🔬 Research Summary by Peizhao Li, a Ph.D. Candidate at Brandeis University working on Machine Learning, with a special interest in Trustworthy and Responsible AI, Algorithmic Fairness, and Deep Learning.
[Original paper by Peizhao Li and Hongfu Liu]
Overview: With the fast development of algorithmic governance, fairness has become a compulsory property for machine learning models to suppress unintentional discrimination. In this paper, we focus on the pre-processing aspect for achieving fairness and propose a data reweighing approach that only adjusts the weight for samples in the training phase. Our algorithm computes individual weight for each training instance via influence function and linear programming, and in most cases, demonstrates cost-free fairness through vanilla classifiers.
Introduction
For artificial intelligence technology deployed in high-stakes applications like welfare distribution or school admission, it is essential to regulate algorithms and prevent unaware discrimination and unfairness in decision-making. Even though general data-driven algorithms are not designed to be unfair, the outcomes can unintentionally violate the AI principle of equality. Typically learning from historically biased data, the learner can retain or amplify the inherent bias if there is no proper constraint on data or algorithms. As a consequence, the decisions from these algorithms may disadvantage users in certain sensitive groups (e.g., women and African Americans), therefore raising societal concerns.
To mitigate unfairness algorithmically, the authors focus on the pre-processing aspect of fair learning algorithms, i.e., only transform the input training data for a machine learning model to make its decision to be fair. The pre-processing category directly diagnoses and corrects the source of bias and can be easily adapted to existing data analytic pipelines. The proposed method is to granularly compute a weight for every sample in the training set, and a vanilla model without any additional constraints can be trained on this reweighed training set and deliver fair predictions.
Key Insights
Characterizing Sample Influence
Consider a classifier trained on some training set. We can assess the contribution from one individual training sample by training the classifier two times with and without that training sample, respectively. By doing so, it is straightforward to know in a counterfactual how the model will change with regard to some typical measurements, e.g., fairness or predictive utility.
However, it can be prohibitively expensive to retrain the model many times. Influence function from robust statistics provides a first-order approximation to the contribution of a training sample. It measures the effect of changing an infinitesimal weight in a sample, then linearly extrapolates to a whole removal of a sample. The influence function offers us an option to estimate the change of a model toward any evaluative function by removing some training samples from the training set.
Fairness at No Utility Cost
Having the influence function at hand, and by writing the fairness goal into a differentiable equation, we can quantify the impact of one training sample on two objectives: predictive utility and fairness. We use the influence function to calculate how the model will change and how the utility and fairness will change by reweighing a training instance. Theoretically, under the assumption of training data sufficiency and diversity, and with a proper model dimension, the authors prove that some training data reweighing strategy always exists to improve the fairness while at least keeping the utility not decreasing.
Having that theoretical finding, the authors solve the reweighing through linear programming and compute the individual weight for each training instance. The linear programs have constraints from both utility and fairness to achieve a fair model prediction without harming the prediction performance. Solving the linear programming problem can be very fast, i.e., within a few seconds, for a tabular dataset with 30k+ training samples.
The authors demonstrate the proposed reweighing approach on multiple tabular datasets compared to several pre-processing and in-processing methods. Experimental results show that such reweighing method can achieve cost-free fairness in most cases. In contrast, other competitive methods usually get fair results by introducing non-negligible harm to the model’s utility.
Between the lines
In recent years, many algorithms have been developed to make ML models meet criteria from AI ethics. However, many algorithms bring non-trivial change to the original model, greatly changing the utility performance. Fairness at no utility cost is a favorable property since it could help to popularize fair algorithms for extensive utility-driven products and alleviate the concerns from the deployment of fair algorithms.