Affective Color in Visualization

Lyn Bartram Abhisekh Patra Maureen Stone

Simon Fraser University Simon Fraser University Tableau Software
Surrey, BC, Canada Surrey, BC, Canada Seattle, WA, USA
[email protected] [email protected] [email protected]

ABSTRACT in professional craft and is qualitatively rather than empiri-

Communicating the right affect, a feeling, experience or emo- cally validated. Color psychology has proven the connection
tion, is critical in creating engaging visual communication. between individual colors and affect, but to date there are
We carried out three studies examining how different color no studies in how combinations of colors (palettes) may con-
properties (lightness, chroma and hue) and different palette vey different affect in the limited scope used in categorical
properties (combinations and distribution of colors) contribute information visualization.
to different affective interpretations in information visualiza-
Unlike the rich palettes available to design applications, map-
tion where the numbers of colors is typically smaller than
ping colors to categorical data (the most common use of color
the rich palettes used in design. Our results show how color
in visualization) introduces two important constraints. First,
and palette properties can be manipulated to achieve affective
the usable scale of the palette is small: typically, 5-10 col-
expressiveness even in the small sets of colors used for data
ors [42]. Second, the colors need to be strongly perceptually
encoding in information visualization.
distinct [39]. We are interested in validated computational
models of how color relates to desired affect in these con-
ACM Classification Keywords
strained palettes.
H.5.m. Information Interfaces and Presentation: Misc-Colors
We had two basic research questions. First, can we relate
Author Keywords certain affective impressions to properties of a color palette in
Affective Visualization; color Perception; Design abstract images such as simple visualizations? And second,
what might this mean operationally for the design of affective
INTRODUCTION color palettes that are useful in information visualization?
We react emotionally as well as cognitively to visual imagery, We note an important point here. This work asks: given a
and those emotions influence both how we use the information desired affect, what palette colors should you choose? It does
presented to us and how we are affected by its presence in not answer the broader and much harder question: given an
our visual environment [34]. These systems are not indepen- arbitrary palette, what affect does it convey? Our goal is not
dent; emotion can result from cognitive reasoning, and affect to absolutely link color and affect for all cases. Instead, we
influences cognition [34]. Affect matters in visualization for propose that simply being able to recommend statistically
communicative intent, engagement, and storytelling [6, 15]; likely palettes for a particular affect is useful to visualization
there is evidence it supports problem solving [13]. While there practitioners.
is a long history of research and practice in how certain visual
elements relate to affect, there is yet no framework of affective In this paper, we report three studies examining how differ-
principles for visualization. The goal of our research is to ent properties of color palettes were associated with specific
examine the affective capacity of visual features such as color affective impressions. Our results show relations between per-
as part of this emerging framework. ceptual color properties (hue, chroma and lightness), palette
composition (hue clusters, color frequency) and certain types
Color palettes play a central role in data visualization where of affect. This research makes the following contributions.
they are frequently used to map categorical attributes for ef- First, our results affirm the potential of color for conveying
fective discrimination and identification [39, 5]. Principles for meaning and identify initial palettes that enhance these af-
using color to represent data in visualization are well estab- fective impressions even in the limited space of information
lished in the field and empirically validated [39, 5]. Designers visualization palettes. Second, our findings support initial
and artists manipulate color to communicate affect but their guidelines based on the distribution of hue, chroma and light-
knowledge of how to design affective palettes is largely rooted ness for the affects we studied. Some are very distinct, while
others overlap. Third, we propose a new method for modeling
Permission to make digital or hard copies of all or part of this work for personal or palettes using social network analysis that shows some inter-
classroom use is granted without fee provided that copies are not made or distributed esting relationships between colors in palettes worth further
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM exploration. These results extend research in design practice
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, and introduce new dimensions of expressivity to visualization.
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from [email protected].
CHI 2017, May 6-11, 2017, Denver, CO, USA.
Copyright is held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-4655-9/17/05 ...$15.00.
DOI: http://dx.doi.org/10.1145/3025453.3026041
AFFECTIVE VISUALIZATION lightness and chroma can be altered without contravening
The role of affect in information visualization applications is categorical meaning or desired “personality.” [32]
emerging, as researchers identify its importance in narrative
[6, 36], problem solving [13, 33, 16] and contextual framing. COLOR, AFFECT AND AESTHETICS
The fundamental difference between data and affective visu- Aesthetic principles for palette design are typically based on
alization is in communicative intent. Information visualiza- color harmony, an attempt to model which colors work to-
tion seeks to represent data with visual features for effective gether visually. This is commonly expressed as geometries
cognitive interpretation. Affective visualization uses visual with respect to hue wheel, as well as careful control of light-
features to evoke a mood, feeling or impression. These are ness variation [17, 41]. Itten’s qualitatively grounded contrast
commonly described by factor-based classifications such as model formalized the concepts of warm and cool colors, and
the well-known PAD model of affect [35] that plots them in a postulated that tints (light colors) represent the brighter and
dimensional space defined by pleasure (valence) and arousal better aspects of life, while shades (dark colors) represent the
axes. Valence covers hedonic range, from positive (happiness, darker, sad, and negative forces [17]. Itten contrasts have been
