate same-time
collaboration, CLIP supports awareness through user-controlled shar-
ing of work amongst team members, with colour coding to indicate
who did what. The specific scenario for which we designed CLIP is
to support team-based analysis of a document collection for solving
a mystery task. However, we envision that the design ideas in CLIP
could be applied to other analysis scenarios. Rather than directly vi-
sualizing the document collection, CLIP visualizes and links the team
members’ externalizations relevant to their analysis. Interesting enti-
ties (e.g., people, locations, or events described in the documents) can
be externalized as nodes, and relationships between events as links.
Each recorded entity can be optionally linked to free form text notes
and to a timeline. In our study, participants took the role of intelligence
analysts solving a mystery task from the VAST 2006 challenge. In the
following sections, we begin with a scenario related to our study task,
illustrating CLIP’s use. This is followed by a description of CLIP’s
features related to externalization and awareness support.
4.1 Scenario
Laura, Alex and Mary are reviewing a set of documents to solve a
mystery task. Laura has been focusing on a suspicious event at the
‘Silo’. From the article she finds that ‘George Prado’ is up to some il-
legal activity, maybe running a meth lab. She suspects ‘George Prado’
is the key person, so she records him as the main suspect and cre-
ates a node with his name. In addition, she creates a note containing
her hypothesis that “George is probably running a meth lab”. This
could also bring George Prado to her collaborators’ attention. Then
she starts gathering evidence to support or refute her hypothesis about
him. Later, she finds an interesting article about the ‘Silo’ and she cre-
ates a node for ‘Silo’. As soon as the node is created, the visual glyphs
on the node inform her that that Mary has also been investigating the
‘Silo’ event and has recorded information about it. From there, Laura
gets interested to see what else Mary has found so far. She opts to
merge Mary’s entire graph into her own view. She discovers that Mary
has found interesting relationships involving the ‘Silo’ event. Tracing
Mary’s work, she finds out that both Alex and Mary have collected
substantial evidence of the terrorist group’s links to George Prado.
Laura decides to also view Alex’s full work by merging it with her
own. Looking at notes made by Alex, she realizes that ‘George Prado’
is up to something bigger than running a meth lab. She opens a dis-
cussion about ‘George Prado’ running the ‘FFE farm’ and from there,
they connect the facts and validate their hypotheses. At the same time,
Mary records a new finding that shows George’s brother is a security
guard at the ‘Annex’, a chemicals warehouse. As soon as Laura and
Alex see this new finding in their own views, they start talking to Mary
and sharing all their findings so far; in this discussion they realize that
the Prado family are probably supplying chemicals to the terrorists.
4.2 Externalization
The importance and benefits of recording externalizations (notes, find-
ings, etc.) have been emphasized by many researchers (e.g., [14, 22,
27]). CLIP provides space for recording and visually representing im-
portant entities and relationships, the time order of events, and free
form text notes. Each recorded entity takes the form of a node in a
node-link graph (Figure1A). Each item is indicated by a unique colour
corresponding to the owner. Initially, each user logs into their instance
of CLIP by selecting a username and a colour. Evidence can be at-
tached to each node. According to previous research [25], returning to
original sources and checking the references can be very tedious and
time consuming. Attaching evidence to nodes in CLIP helps analysts
easily return to original sources to verify accuracy of reported find-
ings. Each node represents an entity and has six main attributes: text,
type, note, image, date, and evidence list. Only the text and evidence
list (at least one evidence document) are required for node creation;
other node attributes are optional and can be updated any time.
To add a new node to the graph, a user enters information in a popup
dialog (Figure 2). Each list in the dialog is pre-populated by the val-
ues that were extracted from the document corpus. To enter values
Fig. 1: Screenshot of CLIP. A) Graph pane, to create a network diagram of people, locations, and events, B) Timeline, to see the timeline of
events, C) List of notes, easy review of all notes in one location, D) Tabs, to see collaborators’ views, E) Merge option, to choose a collaborator’s
work to be merged with your own, F) Evidence cloud, to see the list of evidence and their frequencies, G, H) Filtering and sorting options.
that do not belong to any of these categories, there is a text box la-
beled “Other”. To assist users with data entry and improve the discov-
ery of common work, we implemented an auto-fill feature for this text
box that provides suggestions (name values such as chemicals, events,
etc.). A date stamp (null by default) can be attached to the node (Figure
2B). Any node with a non-null date stamp will automatically appear
on the timeline. A date stamp can be attached to any node type (i.e.
Person, Location, Organization, or Other). The rationale behind this
design was to enable users to associate entities of all types with time
if needed. Interestingly, in our study we observed that many partici-
pants associated people and location names with dates, perhaps as a
shorthand way to represent an event. To attach evidence to a node,
a user can select documents from the evidence list (Figure 2C). The
current design permits attaching up to twelve documents to a node as
evidence. The dialog also contains a text area for recording a note
(Figure 2D). Note content type (Hypothesis/Question/Finding/Other)
and scope (Public/ Private) can be set as well. To assist a user to easily
identify private versus public notes, a lock icon appears on the private
notes. Finally, an image can be attached to a new node (Figure 2E).
Fig. 2: Dialog for creating a new node. A) Select node’s text, B) Add
date, C) Add evidence, D) Free form note, E) Add image.
Each graph node has a toggle button (+ or -) that controls the vis-
ibility of the node’s children (if any). Collapsing a node collapses all
the branches that stem from the node. This feature makes it possible
to collapse/ expand the graph from any given node, improving scala-
bility. Users can filter nodes based on type (Figure 1G). This enables
them to hide parts of graph if required (improving the scalability) or
quickly locate nodes of specific types.
