Daehyun Kwak
Youn-Kyung Lim
Abstract
Designing user experiences for group recommendation systems (GRS) is challenging, requiring a nuanced understanding of the influence of social interactions between users. Using Spotify Blend as a real-world case of music GRS, we conducted empirical studies to investigate intricate social interactions among South Korean users in GRS. Through a preliminary survey about Blend experiences in general, we narrowed the focus for the main study to relationships between two users who are acquainted or close. Building on this, we conducted a 21-day diary study and interviews with 30 participants (15 pairs) to probe more in-depth interpersonal dynamics within Blend. Our findings reveal that users engaged in implicit social interactions, including tacit understanding of their companions and indirect communication. We conclude by discussing the newly discovered value of GRS as a social catalyst, along with design attributes and challenges for the social experiences it mediates.
1 INTRODUCTION
Recommender systems continue to permeate people’s everyday lives. While those systems have traditionally been designed for individual users, recent attempts have emerged to provide recommendations for multiple users. Named as group recommendation systems (GRS) [3, 41], these systems offer personalized recommendations to multiple users based on each user’s unique preferences. For instance, GRS can provide recommendations on a list of activities that users can enjoy collectively, such as music [14, 27, 37], movies [1, 15, 46], and TV programs [9, 52, 53].
A majority of past research on GRS has revolved around the issue of accuracy and efficiency of shared personalized recommendations, championing the value of incorporating multiple GRS users’ preferences as a priority. For instance, researchers have been spurred into studies on recommendation mechanisms from a rather technical perspective, which includes what data GRS should collect [27, 47, 48] or how GRS should aggregate user preferences [24]. Research into interaction design for GRS has also been much dedicated to more efficient attainment of a group goal, such as designing interactions that support multiple users’ decision-making processes [40, 45, 54] or support users’ understanding of recommendation algorithms [13, 49].
While achieving the accuracy of personalized recommendations is important, we suggest that researchers should orient their attention more towards designing GRS in consideration of its social contexts, given its nature as a technology shared and interacted by groups. Indeed, prior works suggest that interpersonal dynamics occurring in such social contexts can also determine the qualities of the user experiences of a shared technology [42, 43, 58]. For example, Tibau et al. [58] identified that a music-sharing system for geologically distributed family members enabled serendipitous informal communication among them, which enhanced their sense of connectedness and eventually positively influenced the user experiences of the system. While not explicitly conducted as research on GRS, this prior work implies the potential of multi-user content-sharing technology as a facilitator of social interactions, and therefore an important research frontier in the design of GRS from the social perspective. For example, we can imagine situations where users may engage in collaborative playlist curation, fostering a sense of shared musical identity, and strengthening social bonds. Also, considering the potential for diverse preferences within a group, GRS could implement features that facilitate negotiation and compromise, ensuring a more satisfying collective experience.
Motivated by this need for a renewed conceptualization of GRS from a social lens, we aimed to explore what social interactions can occur in and be facilitated by GRS. In this paper, we used Spotify Blend, which is a real-world GRS service actively used in social contexts. Through a preliminary survey to gain insights into South Korean users’ general experiences with Blend, we revealed the need to investigate social interactions in Blend, particularly between two people who are acquainted with or close to each other. Building upon these findings, we conducted a 21-day diary study and interviews with 30 participants, consisting of 15 pairs. Delving more in-depth into the various interpersonal dynamics that unfolded within the context of Blend, we observed a variety of implicit social interactions that participants engaged in while using Blend, which encompassed unspoken mutual understanding of their companions as well as indirect forms of communication, such as offering non-coercive suggestions, confirming an expansion of common interests, and demonstrating prosocial behaviors.
Based on the findings, we discuss the newly found value of GRS as a social catalyst, suggesting two key attributes that determine the experiential quality of such a system: the level of ambiguity of recommendation algorithms and the visibility of user activities. We conclude by discussing important design challenges that may arise in GRS due to the co-existence of individual and collective recommendation experiences.
2 RELATED WORK
We motivate our focus on the understanding of social interactions in GRS by discussing previous literature on the user experience design of GRS. Also, we review existing work on how personalized recommendations are experienced in various social situations, identifying the significance of studying social interactions that are uniquely captured in GRS.
2.1 User Experiences of Group Recommendation Systems
Current research on GRS has been largely dedicated to providing a final decision that would satisfy all group members. Toward this goal, researchers have placed a heavy emphasis on recommendation accuracy and sought to address relevant key agendas [24] such as aggregating preferences, generating recommendations, explaining the rationales of recommendations, and helping users to settle on a final decision. Most research is on technical solutions [5, 6, 35, 36, 39, 60] developed to inform the process of effectively aggregating various preferences and accordingly generating recommendations. Also suggested is a design strategy [49] that explains the basis on which recommendations are generated by displaying the user profiles associated with each recommended item. Plus, there have been suggested systems that facilitate group members’ democratical voting on preferences [45], make a single user an ultimate decision-maker [54], or provide a chat box for negotiation among group members [40].
While the provision of accurate recommendations is crucial in ensuring pleasant GRS user experiences and satisfaction, other works have started rethinking how the very experience of social interactions in GRS can determine its experiential qualities [59]. For instance, a prior work [47] has identified that social factors, such as tie strength, can affect GRS experiences. Also, they noted that users may follow the recommendations that other members find appealing, even if the recommendations do not align with their own preferences. User personalities seem to matter as well; for example, Recio-Garcia et al. [50] and Quijano-Sanchez et al. [47] suggested that cooperative users tend to be more receptive to the preferences of other members, while assertive users prioritize their own preferences relatively more. Also, users may lower their preferences for a particular content when they become aware that others do not enjoy it [18, 36].
Collectively, these prior studies suggest that recommendation accuracy is not the sole contributor to good GRS user experiences, highlighting the importance of social factors among users in affecting those experiences. Building on those studies that foreground the social nature of GRS, our research broadens the conceptualization of GRS beyond its conventional role as a group decision-making tool. In this paper, we investigate the potential of GRS as a social technology, which mediates social interactions and shared experiences among users.
2.2 Social Experiences Mediated by Personalized Recommendations
As the term ‘personalization’ implies, algorithmic recommendations have commonly been perceived as tailoring content and experiences within personal contexts. However, emerging research indicates that people experience personalized recommendations in various social situations, beyond merely consuming them on an individual level. For example, studies have investigated the experiences of account sharing in personalized services, discovering that people want to share personalized Netflix recommendations [7, 51] with trustworthy partners [23]. These studies underscore the importance of understanding how recommender systems should be designed in consideration of the practice of sharing within interpersonal relationships. Also, a recent investigation by Bhuiyan et al. [8] illustrated that design for the sharing of personalized YouTube recommendations among unknown users could foster social ties among them. Another body of literature has investigated the user experiences of disclosing personalized recommendations in social situations, such as YouTube co-watching [56] or using one’s Spotify account for listening to music at a party [22]. These works offered a nuanced understanding of how to support people in managing their identities revealed through curated content in real-world shared environments.
The above collection of work demonstrates that personalized recommendations, when experienced in social situations, can affect social interactions and relationships among users. Building on this, we capture the potential of similar phenomena in GRS, a system intentionally designed for a group of users to experience shared recommendations. Particularly, by adopting Spotify Blend as a context of our study, we aimed to explore nuanced social interaction practices in a more intricate social situation where users experience recommendations based on both individual and collective preferences. With this motivation, our work aims to illuminate what social interactions and experiential values are entangled with such a system.
3 STUDY METHODS
We aimed to explore what social interactions among users would occur while using GRS, as well as how GRS could facilitate such interactions. To achieve this aim, we initially conducted a qualitative survey with 38 respondents, as well as 21-day diary studies and semi-structured interviews with 30 participants, throughout which we delved into user experiences of the Blend service in Spotify, a type of GRS that we considered to be a suitable context of our inquiry. All study protocols were approved by the Institutional Review Board (IRB No.KH2023-101). In the following, we first provide a brief description of the Blend service and the rationale behind our choosing it as the context of inquiry. We then turn to describe the study methods and participants.
3.1 Research Context: Spotify Blend
Blend (Figure 1) is a shared playlist service launched in 2021 by the well-known music streaming platform Spotify, designed to create a “personalized playlist that automatically combines two users’ audio tastes together” [30]. Blend creates a shared music playlist with song recommendations for multiple users in the same Blend group, by learning about all the group members’ tastes inferred from logs such as recent tracks played and liked songs. Users who belong to the same group get to see the same playlist all the time, i.e. the same items are recommended in the same order for everyone. Considering these, we thought that Blend aligned well with the core characteristics of GRS, choosing it as an appropriate context for exploring our research question.