pleasure, love) to negative (pain, anger, sadness, fear). Arousal used in image analysis of affective colors [30] and aesthetic
reflects intensity from calm (unaroused, relaxed, sleepy, etc.) color selections [29].
to excited (high arousal, stimulated, nervous, alert, etc.). Typ-
ical emotions such as surprise, disgust or compassion can Relatively little research has examined color palettes and affect.
be placed in this 2D space; extensive emotion research has Madden identified two strategies of color association in logo
defined many more nuanced affects (such as affection or bore- design (which colors are used together): consistency (colors
dom) in this model as well [23]. have a similar meaning) and complementarity (each color is
associated with a different meaning) [31]. He found that when
blue was used as the designated logo color, it was most often
RELATED WORK paired with white, green and yellow (largely a consistency
Color psychology examines the interplay between color, cog- strategy). In contrast, when red was the designated main
nition, affect and behaviour, considering factors of culture [8, color, respondents preferred a complementary strategy, with
31], emotional response [40, 21] and behavioural influence the secondary color different across cultures.
(particularly around consumer response) [24, 22, 3]. Various
studies show that “warm” colors (red, yellow and orange) are PALETTE GENERATION
more arousing than “cool" hues of blue and green [32, 3]. Foundational work in color naming [4] showed that all cul-
Red is considered hot, vibrant and intense across cultures [22, tures have common concepts of a small set of basic colors and
31, 9], and most likely to induce arousal and anxiety [21]. It their associated verbal names. Heer and Stone built a proba-
is also tightly coupled with semantic cultural meaning, and bilistic model of color naming [14], showing how it can map
thus varies in its association with valence. Yellow, orange and between colors and names and measure color similarity. They
brown have less consistency of affective response, although applied this to compare and evaluate palette design, where
yellow is also considered exciting. A study of website color for minimizing name overlap and maximizing saliency are impor-
e-commerce across cultures found that respondents disliked tant for comprehension and memorability [39]. This work was
the yellow scheme, terming it “too showy” and not “appropri- expanded to semantic color design for visualization by Lin
ately professional” [8]. Studies also associate brown with “sad” et al. [25] and Setlur and Stone [37] for finding appropriate
and “stale” ratings [31]. Blue, and to a lesser extent green, colors associated with categories. These algorithms explore
have positive links to the natural world and are associated with the rich associations grounded in the concept-color relations
positive content [31, 32, 9, 10]. In particular, blue is strongly of objects and identities, but have not been applied to more
associated with peacefulness and calm across cultures [8, 31]. nuanced terms of emotion, atmosphere or affect.
In advertising, blue is associated with trust [8] and compe-
tence [22]: it evokes stronger buying impulses than red [3]. Color-word associations have long been explored as the basis
Alternately, in room color, blue is more likely to be associated for palette design using different strategies: manual selection
with depression, and red with anxiety [21]. Clearly context is from a predesigned set of colors linked to a particular word or
a critical moderating factor in these findings. concept, such as the Kobayashi color Image [20] or the Adobe
Kuler website; or automatic color extraction and combina-
Most color psychology has focused on hue rather than chroma tions given a set of rules using concepts of color harmony [28,
and lightness, although recent studies show the influence of 7]; perceptual contrast [19], keyword-color scales [38] and
lightness and chroma on affective response [24]. High chroma user pair preferences [11]. Kobayashi’s empirically grounded
color is exciting and intense; low chroma colors are calmer color Image Scale [20] provides 130 basic colors combined in
and less dominant. Greys have been described as “serious” 1170 three-color combinations indexed by 180 keywords such
and “professional”. Lightness is associated with calm: lighter as “provocative” or “romantic,” a complex set of expressive
colors are considered more pleasant, less arousing and less concepts. Lindner et. al. used color-word associations and
dominant than dark colors [22, 24]. Black is negative and harmonic color themes based on the color wheel for automatic
dominant [31, 40]; dark browns are sad. In one study, subjects palette generation in which the user described the semantic
were less upset when they read about murder on light pink content and the color extraction algorithm determines the ap-
paper than on white [40]. These findings may be useful when propriate colors from a precomputed color thesaurus [27] that
hue choice is limited by branding or other assignment, because matches a word to its underlying distribution in HSV[28]. In a
small user study they compared their palettes to Adobe Kuler’s affects: Serious, Playful, Trustworthy and Disturbing. We
user-generated palettes and found no significant improvement, chose these as examples in a much larger space rather than as
concluding that color palette preferences are highly influenced definitive “best” options in quantifying exact mappings. These
by personal taste. Jahanian et al. [18] took a different ap- can be considered as combinations of the core affects. Dis-
proach, using a color extraction method for magazine design turbing is Exciting+Negative; Playful is Exciting+Positive;
palettes based on word association with the concepts in the Trustworthy is Calm+Positive; Serious is Calm+ Negative
Kobayashi scale [18]. More recently, researchers have ex- (although the valence may be less pronounced). We realize
plored automated palette generation for the sparser palettes that these might not produce exactly equally weighted loca-
in visualization. Wijffelaars et al. developed a generalized tions in the affect space, where other emotions such as “Happy”
method based on sampling of a continuous path through color might. But we have a second important design goal: ecolog-
space at uniform intervals [43] with no user input into the ical validity. Emotions such as happy or sad may not be of
generation. In contrast, Colorgorical is an interactive tool interest in visualization applications, but trust and serious are
for automatic palette generation that uses a sampling based important affects in business communication [8]. Playful, se-
on user discriminability and preferences in which users can rious and disturbing are relevant to storytelling, an emerging
customize the sampled selections. [11]. field in data visualization [6, 15].

