Houda Lamqaddam is a PhD candidate in Information Visualization and Art History in KU Leuven, Belgium. Her research interests revolve around the usability and relevance of information visualisation techniques for art historical data. She explores the use of humanist theory and practice to inform visualisation design, starting from data interrogation and critical source evaluation, to existing traditions of visual design languages and conventions.
Inez De Prekel (1991) received her MA in Art History in June 2017 at the University of Leuven (KU Leuven). She then joined Project Cornelia as a PhD student. Inez studies the interplay between social dynamics and artistic developments in the seventeenth-century Antwerp art world, with a focus on a group of Antwerp tenebrist painters.
Koenraad Brosens is a Professor in the Art History Department of the University of Leuven (KU Leuven) and Vice-Dean of Education at the Faculty of Arts. He has published widely on Flemish and French tapestry and is PI of Project Cornelia (projectcornelia.be). Koen was visiting professor at the University of Pennsylvania Philadelphia (2007), the Peter Paul Rubens Chair at the University of California Berkeley (2013), and The J.P. Getty Museum Scholar (2019). In 2020 he received the Pioneer Award of the Humanities & Social Sciences Group of the KU Leuven
Katrien Verbert is an Associate Professor at the HCI research group of the KU Leuven, Belgium. Her research interests include recommender systems, visualization techniques, visual analytics, and applications in healthcare, learning analytics, precision agriculture and digital humanities.
This is the source
Network representation is a crucial topic in historical social network analysis. The debate around their value and connotations, led by humanist scholars, is today more relevant then ever, seeing how common these representations are as support for historical analysis. Force-directed networks, in particular, are popular as they can be developed relatively quickly, and reveal patterns and structures in data. The underlying algorithms, although powerful in revealing hidden patterns, do not retain meaningful structure and existing hierarchies within historical social networks. In this article, we question the foreign aspect of this structure that force-directed layout create in historical datasets. We explore the importance of hierarchy in social networks, and investigate whether hierarchies - strongly present within our models of social structure - affect our perception of social network data. Results from our user evaluation indicate that hierarchical network representations reduce cognitive load and leads to more frequent and deeper insights into historical social networks. We also find that users report a preference for the hierarchical graph representation. We analyse these findings in light of the broader discussion on the value of force-directed network representations within humanistic research, and introduce open questions for future work in this line of research.
We explore the importance of hierarchy in social networks, and investigate whether hierarchies - strongly present within our models of social structure - affect our perception of social network data
Networks are some of the most commonly found visualizations in digital humanities research. Used to represent relationships between persons or concepts, this type of graphs has also been implemented through various layouts to better support specific tasks and users.
The wide availability of layout algorithms has multiple advantages for researchers who aim to analyze social networks. Force-directed layouts such as the Gephi Atlas drastically simplify the visualization – therefore analysis – of large network data. They reveal patterns and clusters by using complex physical models, all with limited need for user intervention or mathematical expertise. These models excel in domains where datasets are large, and where the goal is to identify the structure – or shape – of the dataset.
However, the widened access to these automated graph drawing algorithms drives us to
interrogate their fitedness for the analysis of humanistic material. On the one hand,
force-directed layouts visually remove existing structures in the data, even when
these structures are meaningful to users and their daily tasks. On the other,
humanistic datasets are often small, but do consist of meaningful inherent clusters,
distances and relationships. Such a mismatch leads to representations that can be at
odds with users' internal perceptions of the data structure. For instance, family
members in a correspondence graph are likely to be spread apart. Similarly,
non-contemporary actors can be placed side to side if they share a common
correspondent. This inconsistency between mental models and visual representation can
become a challenge for users. More critically, it has the potential to impact the
adoption of visualization tools as a whole as research suggests usability and user
trust, rather than scalability, are main obstacles to adoption of digital tools in
the humanities
In this paper, we address this challenge by introducing
Our initial results reveal that force-layered graph layout outperforms force-directed layout in terms of number and depth of insights, cognitive load reduction, and user preference. We discuss these findings along with implications for design and future directions for this line of research. Specifically, we interrogate the artificial structure these layouts create, and evaluate whether their disconnect with existing data hierarchies affects user understanding, preference, and experience.
