skip to main content
10.1145/3678698.3678701acmotherconferencesArticle/Chapter ViewFull TextPublication PagesvinciConference Proceedingsconference-collections
research-article
Open access

Color Recommendations Based on Individual Differences

Published: 11 December 2024 Publication History

Abstract

Color is an important element in the appealing presentation of a product or advertisement that greatly influences the quality of visual perception. The selection of preferred colors differs among visual media creators according to their individual preferences, which may be influenced by their personality traits and cognitive abilities, as well as their gender, age, and cultural background. While numerous theoretical, psychological, and computational approaches have been explored to recommend colors to individual users, they have typically focused on a limited number of factors without exploring their mutual synergistic effects. In this paper, we present a color recommendation model based on multiple fields of variables that parameterize the space of individual differences. In particular, we explicitly account for synergistic effects between such feature variables in our formulation using factorization machines that describe relationships between feature variables from different fields. We collect data on these relationships through crowdsourced user studies, which require a careful questionnaire design to effectively capture the distribution of individual differences among participants. The implementation of a prototype color recommendation interface and associated results demonstrate the feasibility of the proposed approach.
Figure 1:
Figure 1: Snapshot of the prototype system for color recommendations.

1 Introduction

Optimization of visual media design has been a persistent and significant problem in the field of multimedia for quite some time. The most challenging aspect is the formulation of the aesthetic criteria of such visual design, which is usually achieved by defining an objective function and optimizing it under necessary conditions. In recent years, we have seen a significant leap in the complexity of visual design optimization due to the insightful consideration of aesthetic criteria.
However, conventional approaches usually satisfy the preferences of users as a whole but do not adequately consider the preferences of individual users. This is because each user has different preferences based on their personality traits and cognitive abilities, resulting in a wide variety of expected visual designs. According to [25], even if the overall visual design aesthetic for information comprehension is enhanced, it is not compelling enough to support the decision making of individual users due to their different design requirements. In addition, several studies, including recent ones (e.g. [3, 11, 14, 16, 34]), have shown that people with certain personalities prefer colors that match them, suggesting the potential for innovative approaches to visual media design based on individual differences.
Our ultimate goal is to find a new formulation that can account for individual differences among users in visual design optimization problems. More specifically, as the primary technical challenge of this work, we target a color recommendation problem that takes into account intricate individual differences. To address this challenge, we collect data examples that show how color choices are mapped to a space of individual differences. Our approach introduces multiple fields of feature variables to parameterize the space, including personality traits, cognitive abilities, gender, and age, in an attempt to thoroughly capture individual differences. We meticulously collect data samples of such a mapping through carefully designed crowdsourced user studies in which we solicit thoughtful responses from participants. Our study also formulates possible synergistic effects between the feature variables that make up the space of individual differences, taking advantage of sophisticated recommendation models.
We summarize the technical contributions of this paper as follows:
We present a new color recommendation model that explicitly introduces individual differences into the solution space when optimizing visual media design problems.
We establish a robust correspondence between color choices and individual differences through careful data collection via crowdsourced user studies.
We implement the color recommendation model by explicitly characterizing synergistic effects between feature variables that constitute the space of individual differences.
The remainder of this paper is organized as follows: Section 2 presents a brief review of previous research relevant to this study. Section 3 describes the method of collecting the data samples representing the relationship between color preferences and individual differences through crowdsourced user studies. Section 4 describes the main framework of our color recommendation model along with our techniques for extracting synergistic effects between feature variables that parameterize the space of individual differences. We present the experimental results and discussion in Section 5, followed by the conclusion in Section 6.

2 Related Work

This section briefly reviews related research on color preferences, color design tools, and recommendation systems.

2.1 Color Preferences

Researchers have studied color preferences for many years, each trying to extract specific factors that influence color choice and analyze the degree of their influence. Factors traditionally thought to influence color preferences include culture, gender, age, and country of residence.
Hurlbert and Ling [14] suggested that experiments with British and Chinese subjects revealed distinct color preferences according to gender due to biological differences in the human visual system. Bonnardel et al. [3] examined gender differences in color preferences between British and Indian people, claiming that women showed an even stronger preference for pink and purple. In terms of age-related differences, Gong and Lee [11] reported that an older group tended to prefer colors with low saturation and achromatic colors. A survey of 1,290 Chinese adults aged 17 to 87 by Zhang et al. [34] showed that preference for short-wavelength colors (e.g., cyan, blue, and purple) decreases with age.
Personality traits are a critical factor in shaping individual differences in thought and behavior [28], and their tendencies influence behavior and decision making [18]. In general, an individual’s personality traits can often be expressed in terms of the Big Five personality traits, also referred to as the Five Factor Model, in which the five traits are defined as Extraversion, Agreeableness, Conscientiousness, Neuroticism, and Openness [10]. Recently, Jue and Ha [16] introduced the Big Five personality traits to analyze the correlation between personality traits and color preferences.
Cognitive abilities are the mental skills used for problem solving and visual reasoning. To our knowledge, no direct relationships between cognitive abilities and color preferences have been reported. However, an individual differences study by Liu et al. [18] suggests that abilities in five cognitive tasks, namely spatial ability, spatial memory, working memory, associative memory, and perceptual speed can affect performance in various visualization tasks. In this study, we intend to use the scores from each of the cognitive ability tests [8] as feature variables of individual cognitive abilities and further explore their interrelationship with color preferences.