Links represent relationships between captured entities. Each link
has three main attributes: text, note and evidence list. These attributes
mirror those of the nodes. Unlike a node’s evidence list, a link’s evi-
dence list can be null. Figure 3 shows a node’s design. Node text is
placed in the middle of the circle. A yellow note icon above the text
indicates that a note is attached to the node (Figure 3A). If the evidence
list is not empty, there are segments drawn on the outside of the node
circle, one for each evidence document attached to the node. These
visual cues provide a quick overview of each node. Segments in a col-
laborator’s colour (Figure 3B and 3C) represent evidence found by a
collaborator, one of the ways we reveal LCW. By default, all notes are
Fig. 3: Node details. A) View of a node before a collaborator creates
the same node and B) after a collaborator creates the same node. Vi-
sual cues (colour coded segments) indicate it is a common node and
reveal the common and different evidence. C) Enlarged node to see the
details. Enlarging a node also highlights related items in other views,
such as notes in the note list, timeline items, and evidence.
placed in the note panel (Figure1C). Each note can be closed by the
user later. When closed, the yellow note icon (inside the related graph
node) changes to red as a visual indication. Users can sort and filter
notes by note type, time, and owner (Figure 1H). When a node with a
Date-stamp attribute is created, the system automatically places a box
with the node’s text on the timeline (Figure1B). Items on the timeline
are ordered chronologically from left to right and items with similar
date stamps are grouped.
The evidence cloud (Figure 1F) is a tag cloud of the documents
attached to nodes as evidence. Font size indicates the frequency of
attachment as evidence. Content of this view is based on information
included in the workspace. If a user includes all collaborators’ graphs,
the evidence cloud includes all evidence items across the group. This
view identifies documents that have been noted as relevant and reveals
document importance based on frequency.
The implementation of CLIP supports full coordination of all views.
When a node is enlarged to view details, the corresponding note and/
or timeline item (if present) are highlighted by fading out other items,
enabling the user to quickly identify related items. Similarly, selecting
a note or timeline item highlights corresponding items in other views.
Clicking on a document name in the evidence cloud highlights nodes
that contain the selected document as attached evidence.
4.3 Awareness Support
To support awareness of collaborators’ activities, instances of CLIP
that are running on different machines communicate in real-time to
share information. Distinctive colours are used to distinguish work by
different people, similar to Cambiera [21].
Partial merging: If another user has a node with the same name,
then the local node is changed to notify the user that there is similar
work. To keep changes in the local node subtle and yet noticeable,
the only visual alteration is in the evidence list decorated around the
node (colour coded segments). Evidence lists of the local node and the
collaborator’s node are combined, and repeating evidence segments
are stacked up. Figure 3 shows a node ‘George Prado’ before and after
a collaborator adds evidence. The colour of the local node is green and
the colour of the collaborator is pink. Figure 3A shows the node before
partial merging. Figure 3B depicts the same node after partial merging.
In this case, the collaborators have two evidence items in common
and other evidence items identified by only one person or the other.
In addition to these visual cues, CLIP automatically combines all the
collaborators’ notes related to the node (ordered chronologically by
default). By right clicking on a node, users can enlarge it (Figure 3C)
to reveal more detail. Enlarging a node automatically highlights all
related items (i.e., timeline items, notes, and evidence).
Tabs: Each tab in CLIP (Figure1D) encompasses a view of the anal-
ysis work in progress in another copy of CLIP. Tabs are labeled with
the collaborator’s colour and username to enable fast recognition of
who owns the work. Tabs show a node-link layout that is identical to
the node-link layout created by the owner of that information.
Full merging: Figure 1 depicts an example of a fully merged view.
The merged design enables the viewer to easily gain an understanding
of how their collaborators’ work relates to their own (e.g., what entities
their collaborators are interested in and why, and what evidence they
have found). Figure 1E is a list of collaborators’ names that can be
used to decide whose work to merge with your own. Checking the box
next to a collaborator’s name merges all of the collaborator’s work into
the local view. CLIP re-computes the graph layout and unites nodes
with the same name. The primary user’s layout is maintained as much
as possible in order to preserve their mental map.
5 USER ST UDY
We conducted a user study to gain a better understanding of LCW’s ef-
fects on collaborative sensemaking. We employed a between-subjects
experimental design to compare CLIP to a baseline tool (BT) with the
LCW features removed.
5.1 Participants
We recruited 48 participants (16 groups of three, 8 groups / condi-
tion), who were graduate and senior undergraduate students from var-
ied disciplines. To simulate work situations and create a comfortable
environment, group members were required to know each other and
have previous joint teamwork experience in a school or work project.
To mitigate the impact of using students, we targeted participants who
had some experience with data analysis. Participants’ age ranged from
20-60 (Avg=28). There were 37 grad students and 28 males out of 48.
We assigned groups randomly to each condition. Participants were
compensated with $20 each. To encourage active participation, we
provided a small financial reward for the team with the highest score.
5.2 Dataset and Scenario
We employed the “Stegosaurus” dataset from the VAST 2006 chal-
lenge [40]. This synthetic dataset contains approximately 240 docu-
ments, mostly news articles plus a few maps and supporting pieces
of information. The documents describe approximately 3000 entities.
The scenario involves finding a hidden chemical weapon production.
We chose Stegosaurus because it is a standard task to evaluate visual
analytics tools, and has been used in other studies [1,22, 35]. We were
also careful to select a task that represents a real life scenario but can
be solved by non-experts. The dataset contains a scenario that reveals
the first clue. From there, analysts are challenged to work through the
dataset and iteratively search and filter to find the ten most relevant
documents. Similar to real life scenarios, there are distractors that
could point analysts in the wrong direction. While dataset authors es-
timated that the plot could be solved in about 2-6 hours with standard
tools [40], we ended all of our sessions at 90 minutes as in [22].
5.3 Apparatus
Our experimental setup included two iMacs and one 17” MacBook
Pro, arranged as shown in Figure 4. Participants were collocated and
therefore could speak to each other and look at each other’s screen if
they wished. The physical arrangement was determined through pilot
studies where we experimented with several different options to find
an arrangement that was comfortable for participants. First we tried
arranging the group members within a U-shaped table, so they could
easily look at each other’s screens; however we noticed that they did
not have much discussion. Then we arranged them around a table with
three laptops to simulate most current work practices, but we received
many complaints about the small display size. This led us to the final
larger-screen setup. We expected participants to complain about the
screens blocking their views in this configuration, but they reported
that it was very practical. We compared CLIP against a Baseline tool
(BT). All participants in a group used the same version, either CLIP
or BT. BT was identical to CLIP, except that we removed the LCW
features (i.e., partial and full merging, as defined in section 4.3). We
emphasize that BT still contained some awareness features; specifi-
cally, collaborators could still examine each other’s work through tabs.
