Practical Strategies for Applying the Disparate Impact Remover at Inference Time

Bias mitigation in machine learning often relies on pre-processing techniques such as the Disparate Impact Remover (DIR). However, applying DIR at inference time remains underexplored. We investigate methods of reusing data from the training phase during the inference phase when data comes in single data instances, in mini-batches of ten, and as full test sets.We propose two new methods and compare them against the original approach based on the quantile function. The first method, dictionary-based DIR, stores the mappings learned during training and applies them at inference time using nearest-neighbour matching for unseen values. The second method, merge-based DIR, dynamically merges incoming instances with a subset of training data before reapplying DIR.We evaluated these methods on three benchmark datasets: UCI Adult Income, Ricci v. DeStefano, and German Credit Data, measuring both fairness (Disparate Impact) and predictive performance (F1-score). The results show that the dictionary-based approach achieves accuracy comparable to the original quantile-based DIR with outcomes that are independent of the data input size. In contrast, the merge-based approach can produce more fair but less stable results that vary depending on the data size.