Also, in a Blend playlist, users’ Spotify account profile pictures are shown alongside each recommended item. For instance, if two users use Blend together, a playlist can be created as in Figure 1: recommendation for User A (Figure 1a) and for User B (Figure 1b) based on his or her individual activity, as well as recommendation that caters to both users (Figure 1c). This feature serves to explicitly indicate which user’s tastes each recommended item is based on, thereby expected to help participants have a clearer mental model of Blend as GRS and accordingly report their experiences with that perspective in mind.
As for additional description on the service, a Blend playlist is updated on a daily basis to incorporate the drifting and evolving tastes of users so that recommendations are kept up to date. Spotify’s official engineering article [30] explains that Blend’s recommendation algorithm is designed to ensure that recommendations are relevant (i.e., how well the recommended items represent each user’s preferences), equal (i.e., whether the recommended items for users are evenly distributed), and democratic (i.e., whether items that users like prominently appear at the top). At the time our study was conducted (i.e. from June to August 2023), a single Blend playlist could consist of a minimum of two users and a maximum of 10 users.
Figure 1: A sample screenshot of Spotify Blend (as of 2023)
3.2 Preliminary Study: Qualitative Survey
Considering the nascent nature of our research agenda, which seeks to understand user experiences of GRS from a social technology perspective, we recognized the need to further sharpen the scope of our investigation as an initial step. This required us to first collect people’s accounts on various social experiences in Blend so that we could identify the key issues pinpointing what we should further explore through diary studies. In doing so, we conducted a qualitative survey with current and previous users of Blend, as described below.
3.2.1 Survey Design and Process.
The survey was designed using Google Forms and conducted in South Korea. It started with questions about basic demographics and overall Blend usage (e.g., with whom they used Blend, why they used it, and how often they used it). We then inquired about whether they were satisfied with using the service, whether they could get to know Blend member(s) better, whether they felt closer to other member(s) while using Blend, and whether they felt conscious of the other member(s) while using Blend. Lastly, we asked open-ended questions about any memorable moments while using Blend, and opinions on Blend’s role in the relationship with other member(s).
The survey link was spread via adverts on the university campus, online communities of South Korean Blend users, social media, and word of mouth. All respondents each received 10,000 KRW (approx. 8 USD) for completing our survey. A total of 38 respondents (17 females and 21 males) took part in our survey, with 12 falling within the age range of 18-22, 17 in the 23-29, 7 in the 30-34, and 2 in the 35 and older. In terms of usage duration, 11 respondents had used it for less than one month, 16 for 1-3 months, 4 for 3-6 months, and 7 for more than six months. Also, 17 respondents were currently using Blend, while the remaining 21 had used Blend in the past but were not currently using it anymore.
The former-user respondents stopped using Blend because Spotify’s plan was expensive (n=6), they wanted to listen to only their own favorite music (n=3), other member(s) stopped using Spotify (n=3), and they were more comfortable sharing music in different ways (n=3). Also, some cited reasons such as finding Blend not fun (n=2), breakup with partners (n=2), knowing enough about others’ musical tastes (n=1), and low music consumption (n=1). As all of these reasons were primarily related to either the respondents’ practical issues or the perceived usefulness of the Blend, we concluded that the data from these former-user respondents were still valuable if we aimed to explore their social experiences while using Blend.
Also, while the survey did not collect information on how long ago the former-user respondents stopped using Blend in the survey, we deemed their responses to be vivid, authentic experiences based on the detailed descriptions. (On average, former-user respondents wrote 21 words and current-user respondents wrote 19 words for each open-ended question.)
For data analysis, we compiled the open-ended survey responses by question into a spreadsheet and subsequently performed affinity clustering. The first author developed the initial insights based on the clustering, and the other two researchers discussed the intermediate deliverables together to build consensus. In the analysis process, we focused on what and how social interactions are mediated by GRS.
3.2.2 Results: The Potential of Investigating Social Interactions in Blend between Two Real-World Relationships — Acquaintances or Close Connections.
First, we discovered that most respondents used Blend with someone whom they have known in the real world, such as a friend or acquaintance (n=30), a significant other (n=9), or a family member (n=5). For example, a respondent said: “[Blend] is like the exchange diaries that I used to keep with close friends or people I dated when I was younger. Sharing the songs that we’re currently into in our own space creates a sense of connection, making us feel like a special relationship like we’re meant to be together.” Also, more than half of the respondents said that they would like to use Blend with someone other than the current member(s) as well, especially someone they want to become closer to (n=28) and get to know more (n=26). Further, the occurrence of social interaction between them was glimpsed from their responses mentioning the role of Blend, such as helping them understand other member’s tastes or triggering conversation. These insights led us to narrow down our focus to an exploration of social interactions mediated by GRS, particularly within the relationships between people who are already acquainted with or close to each other in the real world.
Also, we noticed that most respondents (n=35 out of 38 respondents) reported that they had been using or had used Blend with someone in pairs. While Spotify’s update that expanded the maximum number of Blend members from 2 to 10 was in March 2022, we did not collect information about precisely when respondents started using Blend in our survey. The absence of data regarding precise periods for respondents’ Blend usage in our survey limits our ability to ascertain how many users consciously opted for the pair-sharing experience, irrespective of the enhanced capacity for additional members in a Blend group. Despite this limitation, our deliberate choice to focus on pair users is meaningful. Given the scarcity of existing research on investigating GRS experiences from a social lens, we aimed to delve more deeply into the experiences of a clearly targeted user group first, expecting to uncover specific and tangible findings that mark a starting point in this research domain. Upon our choice to focus on pair users, in the rest of this paper, we will indicate each person in a pair as a participant and companion respectively.
3.3 Main Study: Diary Studies and Interview
Based on the insights from the preliminary survey, we established the specific focus for the main study: exploring how two real-world acquainted or close relationships socially interact through Blend. We conducted 21-day (3-week) diary studies and interviews, investigating how two users who are acquainted or close socially interact to manage their relationships while using GRS — i.e., Blend in our case. We used diary studies to capture in rich, nuanced detail the narratives related to people’s social interactions and relationships in their everyday lives.
3.3.1 Study Participants.
We determined the participation requirements of this study considering the specified scope of the investigation. In the recruitment post, we announced that pairs of two individuals interested in getting to know each other’s musical tastes were eligible for participation. Also, in order to observe natural experiences over time about how participants perceive recommendations and interact with their companions while using GRS, we recruited all participants with no previous experiences of using the Blend feature. However, it was important to make sure that we recruited people who enjoyed listening to music often so that they would actively engage in using the ‘music’ recommendation system during our study period. Therefore, we included a question asking how often they listen to music so that we could recruit people who self-reported that they frequently do so. We disseminated a recruitment post via the university’s online communities, social media, and word of mouth.
Taking these into consideration, we recruited 30 participants (17 females), i.e. 15 pairs, whose detailed demographics are referred to in Table 1. The average age of the participants was 28.9 (SD=8.2, MIN=22, and MAX=56), and they encompassed five relationship types: parent-child (2 pairs), siblings (4 pairs), spouses or romantic couples (4 pairs), friends (3 pairs), and co-workers (2 pairs). Our effort to ensure the variations in age groups and relationship types allowed us to enrich the insights into various types of social interactions, yielding a more comprehensive understanding of social interactions and behaviors that GRS can facilitate.