2.2 Color Design Tools

Respecting color harmony is an important mechanism for coordinating consistent color palettes. From a theoretical perspective, Ou et al. developed a color harmony model for two colors [27] and then extended it to three colors [26]. Elsewhere, Cohen-Or et al. [6] developed a computational approach to extending color harmony specifically for computer graphics applications. This pioneering work was followed by the development of a knowledge-based approach for retrieving predefined color design rules [33] and semantically resonant color selection in data visualization [17].
The implementation of color palette design tools has been another important topic for the graphics and visualization communities. ColorBrewer [12] has been a popular color design interface for thematic map design. Linguistic information facilitates interactive exploration for better association between concepts and colors [31]. A computational approach for preserving sufficient perceptual distances between representative colors for labeling purposes has also been formulated [9]. A recent advance has been made by Misue, who has taken advantage of cross-sectional contours in color space to explicitly visualize color differences [21] and has developed the exploration of color ramps for quantitative data [22].

2.3 Recommendation Systems

Collaborative filtering is one of the most widely used techniques in recommendation systems. It simulates the underlying trends in users’ past preferences and thus allows us to predict users’ preferred colors. A number of improvements have been made to this technique [4]. For example, Adamopoulos and Tuzhilin [1] invented a probabilistic approach to avoid unwanted bias in recommendations, and Hassan and Hamada [13] incorporated a neural network model to implement multi-criteria recommendation systems.
The collaborative filtering process typically begins by creating a data matrix of preference scores for users and items (i.e., colors) as a table (left side of Figure 2). Matrix factorization (MF) facilitates the prediction of scores for unknown user-item pairs by decomposing the data matrix into two factor matrices for users and items, each consisting of a set of essential factor vectors (right side of Figure 2). In practice, the next items for specific users can be recommended by comparing the unknown data entries to those in the matrix product of these factor matrices. O’Donovan et al. [24] predicted aesthetic color preferences for individual users using probabilistic MF. Although MF is limited to describing linear relationships between users and items, it can provide highly interpretable recommendations without the time-consuming training of the predictive model typically used in machine learning approaches.
Among MF techniques, non-negative matrix factorization (NMF) facilitates the interpretation of factor matrices because each entry in the matrix product is represented as a weighted sum of score values, which are usually non-negative (Figure 2). Furthermore, to explore meaningful trends with respect to another condition together with users and items, we need to factorize 3D tensors instead of 2D data matrices. Morita et al. [23] implemented a color recommendation system using non-negative tensor factorization (NTF) [32], where they decomposed a 3D data tensor in terms of user, color, and affective expression into three factor matrices.
Figure 2:
Figure 2: NMF lets us factorize a data matrix V into matrices W and H. \(\hat{V}\) is the reconstructed matrix as the product of W and H and approximates the original matrix V [23].
Although tensor factorization (TF) models can make predictions regarding additional conditions by increasing the dimensionality of the data representations, they suffer from computationally expensive tensor decomposition as the dimensionality increases. This problem can be alleviated by the formulation of factorization machines developed by Rendle [29], which is a generalized version of conventional factorization models that is independent of the dimensionality of the data representations. In this study, we use this formulation and its variants because we parameterize the space of individual differences by multiple domains of feature variables. We will describe this formulation in detail later in Section 4.1.

3 Data Collection Through Crowdsourced User Studies

As described in Section 1, our first step is to collect data samples of the mapping between preferred color choices and the space of individual differences. We parameterized this space with several categories of variables, including gender, age, personality traits, and cognitive abilities, in addition to the background conditions of color choice. In particular, collecting samples of personality traits and cognitive abilities is crucial for studying the impact of individual differences in color choice. To this end, we introduced a crowdsourced user study with personality and cognitive tests, where we collected the Big Five personality traits and cognitive ability scores to explicitly form the parametric space of individual differences. At the same time, we surveyed participants to measure their individual color preferences along with their rating scores. The actual questionnaires were created using a questionnaire tool provided by QuestionPro, and participants were recruited using the crowdsourcing service Amazon Mechanical Turk (MTurk).

