Research map / Learning theory and algorithms / Prediction and decision losses
Algorithmic Collective Action in Recommender Systems: Promoting Songs by Reordering Playlists
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes two easy-to-implement, authenticity-constrained strategies (InClust and DirLoF) for a fan collective to insert a target song at strategic positions in playlists they control, boosting its test-time recommendations from a transformer playlist-continuation model.
Problem
Underrepresented artists have little visibility in popularity-biased music recommender systems; the paper asks whether a small collective of users can use strategic playlist editing to promote a new song in a transformer-based automatic playlist continuation model.
Approach
Insert a target song (edit distance 1) at a strategically chosen position in each collective-owned playlist: InClust places it before frequent songs to concentrate effort on similar contexts, while DirLoF places it after low-frequency songs to exploit long-tail statistics; both are derived from a statistical optimality assumption and need no model weights.
Key results
- DirLoF reaches amplification of up to 25x for collectives of about 0.025% of training data, roughly 40x more than an average song at the same training frequency.
- InClust becomes effective for collectives at or above 0.1% and reaches amplification of about 10 at alpha of roughly 2%.
- Performance loss on the firm's recommender from the hybrid strategy is marginal, and recommendations of other songs and collective participants are largely preserved.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DirLoF's low-frequency anchor placement instead of random insertion | Deezer transformer APC · Collective sizes α=0.0002, α=0.001, and α=0.002 | amplification: 22.82 (6.21) vs 0.00 (0.00); 16.38 (7.37) vs 3.36 (4.45); 10.73 (2.66) vs 4.32 (2.90) | Helped |
| Replace strategic placement with fixed insertion at positions 0, 1, 3, or 5 | Target-song insertion · α=0.0002, α=0.001, and α=0.002 | amplification: 0.00 (0.00) | Hurt |
| Replace strategic placement with fixed insertion at position 7 | Target-song insertion · α=0.0002, α=0.001, and α=0.002 | amplification: 0.00 (0.00), 0.02 (0.04), 0.07 (0.08) | Hurt |
| Use 4 attention heads instead of the hp* setting (ablation) | DirLoF · α=0.001 | amplification: 16.38 (7.37) -> 18.47 (9.03) | No clear effect |
| Set learning rate to 0.5 instead of the hp* setting (ablation) | DirLoF · α=0.001 | amplification: 16.38 (7.37) -> 21.77 (1.67) | No clear effect |
| Set dropout to 0.2 instead of the hp* setting (ablation) | DirLoF · α=0.001 | amplification: 16.38 (7.37) -> 16.13 (10.07) | No clear effect |
| Set weight decay to 5e-05 instead of the hp* setting (ablation) | DirLoF · α=0.001 | amplification: 16.38 (7.37) -> 26.32 (4.38) | No clear effect |
| Change attention heads, learning rate, dropout, and weight decay together (ablation) | DirLoF · α=0.001 | amplification: 16.38 (7.37) -> 27.92 (5.53) | No clear effect |
| Train for fewer epochs rather than 18 (ablation) | DirLoF · α=0.001 | amplification: 1 epoch: 0.02 (0.04); 2 epochs: 7.97 (5.86); 4 epochs: 9.88 (16.10); 8 epochs: 19.45 (12.82); 12 epochs: 25.79 (10.30); | Mixed |
| Use scraped 2024 Spotify stream counts as proxy song statistics | DirLoF · α=1% and α=0.1% | amplification relative to full-information setting: over 85%; about 50% | Helped |
| Insert the target song at random positions rather than target contexts (ablation) | InClust · Targeted context clusters | target song similarity rank | Hurt |
| Add the InClust context-targeting placement rather than random placement (ablation) | InClust · Unseen test playlists | amplification | Helped |
Surprising findings
- Placing the target song at a fixed position in every playlist is largely ineffective (amplification near 0), while random placement is ineffective for tiny collectives.
- Scraped 2024 Spotify stream counts serve as effective proxies for historical training-data popularity, letting a 1% collective reach over 85% of full-information amplification.
Limitations
- Experiments use a single public Deezer model trained on the 2010-2017 US Spotify MPD dataset, not on current platform data.
- Limitations and potential for misuse: individuals or popular artists could exploit the same strategies to reinforce inequalities.
- Actual implementation of an orchestration app is left for future work.
Experimental setup
Datasets: Spotify Million Playlist Dataset scraped Spotify stream counts (2024)
Benchmarks: Spotify Million Playlist Dataset test split
Metrics: amplification success rate R-precision NDCG number of clicks
Baselines: Random AtTheEnd Insert@i Random@i-j adversarial baseline (conservative)
Builds on: algorithmic collective action shilling attacks
Closest papers by meaning
- Echoes in Filter Bubble: Diagnosing and Curing Popularity Bias in Generative Recommender Systems (NeurIPS 2026 · similarity 0.83)
- A Statistical Framework for Algorithmic Collective Action with Multiple Collectives (NeurIPS 2026 · similarity 0.83)
- MuseChat: A Conversational Music Recommendation System for Videos (CVPR 2024 · similarity 0.83)
- Causal Representation Learning for Generalisable Recommendation (NeurIPS 2026 · similarity 0.83)
- Recommender Ecosystems: A Mechanism Design Perspective on Holistic Modeling and Optimization (AAAI 2024 · similarity 0.82)
- Would You Like Your Data to Be Trained? A User Controllable Recommendation Framework (AAAI 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.