Each participant took part in a series of orientations for the study, a 21-day diary writing activity, and a post-hoc individual interview session. All participants received compensation of 150,000 KRW (approx. 115 USD) in recognition of their time and efforts upon completion of their participation.
| ID | Gender | Age | Relationship type | Participants’ self-reported details about their relationship (before beginning the study) |
| P1 | Female | 53 | Parent-child | Have been living separately for 8 years |
| P2 | Male | 26 | ||
| P3 | Female | 56 | Parent-child | Have been living together all the time |
| P4 | Female | 25 | ||
| P5 | Male | 26 | Siblings | Have been living separately for 8 years |
| P6 | Male | 23 | ||
| P7 | Female | 34 | Siblings | Have been living separately for 4 years |
| P8 | Female | 22 | ||
| P9 | Male | 27 | Siblings | Have been living separately for 10 years |
| P10 | Female | 24 | ||
| P11 | Female | 28 | Siblings | Have been living separately for 8 years (P11 residing overseas) |
| P12 | Female | 26 | ||
| P13 | Female | 40 | Spouses | Have been living together for about 8 years |
| P14 | Male | 37 | ||
| P15 | Male | 32 | Romantic couple | Meeting each other in person almost every day |
| P16 | Female | 27 | ||
| P17 | Male | 26 | Romantic couple | Meeting each other in person once every 1 to 2 weeks |
| P18 | Female | 25 | ||
| P19 | Male | 24 | Romantic couple | Meeting each other in person almost every day |
| P20 | Female | 22 | ||
| P21 | Male | 24 | Friends | Meeting each other in person once every 2 to 3 weeks |
| P22 | Female | 23 | ||
| P23 | Male | 32 | Friends | Meeting each other in person once every month or two |
| P24 | Male | 30 | ||
| P25 | Female | 23 | Friends | Meeting each other in person once every 2 to 3 months |
| P26 | Female | 22 | ||
| P27 | Male | 29 | Co-workers | Meeting each other in person almost every day |
| P28 | Male | 26 | ||
| P29 | Female | 29 | Co-workers | Meeting each other in person almost every day |
| P30 | Female | 27 |
Table 1: Participants’ demographics and description of their relationships
3.3.2 Diary Study.
The diary study was conducted for 21 days. We chose Notion, a popular note-taking application, as a digital diary format considering its user-friendly interface and cross-device compatibility that were expected to support comfortable diary writing. Each participant was asked to write their own diary, and no participants were granted access to their companion’s diary entries so that we could make it more comfortable for each participant to record candid personal accounts about their Blend experiences and relationships.
A week before starting the diary writing period, participants were encouraged to incorporate Spotify into their daily routines for music listening. This served two purposes. First, because most participants except two (P24, P27) had no previous experience using it, we wanted to familiarize our participants with the application. Second, we wanted to provide Spotify’s recommendation algorithms with some time to learn about participants’ preferences and tastes.
Throughout the diary writing period, participants were guided to: 1) use the Spotify application by priority for listening to music during the study period, even if they used to use other music streaming services, 2) check the Blend playlist at least once a day, 3) listen to songs in the Blend playlist, and 4) record their thoughts on their Blend experience in their diaries. Regarding the diary entries, we instructed participants to complete a daily diary template consisting of six questions and an optional form for submitting screenshots. All diary entries were designed to delve deeper into the quality of users’ social interactions mediated by Blend in detail, on the basis of the insights identified in our preliminary survey. The detail entries are the following:
(1) | In what situation did you use Blend today? | ||||
(2) | What was the most memorable song recommendation today? Whose profile picture(s) accompanied it? Why do you find it the most memorable? | ||||
(3) | Did you receive any song recommendations accompanied by both profile pictures? If so, how did you feel when you saw and/or listened to the song(s)? | ||||
(4) | In Blend, you get music recommendations based on each person’s preferences. Was there any moment today when this fact influenced your experience of using Blend? | ||||
(5) | Did you have any conversation with your companion about the music recommended by Blend today? If so, what was the conversation about? If not, what kind of conversation would you like to have with your companion in the future? | ||||
(6) | Do you think that using Blend is affecting your relationship with your companion? Why do you think so? [This question was asked on DAY 1, 7, 14, and 21 only, considering noteworthy changes in a relationship may not occur on a daily basis.] | ||||
(7) | (optional) Please upload a screen capture of Blend that is related to today’s diary if you want to share it with the researchers. | ||||
The researchers sent daily reminders at 8 p.m. using a mobile instant messaging application to encourage writing a diary reflecting each day’s experience of Blend. The application was also used as a channel for participants to freely ask the researchers any questions that arose during the study period.
A total of 571 diary pages were collected from the diary study, excluding cases where participants occasionally skipped writing diaries on certain days, such as when they did not use Blend during the study period. We present an example diary page in Appendix A. The diary consisted of 69 Korean words per day on average, with a minimum of 10 words and a maximum of 667 words.
3.3.3 Post-hoc Interview.
Upon finishing the diary study, we conducted in-depth one-on-one interviews with each individual participant. The interview questions were designed to understand the detailed contexts and meanings of the diary data. Just as diary entries, we did not conduct interviews in pairs so that each participant could discuss their personal accounts more freely. Interviews were conducted online using Zoom, lasting approximately 42 minutes on average, with the shortest and longest sessions taking 25 and 63 minutes respectively. All the interviews were audio-recorded with participants’ consent.
Beginning an interview session, we first asked participants 1) their relationship with the companion, as well as their prior experiences of sharing media (e.g., how often they typically communicate, how well they knew each other’s preferences, whether they usually engaged in sharing media like music or videos). Then, we inquired about 2) the detailed social contexts related to the Blend usage (e.g., why they wanted to use Blend with their companions, what role Blend played in their relationships, whether they felt that they were having an impact on each other while using Blend, and whether there were any perceptions or behaviors that arose due to being conscious of the companions while using Blend). Finally, we inquired about 3) the specific contexts and details regarding several noticeable diary entries.
3.3.4 Data Analysis.
We collected participants’ diary responses and transcribed audio recordings of interviews. All data from the diaries and interview transcripts were qualitatively analyzed through the thematic analysis method [10]. As interview transcripts were supportive data that provided an in-depth context of the diary data, we simultaneously reviewed both diary and interview data to gain a thorough understanding of user experience with GRS. Two researchers identified key quotes from the diary and interview data by discussion based on memoing, then engaged in an iterative process for open coding. The analysis focused on identifying patterns of users’ perceptions and behaviors regarding social experience mediated by GRS. The developed codes encompassed categories such as how participants noticed or inferred about their companion through recommendations, how they were conscious about their companion, how they utilized recommendations as a means of communication, and how they perceived recommendations for both. All researchers regularly reviewed the analysis progress and refined themes until we achieved a consensus. As a result, we organized two primary themes and six sub-themes about social behaviors that participants engaged in to lubricate their social relationships with GRS.
4 FINDINGS
Our study uncovered various social behaviors that participants exhibited to lubricate their relationships through using Blend. Our key finding was that participants engaged in implicit social interaction, including making tacit understanding (§4.1) of each other, as well as engaging in indirect communication (§4.2) with their companions by offering non-coercive suggestions, confirming an expansion of common interests, and demonstrating prosocial behaviors for their companions’ good. We report the details of our findings below.
4.1 Making Tacit Understanding of a Companion
First, we found participants’ attempts for implicit inference of their companions through the medium of GRS, which served to nurture their relationships. In particular, participants inferred unknown and unexpected tastes as well as recent emotions and circumstances of their companions through shared recommendations. They also simultaneously became conscious of how their companions would understand them and accordingly managed their ‘impression’ to control their self-image.
4.1.1 Noticing Unknown and Unexpected Tastes of a Companion.
We observed that participants (n=23) discovered clues about their companions’ musical tastes from the songs that accompanied their profile pictures. Blend particularly worked as an opportunity for preference understanding for participants with limited information about their companions (n=10; parent-child, siblings, co-workers). For instance, P2 reported that using Blend was a “nice experience that helped him discover his mother’s music tastes for the first time in their life,” which made him communicate with his mother and feel like they became closer to her.
This kind of implicit understanding was especially perceived as helpful by participants who could hardly engage in direct and explicit conversations with their companions. As an example, P3 (mother) and P4 (daughter) were living together, and P4 was too busy to have time for casual chit-chat at home. They both reported that they could discover what songs the other person was currently enjoying by regularly checking Blend. Also, participants mentioned that they could better understand their companions’ tastes that were difficult to be precisely articulated. For example, P2 and P9 often directly asked their companions what kinds of songs they liked, but they felt that it was always difficult for their companions to clarify their tastes. Considering this, they perceived Blend as a chance to grasp such tastes “in a roundabout manner” (P2) yet “for sure” (P9) than before:
“Compared to my past experiences of trying to get to know my sister’s preferences, using Blend was much more useful than, for example, asking ‘What do you like?’ when you’re buying a birthday gift. Because [Blend] definitely shows what kind of songs she usually listens to.” (P9, Interview)
In particular, participants (n=13; siblings, couples, friends) reported that they noticed their companions’ new, unexpected tastes. This was a surprising experience for our participants, as they thought that they had already been quite close to their companions and knew almost everything about them. For instance, P6 said that he was surprised to learn that his older brother, like himself, listened to hip-hop music. Also, P16 was pleased to find out for the first time that her boyfriend also enjoys New Age music. We found through our diary entries that the couple came to enjoy New Age music together thereafter:
“I thought, like ‘Did my boyfriend really like this kind of music?’ when I saw that a New Age song was recommended.” (P16, Diary DAY 2)
“We’ve been dating each other for 2 years, and to be honest I thought I knew almost everything about each other, but using Blend and sharing our preferences made me realize that there were still some aspects of each other that I wasn’t aware of.” (P15, Interview)
In particular, participants described that they felt closer to their companions when newly discovered tastes of their companions were actually shared interests. As an example, P13 had lived with her husband for 8 years without ever considering that they might have similar musical tastes. However, while using Blend, she discovered that they had so much in common that she even “mistakenly thought of Blend as a playlist just for herself,” which consequently made her pleasant.