3.1 Collecting Color Preferences

In this study, we used a set of 41 colors created by Bartram et al. [2], who originally used them to investigate color selection tendencies across eight affective categories (i.e., positive, exciting, disturbing, calm, negative, playful, serious, and trustworthy). In the survey we administered, participants were asked to select their top five colors from 41 color swatches in order of preference for the eight affective conditions proposed by Bartram et al. and two additional conditions (most and least favorite colors). Previously selected colors were removed before the next choice to ensure that the same color could not be selected for each condition. We assigned scores from 5 (first preferred) to 1 (fifth preferred) to the top five colors chosen in this order of preference.

3.2 Collecting Personality Traits

In our setup, the Big Five personality traits were measured using the Mini-IPIP [7], a personality test based on the Five Factor Model. The Mini-IPIP is an abbreviated version of the full personality test, consisting of only 20 questions, which reduces the risk of attention loss and dropout due to the large number of questions. The Mini-IPIP quantifies the levels of the five primary factors that make up human personality (i.e., extraversion, agreeableness, conscientiousness, neuroticism, and openness). Each item in this questionnaire is described by a phrase that describes a person’s behavior. Participants rate how well each item describes them on a 5-point Likert scale (“very inaccurate” = 1, “moderately inaccurate” = 2, “neither inaccurate nor accurate” = 3, “moderately accurate” = 4, “very accurate” = 5). There are four items per personality factor and two types of phrases representing the extremes of each factor. For example, the statement “Am the life of the party” in Table 1 reflects a positive aspect of the extraversion factor. In contrast, “Keep in the background” reflects a negative aspect of the extraversion factor. Therefore, items representing the negative aspect of each factor are given inverted scores. For example, if the participant selects “very accurate” for “Keep in the background,” the score is reversed to 1 point. The minimum and maximum scores for each factor are 4 and 20 points, respectively, and the scores are normalized to the interval [0, 1] for later use.
Table 1:
ItemFactorText
1EAm the life of the party.
2ASympathize with others’ feelings
3CGet chores done right away.
4NHave frequent mood swings.
5OHave a vivid imagination.
6EDon’t talk a lot. (R)
7AAm not interested in other people’s problems. (R)
8COften forget to put things back in their proper place. (R)
9NAm relaxed most of the time. (R)
10OAm not interested in abstract ideas. (R)
11ETalk to a lot of different people at parties.
12AFeel others’ emotions.
13CLike order.
14NGet upset easily.
15OHave difficulty understanding abstract ideas. (R)
16EKeep in the background. (R)
17AAm not really interested in others. (R)
18CMake a mess of things. (R)
19NSeldom feel blue. (R)
20ODo not have a good imagination. (R)
Table 1: Mini-IPIP scale items created by Donnellan et al. [7]. E: Extraversion; A: Agreeableness; C: Conscientiousness; N: Neuroticism; O: Openness. (R): Reverse Scored Item.

3.3 Collecting Cognitive Abilities

Individual cognitive ability was measured using the Kit of Factor-Referenced Cognitive Tests created by Ekstrom et al. [8]. We employ the five cognitive abilities identified by Liu et al. [18] (i.e., spatial ability, spatial memory, working memory, associative memory, and perceptual speed). While the original version of this kit consists of pencil and paper tests, we have carefully selected questions for each of the five cognitive abilities that can be administered online. Our pilot study, conducted in the early stages of this research, revealed a higher incidence of careless responses toward the end of each test. Our thoughtful adjustments, including the elimination of some questions, were made to keep the entire set of cognitive ability tests within a 30-minute timeframe to reduce the burden on participants. We ensured a fair testing experience for all participants by adjusting the time limit for each test in proportion to the number of questions.