COLOR METRICS APPROACH

Perceptually-based color metrics can be computationally rep- We explored color-affect relations in three studies. First (S1),
resented as geometries in a three dimensional color space. we analyzed the color palettes of tagged images in two large
The common CIELAB representation offers a color distance social network databases, Flickr and deviantArt.com. From
metric ∆E. This is simply the Euclidean distance between these results, we generated a set of colors for a design study
two colors, expressed by their coordinates, L∗ , a∗ , b∗ . Using (S2) in which users selected 5 colors for categorical coloring
polar coordinates to represent a∗ , b∗ creates a more intuitive in simple visualizations to best communicate a specified affect.
representation. The radial distance, C∗ defines chroma or rel- The results of S2 produced both distinct patterns for color
ative saturation, hue is defined by the hue angle, h°. In this selection and grouping by affect and a new metric of palette
paper, we simplify this notation to L,C,H. We also use the weight. We then ran a validation study (S3) in which we
terms ‘lightness,’ ‘chroma’ and ‘hue’ for these quantities. We generated palettes based on this metric and asked users to rank
also use the term ‘saturated’ and ‘desaturated’ descriptively to them for the specified affect. We discuss each in turn.
describe high and low chroma colors respectively.
STUDY 1: IMAGE ANALYSIS
Quantitative measures to represent color palettes are more Following [25], we analyzed the palettes of 8608 images. We
complex. Lin et al. analyzed them using color distance metrics selected these by searching for images tagged with terms re-
to extract distinct palettes from images [26]. These included lated to our eight affects, and then eliminating all images
mean, min and max distances between colors in themes. Two with humans or human-influenced scenes (such as war im-
quantitative approaches to measuring hue variation are entropy ages) to avoid conflation with content [23]. Roughly 2/3 of
(how distinct the colors are) or saturation-weighted hue disper- these images were abstract; the remaining 1/3 were landscapes,
sion [WHD], which calculates the angular dispersion between still-life/object, or nature scenes.
hues along the hue wheel, taking saturation into account[12].
When WHD is higher, hues are more concentrated (less dis- For each image, we calculated mean lightness and chroma. We
persed). Methods such as k-means clustering are commonly then computed distribution histograms with a bin size of 10,
used to model color associations [25]. calculating the percentage of pixels in the image that was in the
lightness or chroma range of the bin. We generated 8 sets of
40 distinct hues each using k-means clustering, plus lightness
MOTIVATION: AFFECT AND COLOR
and chromaticity ranges obtained from the image analysis. For
We were curious whether color and affect would be linked in
simple 5-color categorical palettes applied to common visu-
Calm Exciting Serious Playful Positive Negative Disturbing Trustworthy
alizations. Are there different groupings of colors that align
more consistently to areas in that space? Can the relative dif-
ferences suggest palette selection based on the desired affect’s
“location” in the PAD affect space? In other words, if an affect
combines Calm and Positive, would the colors selected for
it be similar to colors in palettes selected in both? Our work
differs from other concept-color association research in two
ways. First, we focus on palettes rather than only single col-
ors. Second, we explore whether color associations along the
valence and arousal dimensions map to PAD space such that
color selection for other affects reflects their relative influence.
We began by selecting 8 affects. The 4 “core affects” of Calm,
Exciting and Positive, Negative represented the valence and Figure 1. Most common colors by affect in image analysis study (S1)
arousal dimensions of the PAD model. We added 4 “pragmatic”
each hue, we calculated the probability of a candidate color
c from the set given an affect value and corresponding image
histogram using Lin’s algorithm [25]. Each hue was measured
against the set of different images that were categorized by
affect. We then selected the most commonly used colors in
each of the affective image sets (Figure 1).

Results
As our data distribution was marginally non-normal, we used
the Kruskal-Wallis test for significance (the rank-based non-
parametric equivalent of an ANOVA for multi-factor data).
Affect had a significant effect on both lightness X2 (7, 8607)
=199.6250, p<.0001 and chroma X2 (7, 8607) =387.7106,
p<.0001. Calm, Playful and Exciting images were lighter
than Disturbing and Negative. Negative and Calm were less
saturated than Playful and Exciting. We see patterns in Calm
palettes containing a larger concentration of light and desatu-
rated blues and greens. Trustworthy also has blues, purples
and some greens. Playful and Exciting use highly saturated Figure 2. Study 2 interface
colors like reds, vibrant greens and blues with Exciting having
relatively more dark reds. Disturbing has a larger distribution
of dark browns, blues, reds and black. Negative used more
grey and muted browns. With the exception of Positive, these
overall balance by adjusting a slider to control alpha. When
results mirror what color psychology would predict. We were
the participant was finished with the color selection, s/he rated
surprised by the amount of browns and dark colors in Positive.
satisfaction using a slider between 0 and 10 to indicate how
successfully s/he felt the colors expressed the concept. At the
STUDY 2: USER-DESIGNED PALETTES
beginning of each trial the initial alpha value was set to 1 and
We used these results to design colors for a user study. We
the visualizations were colored in neutral gray.
asked a visualization color expert to design a smaller set for
use in our next study. She reduced the set to 36 colors, using a Because we were interested in the relative locations of color
combination of k-means clustering to combine similar colors, groups in the PAD space, we paired endpoint tasks of the
and filtering to remove colors that were either too light to offer affect model (Calm-Exciting, Positive-Negative) to establish
sufficient contrast with the background or too dark for the hues differences (if any), drawing from the core theoretical work
to be easily visible when small. She then used visual tools and in mapping affect. Similarly, we considered these differences
visualizations of the associated color metrics to establish that in the pragmatic case (Serious-Playful) with clear semantic
the resulting colors offered a balanced mix of CIELAB hue, opposites. Where there was not a clear opposite (Trustworthy,
chroma and lightness values. Disturbing) we did not pair the tasks but presented them
singly. This provided a relative rather than an absolute measure
We ran a pilot study where people with design experience
of how palettes might differ with respect to comparative affect.
created palettes based on affect. The goal of our study was
We used this design because we are not expecting to be able
to see whether they would consistently assign different color
to absolutely quantify all possible palettes with respect to
sets for each of our 8 affective categories in simple visualiza-
affect, but rather to understand what might make a desired
tion tasks. The results from the pilot showed affect strongly
affective impression comparatively stronger or weaker. In
influenced colors chosen, similar to those in Study 1. We then
the cases of Trustworthy and Disturbing (both affects with
expanded our study to a wider population, to see if the results
strong ecological validity that are not clear opposites), we
were specific to designers and to get a larger data sample. We
simply sought to see how they might relate to the other affect
extended the color set to 41 colors, adding more dark, because
groupings.
our results indicated that the original palette did not provide
quite enough different colors for some of the darker affects. Each affect task was independent and required separate color
We show the detailed results for this larger experiment. assignments, alpha and satisfaction rating. Participants could
reassign the 5 colors in the palettes until satisfied: they were
Method not constrained to mapping a color to a particular spatial loca-
Participants completed categorical coloring tasks for two sim- tion or size in the visualization. The experiment began with
ple visualizations (bar chart or map) using the interface shown a training task in which participants colored Happy and Sad
in Figure 2 to select palettes of 5 colors for each task from a affects. No time limit was set on training; participants pro-
set of 41. We chose a map and a bar chart as they are familiar ceeded to the main study when ready. The experiment was
information representations. Participants were instructed to hosted on our university web server. People could stop at any
select colors to convey an affect from our eight affective cate- time, and login to the system later to complete the study: the
gories. They were not told what the data represented other than system retained state. We report only on results where all tasks
affective intent. They could additionally modify the palette were completed.
Factors and metrics
We had two independent variables: Affect (8) and Visualiza-
tion (2: bar chart, US map). Our raw dependent variables were
color metrics (lightness, chroma, hue), alpha, and satisfaction
rating.