We kept the tabs because they are similar to what many systems pro-
vide currently; for instance, Jigsaw’s Tablet view [13] allows analysts
to take notes in schematic form but offers no collaborative options
to share between concurrent users. This approach also allowed us to
specifically investigate the effects of the LCW technique.
Fig. 4: User study setup and physical arrangement.
5.4 Procedure
We assigned groups to conditions at random. The procedure for both
versions was the same. We began with a tutorial on the system’s fea-
tures (15 minutes for BT, 20 minutes for CLIP). We asked participants
to try out the system’s features with a different sample data set. An
observer was present to answer their questions and help them to ex-
periment with all the features. Then, participants received background
information about the task and started by reading the scenario, which
provided the first clue. Documents were all digital, and all partici-
pants had access to all documents. Participants used Mac’s Spotlight
to search the text corpus. To search within a document, they used Mi-
crosoft Word’s search functionality. They recorded their results into
CLIP or BT. We ended the study whenever the teams were confident
and ready to present their results, or at 90 minutes, whichever came
first. Then we asked groups to write a report of their findings and hy-
potheses. Following the task, we conducted an open-ended interview
with each group to discuss the system’s features, their challenges, and
suggestions to improve the system.
5.5 Measures and Hypotheses
In this section we summarize measures and hypotheses related to
each of our research questions. We gathered data from five differ-
ent sources: videos, interaction logs, the final written report submitted
by each group, screen shots of visual elements created in CLIP or BT,
and notes taken by the observer. All the sessions, including debriefing
and interview, were audio and video-recorded. In total, we gathered
96 hours of video that includes the 16 groups’ analysis and follow-up
interviews. We used Transana [11] to analyze the videos and measure
the total conversation time for each group.
5.5.1 Performance
To measure performance (RQ1), we analyzed groups’ written reports.
Using the same scoring scheme as Isenberg et al. [22], groups received
positive points for facts they had connected (maximum of 11) and
negative points for wrong hypotheses. 11 was the maximum possible
score (i.e., all the facts were successfully discovered and connected)
and a negative score means that the group uncovered few facts and pro-
duced incorrect hypotheses. In addition, and similar to [22], we also
counted the number of discovered relevant documents as an indicator
of performance. Successful completion of the task was partly related
to participants’ ability to find the 10 most relevant documents in the
corpus and connect the facts within them. We analyzed the screen
shots and logs to obtain the number of relevant documents discovered
by each group. We hypothesized that CLIP groups would have better
results on performance measures, as follows:
•H1: CLIP groups will have higher task scores and find a greater
number of relevant documents than BT groups.
5.5.2 Communication, Coordination, and Awareness
We transcribed all the conversations to quantitatively measure com-
munication effectiveness (RQ2), coordination (RQ3), and awareness
(RQ4). Using an iteratively built coding scheme, we categorized each
instance of conversation. We define an instance of conversation as one
or more consecutive statements by a single individual. We chose to
code instances of conversation because other possible units, such as
sentences, are difficult to clearly delineate in oral conversation. The
coding scheme was comprised of seven different categories (DH, RV,
CO, SA, VF, QF, and RU). Table 1 depicts each code, along with
its definition and example. Conversations were coded as DH when-
ever group members were engaged in a discussion trying to connect
the facts and generate hypotheses. This was different from VF (ver-
balizing findings) when they were not actually connecting facts, they
were only stating findings that they found interesting. This usually
involved reading parts of a document out loud or reporting a sum-
mary of a finding. Referring to the visualization tool (RV) represented
instances where participants orally referenced visualization elements
such as nodes or notes. The seventh code, RU (Relevant but otherwise
uncategorized), was used for any instance of conversation that was re-
lated to the case but did not fit within any of the former six codes.
Sometimes a single instance reflected more than one code. For exam-
ple, there were instances when participants were referring to the visu-
alization and then they started to have a discussion about their findings
and tried to connect them together. We coded these instances as both
RV and DH. Other instances of double coding included RV and CO.
Therefore, counts of the codes are not mutually exclusive. Over 2800
instances of conversation were coded using the scheme. We did not
code conversations between group members and the experimenter.
Two independent coders coded the conversation data. We assigned
groups randomly to each coder. Each coded 10 groups (5 CLIP and 5
BT groups), with 4 overlapping ones. Inter-coder reliability was 0.91,
calculated using Krippendorff’s alpha.
Research [2] has shown that fully sharing the work across the group
can trigger discussions that are focused on solving the problem. Re-
ferring to the visualization also can enhance communication [17]. Be-
cause LCW should enable collaborators to more easily integrate their
findings and discuss a shared view of their externalizations, we ex-
pected that CLIP groups would discuss more facts and hypotheses
(DH) and refer to the visualization more often (RV):
•H2: CLIP groups will have more instances of DH and RV than
BT groups.
We coded coordination (CO) utterances as those where collaborators
tried to coordinate the group activities by dividing the task, documents,
the search (e.g.,“You search for flowers and I will search for apples”),
etc. According to prior research [12], we expected CLIP groups to
better coordinate their work. We argue that if the tool supports better
awareness, collaborators will be able to coordinate their actions at a
much lower level of granularity. That is, instead of simply doing a
high-level division of work at the beginning and then sharing findings
at the end, collaborators will be able to continually adjust their task
division as the work progresses. Therefore, we expected to see more
CO instances with CLIP than with baseline:
•H3: CLIP groups will have more instances of CO than BT groups
(because they will coordinate at a lower level of granularity).
In order to measure awareness, we coded conversations that were ba-
sically for seeking or sharing awareness about each other’s activities
and findings. For example, questions such as “Are you guys going
forward?” or “What have you found so far?” were coded as seeking
awareness (SA). Questions about another group member’s finding(s)
were coded as (QF), and verbalizing one’s own findings as a way of
sharing was coded as (VF). The rational behind this coding was that we
noticed baseline groups spent more time interrupting other members
to ask questions about findings or activities, and more time announc-
ing their findings out loud. These questions and verbalizations could
be easily eliminated if they could see each others’ findings at a glance
(the way they could see everyone’s results in a merged view in CLIP).
