1 Introduction
Today, the lifestyle of people is moving toward more online-based activities. These changes have been accelerated due to the outbreak of COVID-19, leading to an increase in additional activities in the online environment. Several industries were affected by this change, with the art industry being one of the most influenced areas.
From a revenue perspective, more than 90% of museums and galleries were closed in March 2020 [
11]. Furthermore, based on the Network of European Museum Organisations (NEMO) survey, which included 1,000 responses in 48 countries from March to April 2020, most museums reported an income loss from 75% to 80%. This situation forced museums to reallocate staff to strengthen their online exhibition activities, and half of the museums responded that they offer at least one online service [
24].
For instance, more than 2,500 museums and galleries around the world teamed up with Google Arts & Culture aiming to provide the facility for anyone and everyone for various virtual tours and online exhibits to meet the increased demand [
22]. King et al. [
11] investigated 88 temporary exhibitions across the UK during the COVID-19 crisis and found that 21 exhibitions were offering online content showcasing their displays. This content was available through multiple channels, including their websites, social media, and YouTube. In other words, as investment in online exhibitions is inevitable, attracting visitors to online exhibitions has become a more important issue for the success of art museums and galleries. In this research, we focus on a preference-based recommendation system as a tool to attract audiences online by applying it to museums’ marketing strategies and consequently raising the satisfaction of audiences on online art exhibitions.
Personalized recommendations by a recommendation system, which is one of the information filtering systems, reflect users’ preferences in presenting such items. A recommendation system collects data about the user’s preferences on different items like music, movies, and shopping [
3]. Liang et al. [
15] revealed that personalized services help increase user satisfaction by accurately recommending relevant content to users. In the film and music industries, recommendation systems serve as vital tools for delivering online services and enhancing profits. For instance, Netflix integrated a recommendation system into its platform to reduce subscription cancellations. According to Gomez-Uribe and Hunt [
9], a combination of recommendations such as ‘Trending Now,’ ‘Continue Watching,’ ‘Because You Watched,’ and others could potentially save up to $1 billion annually. Additionally, Schedl et al. [
28] highlight the significance of recommendation systems in the music industry, noting that in the United States, streaming platforms have become the primary revenue source for music consumption. However, in comparison to other cultural domains where diverse recommendation system techniques are actively researched and implemented, the application of recommendation systems in the art field has garnered less attention [
8].
Herein, we explore whether users’ satisfaction increases when artworks are recommended based on personal taste. If artwork recommendations that suit personal tastes increase the satisfaction of audiences, then it can be confirmed that introducing an artwork recommendation system in online exhibitions has a practical effect on attracting the audience and leads the exhibition to success. Thus, we attempt to address the following research question (RQ):
RQ: Can user satisfaction be affected by a preference-based artwork recommendation system?
2 Related Work
The most commonly used filtering methods for recommending specific items are collaborative- and content-based filtering approaches. The collaborative filtering approach recommends new items to a target user according to their previous preference information and similarity with other users [
7]. Content-based filtering recommends items to a user based on item information and the profile of the user’s interests [
23]. While collaborative- and content-based filtering are the two most commonly used recommendation system methods, clustering, a method of grouping a set of inputs based on their feature similarities, is also a commonly used supporting technique [
7].
In the fields of music and video, several clustering-based recommendation solutions have been addressed, whereas commercial services utilizing the solutions are proposed and presented. For instance, Spotify, one of the most popular music streaming services, utilizes music features such as users’ rating, how frequently a track has been played, and music genre to generate automatic playlists for its users [
5]. King and Imbrasaitė [
12] proposed a recommendation model, which creates six unique clusters developed by
k-means and Q-learning, considering users’ listening history and activities (e.g., skips, search, jumps). Roy et al. [
25] introduced a multi-context-aware user model, delineating various clusters derived from situational contexts extracted from playlist names and music characteristic features of Indian classical music, specifically Raga. Given Raga’s distinctive segments within a single composition, clustering techniques were employed to suggest a session that aligns with a user’s preferences.
