Algorithmic Collective Action in Recommender Systems: Promoting Songs by Reordering Playlists

Joachim Baumann, Celestine Mendler-Dünner

Introduction

In the modern music industry, streaming platforms like Spotify and Deezer act as the primary gatekeepers of discovery. Their recommendation algorithms—specifically Automatic Playlist Continuation (APC) models—largely determine which artists reach new listeners. However, these systems are frequently criticized for "popularity bias," a phenomenon where a small fraction of established artists receives the vast majority of exposure, while independent or "long-tail" artists remain invisible.

This research explores a novel approach to addressing this imbalance: algorithmic collective action. Unlike traditional "shilling attacks" that rely on fake profiles or fraudulent ratings, this study investigates how a coordinated group of real fans can leverage their own data to influence industry-scale transformer-based models without degrading the platform's quality or compromising their own user experience.

The Problem: The Popularity Trap

APC models are designed to predict the next song in a sequence based on historical user behavior. Because these models are trained on massive datasets where popular songs appear more frequently, they naturally develop a bias toward recommending those same hits. For an emerging artist, breaking into these recommendation loops is nearly impossible without significant external marketing spend.

Previous attempts to manipulate these systems often involved malicious botnets. This paper shifts the focus toward "organic" influence, asking whether a small, coordinated collective can act as a lever to boost an artist's visibility while maintaining "authenticity"—defined here as making only minimal, realistic changes to existing playlists.

Strategic Insertion Methods

The researchers focused on transformer-based architectures, which learn the statistical relationships between songs in a sequence. They proposed two primary strategies for song insertion, governed by an authenticity constraint: fans only add the target song ($s^*$) once to their existing playlists (a Levenshtein distance of 1).

  1. InClust (Concentrating Effort): This strategy targets specific clusters in the model's embedding space. Fans coordinate to place the target song directly before a highly popular "anchor" song. By doing this across multiple playlists, the collective creates a strong statistical signal that the target song is a precursor to a hit, effectively forcing the model to associate the two.
  2. DirLoF (Directly targeting Low-Frequency contexts): This strategy exploits the "long-tail" of the dataset. Fans place the target song after a very rare song. Because the model has limited data on what follows rare tracks, even a small number of coordinated insertions can "overpower" the existing signal, making the target song the definitive recommendation for that specific context.
  3. Hybrid Strategy: This approach combines both methods, using InClust for popular anchors and falling back to DirLoF for others to maximize reach.

Crucially, these strategies do not require access to the model's internal weights. They can be executed using aggregate song statistics, such as play counts, which are often publicly available or can be estimated via scraping.

Experimental Framework

The study utilized the Spotify Million Playlist Dataset (MPD) to simulate user behavior and evaluated the strategies against an industry-scale APC model released by Deezer. This transformer-based model reflects the architectures currently deployed to millions of users, providing a realistic testbed for the collective action strategies.

Key Insights and Results

The findings demonstrate that recommender systems are surprisingly susceptible to small-scale coordinated efforts:

  • The Power of Small Collectives: A group controlling less than 0.01% of the training data—as few as 60 playlists—can trigger a massive boost in recommendations.
  • 40x Amplification: Strategically placed songs received up to 40 times more test-time recommendations than an average song with the same initial frequency in the training set. For very small collectives, the DirLoF strategy achieved an amplification factor of 25x.
  • Strategic Placement is Vital: Simply adding a song to a playlist is insufficient. The research found that the specific "where" (the context) matters as much as the "if." Placing a song at a fixed position (e.g., always at the end) was largely ineffective compared to the context-aware InClust and DirLoF methods.
  • System Stability: Remarkably, these actions do not "break" the recommender for other users. Metrics like NDCG (Normalized Discounted Cumulative Gain) and R-precision remained stable. The target artist's gains were "stolen" from a wide variety of other artists rather than disproportionately harming a single competitor.
  • Robustness: Even when fans lacked perfect data and relied on scraped 2024 Spotify stream counts to estimate 2017 training statistics, they still achieved over 85% of the potential recommendation boost.

Ethical Implications and Limitations

While algorithmic collective action offers a path for indie artists to bypass traditional gatekeepers, it presents a significant ethical dilemma. The authors note that the same tools could be co-opted by major labels or already-popular artists to further entrench their dominance, potentially exacerbating the inequality the paper seeks to mitigate.

Furthermore, while the mathematical framework is sound, the "orchestration" of such a movement remains a hurdle. Building a platform to coordinate thousands of fans to place songs at specific anchor points requires significant community management. Finally, the long-term response from streaming platforms—who may develop counter-measures against such coordination—remains an open area for future research.

Summary updated

BibTeX
@inproceedings{baumann2024algorithmic,
  author = {Joachim Baumann and Celestine Mendler-Dünner},
  title = {{Algorithmic Collective Action in Recommender Systems: Promoting Songs by Reordering Playlists}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2024},
  url = {https://openreview.net/forum?id=wGjSbaMsop},
}
Figure