•H4: CLIP groups will have fewer instances of SA, VF, and QF
than BT groups (because they will be less reliant on the verbal
channel for awareness).
To further explore awareness, we analyzed responses to the inter-
view question about the extent to which participants were aware of
each others’ work. We also considered checkpoints, when in the mid-
dle of the session we stopped them and asked each individual to ex-
plain their findings and hypotheses. Then we asked them whether find-
ings of one group member were surprising to others.
6 RE SU LTS
In this section, we present both quantitative and qualitative findings of
the study including the usage statistics of CLIP’s main features.
6.1 Quantitative Findings
Table 2 presents scores achieved by CLIP and baseline groups, as well
as results of the communication analysis. It reports the number of
instances of discussion of hypotheses (DH), referring to the visual-
ization (RV), coordination (CO), seeking awareness (SA), verbaliz-
ing findings (VF) and asking questions about another group member’s
findings (QF).
6.1.1 Task Performance
CLIP groups achieved considerably higher scores than baseline
groups, strongly supporting H1. As shown in Table 2, scores of
CLIP groups ranged from 5 to 11 (Avg=8.25, SD=2), whereas baseline
Code Description Example
DH Having discussion or generating hypotheses “US government is supplying the rebels with Lewisite.”
RV Referring to the visualization tool “Link that with your apples. I will make a new node, linking Parazuelan.”
CO Coordinating the group “Let’s divide the work now, I will search for apples you look for flowers.”
SA Seeking awareness “What do you guys got?”
VF Verbalizing findings “Former farm worker Francisco Dorado formed Shining Future in 1988.”
QF Questions about findings of another group member “What did you find about apple bursting?”
RU Relevant but otherwise uncategorized “Oh okay found that article.”
Table 1: Communication coding scheme.
groups were from -2 to 7 (Avg=2.75, SD=2.8). The maximum possi-
ble score was 11. A two-tailed t test showed a statistically significant
difference between the average performance of CLIP and BT groups
(p<0.001). With the exception of group 3, all CLIP groups achieved
7 or higher. We believe the subpar performance of group 3 resulted
from their strategy: they spent considerable time organizing the data
chronologically before engaging in analysis.
With only one exception (G3, found 9 out of 10) all CLIP groups
successfully found the 10 most relevant documents. On the other hand,
only two baseline groups were able to find all of the relevant docu-
ments (G9 and G16). Even the top three ranked BT groups (9, 13 and
16) who found 10, 9 and 10 relevant documents respectively were not
able to connect all the facts. A two-tailed t test showed a statistically
significant difference (p<0.001) between the average number of rel-
evant documents found by CLIP (Avg=9.9, SD=0.4) and BT groups
(Avg=6, SD=3). Task time was not an important factor. We found no
correlation between scores and time (r2= 0.028), and no difference in
average time between the conditions (CLIP Avg= 87.6 min, SD=88,
BT Avg=86.8 min, SD=87), probably because the task was quite long
and most groups used up nearly all the available time.
Tool
CLIP
Avg
Group
12
5
8
6
15
1
11
3
-
Score
11
10
10
8
8
7
7
5
8
DH
185
127
124
131
123
116
102
88
116
RV
178
76
26
37
15
20
20
65
30
CO
57
15
23
16
10
10
20
11
15
SA
0
0
1
2
4
2
1
1
2
VF
15
10
1
11
4
2
6
7
5
QF
5
7
6
6
3
5
4
2
4
Baseline
Avg
9
13
16
10
4
14
2
7
-
7
6
5
2
2
2
0
-2
3
116
19
114
23
20
13
11
25
43
17
5
6
5
5
8
4
9
7
10
5
14
8
9
6
5
4
8
9
8
7
7
11
9
4
1
7
38
19
10
18
21
14
5
3
16
27
9
5
13
15
16
2
5
12
Table 2: Comparison of performance, communication and coordina-
tion of CLIP versus Baseline groups.
6.1.2 Communication
H2 predicted that CLIP would foster discussion of facts and hypothe-
ses (more DH). Our results strongly support this hypothesis (see Ta-
ble 2). A two-tailed t test showed a significant difference in the num-
ber of DH utterances between CLIP and BT groups (CLIP Avg=116,
SD= 28, BT Avg=43, SD=45, p<0.001). Although there was no sig-
nificant difference in the overall talking time between conditions, the
difference in DH means that CLIP groups had significantly more dis-
cussions about hypotheses and connections between facts.
H2 also predicted that CLIP groups would refer to the visualiza-
tion more often, and our results also confirmed this prediction. CLIP
groups extensively referred to the visualization tool (RV), significantly
more often than BT groups (CLIP Avg=30, SD=55, BT Avg= 7, SD=4,
p<0.001). We also observed that there was more discussion triggered
by the system in CLIP groups. This was mostly when participants re-
alized their teammate had done some related work. CLIP groups also
had fewer awareness seeking conversations (see section 6.1.4).
6.1.3 Coordination
H3 predicted more instances of CO in CLIP groups than in BT groups
(reflecting more detailed task division). We found a significant differ-
ence in the number of CO instances (CLIP Avg=15, SD=16, BT Avg=
8, SD=3, p<0.01). In relation to this, we also noticed many instances
where CLIP groups coordinated their work via the tool. To further an-
alyze the effect of the visualization tool on coordination, we looked
into RV examples that were double coded with CO. For instance, we
coded this as RV and CO: “Link my node with your apples and I will
make a new node to link Parazuela”. This is an example of coordina-
tion where collaborators deliberately connected their results through
the tool in order to solve the problem. We observed and recorded
many of these instances for CLIP groups.
6.1.4 Awareness
H4 predicted that using LCW would help collaborators to maintain
awareness of each other’s work with less reliance on verbal com-
munication. Conversation analysis strongly supported this hypothe-
sis. CLIP groups had significantly fewer awareness seeking utterances
(SA) (CLIP Avg=2, SD=1, BT Avg=7, SD=3, p<0.001). CLIP groups
reported that it was much easier to figure out who was doing what by
looking at the merged view. CLIP groups also verbalized their findings
significantly less than BT groups (CLIP Avg=5, SD= 5, BT Avg=16,
SD=11, p<0.04). There was a marginally significant difference in the
number of QF (CLIP Avg=4, SD=2, BT Avg=12, SD=8, p<0.06).