Another noteworthy finding is that participants (n=8) noticed the differences in their own and their companion’s tastes. For instance, P1 and P3, who shared Blend with their son (P2) and daughter (P4) respectively, mentioned that they had different tastes due to a generation gap. Participants who shared Blend with friends (P21, P22, P23, P26) or co-workers (P29, P30) also reported their experiences of noticing taste differences. However, we note that participants did not consider such experiences unpleasant. For example, the four participants in parent-child relationships were satisfied that their differences in musical tastes sparked frequent conversations between them. Consequently, P1 mentioned that she “had become interested in Gen Z’s music tastes” using Blend with her son. Also, although not as affectionate as the family, the friends and co-workers participants also described that these experiences eventually helped them become more intimate with each other, as illustrated by P26:
“I used to think that our music tastes were quite similar, but when I saw the songs recommended by Blend, I actually noticed clear differences from my own preferences. It made me feel like I got to know the person [P25] even better. As a result, I feel like we’ve become closer.” (P26, Diary DAY 14)
Collectively, participants could indirectly discover their companions’ music tastes, especially unknown and/or unexpected ones, on the basis of the songs that appeared as shared recommendations. Regardless of whether their companions’ tastes were in common or different from theirs, participants appreciated that using Blend became an opportunity to understand their companions on a deeper level.
4.1.2 Inferring Recent Emotions and Circumstances of a Companion.
While music tastes are the more intuitively inferable information from Blend, we observed that participants (n=11) inferred other types of information from shared recommendations as well. First, participants inferred their companions’ recent emotional statuses from Blend songs recommended along with their companions’ profile pictures. For instance, P6, upon noticing an unusual recommendation of love songs in his Blend, thought that maybe his brother had recently started dating someone. Also, P13 reflected on her experience of being concerned whether her husband might be feeling weary, as she noticed many sentimental songs recommended. When we asked the details of this diary entry, she described:
“You can just tell in everyday life if someone has worries just by watching them eat or observing his actions while walking. And when you look at the titles or lyrics of the songs that you get from Blend, they really match those behaviors. So for example, if he seemed a bit down and looked a bit sad yesterday, then the next day, sad songs tend to be there [in Blend].” (P13, Interview)
Also, participants tended to infer their companions’ current circumstances while experiencing Blend. For example, P1 and P17 inquired about whether their companions were busy based on the number of profile pictures that appeared alongside recommendation items; they thought their companions were too busy to listen to music, which resulted in fewer songs recommended for the companions. Also, P7 and P8, who were siblings, worried about each other when they began to get more calm songs recommended; each wondered whether it was because her sister had trouble sleeping and listened to sleep-inducing songs. P12 also reported how she would guess what her sister had been up to lately through Blend:
“Using Blend is like a more indirect yet deeper form of communication, if I could put it that way. It’s a bit paradoxical and hard to explain... but I mean, other people may not directly say things like ‘I’ve been thinking about this lately’ or ‘I believe this kind of news is important these days,’ but you start to think things like ‘They are just hanging in there’ or ‘They must be listening to these kinds of songs these days.’ It feels like a different kind of communication compared to directly engaging in everyday conversations. In fact, you might not be able to find out those aspects when you were to ask so directly.” (P12, Interview)
Collectively, participants inferred information other than musical tastes from shared recommendations, including the emotional statuses and current circumstances of their companions. This type of inference occurred in an even more indirect manner than an inference about music tastes. Also, it is interesting that this kind of inference was mostly observed among family members and romantic couples, that is, those who were particularly intimate and eager to know more about each other.
4.1.3 Engaging in Impression Management.
Because participants made inferences about their companions, they also thought that their companions would infer information about themselves based on the same reasoning process. Accordingly, several participants (n=9) consciously managed their ‘impression’ [21] so that they could control how their self-image would be perceived by their companions.
In particular, participants attempted to conceal certain tastes of themselves. They did not want particular songs to appear in Blend even if they liked those, because they did not want their companions to know that they enjoyed listening to those kinds of music. For example, P30 felt embarrassed when songs by a ‘controversial’ singer were recommended on Blend along with her profile picture, saying that the song was her “guilty pleasure” and she did not want her co-worker to know that. Also, P21 mentioned that he personally enjoyed old anime songs, but he was concerned that if those songs appeared in Blend, his friend might think he was stuck in the past with music. P17 and P19 refrained from listening to songs by rappers who write somewhat ‘improper’ lyrics during our study period, even though they used to enjoy those songs, because they were concerned that their girlfriends might perceive these songs as problematic.
Also, participants tried to avoid their taste being misunderstood. Accordingly, they were dissatisfied when songs that they did not like were recommended in Blend alongside their profile pictures, as P26 reported:
“I don’t like ‘DADDY’ by Psy, and so I didn’t want this song to be shown to [P25]… I didn’t want her to think that I liked that song. So I immediately removed it from the playlist when I found a button for doing so.” (P26, Diary DAY 5)
Furthermore, some participants utilized Blend as a means of more active self-expression. For example, P19 believed that his girlfriend would gauge his current emotional status from their Blend, and therefore he attempted to make emotionally-driven songs to appear as shared recommendations, expecting that his current mood would be indirectly expressed to her.
“I was having a kind of emotional suffering because some things were happening at that time, and I kind of wanted my girlfriend to become aware of it from those songs.” (P19, Interview)
Of note is the fact that this behavior of impression management may change over time. For example, P29 was initially hesitant to use Blend in a ‘natural’ way because she thought she was not so intimate with her co-worker enough to share her personal tastes. This made her worry about her tastes being exposed through Blend, describing this concern as “revealing a diary.” However, over time she interacted with her companion, especially through the use of Blend, and they became closer than they initially were. Accordingly, we observed from her diary that she no longer had concerns about her tastes revealed by recommendations from the third week onwards:
“We talked a lot and got to know each other better, and I didn’t feel too ashamed anymore. After talking with her a lot, I realized that she wasn’t the kind of person who would judge me or think poorly of me just because I liked a particular kind of music. I started to listen [to the songs I like] more naturally afterward.” (P29, Interview)
The above accounts collectively illustrate participants’ attempts to manage their self-image that might be inferred by their companions from shared recommendations. Whether participants would engage in these practices seemed to depend on whether they perceived the need for impression management in relationships with their companions. For example, the above practices were often observed in relationships such as romantic couples, friends, and co-workers, but not in more ‘candid’ relationships like family.
4.2 Engaging in Indirect Communication with a Companion
We also observed the participants’ attempts to indirectly communicate with their companions. Especially, participants appreciated Blend as a medium of offering non-coercive suggestions, confirming an expansion of common interests, and demonstrating prosocial behaviors.