3.4 Filtering Out Unwanted Responses

Unwanted responses are typically characterized by random answers, failure to read the text of the questions and options, or repeating the same answer choices without reading the questions. Such carelessness by participants has caused unwanted problems in previous studies using reward-based online surveys, such as crowdsourcing, because their responses can negatively affect the reliability and usefulness of the data collected. MTurk offers a service called “Master” that allows us to assign tasks only to skilled workers. Even so, a survey completed only by master-registered workers will still contain inappropriate or unintended responses, as reported in [30]. In practice, we encountered such problems in our pilot user studies and recognized the need to carefully monitor response data and filter out careless responses from our data samples.
To address this issue, we implemented a two-step data-sampling process for effective participant screening. In Step 1, participants were asked to respond to measures of color preference and personality traits as in the first user study, where we used screening criteria consisting of directed qualification scales (DQSs) [19], consecutive identical responses, and overall response time. Here, DQSs serve as a technique for inserting instructions within the questionnaire that direct participants to explicitly select a particular item. For example, we have included “Choose the item closest to the right” and “Please select very inaccurate” as DQS items in our personality test. This makes it easier to eliminate unwanted participants by trapping them. In Step 2, participation in the cognitive ability tests of the second user study was limited to respondents carefully selected based on their responses in the first user study. This approach was carefully developed based on our experience with pilot user studies and helped reduce the burden of lengthy tasks on participants. In addition, since it is inherently difficult to determine the authenticity of responses in cognitive tests, this screening process also served as a preliminary investigation to identify workers on MTurk who were likely to provide serious responses.
In summary, we used the following procedures to recruit participants for crowdsourced user studies. We recruited MTurk workers by announcing the outline of our survey, including the reward amount and the expected allotted time. Workers who agreed to participate in the study were directed to a web page with a survey and test designed by QuestionPro, and then the survey began. The reward amount was calculated based on the US federal minimum wage of US$7.25 per hour: Step 1 was paid US$2.40 because it took 20 minutes to complete, and Step 2 was paid US$4.00 because it took 30 minutes to complete and was a highly burdensome task. We recruited 300 workers for Step 1. To ensure response quality, we qualified only 186 workers for Step 2. Since taking the tests in Step 2 was also voluntary, we ended up with a total of 135 workers who completed all of the personality tests, color selection, and cognitive ability measures. Note that the workers consisted of 58 females and 77 males, and were asked to indicate their corresponding age groups: under 20, 20-24, 25-29, 30-34, 35-39, 40-44, 45-49, 50-54, 55-59, 60-64, and 65 years and older, where we had 0, 11, 24, 28, 25, 16, 5, 12, 5, 6, and 3 participants in the groups, respectively.
Figure 3:
Figure 3: FM data structure. The feature vector consists of one-hot encoding of IDs of the user, color, gender, age group, quantities of personality traits and cognitive abilities, and color selection condition.

4 Recommendations with Multiple Types of Variables

Data collection from crowdsourced user studies allows the prediction of preferred colors based on several domains of variables, including gender, age, personality traits, and cognitive ability, in addition to color selection criteria. As explained in Section 2.3, traditional collaborative filtering approaches for recommendation systems provide color choices based on the past preferences of users, usually represented as two-dimensional matrices (Figure 2). In our recent study [23], we improved the performance of color recommendation systems by introducing affective expressions as an additional condition and representing the color choice data as 3D tensors. However, in the present case, we need to consider four or more fields of feature variables to predict preferred colors, which results in expressing data samples with tensors of four or more dimensions. This is difficult to implement because representing data samples with high-dimensional tensors is tedious in itself, and decomposing such tensors into lower-dimensional matrices requires special efforts. In this study, we solved the problem by introducing factorization machines (FMs), which allow us to handle multiple fields of feature variables in predicting favorite colors.

4.1 Factorization Machines

An FM, proposed by Rendle [29], is a generic predictor that combines features of the support vector machine (SVM) and the MF. Although SVMs are expected to perform well in classification tasks, they may still face challenges when learning sparse data, such as those handled in collaborative filtering. Various factorization models have been considered adequate for learning sparse data, but they are limited by the dimensionality of the data representation.
FMs allow us to solve problems arising from the high dimensionality of data tensors. Let us first consider a simple case where three users have rated some of the four colors, as represented by the 3 × 4 score matrix on the left of Figure 3. FMs transform each entry of the score matrix into a feature vector to explicitly formulate an interaction between the corresponding user and the color. For example, in FMs, the score of 1.0 that the first user gives to orange in the score matrix is represented as the top row vector of the table on the right (outlined in red) in Figure 3. In addition, as shown in the table on the right, we can naturally encode the gender, age group, personality traits, cognitive abilities, and color selection condition of each user, which of course introduces additional variable fields into this representation. More specifically, FMs introduce a convenient one-hot encoding of feature variables to reduce the high-dimensional representation to its low-dimensional approximation.
To compute the actual score for each color, we employ 2-way FM, a simple FM model that characterizes all pairwise interactions between feature variables as well as single ones. The score y for this type of FM is defined for n feature variables xi (i = 1, …, n) as:
\begin{equation}y = w_0 + \sum _{i=1}^n w_i x_i + \sum _{i=1}^{n-1} \sum _{j=i+1}^n ({\boldsymbol $v$}_i, {\boldsymbol $v$}_j) x_i x_j,\end{equation}
(1)
where the bias \(w_0 \in {\boldsymbol $R$}\), weight vectors \({\boldsymbol $w$}= (w_i) \in {\boldsymbol $R$}^n\), and latent vectors \({\boldsymbol $v$}_i \in {\boldsymbol $R$}^k \, (i = 1,\ldots ,n)\) are estimated by training the model. Note that the latent vectors are introduced in such a way that the inner product of \({\boldsymbol $v$}_i\) and \({\boldsymbol $v$}_j\) (i.e., \(({\boldsymbol $v$}_i, {\boldsymbol $v$}_j)\)) explicitly describes the synergistic effects between the two variables xi and xj, which can be considered as another advantage of the FMs. The dimension of the latent vector, k, is adjusted according to the required degree of precision in simulating the underlying color selection trends to be extracted; the larger the value of k, the more complicated the trends that can be captured. In our recommendation scenario, we use this formula to calculate the scores for the 41 color samples and present the color sets with good scores as candidates for selection.

