Abstract
This paper bridges the gap between Bayesian statistics and critical theory, arguing that "subjective Bayes"—which explicitly incorporates prior beliefs and uncertainty—resonates with feminist and decolonial concepts of "situated knowledge." By moving from a single "authoritative" model to a distribution of possible functions, the authors propose a more humble, politically aware approach to machine learning and data science that acknowledges the positionality of the researcher.
Introduction: Challenging the 'View from Nowhere'
Modern machine learning often operates under a "function-fitting paradigm." This approach assumes there is a single, universal mathematical function ($f$) that perfectly describes the relationship between inputs and outputs. The authors argue this creates a "view from nowhere"—what feminist scholar Donna Haraway calls the "God trick"—which masks the social, political, and colonial biases inherent in data collection and algorithmic design.
The challenge is that standard statistical methods often ignore the "situatedness" of knowledge—the idea that all knowledge is produced from a specific position, influenced by power dynamics, culture, and history. This paper explores whether the mathematical structure of Bayesian inference can be used to challenge this false universalism and create "pluriversal" models that acknowledge multiple, valid perspectives.
Methodology and the Bayesian Framework
The authors employ a mix of mathematical demonstration and theoretical synthesis. To illustrate their point, they use a "toy example" comparing standard frequentist linear regression (which produces one "best" line) with Bayesian inference.
The core of the approach relies on the Bayesian update process, where prior beliefs are combined with observed data to produce a posterior distribution.
The study utilizes a small, self-referential dataset consisting of the 10 most frequent words across the articles in the "Situated Bayes" special issue of Computational Culture. Using Markov Chain Monte Carlo (MCMC) algorithms via the PyMC library, the authors generate 16,000 samples to visualize an "ensemble" of 500 possible regression lines rather than a single line of best fit.
Key Insights: Subjectivity and Pluriversality
The paper offers several critical reframings of statistical concepts:
- Subjective Priors as Positionality: In Bayesian statistics, the "prior" represents an agent’s degree of belief before seeing data. The authors reframe this as a mathematical way to account for a researcher's "positionality" or "situatedness."
- Pluriversal Functions: Unlike standard models that emit a single answer, Bayesian inference learns a distribution of functions. This multiplicity mirrors the "pluriverse"—a world where many worlds (and many valid interpretations) coexist.
- Math as Media: Mathematics is treated not as a neutral language but as a "mediatic entity" that carries historical and colonial weight.
- Epistemic Humbleness: By explicitly modeling uncertainty, Bayesian methods can be "humbled," acknowledging they are one way of mapping reality rather than a totalizing truth.
Results: Visualizing the Situated Bayes
The Bayesian model successfully undermined the "authority of the single line." While standard regression suggested a fixed relationship, the Bayesian posterior showed a wide spread of plausible slopes and intercepts, visually demonstrating that many different "truths" could explain the same data.
The authors produced a significant conceptual map (shown above) that aligns data science elements (like "Objectivism") with their feminist/decolonial critiques ("Reductivist translatability") and proposed alternatives ("Relational empiricism"). This framework establishes a research agenda for applying these perspectives to software studies and algorithmic design.
Limitations and Future Directions
Despite the promise of "situated Bayes," the authors acknowledge significant caveats. Even "multiple lines" remain trapped within a Cartesian coordinate system, which is itself a Western imposition. Furthermore, the "toy example" used is intentionally simple; scaling these "situated" approaches to industrial-scale machine learning models remains a challenge.
Future work may explore "non-parametric" Bayesian models, which adapt in complexity to the data, potentially offering even more room for "pluriversal" knowledge and marking a path for future technical research in human-computer interaction (HCI).
Summary updated