Fairness-Aware Online Positive-Unlabeled Learning

Hoin Jung, Xiaoqian WangEMNLP 2024 · OpenReview

TL;DR

This paper introduces a novel approach called Fairness-Aware Online Positive-Unlabeled Learning (FOPU) to address the critical issue of fairness in text classification tasks where only positive and unlabeled data are available, especially in dynamic, online environments. FOPU integrates a convex fairness constraint into the learning process, significantly improving fairness metrics without sacrificing classification accuracy.

What Problem Does This Paper Address?

Many real-world text classification applications, such as detecting toxicity or misinformation, face a significant challenge: obtaining perfectly labeled data is difficult because not all problematic instances are reported. Positive-Unlabeled (PU) learning offers a solution by training on labeled positive examples and a large pool of unlabeled data. However, a major concern with PU learning, particularly in online settings where data arrives sequentially, is fairness. Certain demographic groups might be disproportionately flagged as problematic due to inherent biases in the data. Existing research has largely overlooked this fairness issue in PU learning, especially in the context of online learning, leaving a critical gap in ensuring equitable AI systems.

Methods

The authors propose a novel approach, FOPU, that extends existing PU learning methods to both offline and online learning scenarios for linear and non-linear classifiers. The core of their method is the incorporation of a convex fairness constraint directly into the training objective. This constraint is designed to ensure Equalized Odds fairness, a metric that aims for equal true positive rates (TPR) and false positive rates (FPR) across different demographic groups.

Key aspects of their method include:

  • Convex Fairness Constraint: They adapt existing fairness constraints (like Demographic Parity and Equalized Odds) into convex forms, ensuring a unique optimal solution during optimization.
  • Online Learning Integration: They extend their approach to online learning by using techniques like Online Gradient Descent (OGD) for linear classifiers and Hedge Backpropagation for deep neural networks (MLP, LSTM). This allows the model to adapt to continuously arriving data.
  • Positive Rate Penalty: To prevent the fairness constraint from negatively impacting the crucial True Positive Rate (TPR), they introduce a penalty term that discourages a decrease in TPR or an increase in FPR compared to previous iterations.
  • Theoretical Guarantees: The paper provides theoretical analysis, including cumulative fairness regret bounds for both linear and deep neural network classifiers in online settings.

Datasets

The authors evaluate their approach on three publicly available datasets:

  • Wikipedia Talk: Used for toxicity classification.
  • Chat Toxicity: Also used for toxicity classification.
  • NELA-2018: Used for misinformation detection.

These datasets are standard benchmarks in their respective domains. The authors note that for NELA-2018, political leaning is considered the sensitive attribute.

Key Insights and Ideas

  • Fairness is Crucial in Online PU Learning: The paper highlights that the combination of PU learning and online learning exacerbates fairness issues, making it a pressing concern that needs explicit attention.
  • Convexity is Key for Robust Fairness: By formulating fairness constraints as convex optimization problems, the authors ensure stable and reliable convergence to fair solutions.
  • Online Learning Can Worsen Fairness: The experimental results demonstrate that online learning, without specific fairness interventions, can indeed amplify existing fairness disparities compared to offline learning.
  • FOPU Balances Fairness and Utility: The proposed FOPU method effectively improves fairness metrics (like ΔEOd) without significantly compromising the classification accuracy (F1-score).
  • Pre-trained Models Can Be Integrated: The FOPU framework is shown to be flexible enough to be integrated with pre-trained language models like BERT and DistilBERT, further enhancing its applicability.

Key Results

The experimental results consistently show that FOPU significantly improves fairness metrics (specifically, reducing ΔEOd) across various datasets and classifiers (linear, MLP, LSTM) in both offline and online settings. Crucially, this improvement in fairness is achieved while maintaining comparable F1-scores to the baseline methods without fairness constraints. For instance, Table 2 shows that FOPU consistently lowers ΔEOd while keeping F1-scores competitive. The ablation study in Table 4 also demonstrates the positive impact of the "positive rate penalty" in maintaining recall.

Limitations and Future Work

  • Limitations of Existing PU Methods: The authors note that some state-of-the-art PU learning methods (like Dist-PU and Robust-PU) require large datasets for training, making them less suitable for online settings where data is incrementally available.
  • Focus on Specific Fairness Metrics: The current work focuses on Equalized Odds. Future work could explore other fairness metrics and their interplay with PU learning.
  • Adaptability to Other Domains: While demonstrated on text classification, extending FOPU to other domains with similar PU learning challenges and fairness concerns would be valuable.
  • Hyperparameter Tuning: Like most machine learning models, FOPU's performance might be sensitive to hyperparameter choices, and further research into robust hyperparameter optimization for online fairness could be beneficial.

Summary updated

Abstract

Machine learning applications for text classification are increasingly used in domains such as toxicity and misinformation detection in online settings. However, obtaining precisely labeled data for training remains challenging, particularly because not all problematic instances are reported. Positive-Unlabeled (PU) learning, which uses only labeled positive and unlabeled samples, offers a solution for these scenarios. A significant concern in PU learning, especially in online settings, is fairness: specific groups may be disproportionately classified as problematic. Despite its importance, this issue has not been explicitly addressed in research. This paper aims to bridge this gap by investigating the fairness of PU learning in both offline and online settings. We propose a novel approach to achieve more equitable results by extending PU learning methods to online learning for both linear and non-linear classifiers and analyzing the impact of the online setting on fairness. Our approach incorporates a convex fairness constraint during training, applicable to both offline and online PU learning. Our solution is theoretically robust, and experimental results demonstrate its efficacy in improving fairness in PU learning in text classification.
BibTeX
@inproceedings{jung2024fairness,
  author = {Hoin Jung and Xiaoqian Wang},
  title = {{Fairness-Aware Online Positive-Unlabeled Learning}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2024},
  url = {https://openreview.net/forum?id=QVvo9JRd0Z},
}
Figure