4.2 Sophistication of the FM Model

In this study, we also introduced field-aware factorization machines (FFMs) [15] as an extension of FMs to enable high-quality prediction of color scores. The difference between FFMs and FMs lies in the preparation of different latent vectors for each field of feature variables. For the FFM model, the color score Eq. (1) is replaced by
\begin{equation}y = w_0 + \sum _{i=1}^n w_i x_i + \sum _{i=1}^{n-1} \sum _{j=i+1}^n ({\boldsymbol $v$}_{i,f(j)}, {\boldsymbol $v$}_{j,f(i)}) x_i x_j,\end{equation}
(2)
where f(i) is the field ID for the i-th feature variable. More specifically, in our color recommendation based on individual differences, we introduced a total of 209 feature variables, which are classified into the user field (135 variables), color field (41 variables), gender field (2 variables), age field (11 variables), personality trait field (5 variables), cognitive ability field (5 variables), and condition field (10 variables), as shown in Figure 3. This means that we have seven different fields of feature variables and use seven different latent vectors for each field in this formula. Here, except for the five personality trait variables and the five cognitive ability variables, we assign 0 or 1 to the variables because they will eventually become binary variables due to the one-hot encoding method.

5 Results

This section presents the experimental results, which demonstrate the effectiveness of the proposed approach, along with a discussion of its possible limitations and future extensions.

5.1 Training the FM Models

The implementation of our color recommendation system starts with training the FM model with data samples obtained from crowdsourced user studies. From the results of the user studies with the 135 participants, we extracted 6,750 ratings of the top five preferred colors for each of the 10 color selection conditions (i.e., 135 × 5 × 10 = 6, 750). We then randomly divided the color rating samples into a training dataset (80%), a test dataset (10%), and a validation dataset (10%). We employed the xLearn machine learning package to compute two-way FMs and FFMs to construct the prediction model for color selection. Here, we used the Follow the Regularized Leader (FTRL) algorithm [20] as the optimizer, wherein we set the learning rate to 0.01 and the epoch number to 200. The dimensions of the latent vectors were tested with k = 8, k = 32, and k = 128. Table 2 shows the statistics of the errors for each case, for which we used a laptop PC (MacBook Pro) with an Intel Core i7 four-core CPU and 32GB RAM to measure the computation times.
Table 2:
 FMFFM
 k = 8k = 32k = 128k = 8k = 32k = 128
RSME1.7112.5092.5172.4392.6692.641
MAE1.4021.9601.9931.9312.1022.037
MAPE0.73430.96180.97330.93321.0260.9778
Time1.152.989.977.6027.78112.12
Table 2: Root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) of the test data, and calculation time (in seconds) for each case.