4.2.1 Offering Non-coercive Suggestions.
We found that participants (n=14) intentionally attempted to make Blend include the songs that they would like to suggest to their companion. These participants believed that Blend would definitely incorporate their activities, such as play histories, into its recommendations. Consequently, they deliberately searched for the songs that they wanted to introduce to their companions and listened to them by themselves:
“I searched for the song that I like but was not in Blend, and I played it, hoping that it might pop up in Blend so that the other person might listen to that. I was trying a kind of recommendation in an indirect way...” (P26, Diary DAY 2)
“I wanted Blend to include more songs that my older sister doesn’t usually enjoy but fit my tastes. I came to first listen to all the songs in Blend, and deliberately play the songs that only I like, and then check how Blend incorporated that [behavior].” (P12, Diary DAY 7)
Drawing a comparison between this kind of strategy and their previous experiences of directly sharing media content, participants reported how non-coercive suggestions were perceived to be a much more effective approach. For instance, participants reflected on the experiences of feeling somewhat obliged to give feedback on a suggestion when someone offered one in a direct manner (e.g., someone sending a URL of a video clip via online messengers). This was especially the case when suggested content was out of line with participants’ interests, where they felt like they were “forced” (P13) to accept the suggestion. While participants had made a few attempts to directly exchange recommendations with someone else, due to the lukewarm response from each other, they naturally stopped engaging in such behavior. In contrast, participants believed that by using Blend, they could offer suggestions in a more indirect manner, and at the same time, there would be a much greater likelihood that their companions would naturally listen to the songs that they suggested with less pressure regarding feedback. Many used the term “nudging” to describe this practice of offering non-coercive suggestions, highlighting how it could positively impact their relationships. For instance, P12 described her experience of naturally leading her sister to listen to music aligned with her tastes, which, in her perception, contributed to fostering a more balanced relationship between them:
“My sister is older than me, so it was almost always her to discover new songs, and I grew up feeling like I was passively accepting my sister’s tastes. (...) In the past, whenever I sent links to songs that I would like to recommend to her, she often seemed uninterested, so I thought I wouldn’t try to offer suggestions so directly. But when she listened to the songs of my tastes that I put in Blend and she praised me saying ‘I liked these, so I searched for more of these,’ I felt a bit proud, and how should I put it, felt independent.” (P12, Interview)
Similarly, P19 reflected on how his implicit attempt to include one of his favorite songs into the Blend playlist led his reluctant girlfriend to voluntarily listen to the song, subsequently becoming more open to exploring and embracing his musical tastes. Comparing this experience with previous unsuccessful attempts to persuade her to listen to the song directly, P19 reported that such non-coercive suggestion played a role in fostering a more receptive and open-minded attitude toward the preferences of the significant other:
“For example, one of the songs that I put in Blend was a soundtrack from ‘Attack on Titan.’ Before, no matter how much I tried to persuade [her] to listen to that song, [she] never did. But the song sneakily came up in Blend, and I saw that she [voluntarily] listened to it and found it enjoyable. She used to say she would never watch ‘Attack on Titan’ even if I asked her to. But after listening to the song, she proactively suggested we watch it. It was quite surprising. It appeared as though she had become more open to embracing my tastes.” (P19, Interview)
The above accounts collectively illustrate that participants, beyond passively receiving shared recommendations from a system, attempted deliberately to expose desired songs to their companions by themselves in the form of non-coercive suggestions. This significantly reduced the perceived burden on the recipient compared to offering suggestions in a more direct manner, ultimately fostering more pleasant communication experiences between the two parties.
4.2.2 Confirming an Expansion of Common Interests.
We also found the cases of participants (n=22) who appreciated an opportunity to expand their common interests while using Blend. They considered such areas of common interest, which were confirmed through ‘recommendations for both,’ to be particularly important and meaningful, as P26 and P9 describe:
“Those [songs] weren’t for both of us before, but they changed [to be recommendations for both], and I felt good. I feel like our tastes are becoming more alike!” (P26, Diary DAY 2)
“6-70% of the entire 50 songs in the playlist are recommendations for both. I felt a kind of relational cohesion between us.” (P9, Diary DAY 17)
P5 also got to like an artist that his younger brother favored during the study period, which made Blend offer that artist’s song as a recommendation item for both. Accordingly, he reported in his diary that he “felt good because it was like an AI recognizing that my brother and I were developing similar musical tastes.”
Especially, when both of their profile pictures started appearing alongside several recommendations, participants became surprised, thankful, and pleased to start confirming that they were gradually reaching an intersection of tastes, because they believed their companions were also listening to the songs of participants’ tastes. This kind of experience made our participants realize that they were making an actual impact on their companions in expanding their common interests, leading to positive user experiences of Blend, as P20 illustrates:
“What is most memorable is that there’s a song called ‘2002’ by Anne-Marie, and I listened to it, and as far as I know, it’s not my boyfriend’s taste. But I saw that the song appeared in Blend as a recommendation for both of us, and I thought like ‘He tried some songs that I like,’ ‘It’s so kind of him that he listened to those when those were recommended, even if he doesn’t like them,’ and I felt like thinking about my boyfriend quite often.” (P20, Interview)
Collectively, these above accounts highlight that participants appreciated Blend for enabling them to indirectly confirm an expansion of their common interests. They found it meaningful when their musical tastes converged, and they felt a sense of connection and influence on each other’s preferences, which implies the significance of communication for such confirming experiences.
4.2.3 Demonstrating Prosocial Behaviors.
Finally, we also observed the cases of participants who engaged in a variety of prosocial behaviors. For example, several participants (n=7) acted considerately of their companions, caring for the companions’ musical tastes. When they often found songs they enjoyed, they also contemplated whether their companions would like the same songs. Some even asked their companions directly whether the songs appearing in their Blend lately were worth listening to. For example, P5, P19, and P20 intentionally refrained from listening to songs they liked but knew their companion disliked, in order to prevent those songs from appearing in Blend. P25 even went as far as removing songs from their playlist on his own when he found a song that might not align with his companion’s taste. P5’s interview quote well exemplifies this kind of thoughtful attitude:
“I wasn’t really into that kind of music before... but after being discharged from military service, I started occasionally listening to lively songs by girl groups [laughs]. But my younger brother didn’t seem to like those a lot, so I ended up listening [to that kind of music] less. I can always find such songs in another app if I want to listen to them... I didn’t want to make my brother uncomfortable.” (P5, Interview)
Also, some participants (n=5) displayed a willingness to compromise in terms of their tastes for the sake of their companions. They mindfully attempt to listen to recommended songs accompanied by their companions’ profile pictures even if those were slightly outside their own preferences, because they wanted Blend to equally cater to not only their own preferences but also those of their companions. Describing this kind of process as “time for coordination,” P30 said:
“I didn’t want to force each other to listen [to the song in Blend]. For example, I thought, like, I wanted more songs like rock music would be there so that P29 [her co-worker] could enjoy, although they are not my kind. So initially I tried to listen to a lot of songs that I expect to be recommendations for her.” (P30, Interview)
Further, some (n=2) attempted to exhibit a sign of emotional support to their companions. These participants attempted to include songs in Blend that could be supportive of their companions, considering their companions’ current emotions or states. Such behavior itself is seemingly a type of offering non-coercive suggestion described in §4.2.1, but we emphasize the clear difference in participants’ motivation behind — to provide emotional support, rather than to share a musical taste. P6, for example, reported how he deliberately played relaxing, happy songs in an attempt to cheer his brother up:
“He [P6’s brother] recently started living on his own and being financially independent. So I tried to recommend positive and relaxing music through Blend as a way to support him. (...) At some point, I realized that I was putting effort into intentionally including some songs on Blend, hoping that he would listen to those and have positive emotions.” (P6, Diary DAY 14)
P17 also deliberately played uplifting songs, expecting those to be included in the Blend he shared with his busy and tired girlfriend so that he could brighten her spirits:
“Today it seemed like she [P17’s girlfriend] was really busy and tired due to work. So, instead of loud or sad songs, I thought about playing some uplifting ones. I feel like I’m treating Blend recommendations not as an end result, but as some kind of means to do something.” (P17, Diary DAY 2)
These examples as a whole help illustrate the depth of connection and consideration that participants infused into their use of Blend. They reveal its capacity to foster meaningful interactions beyond mere music sharing, emphasizing its role as a conduit for empathy and support among users. These prosocial behaviors demonstrate how Blend transcends its utility as a music recommendation tool, becoming a medium for fostering bonds and offering genuine care among its users.
5 DISCUSSION
Premised on the potential of GRS as a medium of social interactions among users, we have explored the dynamics of social interactions unfolded and facilitated by GRS. Throughout the above findings, we have illustrated that GRS could function as a medium of reciprocal disclosures and communication in a two-person situation. Now, we highlight such power of GRS as a social catalyst and suggest two key attributes for designing GRS from that lens. Also, we discuss the importance of considering the intricacies of individual and collective recommendation experiences, a unique challenge posed by the design of GRS as a social technology.
5.1 GRS as a Social Catalyst
Throughout our study, we observed that GRS recommendations mediated a variety of social interactions and fostered a sense of connection between two users. To date, most research regarding GRS has focused on how to achieve recommendations that can satisfy all users or how to support the decision-making process among users [24]. However, our findings reveal a new viewpoint on GRS: this system can serve as a mediator for users’ social interactions, functioning as shared technology. Upon reflection, we refer to GRS as a ‘social catalyst,’ as these mediators supported mutual understanding, facilitated conversations, and triggered creative communication.
