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

Thank you for the careful reading and the detailed feedback. We have performed the additional investigations in response to your suggestions. Let us elaborate on the individual comments below.

Other baselines. We have implemented baselines where the song is placed in earlier positions in the playlist. The resulting amplification is shown in Table 1 in the supplementary PDF. None of the simple baselines is better than random.

Robustness to hyperparameters. We have modified several hyperparameters of the Deezer model (number of attention heads, learning rate, dropout rate, weight decay). The results, presented in Table 3 of the supplementary PDF, show that the DirLoF strategy consistently outperforms the random strategy across all configurations.

These results are not surprising, as the strategy builds on a generalizable statistical intuition rather than specifics of the model architecture. This makes the strategies robust to model configurations, provided the model approximates the conditional probabilities in the training data sufficiently well.

Rationale behind DirLoF. A sequential recommender is trained to predict the K most likely songs to follow a given seed context. The lever that DirLoF exploits is that certain contexts are overrepresented in the playlists the collective controls. Taking the random baseline as a reference, Fig. 4 shows that for $\alpha<0.001$, randomly inserting $s^*$ is unlikely to lead to recommendations at test time. But by concentrating effort on overrepresented contexts the collective can meet the threshold more often. To identify overrepresented contexts, DirLoF uses that certain songs with global frequency $< \alpha$ still appear in the collective. DirLoF targets contexts that end on such overrepresented songs.

As you mentioned, DirLoF relies on global frequency estimates of songs in the overall training data, and estimating small frequencies from few data points is hard in general. To justify this, we have implemented the strategy with only approximate statistics. And we find that the strategy is still effective; knowing only 10% of the data is sufficient to achieve decent gains (see Fig. 12 in the Appendix). Even scraping public song statistics that do not match the year of the data (we can only scrape current statistics, e.g., total number of streams) provides some useful signal. This convinced us that it is a worthwhile and feasible approach to pursue for small collectives.

Notation. The notation ($P$,$P^*$,$P_0$) serves to frame our work within the original framework of algorithmic collective action by Hardt et al. 2023. There is a dataset $D$ of $N$ playlists sampled from a universe of playlists $P_0$, and a fraction $\alpha$ for these playlists is modified according to the strategy $h$. As a result, the platform does not observe samples from the base distribution, but from a mixture $P$. Training is done using empirical expectations. At test time, the model is evaluated against unseen samples from $P_0$. We will make this more clear in the write-up.

Table 2. We show average recommendations of (anchor) songs at test time and the difference in recommendations ($\Delta R$) with vs. without collective action. These statistics are computed based on results from 5 different train-test fold splits. For each fold, we report the mean and the 95% confidence intervals.

Questions related to framing and fairness:

The focus of algorithmic collective action is to demonstrate that participants can exert control over the algorithm a platform deploys through strategic data reporting. This is a powerful message because it gives users a lever in an otherwise highly imbalanced power relationship. As the title suggests, we show how ‘participants can promote songs through sequence reordering’.

As you point out, how this lever is used can lead to different solutions. The redistribution of recommendations is inherently zero-sum among artists due to the limited budget of recommendations. But it is worth noting that this applies to any updates to a recommendation algorithm, as well as features like ‘Showcase’ on Spotify where one can pay for recommendations.

In an ideal world, this can be used for bottom-up unfairness mitigation in recommender systems, however, a positive effect on overall welfare is by no means guaranteed. But it is not justified to assume that every update a platform implements is beneficial and ethical, or that every update participants promote is harmful. Thus, we think it is valuable to take an approach complementing much of the existing literature and present collective action as a potential opportunity, rather than only a threat, as users often pursue legit interests. Understanding how interactions can be designed to promote positive change while suppressing exploitation is a very intriguing question. What our work contributes is empirical evidence that the fully adversarial model is not sufficiently nuanced to address this question.

Collective action increases system performance. This is only obvious once you take the perspective of participants being economic agents pursuing their own legitimate interests. If incentives were completely misaligned such a gain would not be possible.

Example about revenue. The purpose of this example is to illustrate how recommendations relate to monetary value. A study of collective action is incomplete without discussing these incentives, and it is hard to deny that more recommendations imply more revenue. But we will make this point without explicit examples to avoid misunderstanding.

We hope to have convinced the reviewer of the robustness of our proposed strategies and the valuable perspective for the NeurIPS community. In any case, we will dedicate a separate section in the final version to expand on the results in Appendix D.4 and discuss effects on other artists in more detail, making negative externalities and the need for more work very clear. Thank you for the feedback!

We are glad we could address most of your concerns with the additional experiments and clarifications. Let us follow up on the last one related to the potential for abuse of collective actions strategies. We agree with the reviewer that misuse is a concern and should not be understated. But we still think it is important that the potential for abuse does not prevent the community from working on algorithmic collective action as an interesting and valuable research direction.