Discussion about the semantics of visual design elements is often neglected within
technical scholarship on information visualization. In the case of hierarchy, we
believe this discussion to be critical for two reasons. First, because
Human-Computer-Interaction research has shown the importance of the consistency
between users’ mental models of information and its digital representation
First, we illustrate the difference between the two, through the example of two
visualisations of historical social networks as can be seen in Figure 1. In Figure
1a, a hand-drawn illustration displays the genealogy of the last Byzantine emperor
drawn as a collection of trees. This visualization has a strong spatial organization
based on family lines, roles, and generational hierarchy. In fact, tree structures
can be found in multiple historical representations of networks because they mirror
the conception of lineage as a branching linear connected structure. In Figure 1b, we
see a contemporary representation of a historical network
To describe them formally, hierarchical trees (such as Figure 1a) are a sub-category
of networks that are acyclic (i.e. there are no loop encountered during a traversal
of the tree) and directed (i.e. the relationships described by the edges are
one-directional, rather than reciprocal). The hierarchical aspect signifies a certain
categorization of the nodes so that those placed on the same level hold a similar
value or meaning. Hierarchical trees are a spatial organization whose traces date
back to antiquity, and which can be found in a variety of knowledge domains. As
described by Drucker in her book Graphesis
On the other hand, force-directed networks (Figure 1b) are mathematical products,
where physical simulations ran multiple times on a dataset model attraction and
repulsion between nodes, therefore automatically placing nodes with high connectivity
close to one another, or at the opposite, resulting in disconnected nodes being
placed far from each other, highlighting their disconnect. For Ahnert et al.,
force-directed networks reject hierarchies
, which become hidden and difficult to
notice, even if they were strongly present within the dataset
In the context of humanistic research, however, the intentional organization of network data has an undeniable value, when compared to the visual organization that results of mathematical algorithms. Drucker explains it clearly:
The spatial distribution of network diagrams, topic maps, and other graphical expressions of processed text or intellectual content is often determined by the exigencies of screen real estate, rather than by a semantic value inherent in the visualization. This introduces incidental artifacts of visual information. A point in a graph may be far from another because of a parameter in the program that governs display, rather than on account of the weight accorded to the information in the data set. The argument of the graphic may even be counter to the argument of the information, creating an interpretative warp or skew, so that what we see and read is actually a reification of misinformation.
In that way, network visualizations are no different than other visualizations of
data, as they themselves can reproduce existing biases, and cement them as neutral
knowledge. Porras asks us to keep our eyes open to the dangers of network
visualizations as they can reinscribe historic and contemporary power differentials
In visualization literature, a growing voice argues for the consideration of
semantics and connotation in the visualization design, especially within humanistic
scholarship. In previous work
To build on the previous discussion, we wanted to investigate the effects that hierarchical trees and force-directed graphs have on user perception. Because the literature on mental models in visualization shows link effects on efficiency and user experience, we developed an extensive user evaluation protocol where we compare the two representations and investigate the perceptual impacts of each one. In the next section, we describe the process we followed, and the findings we can extract from it.
Through this process, we aimed to answer the following research questions:
In order to perform a fair comparison, we developed two visualizations showing both hierarchical and standard force-directed visualizations of historical social networks. We developed a prototype for hierarchical tree layout visualization using the Javascript visualization library D3. We chose to develop this algorithm on top of the existing force-directed layout algorithms for multiple reasons. First, the existing force-directed layout algorithms are currently mature and deliver high-quality performance. Perhaps as a consequence, these layouts are supported by major graph drawing libraries and toolkits, which opens future possibilities for wider implementation of hierarchical graph layouts. Force-directed layouts also include parameters to reduce node collision, occlusion, and bring in useful results, such as bringing together highly connected clusters, which, while we do not want it to break inherent hierarchy in the data, still proves to be useful in reducing edge-crossing.
We designed and implemented an algorithm that traverses a force-directed graph and
spatially organizes it based on a user-defined parameter. The full algorithm
description and code can be found on our GitHub page.
Figure 2 shows the same dataset represented as
both a force-directed graph (left), and a hierarchical network (right). We call the
graphs produced by our hierarchical layout
In this section, we describe the evaluation protocol used to answer the RQs described
in section 3. In HCI, user evaluations are a standard way to get feedback and to get
better insight into users’ needs. It is also a common way to evaluate usability and
identify obstacles to adoption of digital tools. Since usability and adoption are a
recurrent issue in DH research, there is real value to be gained from applying such
methods to pinpoint exact usability issues in collaboration with the intended
audience. In this section, we describe each step of the methodology together with the
results found. For each subsection, the full table of results can be found in the
annex. In this section, we sometimes use the notations
Participants. We recruited 15 participants (12 women, 3 men) with background in Art History and Digital Humanities, most of them graduate students in the university’s Digital Humanities MSc program. Participants had on average 3.6 years of experience in their domain, and 1.2 years of experience using digital tools to support their research or study practice. Among participants, most digital tool usage within field of study consisted of digital archives (11 out of 15), spreadsheets (10 out of 15), and graph visualization tools (9 out of 15).
Data. We used data extracted from the Cornelia
Graphs. Figure 3 shows the two network visualization used in this study. We used the same dataset for both graphs not to introduce bias. However, to minimize learning effects, we manually changed the names so that the similarity of the two networks would not be obvious.