5.2 Experimental Results

First, we examine how well the trained FM models simulate the color choices of the subjects who participated in the crowdsourced user studies. Figure 4 compares the color set predicted by the FM models with the five colors selected by the two participants under specific color selection conditions. For the FM models, we used an FM with k = 8 and an FFM with k = 128. The results show that the more complex FFM model with k = 128 outperforms the simpler FM model with k = 8. For example, in the case of Participant A for the Positive condition, the FFM model includes a variety of colors, including greenish and bluish colors, while the FM model does not, as shown in Figure 4(a). Figure 4(b) shows another case in which the simple FM model cannot predict yellowish or brownish colors as the FFM does.
Figure 4:
Figure 4: Comparison of the five colors actually selected by specific participants with the candidate colors predicted using the trained FM models. Colors chosen by (a) Participant A for the Positive condition and (b) Participant B for the Disturbing condition.
We have also implemented a prototype system that suggests preferred colors based on the feature variables of individual differences and conditions for color selection (Figure 1). Our interface consists of two windows. The primary window displays the 41 color samples selected in [2] at the top, ten candidate color samples predicted by our recommendation model at the bottom right, and up to five color bars selected by the user as a palette at the bottom left. The control panel window provides radio buttons and sliders to control the feature parameters that make up the individual difference space and the conditions for color selection. The source code of the system was written in C++ using FLTK for the user interface and OpenGL for drawing color swatches. Our system loads the precomputed weights and latent vectors of the FMs and FFMs on which we base our color score computation. Note that in the design examples below, we specifically use the FFM with k = 128.
Figure 5 shows snapshots of our prototype system used to interactively design the color palette with our recommendation system. Starting from the initial state (5a), we first change the gender, age, and color selection condition, and then adjust the parameters of personality traits and cognitive abilities (5b). After adjusting the parameters, we select color samples one by one from the top ten candidates (5c). We can also cancel some of the selected color samples and introduce another color directly from the 41 color samples and the recommended color samples (5d).
A possible scenario for our color recommendation system is that users first set their gender, age, and color selection conditions, and then receive suggested color options. In addition, they can enter their scores for personality traits and cognitive abilities if they know the scores in advance, e.g. from paper tests. In this way, our prototype system helps users compose their preferred color sets by exploring the feature variables that define individual differences.
Figure 5:
Figure 5: Designing the color palette by adjusting feature variables comprising individual differences. (a) Initial state. (b) Select the gender, age, and color selection condition, and adjust the five personality traits and five cognitive abilities (outlined in red). (c) Add the three color candidates from the pool of recommended colors to the palette (outlined in orange.). (d) Cancel two color swatches and add one directly from the 41 color swatches and one from the pool of recommended colors (outlined in orange).

5.3 Discussion

The first takeaway is that we must be careful to train the FM models to compute the associated weights and latent vectors. Even with a relatively small dataset based on 135 participants, we failed to explore appropriate latent variables for the model such that each feature variable contributed adequately to the scores of individual colors in the early stages of our study. We mitigated this problem by ensuring a sufficient number of epochs for training and by using the FTRL algorithm for optimization instead of simple stochastic gradient descent. The refinement of training setups for FM models may require further investigation.
Our color rating score functions are essentially quadratic polynomials of the feature variables because we use simple two-way FM models in our approach. This formulation does not simulate the nonlinear trends that result from the complex synergistic effects among the feature variables. Replacing the two-way FM models with those with deep learning modules [5] may be a promising approach to solving this problem; it has been left as future work.
Our data samples are based on the responses of only 135 participants and may be too skewed to detect general trends. Increasing the number of participants comes at an additional cost, primarily because we must ask them to complete a time-consuming web-based questionnaire derived from paper-based personality and cognitive tests. This means that we implicitly require users to take these tests prior to designing color palettes if they want to take full advantage of parameter tuning for the space of individual differences. We are currently exploring a new approach to obtain feature variables for such individual differences from users on the fly.
A detailed analysis of possible synergistic effects between different fields of feature variables is a fascinating research topic, especially when investigating the impact of individual differences on color design tasks. By calculating the inner products of latent vectors in eqs. (1) and (2), we can identify synergistic effects between the specific pairs of feature variables. For example, it can be observed that older participants generally tend to choose less saturated colors, as pointed out in [11]. Other trends include consistently positive collaborative effects between extraversion and agreeableness, extraversion and conscientiousness, and agreeableness and conscientiousness, regardless of the choice of FM models (i.e., FMs or FFMs) or the dimension of the latent vectors (i.e., k). Although our initial investigation exposes such meaningful correlations between certain pairs of feature variables, we need to collect more samples to fully validate our findings.
A promising application of the present technique is to recommend the optimal choice among more complicated visualization tasks by exploring the space of individual differences. When understanding the correlation between numerical data samples, for example, it is still challenging to propose a visualization method that respects the individual differences of each user. This is because a particular type of visual graph may require a particular skill in certain cognitive abilities. Broadening the scope of our research is an important future challenge for us.

6 Conclusion

This paper presented an approach to color recommendation that accounts for individual differences among users. The space of individual differences was parameterized by personality traits and cognitive abilities in addition to gender and age. The data samples in the space of individual differences were obtained through crowdsourced user studies in conjunction with selected colors under specific conditions. We used the obtained data samples to train recommendation models that predict the color choices of users based on their individual differences. As recommendation models, we used two-way FMs and FFMs, which can handle the multiple fields of feature variables, along with their mutual synergistic effects, without high-dimensional tensor data representations. We also demonstrated the feasibility of the proposed approach by implementing a prototype color recommendation system and discussed the limitations and potential extensions of the present approach.

Acknowledgments

This work is supported by JSPS KAKENHI grant numbers 24K02981 and 19H04120.

