Fairness-Aware Interactive Target Variable Definition

Dalia Gala, Milo Phillips-Brown, Naman Goel, Carina Prunkl, Laura Alvarez Jubete, Medb Corcoran, Ray Eitel-PorterIJCAI 2025 · DOI

TL;DR

Machine learning models inherit biases from how we define "success" (the target variable), often leading to unfair outcomes. This paper introduces FairTargetSim (FTS), an interactive tool that lets users explore how different definitions of a target variable impact fairness in algorithmic hiring, making the often-hidden ethical implications of this crucial step transparent.

What Problem Does This Paper Address?

When building machine learning systems, defining the "target variable" – what the model should predict or optimize for – is a critical step. This process is often subjective and value-laden, especially in domains like hiring, where "good employee" is not a straightforward metric. The authors argue that biases can be encoded directly into the target variable definition before any data is collected or models are trained, leading to downstream fairness issues. Technical developers often focus on model performance metrics, while non-technical stakeholders may not grasp the fairness implications of these definitions. This gap creates a need for tools that bridge this understanding and foreground the ethical considerations of target variable definition.

Methods

The paper introduces FairTargetSim (FTS), an interactive, simulation-based approach. FTS uses a case study of algorithmic hiring, grounded in real-world psychometric test data. The core of FTS is its interactive interface where users define two distinct target variables for what constitutes a "good" employee. This is done by adjusting sliders that represent the importance of five cognitive traits: memory, information-processing speed, reasoning, attention, and behavioral restraint. FTS then uses these user-defined target variables to train two corresponding Support Vector Machine (SVM) models. The tool provides visualizations that highlight how these different target variable definitions affect the models' training data and overall performance, including fairness metrics. FTS is open-source and available online.

Datasets

The paper uses real-world psychometric test data from Jaffe et al. (2022). This dataset comprises eleven tests, which the authors group into five traits: memory, information-processing speed, reasoning, attention, and behavioral restraint. This data is used to train the models within the FairTargetSim simulation.

Key Insights and Ideas

  • Target Variable Definition is a Primary Source of Bias: The paper strongly emphasizes that fairness issues can originate not just from biased data or algorithms, but fundamentally from how we define what we want the algorithm to achieve.
  • Transparency is Crucial for Stakeholders: FTS aims to make the abstract and often opaque process of target variable definition tangible and understandable for both technical developers and non-technical stakeholders (like hiring managers).
  • Interactive Exploration Drives Understanding: By allowing users to directly manipulate and see the consequences of their choices, FTS fosters a deeper, more intuitive understanding of fairness implications than purely theoretical explanations.
  • Bridging the Technical-Non-Technical Divide: The tool is designed to facilitate communication and collaboration between different groups involved in algorithmic development and deployment.

Key Results

The paper doesn't present traditional quantitative results like accuracy scores for a specific task. Instead, the "results" are demonstrated through the visualizations provided by FTS. These visualizations show:

  • Significant Demographic Differences: Different target variable definitions can lead to substantial differences in the proportion of selected applicants across demographic groups (e.g., gender).
  • Impact on Fairness Metrics: The choice of target variable definition directly influences fairness metrics such as true-positive rates, false-positive rates, and positive/negative predictive values.
  • Model Behavior Variations: Visualizations illustrate how models trained with different target variables evaluate individual applicants differently, as shown in Figure 3.
  • Feature Importance Shifts: Figure 4 demonstrates how user-defined feature importances (target variable definitions) affect model input scores for candidates from different demographic groups.

Limitations and Future Work

Limitations: * The current FTS implementation uses a specific case study (hiring) and a particular dataset. While the framework is extensible, the demonstrations are tied to this context. * The paper acknowledges that in real-world hiring, employers don't directly define cognitive traits but rather identify "good" employees based on existing staff. FTS provides guidance on how to translate this into target variables.

Future Work: * Real-World Application: Applying FTS to actual hiring settings and other domains like college admissions or search engines. * Community Extension: Inviting the broader community to contribute by adding support for different datasets, models, and visualizations. * User Studies: Conducting user studies with both technical and non-technical stakeholders to rigorously evaluate how FTS influences their understanding, development, and use of algorithms.

Summary updated

BibTeX
@inproceedings{gala2025fairness,
  author = {Dalia Gala and Milo Phillips-Brown and Naman Goel and Carina Prunkl and Laura Alvarez Jubete and Medb Corcoran and Ray Eitel-Porter},
  title = {{Fairness-Aware Interactive Target Variable Definition}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/1260},
}
Figure