6.2 Qualitative Findings and Usage Statistics
Three primary awareness channels were available to participants: oral
communication, LCW (CLIP only), and tabs. To complement and
elaborate on our quantitative conversation analysis, in the following
sections we report qualitative observations and the results of our post-
task interviews for each awareness channel. In the interviews, all CLIP
users reported being aware of their collaborators’ work most of the
time. They all attributed this to use of LCW features, especially full
merging. They found partial merging cues to be an interesting notifi-
cation of common work that helped them to understand who else had
related results and evidence. However, all of the CLIP participants at-
tributed their awareness to full merging. Two CLIP groups (G6, G1)
indicated that showing collaborators’ notes was another important fea-
ture that helped them to maintain awareness of each others’ work. In
contrast, many baseline groups mentioned that they were not aware of
each others’ work. Five out of eight groups reported oral communica-
tion as their main awareness mechanism. The rest reported that their
awareness channels were oral communication as well as using tabs.
These results are consistent with our RV, SA and QF findings. Table 3
shows the usage statistics of CLIP’s main features other than LCW.
Tool
CLIP
BT
Node
20(6)
10(7)
Note
22(6)
14(8)
Link
12(5)
7(8)
Timeline
10(3)
6(4)
Tab
52(50)
71(40)
E. Cloud
15(5)
11(12)
Table 3: Usage statistics for CLIP and Baseline (AVG (SD)).
Investigation Externalization Awareness Discussion
Coordination
QH Formulation
Recording LCW increases
Facilitates
Increases
Facilitates
Increases
Facilitates
Leads to new
Initiates new
Directs new
Fig. 5: Collaboration model emphasizing the role of LCW in increasing awareness and discussion among team members. Awareness leads to
better coordination of activities and formulation of new questions/hypotheses, which in turn initiate and direct new investigation.
6.2.1 Oral Communication
CLIP groups’ oral communication focused heavily on discussion of
hypotheses, coordination and referring to the visualization. By con-
trast, for baseline groups, oral communication was the dominant
awareness channel, and without it, participants were not aware of each
others’ activities most of the time. For instance, one member of group
10 asked, “What are you guys doing, why you don’t talk?” and a mem-
ber of group 4 stated, “I felt I did not know what [Participant A] was
doing [she was silent most of the time]”. A similar result was reported
by Wallace et al. [39], who did not provide any form of thinking or
note space for participants in their study. We observed two key prob-
lems associated with communicating only through the verbal channel.
Sometimes sharing out loud disturbed others. One baseline participant
asked her teammate twice to be quiet. Her teammate was trying to
share his findings frequently and make sure that they were all aware
of each others’ work, but she wanted to focus. Instead she decided to
write comprehensive notes to share with the others. Another partici-
pant stated, “I couldn’t read my stuff when others were telling me what
they read, it was grabbing my focus”. The second problem with verbal
sharing was that if the information was not recorded, it could be easily
forgotten. Although speech was the fastest way to get updated about
others’ work, there were a few instances where key facts were shared
verbally, but later on, the group did not report those key facts in their
debriefing (and thus received a lower score than they might have). We
noticed this in particular for two CLIP groups (G1 and G6); session
logs showed that those key facts were never entered into CLIP.
6.2.2 LCW (CLIP only)
Participants reported merging individual work (all CLIP groups, 22
group members), LCW (6 out of 8 groups), and LCW of notes (6 out
of 8 groups) as the three most useful features of the system. CLIP
participants made extensive use of the LCW features, especially the
ability to merge everyone’s node-link graphs together. This could cut
some unnecessary communication, reduce redundancy, and let team
members focus on the task better. According to one CLIP participant,
“[Merging] made it faster because we knew what everyone was look-
ing at. We could go on the same direction or do something else...it
helped us to collaborate more closely, when you’re not paying atten-
tion that much, especially I said, ‘Hey, what about this?’ and someone
else is like, ‘We’ve already done this,’ and you can just look at their
graph. And I can connect my stuff to theirs”.
Looking at the most successful groups (12, 5 and 8, all using CLIP),
we observed that the strategies they had in common were a clear work
division and extensive use of CLIP’s LCW features, suggesting that
these were good predictors of success. Even though these groups
had different leadership styles, they constantly divided the workload.
According to the system log, these groups also made intense use of
CLIP’s features to coordinate their activities. They systematically
merged their partners’ work into their own view to link their work to-
gether. Participants reported that the merged view helped them find
important results in others’ work. It also inspired confidence and
helped them identify relevant keywords. Participants said, “I noticed
that some of my most powerful points... he also had them. I could see
the two colours on it. That gave me confidence,” and, “For common
nodes, I was looking at the evidence. If they were different from mine,
I was checking them as well. Common items made me confident and
helped me to keep going.” CLIP groups became quite dependent on the
shared node-link graph. For instance, in group 5, B was sharing her
findings with A, and A said “Put it down, create a node”. Later when
B complained that the team ignored one of her findings (“I found it
before and I told you!”), A explained why it had gone unnoticed, “Be-
cause you did not connect it to my node, so I did not look at it!”.
We were curious to see how participants in CLIP groups would
choose to use merging. Would they leave the default setting (partial
merging) to keep their workspace uncluttered and avoid the disruption
of constant updates from other participants’ changes? Or would they
choose to see everything? Answer: the latter. Most participants chose
to set merging on from the beginning and kept it visible until the end.
It was interesting that participants reported that oral communication
was disruptive, but CLIP updating the shared view was not. Instead,
participants reported merging to be useful for collaboratively explor-
ing the task, sharing important evidence, exchanging documents, and
reducing redundant work. During the interview participants empha-
sized that merging was one of the most useful features of CLIP.