Identifying a new role of GRS as a social catalyst, our research adds to the knowledge of the traditional research in HCI and design fields that have attempted to facilitate social interactions and relationship development between people. In terms of interaction quality, GRS mediates social interactions that are implicit [4, 25, 38, 55] rather than explicit and instrumental [19, 32, 33]. Indeed, participants perceived the interactions mediated by Blend recommendations as a complementary means that could ease the psychological burden of direct communication [55]. Also, they did not use Blend with the explicit purpose of exchanging information, but they could naturally build a sense of intimacy and connection with their companions through the everyday activity of listening to music [55]. Further, our findings align with previous works that have attempted to support emotional communication [20, 26] or everyday social activities [11, 12, 34, 58].
Given the above implications, we note that many prior studies on social technology, including the ones discussed above, have been conducted by focusing on specific user groups (e.g., families, friends, or romantic partners) or distinct social situations (e.g., long-distance or collocated), revealing that different design strategies are needed to accommodate different dynamics in various social contexts. In light of this, we suggest that future research on GRS also employ more targeted yet diversified contexts of inquiry, shedding additional light on more nuanced characteristics of social interactions that it mediates.
5.2 Key Design Attributes for GRS as a Social Catalyst: Algorithmic Ambiguity and Behavior Visibility
Throughout the study, we noted that the social interactions occurring through Spotify Blend were implicit and indirect. Upon reflection on our study findings, we propose that this distinctive interaction quality is the result of handling two design attributes of GRS, and relatedly, we posit that GRS can also be designed to enable more explicit and direct forms of social interaction by controlling those attributes in different ways. The two design attributes are 1) the level of ambiguity in recommendation algorithms and 2) the visibility of user behaviors.
First, due to the distinct level of ambiguity in recommendation algorithms of Blend, participants creatively inferred a range of information from the recommended songs, including their companions’ tastes, emotions, circumstances, and whether they had listened to a particular song. While most participants were aware that their recommendation algorithm was associated with their music listening history, they did not know precisely how the system learned and made recommendations behind the scenes. This inherent opacity of the recommender engine led participants to interpret the reasons behind the recommendations on their own and develop folk theories [17] for their GRS, thereby creating and enriching personal meanings in them.
Accordingly, we suggest that achieving a high level of transparency in GRS recommendation algorithms may not always be the most desired solution, although prior works have often strived to do so to explain and justify why recommendations are generated [3]. (This argument is grounded in our findings regarding participants’ awareness of their companions, and an example design direction is introduced in the last paragraph of this section.) Rather, we suggest that the level of algorithmic transparency can be carefully managed if one intends to design a particular quality of social interactions between users in GRS. For example, in the case of Blend, the system can reveal more explicitly the details of why each song was recommended (e.g., telling users what percentage a particular song reflects which user’s taste profile), or reversely, it can hide even more information than it does now (e.g., removing users’ profile pictures associated with a particular recommendation). Further research may explore how the degree of algorithmic transparency can be tuned and how the quality of social interaction changes alongside.
Second, it was also the visibility of user behaviors within GRS that influenced the quality of social interactions. For example, in the case of Blend, it was the profile pictures — representing whose preferences each recommendation reflected — that made one’s activities visible to each other and therefore created certain social interactions. Because of the presence of those pictures, people engaged in social experiences such as inferring their companions’ in-app behavioral patterns (e.g., the song with one’s profile picture would mean that s/he had recently played it), making a conjecture about the companions’ music tastes, and concerning how their own tastes would be judged.
We propose that, as with the level of ambiguity in recommendation algorithms, designers can control the way and extent to which user behaviors are disclosed to achieve an intended quality of social interaction in GRS. For instance, allowing users to see how many times the other group member has played a specific song may strengthen a sense of shared musical experiences and promote a deeper understanding of each other’s preferences within the group. We expect further research on how different forms of social interaction in GRS would occur based on different ways and levels of behavior disclosure. Also, while the profile pictures associated with recommended items served as an indicator of user behaviors in the case of Blend, further investigation is needed on what other design elements of GRS will be used to control this attribute as well.
In addition, we note that the quality of social interactions in GRS should be designed by considering the types or status of relationships between users [2, 28]. Although the detailed analysis of how social interactions vary across different types of relationships is out of the scope of this paper, we found some hints about this implication in our findings. For example, co-worker participants were initially more careful with revealing personal information compared to those in more intimate relationships such as family, but as the study progressed, their connections deepened and adopted a more relaxed, open attitude towards self-disclosure. Extending the above discussion, we suggest that the two aforementioned attributes — i.e. the ambiguity of recommendation algorithms and the visibility of user behaviors — can be tailored to align with distinct relationship types and developmental stages. For instance, for those who initially feel uncomfortable about too much mutual disclosure, designing GRS with a more ambiguous explanation of its ‘inner workings’ might help; as the recommendation algorithm is opaque, users might be less concerned with their in-app behavior being shared too directly. However, as time goes by, the system might initiate a gradual shift towards more transparency. We suggest a future research direction toward a more nuanced understanding in the impact of different relationship types on the design of GRS as a social catalyst.
5.3 Balancing between Individual and Collective Recommendation Experiences
While GRS may be a stand-alone service, it can be embedded in a platform whose main service is to provide individualized recommendations for users, just like Blend within Spotify. In that case, GRS would lead users to experience both individually and collectively personalized recommendations within a single platform. Accordingly, additional design challenges arise regarding such social contexts that might not be considered in a platform, which provides individualized personalization service only.
Reflecting on our findings, one such challenge is how to support users in preserving social faces [21] that might be shaped through recommendations. Our participants mindfully avoided listening to specific songs, even if they liked them, to prevent the songs from being shared as recommendations and consequently to safeguard the associated ‘musical persona’ being revealed to their companions. Prior works have identified that such impression management is a notable behavior observed in social interactions in systems like social media [16, 44, 57], and extending this line of work, our study revealed that such behavior can occur in GRS as well. Indeed, people perceive personalized recommendations in algorithmically-driven systems as proxies that reflect their tastes and identities [29, 31], and we further contend that in systems like GRS, it is crucial to consider the impact of social images accompanied by such recommendations.
In light of this, we suggest that platforms employing GRS should empower users with more agency over determining the qualities of individual and collective recommendations respectively. This may be achieved by implementing options such as customizable settings to empower users in nuanced controls over each recommendation algorithm, striking a balance between fostering social connections and preserving self-images that users are unwilling to disclose. For instance, platforms can be designed with features that let users control whether they want specific in-app activities (e.g., listening to a particular song or a playlist) should have an impact on their taste profiles used for GRS and the resulting recommendations for their group. Spotify as of December 2023 offers users an option to choose whether interactions with a Blend playlist should be either included or excluded from their general taste profile, but not vice versa; users cannot choose how much impact their interactions for personal use should have on Blend. In the future, designers might consider offering users an option like ‘Do not recommend this song for the shared playlists’ to better accommodate users’ intentional efforts to manage their impressions. Further, platforms should enhance the visibility of such user-control features. While Spotify has provided a feature called ‘Private Session’ that enables users to shield their in-app activities from influencing recommendation algorithms and therefore utilize it as a tactic for managing Blend recommendations, none of the participants discovered this option during our study. We suggest that platforms should make these controllability-related features more prominent and easily accessible.
6 LIMITATIONS AND FUTURE WORK
Our work presents an initial direction for rethinking GRS as a social medium, and we suggest several avenues for future work, acknowledging several limitations. First, we probed social dynamics between two people as we refined the focus of our study based on survey data. As interpersonal dynamics among three or more people can be more intricate, further studies on user groups of other numbers may unveil various aspects of social interactions through GRS. Additionally, the scope of inquiry in this study was limited to GRS for music recommendation. User experiences with GRS may vary depending on the type of recommended content. For instance, video content might encompass a broader range of personal preferences, such as values or political inclinations. In such cases, users’ tolerance for the influence of other users may differ compared to the GRS of the music domain. Therefore, investigating social interactions in GRS across broader domains can reveal more varied user perceptions and behaviors. Further, all our study participants were South Korean users. Given that social norms are culturally different and sensitive, it would be valuable to explore with users from diverse cultural backgrounds. Also, we note that while our 3-week study has revealed a range of social interactions that Blend mediated, more might unfold during a longer-term study. For instance, a relationship breakdown may be observed, allowing researchers to explore what interactions people attempt to mitigate. Finally, to achieve our research goal, we encouraged our survey respondents and diary study participants to consciously reflect on its role in their social relationships. We acknowledge that if it were not for the questions in our study, people might have been less engaged in such reflection in actual use.