Supplemental Material

MP4 File
Demonstration video

References

[1]
P. Adamopoulos and A. Tuzhilin. 2014. On Over-Specialization and Concentration Bias of Recommendations: Probabilistic Neighborhood Selection in Collaborative Filtering Systems. In Proceedings of the 8th ACM Conference on Recommender Systems (RecSys ’14). 153–160.
[2]
L. Bartram, A. Patra, and M. Stone. 2017. Affective Color in Visualization. In Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems(CHI ’17). 1364–1374.
[3]
V. Bonnardel, S. Beniwal, N. Dubey, M. Pande, and D. Bimler. 2018. Gender difference in color preference across cultures: An archetypal pattern modulated by a female cultural stereotype. Color Research & Application 43, 2 (2018), 209–223.
[4]
R. Chen, Q. Hua, Y.-S. Chang, B. Wang, L. Zhang, and X. Kong. 2018. A Survey of Collaborative Filtering-Based Recommender Systems: From Traditional Methods to Hybrid Methods Based on Social Networks. IEEE Access 6 (2018), 64301–64320.
[5]
H.-T. Cheng, L. Koc, J. Harmsen, T. Shaked, T. Chandra, H. Aradhye, G. Anderson, G. Corrado, W. Chai, M. Ispir, R. Anil, Z. Haque, L. Hong, V. Jain, X. Liu, and H. Shah. 2016. Wide & Deep Learning for Recommender Systems. In Proceedings of the 1st Workshop on Deep Learning for Recommender Systems(DLRS 2016). 7–10.
[6]
D. Cohen-Or, O. Sorkine, R. Gal, T. Leyvand, and Y.-Q. Xu. 2006. Color Harmonization. ACM Transactions on Graphics 25, 3 (2006), 624–630.
[7]
M. B. Donnellan, F. L. Oswald, B. M. Baird, and R. E. Lucas. 2006. The Mini-IPIP Scales: Tiny-Yet-Effective Measures of the Big Five Factors of Personality. Psychological Assessment 18, 2 (2006), 192–203.
[8]
R. B. Ekstrom, J. W. French, and H. H. Harman. 1976. Manual for Kit of Factor-Referenced Cognitive Tests.
[9]
H. Fang, S. Walton, E. Delahaye, J. Harris, D. A. Storchak, and M. Chen. 2017. Categorical Colormap Optimization with Visualization Case Studies. IEEE Transactions on Visualization and Computer Graphics 23, 1 (2017), 871–880.
[10]
L. R. Goldberg. 1992. The development of markers for the Big-Five factor structure. Psychological Assessment 4, 1 (1992), 26–42. Place: US Publisher: American Psychological Association.
[11]
S.‐M. Gong and W.‐Y. Lee. 2017. Colour Preference Model for Elder and Younger Groups. Journal of the International Colour Association 18 (2017), 33–42. https://aic-color.org/resources/Documents/jaic_v18_03.pdf
[12]
M. Harrower and C. A. Brewer. 2003. ColorBrewer.org: An Online Tool for Selecting Colour Schemes for Maps. The Cartographic Journal 40, 1 (2003), 27–37.
[13]
M. Hassan and M. Hamada. 2017. A Neural Networks Approach for Improving the Accuracy of Multi-Criteria Recommender Systems. Applied Sciences 7, 9 (2017).
[14]
A. C. Hurlbert and Y. Ling. 2007. Biological Components of Sex Differences in Color Preference. Current Biology 17, 16 (2007), R623–R625.
[15]
Y. Juan, Y. Zhuang, W.-S. Chin, and C.-J. Lin. 2016. Field-Aware Factorization Machines for CTR Prediction. In Proceedings of the 10th ACM Conference on Recommender Systems(RecSys ’16). 43–50.
[16]
J. Jue and J. H. Ha. 2022. Exploring the Relationships Between Personality and Color Preferences. Frontiers in Psychology 13 (2022).
[17]
S. Lin, J. Fortuna, C. Kulkarni, M. Stone, and J. Heer. 2013. Selecting Semantically-Resonant Colors for Data Visualization. Computer Graphics Forum 32, 3 (2013), 401–410.
[18]
Z. Liu, R. J. Crouser, and A. Ottley. 2020. Survey on Individual Differences in Visualization. Computer Graphics Forum 39, 3 (2020), 693–712.
[19]
M. R. Maniaci and R. D. Rogge. 2014. Caring About Carelessness: Participant Inattention and its Effects on Research. Journal of Research in Personality 48 (Feb. 2014), 61–83.
[20]
H. B. McMahan, G. Holt, D. Sculley, M. Young, D. Ebner, J. Grady, L. Nie, T. Phillips, E. Davydov, D. Golovin, S. Chikkerur, D. Liu, M. Wattenberg, A. M. Hrafnkelsson, T. Boulos, and J. Kubica. 2013. Ad Click Prediction: a View from the Trenches. In Proceedings of the 19th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining(KDD ’13). 1222–1230.
[21]
K. Misue. 2020. Development of a Tool to Help Understand Color Spaces and Color Differences. In Proceedings of the 24th International Conference Information Visualisation(IV2020). 565–572.
[22]
K. Misue. 2020. Tools for Developing Color Ramps for Representing Quantitative Data. In Proceedings of the 13th International Symposium on Visual Information Communication and Interaction(VINCI2020).
[23]
I. Morita, S. Takahashi, S. Nishimura, and K. Misue. 2022. Affective Color Palette Recommendations with Non-negative Tensor Factorization. In 2022 26th International Conference Information Visualisation(IV2022). IEEE, 40–47.
[24]
P. O’Donovan, A. Agarwala, and A. Hertzmann. 2014. Collaborative Filtering of Color Aesthetics. In Proceedings of the Workshop on Computational Aesthetics (CAe’14). 33–40.
[25]
A. Ottley, E. M. Peck, L. T. Harrison, D. Afergan, C. Ziemkiewicz, H. A. Taylor, P. K. J. Han, and R. Chang. 2016. Improving Bayesian Reasoning: The Effects of Phrasing, Visualization, and Spatial Ability. IEEE Transactions on Visualization and Computer Graphics 22, 1 (2016), 529–538.
[26]
L.-C. Ou, P. Chong, M. R. Luo, and C. Minchew. 2011. Additivity of Colour Harmony. Color Research & Application 36 (2011), 355–372.
[27]
L.-C. Ou and M. R. Luo. 2006. A Colour Harmony Model for Two-Colour Combinations. Color Research & Application 31, 3 (2006), 191–204.
[28]
E. M Peck, B. F Yuksel, L. Harrison, A. Ottley, and R. Chang. 2012. Towards a 3-dimensional model of individual cognitive differences: position paper. In Proceedings of the 2012 BELIV Workshop: Beyond Time and Errors - Novel Evaluation Methods for Visualization(BELIV ’12).
[29]
S. Rendle. 2010. Factorization Machines. In 2010 IEEE International Conference on Data Mining. 995–1000.
[30]
A. Saravanos, S. Zervoudakis, D. Zheng, N. Stott, B. Hawryluk, and D. Delfino. 2021. The Hidden Cost of Using Amazon Mechanical Turk for Research. In HCI International 2021 - Late Breaking Papers: Design and User Experience. Vol. 13094. 147–164. Springer LNCS.
[31]
V. Setlur and M. 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.
[32]
A. Shashua and T. Hazan. 2005. Non-Negative Tensor Factorization with Applications to Statistics and Computer Vision. In Proceedings of the 22nd International Conference on Machine Learning(ICML ’05). 792––799.
[33]
L. Wang, J. Giesen, K. T. McDonnell, P. Zolliker, and K. Mueller. 2008. Color Design for Illustrative Visualization. IEEE Transactions on Visualization and Computer Graphics 14, 6 (2008), 1739–1754.
[34]
Y. Zhang, P. Liu, B. Han, Y. Xiang, and L. Li. 2019. Hue, Chroma, and Lightness Preference in Chinese Adults: Age and Gender Differences. Color Research & Application 44, 6 (2019), 967–980.