Five groups (eleven participants) in the baseline condition actually
requested a merging feature that would put everyone’s information in
one view. For example, participants stated, “I was not able to make a
link to someone else’s work, so I could not make a connection,” and “It
is hard to remember what the others have registered by checking the
tabs, so we would like to be able to draw links between nodes created
by different people. It is also good to avoid redundancy.” Only one
participant reported a potential negative side of merging, stating, “It
was interesting, but it was a double-edged sword. It could help me or
push me [in a] correct or incorrect direction.”
6.2.3 Tabs and Notes
In addition to oral communication, most baseline groups also relied
heavily on tabs for awareness. Some groups, however, used tabs only
to quickly check what the others were working on; for example, “I
only looked at their tabs when I was trying to find something that they
have read. I just wanted to refer to their work but not for everything”.
Interestingly, tabs usage in CLIP was not much lower than BT (see
Table 3), even though they also used the merging feature. One CLIP
participant explained that she used tabs to see how other group mem-
bers arranged their nodes (because CLIP’s merge feature recomputed
the layout for each individual to maintain their mental map).
Notes were valued in both CLIP and BT. Collaborators’ notes were
accessible via tabs in both tools, and via merging in CLIP. Participants
stated that notes provided an overview, enabled them to remember why
they had created graph nodes, and allowed them to copy important
information from the documents. Several people reported that notes
helped them to identify interesting information belonging to others.
Participants in group 12 stated, “[C:] The notes on the side. I got most
info from them, to be honest. I would read the notes and go ‘Wow,
that’s cool!’ [B:] Yeah, other people were highlighting things that
you should read.” Similarly, another participant said, “When someone
didn’t write a good note, I didn’t look at what they were doing”.
7 DISCUSSION AND FUTURE WORK
LCW clearly supported groups in this collaborative sensemaking task.
CLIP groups achieved significantly better scores (H1), coordinated
and communicated more effectively (H2 and H3), and relied on LCW
to maintain awareness of each others’ work (H4). CLIP groups had
significantly more discussion about hypotheses and evidence (DH) and
were able to focus their oral communication on discussing the case
and coordination activities rather than using oral communication as the
main awareness channel. This research extends earlier work on LCW,
by establishing its value in the sensemaking loop of collaborative anal-
ysis, not just the information foraging loop, and demonstrating how it
can be applied to externalizations. CLIP also illustrates how the LCW
concept can be employed within a collaborative thinking space.
To better explain the effects of using CLIP on teams’ collaboration,
we derived a collaboration model for CLIP groups based on our re-
sults (see Figure 5). Similar to the model in [12], our model shows
how awareness plays a critical role that enhances communication and
coordination activities. From Figure 5, we can see that recording
externalizations and automatically sharing them via LCW increased
awareness. Increased awareness in turn enabled groups to coordinate
their work at a deeper level. Being able to see others’ results trig-
gered discussion and this assisted teams to better formulate their new
questions and hypotheses (QH Formulation). There are mutual effects
between awareness, coordination and discussion; i.e., each one will
influence the other. QH Formulation and coordination of activities ini-
tiate and direct new investigation. With BT, LCW was missing. In our
model in Figure 5, this means that the link between Externalization
and Awareness was effectively broken. Collaborators still maintained
some awareness through oral communication, but this mechanism was
less effective. Reduced awareness in turn had detrimental effects on
the teams’ coordination, discussion, and investigative activities.
While CLIP provided an effective thinking space for the intelli-
gence analysis task in our study, additional work would be needed to
extend it for more general use. To begin with, CLIP aims to support the
sensemaking loop, and therefore provides no explicit support for in-
formation foraging. Combining CLIP with a complementary tool like
Cambiera [21] may be an effective way to support both phases, which
may be especially crucial when dealing with a larger document set.
Awareness cues related to information foraging could present collab-
orators’ current activities (e.g., revealing that a collaborator is reading
a document or entering a search query). Another limitation of CLIP is
that it does not automatically identify entities or relationships between
documents; we used Jigsaw to extract this information for the purpose
of our study. We would like to integrate CLIP’s thinking space and
LCW features into a document analytics tool such as Jigsaw [37] that
automatically extracts entities and relationships.
Another way to extend CLIP, as suggested by one group, would
be to add visual indications that distinguish past work from current
changes (e.g. colour saturation to indicate node age). Although the
dynamics of the node-link graph show the evolution of the team’s
findings, it is not clear at any given time which changes are the most
current. Another interesting feature suggested by participants was to
create a summary evidence file to help with publication of the results.
Scalability is another important issue. In CLIP, collapsing a node
collapses all the branches that stem from the node, improving overall
scalability and flexibility. To scale this design to large and complex
problems, however, different visual representations might be needed.
The visual structures (e.g., node-link graph) may not scale well even
for individuals, and with multiple analysts, keeping track of collab-
orators’ changes and updates to such a large representation may be
impossible. We predict that the ‘share-everything’ strategy that was
successful for CLIP groups in our study might break down at a larger
scale. A variation of Branch-Explore-Merge [30] might reduce the
number of visual updates since they would only appear upon merging.
For small thinking spaces, this may be a significant disadvantage since
awareness notifications would be delayed. However, in large thinking
spaces, providing awareness notifications in such chunks may cause
less visual distraction and reduce the likelihood of small updates be-
ing missed. There are also other scalability issues in the current de-
sign. First, while the colour coding works well for small groups (our
target), it should be reconsidered for larger groups. Also, for the spe-
cific task used in this study, decorating evidence around a node was
enough. However, the design might need to change for larger datasets
with more evidence items. One possible way to improve scalability
could be to encode the quantity of evidence related to a node as the
node size. Similarly, the size of notes could adapt to their length.
One interesting question that arises with visual thinking spaces is
the potential that they may lead to group-think, a situation where the
group fails to consider possible explanations because they too quickly
follow one avenue of investigation. It is possible that sharing findings
through LCW may discourage a healthy level of independent analy-
sis. We do not have a good way to assess the level of group-think
in our study, in part because avenues of independent thought are nei-
ther easy to categorize nor measure. One possible approach to avoid
group-think could be to design tools that promote discussion of alter-
native hypotheses, perhaps by finding and highlighting disagreements
in the findings. Research into causes of group-think and mechanisms
to prevent it are an important area of future research. Nonetheless, the
much stronger performance of CLIP groups in our study indicates that
the awareness benefits of LCW outweigh costs such as group-think.