Understanding users' existing mental models (i.e. the representation they
internally make of specific information structures or processes) is key if we aim
to support similar structures in building visualizations. Previous work in the HCI
field has demonstrated the value of asking participants to draw their mental
models as a way to externalize them
Before showing participants the graphs, we first started by asking each one to
freely draw a social structure of their choosing, in order to elicit their mental
mapping of social structures. They were invited to sketch out any fictional or
existing social structure network (from stories, movies or books, or their own
personal life
We found that in all
Familiar structure, or expression of an internal mental model? While we can use the results from this elicitation task to induce the existence of a ‘natural’ hierarchical mental model of such data for the users, the exposure to family-tree type representations is an equally likely explanation for the phenomenon. Family trees are an old form of representations that are commonly seen by the public [Mitchell, 2014]. They are also commonly used in (art) historical teaching, which makes our participants even more likely to have built familiarity towards them. In reality, our participants have likely interacted with similar representations in the past, and may have defaulted to recreating a similar structure. We consider that these results do inform us of the internal model within the participants, itself caused by either inherent hierarchization of this data, or unvoluntary training caused by previous exposure.
We asked participants to explore the dataset in each condition while vocalizing
their thought process. Similar to the mental model elicitation approach, this
method allows to make the thought process of participants explicit. We then
captured and transcribed participants think-aloud process and analyzed it using a
thematic analysis to uncover patterns of insight
We found that the force-layered visualization triggered more interpretativehe dies in 1668
or this person has children.
Interpretative insight includes it's just
close family I see,
This person has children, but not with this one I
think.
Finally, in reflective insight, we included comments about the data
quality (I got the feeling this [dataset] is more documented than the other
one
) and emotional engagement with the data (oh they had *many* children
*laughs*
).
It's too bad you can't see more about the person, you can't tell for example
what their profession is,
I don't know the person, but I imagine if I were
researching these people, it can be really interesting to see
[This person]
). Finally, we found a clear divide in the number
of negative emotions such as overwhelmingness or defeat in the condition of
force-directed network usage (There are a lot of persons, and
here a lot of events. I’m trying to see who’s linked to who but it’s really
difficult here.
).
To complement the open nature of the think-aloud exploration, we conducted a more
structured exploration task where we asked participants to answer specific
questions about the social dynamics portrayed in the visualization. We designed
these questions based on the existing literature on graph exploration tasks
,
) and difficult questions (
, Describe the network at a specific year
We assessed the answers to the exploration tasks based on complete and partial correctness. We found that for easy questions, a similar number of participants (F-D: n=11; F-L: n=10) gave correct answers with both conditions. However, in the force-directed condition, half of these answers were incomplete. To the difficult questions, more participants gave correct answers using the force-layered condition (n=6). Notably, however, none of the participants were able to answer the same questions using the force-directed condition completely and correctly. The number of participants who gave incomplete answers was the same for both conditions.
In order to measure how the perception of the graphs compared with their actual
characteristics, we collected participants recollection of some quantitative graph
characteristics such as network size, and number of generations. We also asked
about the perceived complexity of the social structure. Both networks contained
the same number of nodes thirty-eight nodes. We found that participants perceived
the force-layered graph to contain
Finally, a majority of participants reported feeling that the social structure displayed in the force-layered condition was of a simpler nature one than the force-directed. This corroborates the difference in perceived graph size, as participants perceived networks in the force-layered condition to have fewer nodes.
To assess recall, we aimed to see how well participants remembered the graph in
its entirety. We then asked them to draw the graph to their best recollection
after having diverted their attention for a brief moment
We found differences in specific metrics between the recall sketches of the two versions, specifically on the identification of bridge nodes, as well as the retention of the multiple types of relationships. In the force-directed recall sketches, participants have retained the “dog-bone” structure of the graph consisting of two large clusters representing the two main families and tied together through a bridge node. Most participants also retained - at least partially - the lower antennas formed by the two loosely connected pairs nodes. In the force-layered condition, the drawings have retained the left-to-right organization of the nodes, as well as the two distinct subgraphs representing the two families. The bridge node bringing these two groups together was not systematically included. Finally, participants retained the existence of multiple types of relationships better in the force-layered condition, as they recalled more of the godparent type links. Example drawings from participants for each condition can be seen in Figure 5.
While evaluating performance can provide a good understanding of efficiency
achieved using a system, the mental effort required is not answered by these
approaches. We therefore measure participants' cognitive load
As a measure for cognitive load, we used a raw NASA-TLX questionnaire [Hart, 1998]
The NASA-TLX questionnaire is a widely used metric for measuring cognitive load in
the HCI field. It consists of a self-reported scale capturing workload across six
dimensions: Mental, Physical, and Temporal Demands, Frustration, Effort, and
Performance [Hart, 1998]. Participants are asked to score each of these variables
on a 0 — 100 scale where a lower value represents a lower demand
Comparing the raw NASA-TLX scores, we find that the force-layered visualization (average score of 26.3) was found to be less demanding than the force-directed graph (average score of 35.78), the metrics with most variance between conditions being mental demand, effort, and frustration (see Figure 6).