That being said, we have revisited the paper with your comments in mind, and we can see how it could benefit from additional discussion and a broader perspective. We are happy to use the additional page in the final version to move the broader impact statement to the main body, as suggested by the ethics review, to acknowledge the limitations more prominently. While we have mentioned the threat that a single individual could control many playlists, we will also be more specific that a popular artist could have an advantage to gather large collectives. To further mitigate your concern we will also complement this by expanding more on the study of negative externalities on other artists in Section 4.3. We hope this is satisfactory and you consider voting for acceptance after these changes.

I'd like to thank the authors for the very detailed answer to my review. Most of my concerns are now addressed. However, I have to disagree on one of the most important ones: presenting the proposed method as a way to empower users to support their artist is, to me, oversimplistic and a bit disconnected from the reality of streaming services. Collective actions (not linked to recommendation) to divert revenue has been widely used in the music streaming industry, first by legit artists (see the Sleepify album by vulfpeck) but then (much more widely) by fraudsters. Both cases are unfair because they trick the payment system, but it's obviously even worse when done by fraudsters. The proposed method also concerns tricking royalties payment through the recommender system, which raises important fairness issues and is obviously also open to fraudsters (who usually control many accounts) and could also be used by major labels (that have strong marketing power) to divert revenue. I think then the work has critical ethical implications that must be discussed and acknowledged and that the positioning of the proposed method as an opportunity is problematic. I won't change my rating if this aspect is not considered in the paper.

Thank you for your feedback. Let us elaborate on your comments below.

Intuition for the proposed strategies. The design of our strategies relies on the idealized assumption that transformer-based models perform next song prediction by learning to model the conditional probability of songs, given context, in the training data. Achieving that the song is among the K most likely next songs implies a recommendation for similar contexts at test time. With this in mind, it makes sense to select a context $c$ and put the song $s^*$ right after this context to influence conditional probability $p(s^*|c)$. DirLoF and InClust are two different ways to select contexts $c$ to target. DirLoF exploits the overrepresentation of certain contexts in the playlists of the collective to pick those where they can make a larger difference. InClust targets a set of similar contexts that appear multiple times in the collective and targets them in a coordinated fashion. These are two principled levers from a statistical perspective. How we implement them is designed to make them practical.

• DirLoF. DirLoF targets contexts that end on a song that occurs infrequently in the overall training data (typically once among the playlists of the collective). If these contexts have an overall probability smaller than $\alpha$, they are overrepresented. Given the long-tail nature of the song distribution such songs are not that infrequent. Compared to inserting the song at random positions, the top K threshold can be met more often in that way (see Fig. 4).
• InClust. InClust targets contexts that precede a song $s_0$ that frequently occurs in the playlists controlled by the collective. Thereby they implicitly target contexts that are similar from the perspective of the recommender, and close to $\phi(s_0)$ in embedding space, by how the model is trained.

Alternative strategy. There might be other ways to design strategies, but we offer a strong initial baseline for a widely underexplored question. In contrast to the strategy you proposed, it is crucial that our approaches do not rely on knowledge of the embedding function. The collective does not require access to the model parameters to implement our strategies, nor to train a surrogate model to approximate them. It just needs to gather song statistics to identify the contexts worth targeting.

The fact that the strategies are designed based on a statistical intuition of sequential generation, implies that they are not specific to the model architecture. We demonstrate the robustness with additional experiments where we vary the hyperparameters of the model (see Table 3 in the supplementary PDF). With the additional ablation, we hope to have convinced the reviewer of the robust assumption underlying the design of our strategies. It is also in line with our argument that the effectiveness of the strategy decreases if model training is stopped early (see Table 2 in the supplementary PDF).

Theoretical guarantees. End-to-end theoretical guarantees, beyond intuition in an idealized setting, is a lot to ask for, given that related data poisoning strategies, including popular Shilling attacks, rarely come with any guarantees in the context of deep learning-based recommender systems. Instead, we opted for a rigorous empirical evaluation using the example of an industry-scale recommender system that has been deployed in production.

Anomaly detection. We agree that countermeasures could reduce the effectiveness of our strategies. However, they usually come at a cost for the firm. What makes the problem interesting is that it is not a zero-sum game between the platform and the users, in contrast to adversarial attacks. It is possible that agents pursue legitimate interests. In fact, our work is the first to provide empirical evidence that collective action strategies are not necessarily zero-sum. With this in mind – inspecting which use cases the platform would want to protect against is an interesting and open question for future work.

Contribution. We are the first to study the impact of authentic sequence modifications on sequential recommender systems. We show that such strategies are effective and might be a realistic aspect we have to take into account when building future systems. Thus, our work empirically carves out a novel space of questions around machine learning in a broader social-economic context. We provide solid empirical evidence that the typical adversarial threat model is not applicable to algorithmic collective action and call for a more nuanced study of incentives around ML models. This seems relevant for the NeurIPS community and an important contribution. Is there anything else we can do to convince you of this?

