Socially-Motivated Music Recommendation

June 06, 2024 Published by Spotify Research

tldr – The wide variety of reasons people listen to music include both individual motivations, like self-awareness and mood regulation, as well as social motivations, such as demonstrating belonging to a group or feeling connected to friends. Music recommendation systems implicitly tend to focus on individual motivations for music listening. In this work, we explore the possibility of a recommender system geared towards a particular social motivation: the desire to listen to music that is popular within one’s community. We frame a recommendation task around this need and propose a simple method for satisfying it. Using an evaluation metric devised for this task, we examine our approach’s effectiveness across cultures.

Social motivations for music listening

Music is a social activity that connects people globally through concerts, karaoke, and shared listening experiences. Even when individuals listen to music alone, their social motivations influence their choices, often leading them to seek music that fosters a sense of community. This desire for shared experiences is evident when people intentionally listen to music that is popular within their community, which differs from listening aimed at mood regulation or self-awareness. Such communal listening lays the foundation for social bonding through shared musical familiarity. Although these motivations are universal, research shows they are especially significant in collectivist cultures. Collectivism is a common way of characterizing cultures that emphasize duty to in-group and maintaining harmony, in contrast to individualism, which prioritizes independence and uniqueness.

There has been little exploration of how recommendations should account for the social motivations behind music consumption. While algorithmic recommendation systems often draw inspiration from social processes, they seem to ignore important elements of social value. Collaborative filtering, for example, uses the similarity between people’s musical taste as the basis for recommendations. But this common arrangement frequently assumes that the person receiving the recommendation does not care who those other people are and that it does not matter when their listening took place.

A recommendation system framed around music’s social functions might require different assumptions. We explore such a system, focusing on the goal of helping people connect with their communities through shared musical familiarity. To accomplish this goal, we attempt to identify and recommend newly released songs that are “trending” within a user’s community. 

Recommending trending music within a community

To recommend music that is trending within a community, we need two basic components: a method of determining what music is trending, and a model of user communities. To study this problem, we use a dataset of the listening activity of 100,000 Spotify users from 20 countries over a period of two months.

First, to figure out what music is trending, we look at how awareness of a song spreads within a community over time. We measure this by looking at what fraction of a community has listened to a song, day by day, since the song’s release. In the figure below, for an example group of users, we can see how awareness evolves over time for several newly-released songs. Shorter lines reflect songs released closer to the end of our two month sample period. Some songs gain popularity quickly, while others take longer but ultimately achieve greater popularity. Many users do not show awareness of popular songs until over a month after their release. Our recommendation system aims to help these users discover these songs earlier.

We use the idea of the gradual spread of song awareness to implement a heuristic approach to recommendation. On day x after the release of a new song, if the fraction of a group of users who have listened to the song exceeds threshold y, we identify the song as trending; we then recommend the song to all users in the group who have not yet listened to it.

Next, we model the communities where members care about keeping up with each others’ musical preferences. We infer these communities from users’ listening behavior and demographic attributes, using information such as age, gender, country of residence, and the languages and genres in which users listen to music. We use combinations of these attributes to segment users into distinct groups.

We conduct a series of offline experiments to examine the effectiveness of this approach and how it responds to three variables: the attributes used to segment users; and the number of days to wait and the fraction of users required in order to make a recommendation. To evaluate these experiments, we focus on two important qualities of socially-motivated music recommendations: relevance and social value.

To measure relevance, we compute the precision of our recommendations compared to users’ implicit feedback on held-out data. In other words, if we recommend a trending song at day x to a user who would later have listened to the song on their own, we consider the recommendation a success.

Listening to the same music as one’s community implies that the listening must occur around the same time. To capture the social value of this activity, we thus focus on the timeliness of our recommendations. The earlier a user learns about the music trending in their community, the more social value the recommendation provides. We measure timeliness by counting the number of days between day x, when we recommend a song, and when a user would have discovered it on their own according to our sample data. This metric estimates the amount of time our approach “saved” the user from not knowing about a trending song.

By considering both the timeliness and precision of recommendations, we can assess how well our approach meets two key needs associated with socially-motivated music listening: determining the music that is relevant to both an individual and their community, and identifying it promptly enough to provide social value.

Findings

How well does the proposed music recommendation system satisfy users’ interests and provide social value? We consider the results of our offline experimentation from a few perspectives.

First, we find a clear tradeoff between two parameters of our system: the number of days to wait and the fraction of users required in order to make a recommendation. Waiting longer to assess a song’s trendiness leads to recommendations that are more precise but less timely. Similarly, the fraction of users we require to make a recommendation can be understood as a measure of strictness; higher values for this threshold lead to fewer, more accurate recommendations.

Next, we compare different attributes in terms of how well they approximate real-world musical communities. The most effective recommendations emerge from segmenting users by their genre preferences, country of residence, and age. These results correspond to the intuition that geography, culture, and taste all play a role in defining people’s musical communities. It also echoes findings in previous research that people’s relationship to new music changes with age.

Our approach to recommendation was notably more effective in countries characterized by a higher degree of collectivism. This result suggests the particular importance of socially-motivated listening in collectivist cultures, and it is consistent with earlier research on collectivism and listening patterns. These findings also point to the broader importance of cultural differences to music recommendation systems.

The evaluation metric of timeliness offers a useful lens for examining the social value of a music recommendation, emphasizing that when a recommendation occurs can be as important as what is recommended. Studying exactly how the timeliness of recommendations translates into value for listeners presents an opportunity for further research.

For more information, please see our paper:
Socially-Motivated Music Recommendation
Ben Lacker, Samuel Way