Future work should also examine the value of LCW in a field setting
with professional analysts. Our participants were students because it
is extremely difficult to find enough professional analysts for a lab
experiment. We took care to recruit participants with some data anal-
ysis experience and chose a task that did not require domain-specific
knowledge. Nonetheless, student behaviour will undoubtedly differ
from that of experts. For example, professional analysts might have
established coordination strategies and therefore be less reliant on tool
support for coordination. We would also like to explore how LCW
influences collaborative dynamics over a longer analysis period.
Another interesting future direction is to understand how CLIP
could be used on a shared screen (e.g., a wall or tabletop). Finally,
although we examined the value of LCW for collocated work, it might
have even greater value for distributed or asynchronous scenarios.
Maintaining awareness is generally more challenging in these situa-
tions because of the limited communication channels available to col-
laborators. LCW could play a critical awareness role in such situa-
tions, but this will need to be tested in future studies. It is quite pos-
sible that additional features will be needed (e.g., a more extensive
note feature that enables threaded discussions) when verbal and / or
non-verbal awareness communication channels are unavailable.
8 CONCLUSION
CLIP demonstrates how the concept of linked common work can be
employed within collaborative thinking spaces to support the sense-
making loop during collaborative analytics. CLIP provides an envi-
ronment for analysts to record, organize, share and connect results.
Moreover, CLIP extends earlier thinking spaces by integrating LCW
features that reveal relationships between collaborators’ externaliza-
tions to increase awareness among team members. Our user study
compared CLIP to a baseline version without LCW features. Results
demonstrated that LCW significantly improved analytic outcomes at a
collaborative intelligence task. Groups using CLIP were able to com-
municate and coordinate more effectively. They were able to use oral
communication primarily to discuss the task, generate hypotheses, and
coordinate their activities at detailed level, rather than employing it for
disruptive awareness notifications. Most importantly, LCW enabled
collaborators to maintain awareness of each other’s activities and find-
ings and link those findings to their own work.
9 ACKNOWLEDGEMENTS
We thank Leandro Collares and Wanda Boyer for their assistance with
data analysis, and the VisID group at UVic for their helpful sugges-
tions. Thanks also to Rock Leung and our other colleagues at SAP for
their feedback on system design and to Mark Whiting for sharing the
dataset. This research was funded by SAP, NSERC, and GRAND.
REFERENCES
[1] C. Andrews, A. Endert, and C. North. Space to think: large high-
resolution displays for sensemaking. In Proc. SIGCHI Conference on
Human Factors in Computing Systems, pages 55–64. ACM, 2010.
[2] A. D. Balakrishnan, S. R. Fussell, and S. Kiesler. Do visualizations im-
prove synchronous remote collaboration? In Proc. Human Factors in
Computing Systems, pages 1227–1236. ACM, 2008.
[3] A. D. Balakrishnan, S. R. Fussell, S. Kiesler, and A. Kittur. Pitfalls of
information access with visualizations in remote collaborative analysis.
In Proc. Computer supported cooperative work, pages 411–420. ACM,
2010.
[4] E. A. Bier, E. W. Ishak, and E. Chi. Entity workspace: an evidence
file that aids memory, inference, and reading. In Proc. Intelligence and
Security Informatics, pages 466–472. Springer, 2006.
[5] L. Bradel, A. Endert, K. Koch, C. Andrews, and C. North. Large high res-
olution displays for co-located collaborative sensemaking: Display usage
and territoriality. Int. J. Human-Computer Studies, 71(11):1078–1088,
2013.
[6] S. E. Brennan, K. Mueller, G. Zelinsky, I. Ramakrishnan, D. S. Warren,
and A. Kaufman. Toward a multi-analyst, collaborative framework for vi-
sual analytics. In Proc. Visual Analytics Science And Technology (VAST),
pages 129–136. IEEE, 2006.
[7] Y. Chen, J. Alsakran, S. Barlowe, J. Yang, and Y. Zhao. Supporting ef-
fective common ground construction in asynchronous collaborative vi-
sual analytics. In Proc. Visual Analytics Science and Technology (VAST),
pages 101–110. IEEE, 2011.
[8] G. Chin Jr, O. A. Kuchar, and K. E. Wolf. Exploring the analytical pro-
cesses of intelligence analysts. In Proc. Human Factors in Computing
Systems, pages 11–20. ACM, 2009.
[9] M. C. Chuah and S. F. Roth. Visualizing common ground. In Proc. Int.
Conf. on Information Visualization, pages 365–372. IEEE, 2003.
[10] H. H. Clark and S. E. Brennan. Grounding in communication. Perspec-
tives on socially shared cognition, 13(1991):127–149, 1991.
[11] C. Fassnacht and D. Woods. Transana (version 2.20)[computer soft-
ware]. Madison, WI: School of Education at the University of Wisconsin-
Madison, 2007.
[12] V. L. Gava, M. d. M. Spinola, A. C. Tonini, and J. C. Medina. The 3c
cooperation model applied to the classical requirement analysis. JISTEM-
Journal of Information Systems and Technology Management, 9(2):235–
264, 2012.
[13] C. G ¨
org, Z. Liu, and J. Stasko. Reflections on the evolution of the jigsaw
visual analytics system. Information Visualization, 2013.
[14] D. Gotz, M. X. Zhou, and V. Aggarwal. Interactive visual synthesis of
analytic knowledge. In Proc. Visual Analytics Science and Technology
(VAST), pages 51–58. IEEE, 2006.
[15] C. Gutwin and S. Greenberg. Design for individuals, design for groups:
tradeoffs between power and workspace awareness. In Proc. Computer
supported cooperative work, pages 207–216. ACM, 1998.
[16] A. H. Hajizadeh, M. Tory, and R. Leung. Supporting awareness through
collaborative brushing and linking of tabular data. Visualization and
Computer Graphics, IEEE Transactions on, 19(12):2189–2197, 2013.
[17] J. Heer and M. Agrawala. Design considerations for collaborative visual
analytics. Information visualization, 7(1):49–62, 2008.