Hypotheses for Study 2

H1. Affect will influence color choice. Cooler colors will
be used for low arousal (Calm, Serious and Trustworthy).
High arousal affects (Exciting, Disturbing, Playful) will use
warmer colors: more reds, browns and oranges. Positive
affects (Positive, Playful, Trustworthy) will include more
green. Negative will use brown and grey. Figure 3. Frequency distribution of colors by affect (S2).
H2. Affect will have a significant effect on lightness: Calm
colors will be lighter, Negative ones darker. hypotheses H2 that lightness relates to valence (Calm, Posi-
tive, Negative, Playful) Trust and H3 that arousal influences
H3. Affect will have a significant effect on chroma: high colorfulness choices (Exciting, Playful vs Calm, Serious).
arousal affects (Exciting, Playful, Disturbing) will include We also saw a combined effect: both valence and arousal in-
more saturated colors. fluence color choice, notably in Calm which is both lightest
and least chromatic. Satisfaction rating was uniformly high
Design and never significant.
We used a 2-way between-subjects design separated by Visu- Affect was highly significant for hue: X2 (7, 1216) =93.4.
alization type, producing an experiment session of 8 exper- We calculated color dispersion in each palette as saturation-
imental conditions (1 Affect) per group. Trial ordering was weighted hue dispersion (WHD) and saw significant differ-
randomized and block ordering was counterbalanced. A 76x8 ences by affect: X2 (7, 1216) =88.21, p<.0001. Colors in
design yielded 608 trials. Exciting, Playful and Positive were most spread out in hue
space; Calm and Negative palettes were the most tightly clus-
Participants tered.
76 persons, roughly distributed by gender, with normal or
corrected-to-normal vision, were paid to participate. They Looking at the core affects, we note some clear patterns. Calm,
were randomly assigned to one of the 2 groups (Bar or Map). as predicted, makes strong use of cool colors (blues and greens)
and is overall high lightness and low chroma. Exciting em-
phasizes warmer and higher chroma colors: more reds and
Results
yellows. Positive, like Exciting, uses strongly saturated hues
We removed all palettes that were either incomplete or in-
but incorporates more green. Negative, as anticipated, has
cluded duplicate colors, giving us 504 palettes that each had 5
more browns, dark reds, and greys. These patterns combine
unique colors. We then did both visual and statistical analy-
in interesting ways when we consider the pragmatic affects.
sis. Figure 3 shows the resulting palette colors. Each bubble
Playful is both Exciting and Positive in affect: its colors are
represents a color that appeared in a palette in the associated
similar to Exciting, but included more greens and yellows and
affect; the size represents how frequently it was used. Zero
fewer reds than Exciting. Serious shared the greys of Nega-
frequency values are removed, and each bubble cluster has a
tive but used more blue; we would expect that it has aspects
slightly different number of bubbles (colors) in it. However,
of low arousal and Calm. Notably, Serious, which might be
the size scale is the same across all the affects.
considered both calmer than Disturbing and less Negative,
We clearly see that colors and color characteristics varied by used fewer red and yellow hues. Disturbing, which is clearly
affect, confirming H1. We used a a nonparametric Kruskal- Exciting and Negative, reflects the browns and greys of Nega-
Wallis for significance for all tests. Visualization type had tive but incorporates more of the reds and yellows of Exciting.
no effect, so we combine the results in subsequent discus- Calm and Negative palettes were the most tightly clustered.
sion. A statistical analysis on the three perceptual metrics, These findings and the statistical analysis confirm H1.
lightness, chroma and hue, found all were significant. We
found significant effects for affect on lightness X2 (7, 1216) STUDY 3: USER RANKED PALETTES
=426.7211, p<.0001 and chroma X2 (7, 1216) =481.8955, We now wanted to validate these results and to assess an
p<.0001. Calm, Playful and Positive were much lighter than algorithmic design metric. We used insights from S2 to design
Disturbing, Serious and Negative. Calm was the least satu- sets of palettes for each affect and ran a study in which people
rated of the affects; Exciting, Playful and Positive the most selected the best and worst palette for an affect from a set of
colorful (chromatic). While alpha was significant, the effect 5 choices. To avoid any bias potentially introduced by the
was very small: it was only applied in Calm. Since this effect pairings in Study 2, affects were presented singly. For each
was small and simply reinforced the already strong pattern of affect, we calculated the weight of each color as how often that
high lightness and low chroma for this affect, we did not use color was used in a palette in S2. We then calculated palette
alpha in further analysis. These results statistically confirm our weight for each palette in S2 as a sum of its color weights.
Figure 4. Examples of good (high weight) and bad (low weight) palettes