7 CONCLUSION
Conceptualizing GRS as a medium of social interactions, our work investigated how GRS can enable and facilitate social interactions among users. In the context of Spotify Blend as a representative case of GRS, we first conducted a preliminary survey to understand the general potential of Blend as a social medium and narrow the focus of our study. We subsequently delved deeper into its role in fostering two-person relationships through a 21-day diary study and interviews with 30 participants (15 pairs). Within the context of Blend, our participants engaged in implicit reciprocal disclosures and indirect forms of communication, all aimed at managing and developing their relationships. Our findings unveil the value of GRS as a social catalyst, and based on these findings, we discussed two key attributes and challenges for the design of social experiences that GRS can mediate. Our work sheds initial light on a renewed framing of GRS as a social technology, and we hope that our work will inspire further research in this area.
ACKNOWLEDGMENTS
We extend our gratitude to all participants in our studies. Also, we sincerely thank the anonymous reviewers for their insightful comments and suggestions for this paper. This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. NRF-2021R1A2C2004263).
A EXAMPLE DIARY PAGE
Figure 2: Day 14 diary of P6. Diary responses were translated into English. (200 Korean words in the original version)
Supplemental Material
Available for Download
- Akshita Agarwal, Manajit Chakraborty, and C Ravindranath Chowdary. 2017. Does order matter? Effect of order in group recommendation. Expert Systems with Applications 82 (2017), 115–127.Google Scholar
Cross Ref
- Irwin Altman and Dalmas A Taylor. 1973. Social penetration: The development of interpersonal relationships.Holt, Rinehart & Winston.Google Scholar
- Oscar Alvarado, Nyi Nyi Htun, Yucheng Jin, and Katrien Verbert. 2022. A systematic review of interaction design strategies for group recommendation systems. Proceedings of the ACM on Human-Computer Interaction 6, CSCW2 (2022), 1–51.Google ScholarDigital Library
- Elizabeth Bales, Kevin A Li, and William Griwsold. 2011. CoupleVIBE: mobile implicit communication to improve awareness for (long-distance) couples. In Proceedings of the ACM 2011 conference on Computer supported cooperative work. 65–74.Google ScholarDigital Library
- Linas Baltrunas, Tadas Makcinskas, and Francesco Ricci. 2010. Group recommendations with rank aggregation and collaborative filtering. In Proceedings of the fourth ACM conference on Recommender systems. 119–126.Google ScholarDigital Library
- Shlomo Berkovsky and Jill Freyne. 2010. Group-based recipe recommendations: analysis of data aggregation strategies. In Proceedings of the fourth ACM conference on Recommender systems. 111–118.Google ScholarDigital Library
- Steve Berry, Steven Fazzio, Yongyi Zhou, Bethany Scott, and Luis Francisco-Revilla. 2010. Netflix recommendations for groups. Proceedings of the American Society for Information Science and Technology 47, 1 (2010), 1–3.Google ScholarCross Ref
- Md Momen Bhuiyan, Carlos Augusto Bautista Isaza, Tanushree Mitra, and Sang Won Lee. 2022. OtherTube: Facilitating Content Discovery and Reflection by Exchanging YouTube Recommendations with Strangers. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems. 1–17.Google ScholarDigital Library
- David Bonnefoy, Makram Bouzid, Nicolas Lhuillier, and Kevin Mercer. 2007. “More Like This” or “Not for Me”: Delivering Personalised Recommendations in Multi-user Environments. In User Modeling 2007: 11th International Conference, UM 2007, Corfu, Greece, July 25-29, 2007. Proceedings 11. Springer, 87–96.Google Scholar
- Richard E Boyatzis. 1998. Transforming qualitative information: Thematic analysis and code development. sage.Google Scholar
- AJ Bernheim Brush, Kori M Inkpen, and Kimberly Tee. 2008. SPARCS: exploring sharing suggestions to enhance family connectedness. In Proceedings of the 2008 ACM conference on Computer supported cooperative work. 629–638.Google ScholarDigital Library
- Min Zhen Chai, Alessandro Soro, Paul Roe, and Margot Brereton. 2017. Cooking together at a distance: Sustain connectedness for long distance families. In Proceedings of the 2017 CHI Conference Extended Abstracts on Human Factors in Computing Systems. 2437–2444.Google ScholarDigital Library
- Yu Chen and Pearl Pu. 2014. Designing emotion awareness interface for group recommender systems. In Proceedings of the 2014 International Working Conference on Advanced Visual Interfaces. 347–348.Google ScholarDigital Library
- Andrew Crossen, Jay Budzik, and Kristian J Hammond. 2002. Flytrap: intelligent group music recommendation. In Proceedings of the 7th international conference on Intelligent user interfaces. 184–185.Google ScholarDigital Library
- Toon De Pessemier, Simon Dooms, and Luc Martens. 2012. Design and evaluation of a group recommender system. In Proceedings of the sixth ACM conference on Recommender systems. 225–228.Google ScholarDigital Library
- Xianghua Ding, Thomas Erickson, Wendy A Kellogg, and Donald J Patterson. 2012. Informing and performing: investigating how mediated sociality becomes visible. Personal and Ubiquitous Computing 16 (2012), 1095–1117.Google ScholarDigital Library
- Motahhare Eslami, Kristen Vaccaro, Min Kyung Lee, Amit Elazari Bar On, Eric Gilbert, and Karrie Karahalios. 2019. User attitudes towards algorithmic opacity and transparency in online reviewing platforms. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems. 1–14.Google ScholarDigital Library
- Alexander Felfernig, Ludovico Boratto, Martin Stettinger, Marko Tkalčič, 2018. Group recommender systems: An introduction. Springer.Google ScholarDigital Library
- Azadeh Forghani, Carman Neustaedter, Manh C Vu, Tejinder K Judge, and Alissa N Antle. 2018. G2G: The design and evaluation of a shared calendar and messaging system for grandparents and grandchildren. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems. 1–12.Google ScholarDigital Library
- William Gaver and Frances Gaver. 2023. Living with Light Touch: An Autoethnography of a Simple Communication Device in Long-Term Use. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. 1–14.Google ScholarDigital Library
- Erving Goffman. 2005. Interaction ritual: Essays in face to face behavior. AldineTransaction.Google Scholar
- Jesse Haapoja, Airi Lampinen, and Kari Mikko Vesala. 2021. Personalised Services in Social Situations: Principal-Agent Relationships in Account Sharing. Proceedings of the ACM on human-computer interaction 4, CSCW3 (2021), 1–21.Google ScholarDigital Library
- Maia Jacobs, Henriette Cramer, and Louise Barkhuus. 2016. Caring about sharing: Couples’ practices in single user device access. In Proceedings of the 2016 ACM International Conference on Supporting Group Work. 235–243.Google ScholarDigital Library
- Anthony Jameson and Barry Smyth. 2007. Recommendation to groups. In The adaptive web: methods and strategies of web personalization. Springer, 596–627.Google Scholar
- Joseph’Jofish’ Kaye. 2006. I just clicked to say I love you: rich evaluations of minimal communication. In CHI’06 extended abstracts on human factors in computing systems. 363–368.Google Scholar
- Joseph’Jofish’ Kaye, Mariah K Levitt, Jeffrey Nevins, Jessica Golden, and Vanessa Schmidt. 2005. Communicating intimacy one bit at a time. In CHI’05 extended abstracts on Human factors in computing systems. 1529–1532.Google Scholar
- Heung-Nam Kim and Abdulmotaleb El Saddik. 2015. A stochastic approach to group recommendations in social media systems. Information Systems 50 (2015), 76–93.Google ScholarDigital Library
- Mark L Knapp, Anita L Vangelisti, and John P Caughlin. 2014. Interpersonal communication and human relationships. Pearson.Google Scholar
- Daehyun Kwak, Hankyung Kim, and Youn-kyung Lim. 2022. Exploring Users’ Experiences of “Suggested Posts” in Social Media Through the Lens of Social Networking and Interactions. In Companion Publication of the 2022 Conference on Computer Supported Cooperative Work and Social Computing. 37–40.Google Scholar
- Jen Lamere. 2021. A Look Behind Blend: The Personalized Playlist for You…and You. Retrieved September 14, 2023 from https://engineering.atspotify.com/2021/12/a-look-behind-blend-the-personalized-playlist-for-youand-you/Google Scholar