Regarding video-oriented topics, recommendation systems are one of the critical tools to ameliorate user experience and promote sales and services (e.g., Netflix, Disney+, and YouTube) [
33]. For example, 80% of movies playing on Netflix is recommended by its algorithms [
9]. As for YouTube, 60% of video clicks is originated by automatic recommendation solutions [
6]. Moreover, several well-adapted movie recommendation systems developed by
k-means clustering and
k-nearest neighbors algorithms are proposed [
3,
4,
19].
Compared to other culture domains, artwork recommendation systems are still in their early stages. As a representative example, the CHIP project [
31], which implemented a personalized recommendation system for the Rijksmuseum, has shown the potential of personalized recommendations in art. To infer museum visitors’ preferences from a set of scored artifacts, the CHIP project combines both semantic Web technologies and content-based algorithms to recommend other artworks and related content topics.
Most recent research has employed the combined datasets of artworks and their metadata, such as genres, writers, or introduced years. Yilma et al. [
32] proposed a personalized artwork recommendation system by adopting latent Dirichlet allocation based on its representation learning approach. Messina et al. [
17] showed a hybrid approach using artworks’ metadata and visual features, presenting a better recommendation performance than the traditional approach. They employed both artwork images and transaction data from the UGallery3 online artwork store.
TindArt [
30] is an art recommendation system in which the users choose whether they like or dislike the displayed paintings. Then, a set of new paintings is presented with consideration of users’ tastes. Although there are several implications in recommending artworks, there are some concerns and limitations, such as its accountability, because user satisfaction, which is one of the essential parts of the recommendation tasks, is not addressed.
He et al. [
10] used pre-trained deep neural networks for a recommendation task of digital art with data from Behance, a website including millions of digital artworks and corresponding user interactions. However, their approach could not be extended to online exhibition environments. This recommendation method relies on extensive click data from Behance, which constitutes implicit user feedback. Still, implementing it on an art museum platform with relatively fewer interactions from visitors poses a challenge.
Thus, to effectively address it, we propose a preference-based artwork recommendation system, which adopts a clustering approach considering artworks’ visual features. Moreover, to easily apply it to online exhibition environments and investigate both the effectiveness and potentiality of the suggested system, we conducted a within-subjects experiment and tested the system’s practicality.
4 Experiments
To recommend appropriate artworks, it is required to employ a set of artworks that users selected. Thus, a within-subjects experiment was conducted with two types of recommendations. The first set was extracted from a cluster, which was the closest to the artwork that each user liked. The second set was selected from another cluster, which was the farthest from a group of artwork that the user liked. To examine the experiment, we developed a website that collects real-time user preferences on the presented artworks.
4.1 Materials
The website presents 20 artworks selected by 10 artwork styles (2 representative artworks from each style). To select the most representative artworks from each style, Google Trends and Search were employed to validate the representativeness of artworks in each style.
Making Recommendation Sets. To create recommendation sets for each user, user preference had to be observed. For each artwork, “like” or “dislike” buttons were presented and clicked by each user (Figure
3). Then, the website provided two recommendation sets, each including 10 pieces of artwork. After examining all responses, we computed a
Centroid of the Liked Images (CLI). Recommendation set A is organized by 10 artworks from a certain cluster, the closest centroid from CLI, whereas recommendation set B includes 10 artworks from the farthest cluster from CLI.
4.2 Participants
Forty-two undergraduate and graduate students were recruited from one of the large private universities in Seoul, Korea. The participants’ ages ranged from 20 to 39 years. Table
2 shows the demographic profile of participants.
In the pre-experiment survey using a 7-point Likert scale (1: Strongly disagree, 7: Strongly agree), 55% of participants positively responded to the first question (4, 5, 6, or 7; “I like to appreciate artworks”). The responses to the second question (“I have substantial knowledge of artworks”) were distributed (1 or 2: 40%; 3 or 4: 50%; 5, 6, or 7: 10%). Figure
3 shows the proportion of answers to each question in pie charts.