Given that high palette weight represented the cumulative

strongest color combinations for that affect, we conjectured
it would be a good predictor of a palette’s suitability for a
particular affect. We designed palettes for S3 manually based
on weight. For each affect, we automatically generated every
possible combination of 5 colors (749, 398 sets) and calculated
the palette weight of each. We then sorted the list from highest
to lowest and manually selected candidate palettes from three
relative areas: the top weighted palettes, the lowest weighted
palettes and the middle area. We used manual selection to
Figure 6. Palette Ranking by relative weight and affect (Study 3)
minimize color replication in the palettes, as this was difficult
in the highest weighted palettes because there were a few Hypotheses Study 3
very heavily weighted colors in these: we ensured that no two H4. Higher weighted palettes would be more likely to be rated
palettes shared more than 3 colors. Figure 4 shows examples Best than low weighted palettes;
of two good (top weight) and 2 bad (low weight) palettes for
each of Calm and Negative. H5. Lower weighted palettes would be more likely to be rated
as Worst.
Method
In each trial a participant saw 5 spatially identical bar charts, Participants
each colored with a different palette (Figure 5). Colors in the 38 users (and an additional 2 with incomplete data) partic-
palette were randomly assigned to locations in the bar chart. ipated. They were roughly distributed by gender. All had
The participant was asked to identify one as Best and one normal or corrected-to-normal vision. Participants were en-
as Worst. There were 8 affective tasks. For each affective tered into a draw for potential reimbursement. None had taken
category there were 15 possible palettes, binned into 5 “good”, part in the previous studies.
5 “medium” and 5 “bad”. These were only used for random-
ization: in each trial, 2 palettes were randomly selected from Results
the “good” bin, 1 from the “medium” bin and 2 from the “bad” Figure 6 shows the results of palette ratings by weight and
bin. Each participant did 2 repetitions of each affective task (2 affect. As the palette sets were very different in the range of
Best, 2 Worst). Affect tasks were presented in random order. weights, we normalized the weight metric to relative weight
position: that is, the percentage of the overall range that the
weight represented. So, for example, the Calm range was 135
from the lowest weight to the highest weight: a Calm palette
with a weight of 109 had a relative weighted position of 80.7.
We then used a logistic regression analysis to determine if
this would be a significant predictor of Best/Worst choice.
The result was extremely significant:X2 (1, 1218) =1093.32,
p<.0001.
These results confirm that palettes are more likely to be ranked
better as their weight increases. Conversely, palettes with
Figure 5. Study 3 Interface: Users ranked the Best and Worst of 5 low weight were more likely to be selected as Worst. These
palettes for each affect in S3.
results confirm H4 and H5 that palette weight calculated from
Metrics our previous studies proved a strong predictor of how people
The independent variables were affect and palette weight (PW). ranked it for affect expressivity. More generally, they validate
The dependent variable was the rating of Best/Worst for each our essential conjectures that even simple palettes of 5 colors
palette. For each palette we thus had the rating, the individual can convey different affective impressions, though it can been
colors in that palette, the frequency weight for each color and seen in Figures 3 and 7 that is a stronger finding for some
the summed palette weight. affects than others. In particular, we note that the relative
 PALETTES AS SOCIAL NETWORKS
These results show that there are patterns of color prefer-
ence by affect. To explore the relationship between colors in
palettes, we tried modeling palettes as social networks. Social
network analysis provides both a visual and a mathematical
analysis of actors (nodes) and relationships (links) to give in-
sight into the various roles and groupings in a network: where
are the most influential nodes? What are the strongest connec-
tions? Are there clusters? In particular, measures of network
centrality identify the most important nodes in the network as
a function of how well they are connected to other nodes; mod-
els of network structure identify clusters and outliers. While
we have no formal results, our analysis was interesting enough
that we provide a few examples here.

Figure 7. Colors from palettes chosen as “Best” in Study 3

Using the Gephi library [2],we modeled each palette as a
connected subset of 5 colors, where each color Ci was a node
in the network of colors and an edge Eij indicated Ci and Cj
difference between Best and Worst in the more nuanced affect were used in the same palette. Each of the 8 affect groups was
of Trustworthy was smaller, and there were more outliers. We a separate network of our 41 palette colors. This enabled us
surmise this was due to the case that the range of weights in to see not only the frequency of individual color use but also
these palettes were smaller: instead of a few highly weighted structural patterns (combinations of colors).
colors, the preferred colors were more evenly used, and thus We visualized these palettes as networks using a force-directed
the weight ranges were smaller. This raises questions about layout algorithm. Color frequency is mapped to node size;
what else we might need to understand about what contributes pairwise frequency (weight) to edge thickness. For clarity,
to color selection for affective palettes. We calculated the we removed all color combinations with an edge weight less
preferred colors for each affect by counting how many times than 3, showing only pairwise combinations used in 3 or more
the hues occur in the palettes rated Best for that affect. We palettes. Figures 8 and 9 show the resulting networks. Node
then simply selected the colors with the highest count for each size (frequency of color use) reflects the results of Study 2
affect. Figure 7 shows the preferred colors for each of the and 3. The network structures emphasize which colors are
affects in Study 3. These strongly reflect the results from S2. highly connected, which are outliers. Some palettes cluster
If pairing had introduced bias in S2, we would not have seen tightly around a small number of key colors (Calm, Exciting,
these confirmatory results in S3, as there were no comparative Playful). Some are more angular and spread out. And then
paired presentations in this study.

Calm Exciting Positive Negative

Figure 8. Core Affect Combinations

Serious Playful Disturbing Trustworthy

Figure 9. Pragmatic Affect Combinations

there is Trustworthy. In the frequency models,Trustworthy
showed very little structure. It included many colors, fairly
uniformly weighted. Here, we see a distinct structure, with
two groups linked by yellow. We hope to find further insight
by continuing this sort of modeling in the future.