Finally, we performed a preference-based ranking. Participants were asked to subjectively rank the force-directed and the force-layered graphs on a Likert scale on aspects such as perceived suitability to task and data, or visual appeal.
Figure 7 shows the average results of the preference rankings for each category. Globally, participants ranked the force-layered condition higher than the force-directed condition in suitability to both data and tasks. They indicated that the force-layered condition needed less effort to understand than the standard one, and that they had higher confidence in their answers. Finally, the force-layered graphs also scored higher in visual appeal, and overall preference.
Based on a task of mental model elicitation, we found that participants overwhelmingly relied on generational hierarchical structure to draw familial social networks. This can be either because of inherent structure perceived in the data, or association of this type of dataset with traditional ancestry tree representations.
When exploring two different versions of a social network where one is layered
hierarchically and the other one isn’t, we found that participants expressed more –
and deeper – insights in the case of the force-layered visualization. They also
engaged more critically with the dataset, for instance by pointing out
inconsistencies. The force-directed version, however, drew more comments and
descriptions about the visualization features, suggesting that participants engaged
more with the
Concerning recall, we found that participants retained more of the significant clusters with the force-layered condition, although they retained more of the bridge nodes after using the force-directed version. In both cases, the bridge nodes were placed in the center of the visualization, so saliency does not explain the difference in these results. We suggest that can be caused by the topology of each graph. In the force-layered version, the ties among each family create two horizontal substructures that become the major ground of the visual, while the dog-bone structure in the force-directed layout represent the major part of the ground figure, and the bridge node is an essential part to its structure. Finally, using a raw NASA-TLX questionnaire and a preference ranking, we found that participants found the force-layered visualization less demanding cognitively. They also judged it to be more suited to the tasks and data, and more visually appealing.
As we analyze these findings in conversation with the literature on visualization in humanistic research, we identify emerging discussion points:
More research is needed to assess the effect of other types of structure in network visualizations. Once these effects are clearly understood, it would be interesting to evaluate how novel graph layout such as our FL algorithm can be introduced to existing visualization packages. It would also be interesting to make them customizable to specific needs outside of family and generational structures. Indeed, an important implication of this work is that any singular off-the-shelf technique cannot be sufficient for a solid support of DAH research. Rather than promote the use of a specific visualization technique over another, we advocate for researchers access to a variety of representation techniques that can accompany different tasks, data types, and expertise levels.
Finally, the generalizability of our discussion to other humanistic fields is an additional avenue for future research. Since this study was mainly focused on an art historical research, our initial results cannot be readily extrapolated to different domains. However, similar challenges of historical social networks exist for humanist researchers across disciplines. Future research is needed to evaluate how hierarchy in historical social network visualizations affects different research fields depending on their unique needs and constraints.
In this article, we explored the perceptual effect of hierarchy in visualizations of historical social networks. We have found, based on a user-centered approach, that humanist participants had a hierarchical mental model of social networks, indicating that there could be advantages to representing social networks in a similarly hierarchical structure. Our study confirmed this hypothesis, as we found that users reported lower cognitive load, more frequent and deeper insights, as well as a significant preference for the hierarchical representations. Despite these results, we nuance these findings by calling attention to the fact that force-directed graphs can carry meaning in their generated topology, allow users to make conclusions on tightness of networks, clusters, and question the accepted dogmas in terms of community structure. However, because social networks are so critical to certain humanistic research domains, the perceptual benefits to hierarchically structured layouts should not be overlooked. As graph-drawing software is often the most accessible way for researchers to gather quick visual overview, we suggest that these layouts should be more commonly supported as alternatives to allow scholars to have multiple perspectives on their data.
By these findings, we hope to contribute to the necessary critical discussion
nuancing the role of network visualization in humanistic research. In the future, we
would like to broaden the scope of this study by investigating the effects of
hierarchy for other types of historical (social) networks
This research is part of Project Cornelia, a project funded by the University of Leuven (KU Leuven) and the Flemish Fund for Scientific Research-Belgium (FWO-Vlaanderen). We thank our colleagues from AugmentHCI (Department of Computer Science, KU Leuven) and Project Cornelia (Faculty of Arts, KU Leuven) who gave feedback and insight into this study. We are also deeply grateful to the entire NA+DAH team for supporting and encouraging this line of research, and particularly to Dr. John Ladd (Denison University) for reviewing and strengthening this manuscript.