- Angela Y Lee, Hannah Mieczkowski, Nicole B Ellison, and Jeffrey T Hancock. 2022. The algorithmic crystal: Conceptualizing the self through algorithmic personalization on TikTok. Proceedings of the ACM on Human-computer Interaction 6, CSCW2 (2022), 1–22.Google ScholarDigital Library
- Fannie Liu, Mario Esparza, Maria Pavlovskaia, Geoff Kaufman, Laura Dabbish, and Andrés Monroy-Hernández. 2019. Animo: Sharing biosignals on a smartwatch for lightweight social connection. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 3, 1 (2019), 1–19.Google ScholarDigital Library
- Fannie Liu, Chunjong Park, Yu Jiang Tham, Tsung-Yu Tsai, Laura Dabbish, Geoff Kaufman, and Andrés Monroy-Hernández. 2021. Significant otter: Understanding the role of biosignals in communication. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. 1–15.Google ScholarDigital Library
- KuanTing Liu and Roger Andersson Reimer. 2008. Social playlist: enabling touch points and enriching ongoing relationships through collaborative mobile music listening. In Proceedings of the 10th international conference on Human computer interaction with mobile devices and services. 403–406.Google ScholarDigital Library
- Judith Masthoff. 2004. Group modeling: Selecting a sequence of television items to suit a group of viewers. Personalized Digital Television: Targeting Programs to individual Viewers (2004), 93–141.Google Scholar
- Judith Masthoff. 2010. Group recommender systems: Combining individual models. In Recommender systems handbook. Springer, 677–702.Google Scholar
- Joseph F McCarthy and Theodore D Anagnost. 1998. MusicFX: an arbiter of group preferences for computer supported collaborative workouts. In Proceedings of the 1998 ACM conference on Computer supported cooperative work. 363–372.Google ScholarDigital Library
- Florian’Floyd’ Mueller, Frank Vetere, Martin R Gibbs, Jesper Kjeldskov, Sonja Pedell, and Steve Howard. 2005. Hug over a distance. In CHI’05 extended abstracts on Human factors in computing systems. 1673–1676.Google Scholar
- Shabnam Najafian and Nava Tintarev. 2018. Generating consensus explanations for group recommendations: An exploratory study. In Adjunct Publication of the 26th Conference on User Modeling, Adaptation and Personalization. 245–250.Google ScholarDigital Library
- Thuy Ngoc Nguyen and Francesco Ricci. 2018. A chat-based group recommender system for tourism. Information Technology & Tourism 18 (2018), 5–28.Google ScholarCross Ref
- Mark O’connor, Dan Cosley, Joseph A Konstan, and John Riedl. 2001. PolyLens: A recommender system for groups of users. In ECSCW 2001: Proceedings of the Seventh European conference on computer supported cooperative work 16–20 September 2001, Bonn, Germany. Springer, 199–218.Google Scholar
- So Yeon Park and Blair Kaneshiro. 2021. Social music curation that works: Insights from successful collaborative playlists. Proceedings of the ACM on Human-Computer Interaction 5, CSCW1 (2021), 1–27.Google ScholarDigital Library
- So Yeon Park, Audrey Laplante, Jin Ha Lee, and Blair Kaneshiro. 2019. Tunes Together: Perception and Experience of Collaborative Playlists.. In ISMIR. 723–730.Google Scholar
- Ike Picone. 2015. Impression management in social media. The international encyclopedia of digital communication and society 36 (2015), 469–476.Google Scholar
- George Popescu. 2013. Designing a voting mechanism in the GroupFun music recommender system. In HCI International 2013-Posters’ Extended Abstracts: International Conference, HCI International 2013, Las Vegas, NV, USA, July 21-26, 2013, Proceedings, Part II 15. Springer, 383–386.Google ScholarCross Ref
- Lara Quijano-Sánchez, Juan A Recio-García, Belén Díaz-Agudo, and Guillermo Jiménez-Díaz. 2011. Happy movie: A group recommender application in facebook. In Twenty-Fourth International FLAIRS Conference.Google Scholar
- Lara Quijano-Sanchez, Juan A Recio-Garcia, Belen Diaz-Agudo, and Guillermo Jimenez-Diaz. 2013. Social factors in group recommender systems. ACM Transactions on Intelligent Systems and Technology (TIST) 4, 1 (2013), 1–30.Google ScholarDigital Library
- Lara Quijano-Sanchez, Christian Sauer, Juan A Recio-Garcia, and Belen Diaz-Agudo. 2017. Make it personal: a social explanation system applied to group recommendations. Expert Systems with Applications 76 (2017), 36–48.Google ScholarDigital Library
- Juan A Recio-Garcia and Guillermo Jimenez-Diaz. 2019. A Novel Interface for the Explanation of Group Recommendations using Augmented Reality.. In ICCBR Workshops. 133–142.Google Scholar
- Juan A Recio-Garcia, Guillermo Jimenez-Diaz, Antonio A Sanchez-Ruiz, and Belen Diaz-Agudo. 2009. Personality aware recommendations to groups. In Proceedings of the third ACM conference on Recommender systems. 325–328.Google ScholarDigital Library
- Neelima Sailaja and Abigail Fowler. 2022. An Exploration of Account Sharing Practices on Media Platforms. In ACM International Conference on Interactive Media Experiences. 141–150.Google Scholar
- Shunichi Seko, Manabu Motegi, Takashi Yagi, and Shinyo Muto. 2011. Video content recommendation for group based on viewing history and viewer preference. In ITE Technical Report 35.7. The Institute of Image Information and Television Engineers, 25–26.Google Scholar
- Rafael Sotelo, Yolanda Blanco-Fernandez, Martin Lopez-Nores, Alberto Gil-Solla, and Jose J Pazos-Arias. 2009. TV program recommendation for groups based on muldimensional TV-anytime classifications. IEEE Transactions on Consumer Electronics 55, 1 (2009), 248–256.Google ScholarDigital Library
- Martin Stettinger, Alexander Felfernig, Gerhard Leitner, Stefan Reiterer, and Michael Jeran. 2015. Counteracting serial position effects in the choicla group decision support environment. In Proceedings of the 20th international conference on intelligent user interfaces. 148–157.Google ScholarDigital Library
- Rob Strong, Bill Gaver, 1996. Feather, scent and shaker: supporting simple intimacy. In Proceedings of CSCW, Vol. 96. 29–30.Google Scholar
- Emily Sun, Rodrigo De Oliveira, and Joshua Lewandowski. 2017. Challenges on the journey to co-watching YouTube. In Proceedings of the 2017 ACM Conference on Computer Supported Cooperative Work and Social Computing. 783–793.Google ScholarDigital Library
- Yuan Sun, Shuyue Fang, and Zuopeng Justin Zhang. 2021. Impression management strategies on enterprise social media platforms: An affordance perspective. International Journal of Information Management 60 (2021), 102359.Google ScholarDigital Library
- Javier Tibau, Michael Stewart, Steve Harrison, and Deborah Tatar. 2019. FamilySong: Designing to enable music for connection and culture in internationally distributed families. In Proceedings of the 2019 on Designing Interactive Systems Conference. 785–798.Google ScholarDigital Library
- Thi Ngoc Trang Tran, Alexander Felfernig, and Nava Tintarev. 2021. Humanized recommender systems: State-of-the-art and research issues. ACM Transactions on Interactive Intelligent Systems (TiiS) 11, 2 (2021), 1–41.Google ScholarDigital Library
- Zhiwen Yu, Xingshe Zhou, Yanbin Hao, and Jianhua Gu. 2006. TV program recommendation for multiple viewers based on user profile merging. User modeling and user-adapted interaction 16 (2006), 63–82.Google Scholar
Index Terms
- Investigating the Potential of Group Recommendation Systems As a Medium of Social Interactions: A Case of Spotify Blend Experiences between Two Users
Recommendations
Senior users of social networks: Technological experience and use habits analysis
DSAI '18: Proceedings of the 8th International Conference on Software Development and Technologies for Enhancing Accessibility and Fighting Info-exclusionInternet of Everything (IoE) intends to connect people, processes and things in an intelligent way to improve the user's experience and the person's well-being. With increasing life expectancy, the aging population is growing in developed countries, the ...
What makes us two-screen users? The effects of two-screen viewing motivation and psychological traits on social interactions
This study responds to recent calls for data regarding a phenomenon referred to as two-screen viewing (TSV), in which people on social network sites (SNSs) interact with each other regarding TV series they watch. Specifically, the current study aims to ...
Need fulfillment and experiences on social media
With an increasing inflow and outflow of users from social media, understanding the factors the drive their adoption becomes even more pressing. This paper reports on a study with 494 users of Facebook and WhatsApp. Different from traditional uses & ...
Comments