Quaresmini & Primiero (2023) published an article addressing the critical issue of inherent biases in machine learning (ML) models, particularly when applied to people, and proposed a theoretical framework for mitigating these biases to achieve fairer AI classification systems. We will discuss their article to deal with this important issue
Their core argument was that existing bias mitigation techniques, which primarily rely on statistical analyses, are insufficient to address social discrimination effectively. The authors advocate for integrating ethical principles and data quality dimensions, such as consistency, completeness, timeliness, and reliability, into the evaluation and development of AI systems.
ML models trained on vast datasets often perpetuate and amplify existing societal biases. This may be seen in cases such as:
- Facial recognition systems: These systems can generate false positives, misidentifying innocent individuals as suspects due to prejudices related to their social identity.
- Social credit systems: These systems use algorithms to assess an individual’s social credit, potentially leading to discriminatory outcomes based on biased scores.
- Automatic Gender Recognition (AGR): These technologies, which algorithmically derive gender from facial traits, often fail to accurately classify non-binary and transgender individuals due to their reliance on a binary, cis-normative conception of gender.
The authors argue against using facial traits as a proxy for gender identity detection, as it echoes problematic historical practices like phrenology and physiognomy. However, recognising the increasing prevalence of these systems, they focus on making classification strategies fairer to mitigate the adverse impacts on misgendered individuals.
The article critiques the over-reliance on statistical analysis in existing bias mitigation tools, asserting that these methods alone cannot effectively address the complexities of social discrimination. The authors highlight the need for a more comprehensive approach that incorporates ethical considerations and data quality dimensions.
To illustrate these limitations, the article examines Cleanlab, a framework for finding label errors in datasets. Cleanlab uses Confident Learning (CL) to estimate the joint distribution between noisy, given labels and uncorrupted, unknown labels. However, Cleanlab and similar tools often assume the categoricity of data, which proves problematic in cases such as gender labelling, where:
- Non-binary individuals: The label set (male/female) is incomplete, failing to represent their identities.
- Transgender individuals: The dataset is inconsistent, as their gender identity may shift over time.
In these situations, accuracy alone is insufficient to ensure the correctness and fairness of the classification algorithm. The authors contend that other data quality dimensions must be considered to address these limitations.
The authors propose incorporating the following data quality dimensions into bias mitigation strategies:
- Accuracy: The closeness of a value to the correct representation of reality.
- Completeness: The extent to which data has sufficient breadth, depth, and scope for its task.
- Consistency: The coherence of data elements and adherence to semantic rules.
- Timeliness: The freshness of data over time for a specific task.
They argue that timeliness is particularly crucial, as data evolves, and the effectiveness of labels can change over time. By adding the dimension of temporality to completeness and consistency, they address the limitations of static approaches to bias mitigation.
To account for these dynamic factors, the authors suggest considering the labelling task within a defined timeframe: This allows for measuring the probability of a label change over time and associating temporal parameters to labels: This enables the computation of the probability that a given label might change over a given timeframe.
This approach acknowledges that the completeness of a label set, and the perseverance of labels are not static assumptions, and that the correctness of a task depends on accounting for these temporal dynamics.
To implement these concepts, the authors provide mathematical formulations:
They define the completeness of a label set over a timeframe, ensuring that for every data point representing an element in the population of interest, a correct label exists at any two possibly consecutive points in time.
They define reliability of a classification algorithm over a timeframe, ensuring that the algorithm’s change rate (the rate at which its accuracy deteriorates) remains below a certain safe threshold.
Based on these definitions, they argue that fairness for AI classification systems requires the system’s ability to provide reliable and correct outcomes over time, satisfying both the completeness and reliability criteria.
The article notes that existing tools such as BRIO can be adapted to incorporate these temporal dimensions. BRIO already functions as a post-hoc model evaluation tool, allowing users to investigate behavioural differences and measure bias amplification. The time-based analysis of completeness and reliability can be implemented in BRIO by:
- Defining a timeframe to check that any given data point of interest is matched against a desirable label.
- Ensuring that the overall change rate of error for one or more classes of interest does not surpass a certain threshold.
Both features rely on the user for the identification of the desirable label for any data point and for the admissible distance, ensuring that the system remains adaptable and responsive to evolving societal contexts.
The authors conclude by calling for a broader discussion on the data dimensions used in bias mitigation tools and emphasise the critical role of timeliness. By incorporating ethical principles, data quality dimensions, and temporal considerations, AI classification systems can better address the complexities of social discrimination and achieve more equitable and reliable outcomes. They highlight the need for ongoing monitoring and adaptation to ensure fairness in a world where data, identities, and societal norms are constantly evolving. This work provides a theoretical framework that will hopefully be implemented by tools such as BRIO in future studies.
Further reading
Quaresmini, C., & Primiero, G. (2023). Data quality dimensions for fair AI. ArXiv, abs/2305.06967. Read it here