DISCUSSION
What do these results tell us? First and foremost, we are able
to reliably associate color and palette properties with affective
response even in the limited scope of color selection in simple
information visualization representations, extending previous
work in individual color-affect relations. Our studies show
that both the perceptual properties of small color sets, and
the composition patterns of how they are used together, are
affectively distinct. From our studies in image analysis (S1)
and user-designed palettes (S2), we were able to algorithmi-
cally define a simple metric of palette weight based on color
frequency use for each affect that proved a reliable predictor
of preferred affective palettes in a final validation study (S3).
We see consistent patterns in lightness related to affect across Figure 8. The L,C,H distributions for each affect. Hue is rotated by 60°to
all studies. Calm, Playful, Positive and to a lesser extent group warm and cool colors, all values are normalized to a common
Trustworthy are lightest, while Serious, Disturbing and range
Negative are darker. This confirms H1. Similarly, we also
saw consistent use of higher chroma colors for Playful, Ex- about which colors go with others, as we are exploring in the
citing, Positive and Disturbing: where Calm, Serious and network diagrams.
to a lesser extent Trustworthy were less saturated. These
While we cannot reliably profile each affect, even being able to
comparisons are significant, and confirm our hypotheses that
say what colors to avoid for specific affects can be useful. For
lightness and chroma are linked to affect. This indicates that
example, we can say with some authority that highly saturated
any color set intended for use in expressing these affects must
provide enough variation in both lightness and chroma. light colors will NOT be appropriate for Serious or Trust, or
Calm; light blues, beiges and greys are never likely to convey
One of our primary goals in this is to create guidelines for de- Playful; dark red and browns are not Positive; and light colors,
sign and design evaluation simple enough to be programmed, particularly green, do not communicate Negative affect.
similar to work by Bartram et al. [1]. While these will never
The network models help us visualize palette composition
replace the skill of an experienced designer, they can reduce
patterns: groupings that show what colors are likely to be used
the time spent exploring options that simply don’t work and
provide reasonable starting grounds for further refinement. together. Certain affects showed different palette composition
patterns, Calm, Exciting and Playful palettes tightly clus-
Figure 10 summarizes the mapping between affect and color tered around the main colors, confirming high co-occurrence
metrics for each affect. Here, each of lightness, chroma and of these colors in most palettes. In contrast, certain colors in
hue have been mapped to a common range, then plotted. For Positive were only used in combination with bright yellow.
hue, the important metric is whether the colors are warm or Greens form an important cluster in Positive. Serious, Nega-
cool. We make this easier to see by rotating the hue angle by tive and Disturbing palettes show loose sub clusters anchored
60 degrees so that low values are cool, high values are warm. by a central set of core colors. Trustworthy shows a clear
The resulting metrics are labeled (NL, NC, N SH). The box example of two thematic strategies (blue-gray, green-gray)
plots show the distributions. Our hypothesis of hue association bridged by a single common color (yellow). In future work,
with affect is supported by these distributions. we hope to better understand what these patterns mean and
how they might be applied.
Based on the results of study 3, we speculate that if a palette
contains colors near the median of L,C,H for a specific affect, We sought to see if there were general color associations with
then the resulting palette is likely to be judged as matching the more abstract dimensions of valence (Positive-Negative)
that affect, especially if the distribution is tight. Some affects and arousal (Calm-Exciting) that would translate (lend color
have clearer L,C,H “profiles” than others. Of the Core affects, elements) to the specific expressions of Playful, Serious,
Calm is the strongest, requiring high L, low C and cool colors. Trustworthy and Disturbing. We certainly saw evidence
Negative has similar properties for C and H, but dark colors of these crossovers in our limited set of results. Establishing
predominate. Exciting and Positive, however, while distinctly such reliable associations leads to the question of if and how
different than Calm and Negative, are not very different from we might algorithmically determine color selection based on
each other, or from the pragmatic affect Playful. The profiles the desired affect’s location in the PAD affect space. In other
in the other pragmatic affects show similar overlaps, as we words, given a desire to enhance a visualization as more “re-
might expect. This is why we need more nuanced analysis assuring”, for example, can we quantify where “reassuring”
plots between Calm-Exciting and Positive-Negative and de- forms and data contexts. While this is not enough to clearly
termine the relative hues, chroma and lightness values as a distinguish all affects, it helps limit the design space, which
weighted contribution from each? In future work, we plan to can be of significant practical value. For example, one could
explore how the color patterns we see in the core affects may extend Colorgorical[11] to include affect by using L, C and H
generalize to other pragmatic affects. filters.