Thank you for the feedback, we will incorporate your suggestions and adjust the notation. In response to your comment, we also decided to add pseudo code to make the strategies more clear. It can be found in the supplementary PDF.

In the following let us elaborate on your questions:

Our strategies exploit the properties of the attention function in the transformer model, without requiring knowledge of the model parameters. The assumption we operate under when designing the strategies is that the system aims to approximate the conditional probability of the next song, given prior songs within a given context window. We place the target song after a specific context with the goal that the model picks up on these correlations from the training data. This is a probabilistic assumption on the sequence model and it is not specific to the details of the model architecture or the training algorithm used. Thus, it is not surprising that the performance does not change much as hyperparameters of the model are varied (see the new Table 3 in the supplementary PDF). It is worth noting that similar approaches have proven to be very fruitful when designing collective action strategies in classification, where strategies have been designed under a Bayes-optimality assumption (see Ref. [24] in the manuscript).

Intuition for the strategies. Within the context of the last paragraph, our strategies implement two heuristics to select the context $c$ after which to place $s^*$. For comparison, it is useful to have the random baseline in mind: for $\alpha<0.1$, placing the song $s^*$ after a randomly selected context is not sufficient for it to be among the top K for any of the contexts (it leads to no recommendation at test time, see Fig. 4).

• The DirLoF 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 $c$ that end on a low-frequency song (for small collectives, these typically appear only once in the controlled playlists). If these contexts have an overall probability smaller than $\alpha$, it means that they are overrepresented. As a result, targeting these contexts is more likely to effectively result in a recommendation at test time compared to random selection, which is what we see in the experiments.
• The InClust strategy aims to target contexts that appear multiple times among the playlists of the collective in a coordinated fashion. The intuition is that for contexts $c$ that precede a song $s_0$ that is popular within the collective, the context embedding $g(c)$ forms a cluster around the song embedding $\phi(s_0)$, due to the way embeddings are trained. As a result, the collective can achieve large mass for this specific region of the context space by placing $s^*$ before $s_0$. They would hope this is sufficient to meet the top K threshold for similar contexts at test time. Here we find that $\alpha>1$ is necessary for this to be effective. The ablation in Figure 5 illustrates how this strategy increases the similarity between $s^*$ and the targeted contexts, leaving other similarity values widely untouched.

Communication. Both strategies require communication among the members of the collective to coordinate on a target song $s^*$ and coordinate the execution time. In addition, each of the strategies requires one more statistic to be gathered: The InClust strategy requires participants to pool information about their playlists to identify the songs that occur most frequently in the playlists they control – no additional data needs to be collected. The DirLoF strategy requires global song statistics to get a sense of which songs are overrepresented. These statistics are then shared with other members of the collective. We envision the aggregation and communication of song counts to happen through an app or an online forum. Notably, there is a growing literature on developing tools for coordinating platform participants (see e.g., [1,2]) which could be used here.

While both strategies are easy to implement, DirLoF comes with some additional overhead for identifying low-frequency songs. These statistics need to be gathered from external sources but can be shared among the participants. For example, approximate statistics and information based on scraping public playlists and streaming statistics are sufficient for the strategy to be effective (Fig. 12). Beyond this, none of the strategies require access to model weights, model architecture, or the training algorithm used. And none of the strategies require any computations that go beyond simple song counts as they do not use surrogate models.

References:

[1] Do, K., De Los Santos, M., Muller, M., & Savage, S. (2024, May). Designing Gig Worker Sousveillance Tools. In Proceedings of the CHI Conference on Human Factors in Computing Systems (pp. 1-19).

[2] Imteyaz, K., Flores-Saviaga, C., & Savage, S. (2024). GigSense: An LLM-Infused Tool forWorkers' Collective Intelligence. arXiv preprint arXiv:2405.02528.

Thank you for your feedback and the positive assessment of our work.

To respond to your question we relate our work to Bendada et al. (2023). The authors describe a transformer-based recommender system for automatic playlist continuation (APC) that Deezer has deployed in production. The model implements the represent-than-aggregate paradigm to achieve better scalability and the authors demonstrate that the model achieves state-of-the-art performance on the Spotify Million Playlist Dataset. We use their model as a case study for our work. It is one of the few industry-scale APC models that are publicly available, allowing us to retrain it and inspect the effect of playlist modifications on recommendations. In contrast to Bendada et al. (2023), we are interested in the sensitivity of the model to strategic playlist modifications in the training data, and inspect the recommendation of specific artists. This is a dimension that has not been considered in prior work which focused on aggregate performance metrics of the model trained on a fixed dataset. We are the first to focus on algorithmic collective action in this context.

Regarding limitations, we will add a dedicated section in the future version to discuss the limitations of collective action and the negative externalities on other artists in more detail. We refer to the response to Reviewer tKsL for an extended discussion. We agree that this is valuable to put our work in context.