4.3 Procedures
All participants were instructed to enter the art recommendation website. After conducting Making Recommendation Sets (MRS) procedures, recommendation set A was presented to each participant. Then, the participant was asked to respond to the questionnaire items to assess their perspectives on artworks in the set. A few minutes later, recommendation set B was presented to the participant. After watching the set, the same questionnaire items on recommendation set B were presented. Last, the participant was instructed to express their feelings on the experiment in this interview session, which included the reasons for the responses (e.g., textual elements, artwork genre, or artist). The order of recommendation sets was randomized.
The questionnaire items on each set in this study were employed in prior research (similarity, preferences, perceived satisfaction, intention to recommend) [
2,
21]. Additionally, a single question examining visual factors of artworks, genres, artists, and textual information, which can have notable effects on their perspectives on the artworks, was presented. Figure
4 shows the experiment website.
5 Results
A series of
t-tests were conducted to investigate whether the participants have different perspectives on artworks between recommendation sets A and B. Figure
5 shows the summary of the results. (Q1) The results of
t-tests indicated that the participants with recommendation set A (M = 4.95, SD = 1.09) felt a greater similarity of artworks on their artwork selections than those with recommendation set B (M = 3.14, SD = 1.42),
t(42) = 6.43,
p< 0.001. (Q2) Moreover, the participants showed a higher level of perceived preferences for set A (M = 5.12, SD = 1.14) rather than set B (M = 3.12, SD = 1.48),
t(42) = 6.44,
p< 0.001. (Q3) Additionally, the degree of perceived satisfaction on set A (M = 5.02, SD = 1.30) was higher than that on set B (M = 3.33, SD = 1.47),
t(42) = 4.88,
p< 0.001. (Q4) Set A (M = 5.17, SD = 1.41) also allowed the participants to have a greater level of intention to recommend than set B (M = 3.26, SD = 1.60),
t(42) = 6.18,
p< 0.001. To account for the influence of covariate variables, an analysis of covariance (ANCOVA) was performed. The findings indicated that the covariate variables were not statistically significant. Specifically, when examining the difference in perceived satisfaction between recommended sets and gender (
F(1, 81) = 1.17,
p< 0.001), as well as the annual frequency of visits to art museums (
F(1, 81) = 0.003,
p< 0.001), both exhibited
p-values below the significance threshold of 0.5.
In the following interview sessions, most participants (74%) responded that their perceptions of artworks are mainly determined by the artwork presentations, following their genres (19%) and explanations (7%).
6 Discussion and Conclusion
In this article, we proposed an artwork recommendation system, which presents a set of artworks based on users’ preferences. To address it, we employed a combination of VGG16 and k-means clustering algorithms, which categorize similar artworks and display the randomly selected artworks from the cluster for responding to the participants’ preferences. To provide a better understanding of the role of users’ preferences on artwork recommendations, the experiment in this study was designed with the participants’ preferences. Then, we made two recommendation sets, presented them to the participants, and received their responses on the sets. Thus, we confirmed a significant relationship between participants’ preferences and their satisfaction with artworks.
From our practical perspectives, several implications are presented as follows:
—
Because art museums, galleries, and art fairs are consistently moved to online environments, it is required to create online exhibitions that reflect each user’s preferences. Thus, if a user-oriented recommendation system is actively employed in this area, it can promote the sales and marketing of art and make better contributions to art popularization and art industry development.
—
Both utilization and practicality of artwork should be improved. Although there is a notable limitation in commercially using public artworks, stakeholders can find additional commercial points in online environments due to the rapidly improving technologies. One of the notable possibilities lies in comprehending consumer inclinations more comprehensively by connecting preferences for artworks with preferences in other cultural genres, such as music and movies.
Although some implications are presented, notable limitations remain. First, the experiment in this study was conducted with participants, who were mainly in their 20s and 30s. This means that there may be notable generation biases in the results. Moreover, these generations are generally good at using online services. Second, we did not consider any demographic information in our analysis. Last, considering that we solely focused on the WikiArt dataset, it is imperative to extend the application of this recommendation system to other art datasets in future research.