LIMITATIONS A more challenging application would be to include affect in

While these results indicate the potential for using color to the design of palettes where some colors are already defined,
communicate affect, we note several limitations to our studies. especially if they fall outside of the simple L,C,H model of
We have studied exactly one palette size (5 colors) for a limited affect. A simple example would be creating a Trustworthy
set of visualization forms (bar charts and maps), on a white palette for T-Mobile that included its signature bright purple.
background. While we have confirmed that color and affect Here, we would ideally be able to model how to combine that
can be linked even for these simple, functional cases and playful colors with others to create an overall Trustworthy
speculate that the results have broader value than precisely affect.
what we tested, we obviously don’t know how far our results In addition, we introduce network modeling as a method to
can be extended. characterize palette composition. We are concerned not with
Which colors compose an effective palette depends on the size individual color selection per se, but in the much more dif-
of the features they color. In our studies, all of the features ficult to quantify sets of colors and how they are combined
colored were large enough that all of the study colors were easy (the palette). Our network models begin to provide some in-
to perceive and distinguish. The study color set reflected these sight on how single colors that might be individually strong
image constraints. To support fine-grained graphs such as line in one affect may still be combined in a palette more suited
charts and scatter plots, some of the colors in our study palette to another affect (such as the yellow in Trustworthy). While
would be too light, or would be too difficult to differentiate. preliminary, we believe this approach holds promise for ex-
We would need a different set of study colors. We hypothesize ploring richer design options with color combinations based
that we we would see similar palette trends for such graphs, on more sophisticated social network metrics.
even though the detailed color statistics would be different. Our work suggests ways to address this that would be in-
There are many other factors that can influence the affect teresting to explore further. For example, the psychological
induced by an image, including cultural color identities, spa- dimensions of affect (arousal and valence) are shown to be
tial properties such as relative size and juxtapositions, and strongly influenced by lightness and chroma. This indicates
ultimately, the subject matter. Our work deliberately tries to that these properties can be used to “tune” palettes of prede-
minimize these influences, as including them would be much fined hues to these affects, either by modulating the color itself,
more complicated. or by optimizing these values over the palette as a whole. The
work in network modeling suggests, for example, that with
We tested only a limited range of affective impressions in a more complete model there may be ways to combine the
abstract contexts. This speaks to the need for more basic T-Mobile playful purple with other colors to still convey an
research in examining a fuller range of affect, and more applied overall Trustworthy affect.
research in examining the expressive capacity of affective
palettes in actual visualization use. Finally, we stress that we have approached this research from
a perspective of design rather than pure perceptual theory.
Finally, even within our constraints, our results do not of- Our work asks: given a desired affect, what palette colors
fer unique guidance for each affect as was discussed above. should you choose? It does not answer the broader and much
Clearly our current color statistics are not enough. However, harder question: given an arbitrary palette, what affect does
the network models do offer some further directions to pursue. it convey? To make this distinction concrete, if the goal is a
Calm affect, our results suggest light, cool, pastel colors. If
CONCLUSIONS your goal is to create a Disturbing affect, then dark colors,
The first goal of this work was to demonstrate whether color especially reds, are better. Our results cannot reliably tell you
can express affect, even when used as categorical coloring for whether a palette of pastels plus a vibrant blood red (which
simple data visualizations. While some may find our results might well be Disturbing depending on how the red is used)
unsurprising, there are many others who expressed skepticism. is Calm or Disturbing. Being able to absolutely relate color
This study demonstrates clearly that affect is a dimension of to affect for all cases was not our goal: frankly, we don’t
categorical color palette design, and motivates extending work believe this is possible. However, we do believe that simply
in automated palette design to include affective concepts. This being able to recommend validated palettes for a particular
brings new dimensions of expressivity and communicative affect will prove of interest to those who create and employ
scope to visualization. visualizations. Clearly, there is much work to be done.
Our results show a validated relationship between affect, per-
ceptual color properties (hue, chroma and lightness), and Acknowledgments
palette composition for the eight categories we measured. This This work was supported by the Natural Sciences and Engi-
confirms and extends findings in color psychology and design neering Research Council of Canada. We thank Dr. Peter
practice to the context of simple information visualization Gorniak for his assistance in data-mining the image databases.
REFERENCES 15. Jessica Hullman and Nick Diakopoulos. 2011.
1. L. Bartram, B. Cheung, and M. Stone. 2011. The Effect Visualization rhetoric: Framing effects in narrative
of Colour and Transparency on the Perception of visualization. IEEE transactions on visualization and
Overlaid Grids. IEEE Transactions on Visualization and computer graphics 17, 12 (2011), 2231–2240.
Computer Graphics 17, 12 (Dec 2011), 1942–1948.
16. Alice M Isen. 2001. An influence of positive affect on
2. Mathieu Bastian, Sebastien Heymann, Mathieu Jacomy, decision making in complex situations: Theoretical issues
and others. 2009. Gephi: an open source software for with practical implications. Journal of consumer
exploring and manipulating networks. ICWSM 8 (2009), psychology 11, 2 (2001), 75–85.
361–362.
17. Johannes Itten. 1974. The art of color. Reinhold Pub.
3. Joseph A Bellizzi and Robert E Hite. 1992. Corp, New York.
Environmental color, consumer feelings, and purchase