Recommendations

Comments

Information & Contributors

Information

Published In

cover image ACM Other conferences
VINCI '24: Proceedings of the 17th International Symposium on Visual Information Communication and Interaction
December 2024
286 pages
ISBN:9798400709678
DOI:10.1145/3678698

Publisher

Association for Computing Machinery

New York, NY, United States

Publication History

Published: 11 December 2024

Check for updates

Author Tags

  1. Color recommendations
  2. Individual differences
  3. Crowdsourcing user studies
  4. Factorization machines

Qualifiers

  • Research-article

Funding Sources

  • Japan Society for the Promotion of Science

Conference

VINCI 2024

Acceptance Rates

Overall Acceptance Rate 71 of 193 submissions, 37%

Contributors

Other Metrics

Bibliometrics & Citations

Bibliometrics

Article Metrics

  • 0
    Total Citations
  • 196
    Total Downloads
  • Downloads (Last 12 months)196
  • Downloads (Last 6 weeks)64
Reflects downloads up to 15 Feb 2025

Other Metrics

Citations

View Options

View options

PDF

View or Download as a PDF file.

PDF

eReader

View online with eReader.

eReader

HTML Format

View this article in HTML Format.

HTML Format

Login options

Figures

Tables

Media

Share

Share

Share this Publication link

Share on social media