In the ever-evolving landscape of music discovery, the challenge of sifting through the overwhelming number of tracks released daily has become increasingly difficult for both platforms and streamers. This has resulted in a strong dependence on platforms like Spotify, Deezer, or Apple Music, which distribute and promote music through song recommendations. These systems rely on historical data to learn user preferences and predict future content consumption [1,2,3,4,5].

It has been widely documented that music recommendation systems suffer from popularity bias as they tend to concentrate recommendation exposure on a limited fraction of artists, often overlooking new and emerging talent [6,7,8,9,10,11]. As the success and visibility of artists are deeply influenced by the algorithms of these platforms, this can lead to a considerable imbalance in the music industry [12,13] and reinforce existing inequalities [14,15]. As a result, artists have started to fight for more transparency and fairer payments for online streaming services. The "Justice at Spotify" campaign, launched by the Union of Musicians and Allied Workers [16], has been signed by more than 28,000 artists. At the same time, the International Society for Music Information Retrieval has been arguing for promoting the discovery of less popular artists by recommending 'long-tail' items [17], as have other researchers [18,19,20].

We explore algorithmic collective action as a means for emerging artists to gain exposure in ML-powered recommender systems by mobilizing their fan base. Algorithmic collective action [21] describes the coordinated effort of a group of individuals to strategically report the part of the training data they control in order to impact the outcomes of a learning algorithm.

In music recommender systems the training data consists of user-generated playlists 𝑝. Each playlist is composed of an ordered list of songs. We want collective action to preserve user experience. Thus, we design collective action strategies under an authenticity constraint:

where Lev denotes the Levenshtein distance [22], also known as edit distance in information theory, counting the number of operations needed to transform one sequence into another.

We propose a concrete strategy that satisfies this constraint. Our strategy consists of inserting an agreed-upon target song 𝑠 * at a specific position in the playlist, as shown in Figure 2. The position 𝑖 to insert the song 𝑠 * is chosen by identifying the least likely song 𝑠 0 among the songs in the playlist 𝑝 and placing 𝑠 * right after 𝑠 0 . Intuition for the strategy. Sequential recommenders are trained to approximate the conditional distributions of songs. For a given context window, the model then recommends one of the top 𝐾 most likely songs to follow this context. Our strategy aims to exploit contexts that are overrepresented in the data the collective controls to increase the chance of meeting the top 𝐾 threshold. To this end, it selects contexts that end on a low-frequency song (for small collectives, these typically appear only once in the controlled playlists). To find these low-frequency songs participants of the collective can share information about the playlists they own, gather stream counts by scraping Spotify playlists, or use public APIs to gather external song statistics.

Notice that the probabilistic assumption on the sequence model is not specific to the model architecture or the training algorithm used. This makes the strategy robust and easy to implement in practice.

We empirically test the success of our strategy against a recent transformer-based automatic playlist continuation model [4]. The model has been deployed and made publicly available by Deezer-one of the biggest streaming platforms in the world. To train the model we use the Spotify million playlist dataset [23], treating each playlist as a user, and randomly sampling a small 𝛼-fraction to compose the collective. We find that by strategically placing the target song, small collectives can achieve disproportional representation at test time, see Figure 1. The star shows that a collective of size 1% can achieve that the target song is recommended in 6% of the playlist continuations at test time. This corresponds to a factor 6 amplification comparing training time and test time occurrences. In contrast, placing the song at the end of every playlist is largely ineffective. Also interesting to observe in Figure 1 is that a similar strategy does not seem to be implemented by any artist in the investigated data.

We further experiment with collective sizes from 0.03% to 3%, and show amplification in Figure 3. Interestingly, even tiny user collectives, controlling as few as 60 out of 1 million playlists can achieve an amplification of 25×. This is 40× more than an average song occurring at the same frequency in the training data. Notably, this can be achieved by choosing the position at which to insert one song strategically, while leaving the rest of the playlist untouched.

As we have seen, collective action offers an effective lever for platform participants to promote their interests on algorithm-driven platforms. However, strategies can only be effective if they are not creating equally strong incentives for other players in the system to counter them. In Songs are sorted by their training set frequency and aggregated into 50 evenly spaced bins with 95% CI.

Effect on recommendation performance-measured using standard metrics [24]: Normalized Discounted Cumulative Gain (NDCG), R-precision, and number of clicks to find relevant song (#C). the following we study the externalities of our strategy, choosing 𝛼 = 1%. First, we focus on the effect of our strategy on other artists. In Figure 4, we visualize the change in total recommendation counts for individual songs, binned according to their training set frequency. The purple star indicates the song promoted by the collective. We find that the gained recommendations are taken from songs of varying popularity levels, and no artist appears to be affected disproportionally.

Second, we focus on the effect of our strategy on the recommendation performance. Table 4 shows performance along multiple metrics. The loss seems to be very small for the platform. In comparison, the last column shows an alternative strategy with the same success, but replacing a relevant song every time 𝑠 * is recommended. This shows that the gained recommendations often replace irrelevant songs causing relatively little harm. Similarly, we see little performance drop for the platform participants, suggesting that their recommendations are also widely preserved at test time. If the song were to be actually relevant for individuals in the collective such a strategy could even help increase recommendation performance.

We designed an easy-to-implement collective action strategy under a natural authenticity constraint. We demonstrated that it can be effective in promoting a target song even for tiny collectives, while minimally impairing overall user experience. This suggests a widely unexplored design space for effective collective action strategies that differ from typical adversarial data poisoning attacks [c.f. 25,26,27,28]. They offer a powerful data lever [29,30], and an approach to participatory AI [31]. Thus, understanding the role of economic power [32,33], formalizing incentives [34], as well as quantifying long-term payoffs, dynamics and equilibria under collective action promises to be a fruitful direction for future work. 1