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

TL;DR

This work introduces two easily implementable strategies to select the position at which to insert the song with the goal to boost recommendations at test time and demonstrates how carefully designed collective action strategies can be effective while not necessarily being adversarial.

Abstract

We investigate algorithmic collective action in transformer-based recommender systems. Our use case is a music streaming platform where a collective of fans aims to promote the visibility of an underrepresented artist by strategically placing one of their songs in the existing playlists they control. We introduce two easily implementable strategies to select the position at which to insert the song with the goal to boost recommendations at test time. The strategies exploit statistical properties of the learner by targeting discontinuities in the recommendations, and leveraging the long-tail nature of song distributions. We evaluate the efficacy of our strategies using a publicly available recommender system model released by a major music streaming platform. Our findings reveal that through strategic placement even small collectives (controlling less than 0.01\% of the training data) can achieve up to $40\times$ more test time recommendations than an average song with the same number of training set occurrences. Focusing on the externalities of the strategy, we find that the recommendations of other songs are largely preserved, and the newly gained recommendations are distributed across various artists. Together, our findings demonstrate how carefully designed collective action strategies can be effective while not necessarily being adversarial.

Figures (13)

• Figure 1: By strategically choosing the position at which to insert the target song in a playlist, collectives can achieve a disproportionally high recommendation frequency relative to training set occurrences. The dashed red line corresponds to matching frequencies at train and test time, the blue dots correspond to naturally occurring songs.
• Figure 2: Imbalance in recommendation distribution. (left) The Lorenz curve shows that 80% of all recommendations are concentrated among just 10% of artists. (right) The Spotify track frequency distribution shows the long tail of song frequencies in user-generated playlists: close to 50% of tracks in playlists occur only once.
• Figure 3: Strategies: (a) InClust and (b) DirLoF
• Figure 4: Success of our collective action strategies. For tiny collectives DirLoF achieves an amplification of up to $25\times$ while uncoordinated strategies (Random, AtTheEnd) are mostly ineffective. For larger collectives, InClust outperforms DirLoF. Amplification significantly exceeds $1$, implying a disproportional test-time effect due to targeted song placement.
• Figure 5: Information bottleneck. The empirical amplification of the DirLoF strategy decreases with worse song statistics but scraped song streaming counts can serve as a practical solution.

Theorems & Definitions (1)

• Definition 1: Authenticity constraint