[18] J. Heer, F. van Ham, S. Carpendale, C. Weaver, and P. Isenberg. Creation
and collaboration: Engaging new audiences for information visualization.
In Information Visualization, pages 92–133. Springer, 2008.
[19] M. S. Hossain, C. Andrews, N. Ramakrishnan, and C. North. Helping
intelligence analysts make connections. In Proc. Scalable Integration of
Analytics and Visualization, 2011.
[20] P. Isenberg, N. Elmqvist, J. Scholtz, D. Cernea, K.-L. Ma, and H. Hagen.
Collaborative visualization: definition, challenges, and research agenda.
Information Visualization, 10(4):310–326, 2011.
[21] P. Isenberg and D. Fisher. Collaborative brushing and linking for co-
located visual analytics of document collections. In Computer Graphics
Forum, volume 28, pages 1031–1038. Wiley Online Library, 2009.
[22] P. Isenberg, D. Fisher, S. A. Paul, M. R. Morris, K. Inkpen, and M. Cz-
erwinski. Co-located collaborative visual analytics around a tabletop dis-
play. Visualization and Computer Graphics, IEEE Trans., 18(5):689–702,
2012.
[23] E. Kandogan, J. Kim, T. P. Moran, and P. Pedemonte. How a freeform
spatial interface supports simple problem solving tasks. In Proc. of the
SIGCHI Conference on Human Factors in Computing Systems, pages
925–934. ACM, 2011.
[24] Y.-a. Kang, C. Gorg, and J. Stasko. Evaluating visual analytics systems
for investigative analysis: Deriving design principles from a case study. In
Proc. Visual Analytics Science and Technology (VAST), pages 139–146.
IEEE, 2009.
[25] Y.-a. Kang and J. Stasko. Characterizing the intelligence analysis pro-
cess: Informing visual analytics design through a longitudinal field study.
In Proc. Visual Analytics Science and Technology (VAST), pages 21–30.
IEEE, 2011.
[26] K. Kim, W. Javed, C. Williams, N. Elmqvist, and P. Irani. Hugin: A
framework for awareness and coordination in mixed-presence collabora-
tive information visualization. In ACM Intl. Conference on Interactive
Tabletops and Surfaces, pages 231–240. ACM, 2010.
[27] N. Mahyar, A. Sarvghad, and M. Tory. Note-taking in co-located collab-
orative visual analytics: Analysis of an observational study. Information
Visualization, 11(3):190–204, 2012.
[28] N. Mahyar, A. Sarvghad, M. Tory, and T. Weeres. Observations of record-
keeping in co-located collaborative analysis. In Proc. Hawaii Int. Conf.
on System Sciences (HICSS), pages 460–469. IEEE, 2013.
[29] G. Mark and A. Kobsa. The effects of collaboration and system trans-
parency on cive usage: an empirical study and model. Presence: Teleop-
erators and Virtual Environments, 14(1):60–80, 2005.
[30] W. McGrath, B. Bowman, D. McCallum, J. D. Hincapi´
e-Ramos,
N. Elmqvist, and P. Irani. Branch-explore-merge: facilitating real-time
revision control in collaborative visual exploration. In Proc. Interactive
tabletops and surfaces, pages 235–244. ACM, 2012.
[31] D. R. Millen, A. Schriefer, D. Z. Lehder, and S. M. Dray. Mind maps
and causal models: using graphical representations of field research data.
In CHI’97 Extended Abstracts on Human Factors in Computing Systems,
pages 265–266. ACM, 1997.
[32] T. P. Moran, W. Van Melle, and P. Chiu. Spatial interpretation of domain
objects integrated into a freeform electronic whiteboard. In Proc. ACM
Symp. User Interface Software and Technology, pages 175–184. ACM,
1998.
[33] S. A. Paul and M. C. Reddy. Understanding together: sensemaking in
collaborative information seeking. In Proc. Computer Supported Coop-
erative Work, pages 321–330. ACM, 2010.
[34] P. Pirolli and S. Card. The sensemaking process and leverage points for
analyst technology as identified through cognitive task analysis. In Proc.
Int. Conf. on Intelligence Analysis, volume 5, pages 2–4, 2005.
[35] J. M. R. and H. Kasper. Collaborative work on a high-resolution multi-
touch wall-display. Trans. Computer-Human Interaction, 21(2):340,
2014.
[36] A. C. Robinson. Collaborative synthesis of visual analytic results. In
Proc. Visual Analytics Science and Technology (VAST), pages 67–74.
IEEE, 2008.
[37] J. Stasko, C. G ¨
org, and Z. Liu. Jigsaw: supporting investigative analysis
through interactive visualization. Information visualization, 7(2):118–
132, 2008.
[38] K. Vogt, L. Bradel, C. Andrews, C. North, A. Endert, and D. Hutchings.
Co-located collaborative sensemaking on a large high-resolution display
with multiple input devices. In Proc. Human-Computer Interaction, IN-
TERACT 2011, pages 589–604. Springer, 2011.
[39] J. R. Wallace, S. D. Scott, and C. G. MacGregor. Collaborative sense-
making on a digital tabletop and personal tablets: prioritization, com-
parisons, and tableaux. In Proc. Human Factors in Computing Systems,
pages 3345–3354. ACM, 2013.
[40] M. A. Whiting, C. North, A. Endert, J. Scholtz, J. Haack, C. Varley, and
J. Thomas. VAST contest dataset use in education. In Proc. Visual Ana-
lytics Science and Technology (VAST), pages 115–122. IEEE, 2009.
[41] W. Willett, J. Heer, J. Hellerstein, and M. Agrawala. Commentspace:
structured support for collaborative visual analysis. In Proc. Human Fac-
tors in Computing Systems, pages 3131–3140. ACM, 2011.
[42] W. Wright, D. Schroh, P. Proulx, A. Skaburskis, and B. Cort. The sand-
box for analysis: concepts and methods. In Proc. Human Factors in
computing systems, pages 801–810. ACM, 2006.
[43] J. Zhang. The nature of external representations in problem solving. Cog-
nitive science, 21(2):179–217, 1997.
