Algorithmic Collective Action with Two Collectives

TL;DR

The paper introduces a novel framework for algorithmic collective action with two distinct collectives operating on the same data-driven system, examining how differing objectives, strategies, sizes, and homogeneity shape collective efficacy. It formalizes components such as modified data, shared model parameters, objective functions, and a constructiveness score to quantify inter-collective interference, and validates the framework with experiments in language-model classification and recommender systems. Key findings show that interactions between collectives can dramatically alter outcomes—efficacy can drop from near $100\%$ when acting alone to around $25\%$ when a second collective acts concurrently (a loss of up to $75\%$)—and that collective size often dominates the impact of homogeneity in recommender contexts. The work highlights the need for transparency and governance to anticipate and manage multi-collective dynamics, and it sets a foundation for further exploration of more complex social-technical interactions and multi-collective scenarios.

Abstract

Given that data-dependent algorithmic systems have become impactful in more domains of life, the need for individuals to promote their own interests and hold algorithms accountable has grown. To have meaningful influence, individuals must band together to engage in collective action. Groups that engage in such algorithmic collective action are likely to vary in size, membership characteristics, and crucially, objectives. In this work, we introduce a first of a kind framework for studying collective action with two or more collectives that strategically behave to manipulate data-driven systems. With more than one collective acting on a system, unexpected interactions may occur. We use this framework to conduct experiments with language model-based classifiers and recommender systems where two collectives each attempt to achieve their own individual objectives. We examine how differing objectives, strategies, sizes, and homogeneity can impact a collective's efficacy. We find that the unintentional interactions between collectives can be quite significant; a collective acting in isolation may be able to achieve their objective (e.g., improve classification outcomes for themselves or promote a particular item), but when a second collective acts simultaneously, the efficacy of the first group drops by as much as $75\%$. We find that, in the recommender system context, neither fully heterogeneous nor fully homogeneous collectives stand out as most efficacious and that heterogeneity's impact is secondary compared to collective size. Our results signal the need for more transparency in both the underlying algorithmic models and the different behaviors individuals or collectives may take on these systems. This approach also allows collectives to hold algorithmic system developers accountable and provides a framework for people to actively use their own data to promote their own interests.

Figures (12)

• Figure 1: Framework overview. Collectives $c_1$ and $c_2$ form with different objectives. The items they aim to modify (romance vs drama), access to the model (white-box vs black-box access), actions they can do (promote vs demote) may vary. With this, the collectives produces $\hat{X_i}$, their individually modified dataset. Both collectives send their individually modified data into the model, which is then combined to produce $\bar{\mathbf{X}}_{1 \wedge 2}$. The system's objective function produces parameters $\hat{\theta}_{1 \wedge 2}$ which is influenced by the combined actions of $c_1$ and $c_2$. The collectives measure their success by their objective functions ($g_i$) applied on $\tilde{X_i}$, their target dataset, with the common parameters $\hat{\theta}_{1 \wedge 2}$
• Figure 2: Group construction process used for forming two collectives ($C=2$). Here we have $Q= 4$ clusters of users. $q_1$ serves as the seed for $c_1$, and $q_4$s serves as the seed for $c_4$. The collectives are then constructed by sampling with probability $p$ from their seed cluster and then uniformly at random from the remaining clusters $\frac{1-p}{3}$.
• Figure 3: Multiple Collective Action in the Resume Modification Task. Two collectives, each with their own strategy, insert a character to cause resumes with this character pattern to be classified to a given target class. Collectives are labeled [Letter][Number] where [Letter] corresponds to a target class and [Number] corresponds to a specific character used (mapped in \ref{['sec:appendix_classification']}). Each row represents a different character set. The $x$-axis is the $\%$ of population participating in a given collective. The $y$-axis is the efficacy, in this case, the top-one accuracy on predicting the collective's target class. The dashed line represents the baseline efficacy of the collective acting alone on the system. The solid lines show the efficacy of one of the collectives when two of them are acting on the system. We can see behavior when the presence of another collective can be helpful (\ref{['fig:normal_char_two_seperate']} solid lines above dashed lines) where they can be antagonistic (\ref{['fig:normal_char_same_char']}, \ref{['fig:special_char_two_seperate']}\ref{['fig:special_char_same_char']} solid lines below dashed lines) or minor impact (\ref{['fig:normal_char_same_char']}, \ref{['fig:special_char_same_char']}). Most notably, while Figures \ref{['fig:normal_char_two_seperate']} and \ref{['fig:special_char_two_seperate']} are the same scenario, just with different characters, they produce very different outcomes (helpful interaction vs harmful) potentially due to the characters used in \ref{['fig:normal_char_two_seperate']} appear in non-modified data while the characters used in \ref{['fig:special_char_two_seperate']} only appear in the modified data.
• Figure 4: Role of varying sizes. Here, there are two different strategy sets where, for each scenario, the strategy and the target used by each collective is distinct. The $x$-axis shows the percentage participation of the first collective, while the $y$-axis is the percentage participation in the second collective. Each square represents a single collective's efficacy (denoted in the title) (averaged across 5 trials). We see that, while there is some effect of collective sizes in the lower levels of participation in \ref{['fig:varying_normal_char_two_seperate']}, with sufficient amount of participation, both collectives get nearly full efficacy. This is in contrast to \ref{['fig:varying_special_char_two_separate']} where the two collectives act antagonistically. In particular, B101 requires nearly 2x participation compared to A100 to achieve its objective.
• Figure 5: Impact on collective size and homogeneity in changing in HIT ratio for a single collective scenario. The means and the standard deviations of the relative HIT ratio are plotted. The $x$-axis represents homogeneity as measured by the sampling propensity ($p)$, a higher $p$ means that the members of the resulting collectives are more similar. The $y$-axis is the relative HIT ratio, in other words, how much higher/lower are the rankings of a collective's item when acting on the system vs no action. Blue represents $n=10$, orange is $n=20$ and green is $n=50$. The solid lines represent demoting collectives, while dashed lines represent promoting collectives. Collective size plays a much larger influence than homogeneity, especially for demoting groups; homogeneity plays a secondary influence.