likelihood. Psychology & marketing 9, 5 (1992), 18. Ali Jahanian, Jerry Liu, Qian Lin, Daniel Tretter, Eamonn
347–363. O’Brien-Strain, Seungyon Claire Lee, Nic Lyons, and Jan
Allebach. 2013. Recommendation system for automatic
4. Brent Berlin and Paul Kay. 1969. Basic color terms: design of magazine covers. In Proceedings of the 2013
Their universality and evolution. Univ of California Press, international conference on Intelligent user interfaces.
Berkeley. ACM, 95–106.
5. Cynthia A Brewer. 1994. Color use guidelines for
19. Hye-Rin Kim, Min-Joon Yoo, Henry Kang, and In-Kwon
mapping. Visualization in modern cartography (1994),
Lee. 2014. Perceptually-based Color Assignment. In
123–148.
Computer Graphics Forum, Vol. 33. Wiley Online
6. A. Cairo. 2013. U.S. Gun Deaths and the Challenge of Library, 309–318.
Uncertainty. Peachpit. (2013). http:
//www.peachpit.com/articles/article.aspx?p=2036558
20. Shigenobu Kobayashi. 1981. The aim and method of the
color image scale. Color research & application 6, 2
7. Daniel Cohen-Or, Olga Sorkine, Ran Gal, Tommer (1981), 93–107.
Leyvand, and Ying-Qing Xu. 2006. Color harmonization.
In ACM Transactions on Graphics (TOG), Vol. 25. ACM, 21. Nancy Kwallek, Carol M Lewis, and Ann S Robbins.
624–630. 1988. Effects of office interior color on workers’ mood
and productivity. Perceptual and Motor Skills 66, 1
8. Dianne Cyr, Milena Head, and Hector Larios. 2010. (1988), 123–128.
Colour appeal in website design within and across
cultures: A multi-method evaluation. International 22. Lauren I Labrecque and George R Milne. 2012. Exciting
journal of human-computer studies 68, 1 (2010), 1–21. red and competent blue: the importance of color in
marketing. Journal of the Academy of Marketing Science
9. Andrew J Elliot and Markus A Maier. 2014. Color 40, 5 (2012), 711–727.
psychology: Effects of perceiving color on psychological
functioning in humans. Annual review of psychology 65 23. P Lang and Margaret M Bradley. 2007. The International
(2014), 95–120. Affective Picture System (IAPS) in the study of emotion
and attention. Handbook of emotion elicitation and
10. Timo Elliott and Business Objects. 2008. Why business
assessment 29 (2007).
intelligence projects fail-And what to do about it. In Bus.
Objects User Conf. Europe, Vol. 2009. 24. Marie-Christine Lichtlé. 2007. The effect of an
11. Connor C Gramazio, David H Laidlaw, and Karen B advertisement’s colour on emotions evoked by attitude
Schloss. 2016. Colorgorical: Creating Discriminable and towards the ad: The moderating role of the optimal
Preferable Color Palettes for Information Visualization. stimulation level. International Journal of Advertising 26,
IEEE Transactions on Visualization & Computer 1 (2007), 37–62.
Graphics 27, 1 (2016), 521–530. 25. Sharon Lin, Julie Fortuna, Chinmay Kulkarni, Maureen
12. Allan Hanbury. 2003. Circular statistics applied to colour Stone, and Jeffrey Heer. 2013. Selecting
images. In Proceedings of the 8th Computer Vision Semantically-Resonant Colors for Data Visualization. In
Winter Workshop, Vol. 91. Citeseer, 53–71. Computer Graphics Forum, Vol. 32. Wiley Online
Library, Blackwell Publishing Ltd, 401–410.
13. Lane Harrison, Remco Chang, and Aidong Lu. 2012.
Exploring the impact of emotion on visual judgement. In 26. Sharon Lin and Pat Hanrahan. 2013. Modeling how
Visual Analytics Science and Technology (VAST), 2012 people extract color themes from images. In Proceedings
IEEE Conference on. IEEE, 227–228. of the SIGCHI Conference on Human Factors in
Computing Systems (CHI ’13). ACM, 3101–3110.
14. Jeffrey Heer and Maureen Stone. 2012. Color naming
models for color selection, image editing and palette 27. Albrecht Lindner, Bryan Zhi Li, Nicolas Bonnier, and
design. In Proceedings of the SIGCHI Conference on Sabine Süsstrunk. 2012. A large-scale multi-lingual color
Human Factors in Computing Systems. ACM, thesaurus. In Color and Imaging Conference, Vol. 2012.
1007–1016. Society for Imaging Science and Technology, 30–35.
28. Albrecht Lindner and Sabine Süsstrunk. 2013. Automatic 43. Martijn Wijffelaars, Roel Vliegen, Jarke J Van Wijk, and
color palette creation from words. In Color and Imaging Erik-Jan Van Der Linden. 2008. Generating color palettes
Conference, Vol. 2013. Society for Imaging Science and using intuitive parameters. In Computer Graphics Forum,
Technology, 69–74. Vol. 27. Wiley Online Library, 743–750.
29. Peng Lu, Zhijie Kuang, Xujun Peng, and Ruifan Li. 2014.
Discovering Harmony: A Hierarchical Colour Harmony
Model for Aesthetics Assessment. In Asian Conference
on Computer Vision. Springer, 452–467.
30. Jana Machajdik and Allan Hanbury. 2010. Affective
image classification using features inspired by
psychology and art theory. In Proceedings of the 18th
ACM international conference on Multimedia. ACM,
83–92.
31. Thomas J Madden, Kelly Hewett, and Martin S Roth.
2000. Managing images in different cultures: A
cross-national study of color meanings and preferences.
Journal of international marketing 8, 4 (2000), 90–107.
32. Ravi Mehta and Rui Juliet Zhu. 2009. Blue or red?
Exploring the effect of color on cognitive task
performances. Science 323, 5918 (2009), 1226–1229.
33. Don Norman. 2002. Emotion & Design: Attractive
Things Work Better. interactions 9, 4 (2002), 36–42.
34. Donald A Norman. 2005. Emotional design: Why we love
(or hate) everyday things. Basic books, New York.
35. Jonathan Posner, James A Russell, and Bradley S
Peterson. 2005. The circumplex model of affect: An
integrative approach to affective neuroscience, cognitive
development, and psychopathology. Development and
psychopathology 17, 03 (2005), 715–734.
36. Edward Segel and Jeffrey Heer. 2010. Narrative
visualization: Telling stories with data. IEEE transactions
on visualization and computer graphics 16, 6 (2010),
1139–1148.
37. Vidya Setlur and Maureen C Stone. 2016. A linguistic
approach to categorical color assignment for data
visualization. IEEE transactions on visualization and
computer graphics 22, 1 (2016), 698–707.
38. Martin Solli and Reiner Lenz. 2010. Color semantics for
image indexing. In Conference on Colour in Graphics,
Imaging, and Vision, Vol. 2010. Society for Imaging
Science and Technology, 353–358.
39. Maureen Stone. 2003. A field guide to digital color. AK
Peters, Natick, Mass.
40. Patricia Valdez and Albert Mehrabian. 1994. Effects of
color on emotions. Journal of experimental psychology:
General 123, 4 (1994), 394.
41. Lujin Wang and Klaus Mueller. 2008. Harmonic
colormaps for volume visualization. In Proceedings of
the Fifth Eurographics/IEEE VGTC conference on
Point-Based Graphics. Eurographics Association, 33–39.
42. Colin Ware. 2008. Visual thinking for design. Morgan
Kaufmann, Burlington, Mass.