Bias Mitigation Techniques: Decontaminating the Sample

Opening Scene

Once a lab confirms contamination, the fix isn’t one universal procedure — it depends entirely on where the contamination came from and how deep it runs. Sometimes the reagent itself needs replacing before any new sample is prepared. Sometimes the existing sample can be treated and salvaged. Sometimes it’s the final reading that needs recalibration against a known clean standard. A biased model has the same range of possible interventions, and picking the wrong one for the problem at hand wastes effort without actually fixing anything.

In Plain English

Bias mitigation covers a family of distinct techniques, generally grouped into three stages. Pre-processing techniques fix the training data itself before modeling begins — reweighting underrepresented groups, or removing correlations between protected attributes and other features. In-processing techniques build fairness constraints directly into the model’s training process, so the model optimizes for both accuracy and a fairness objective simultaneously. Post-processing techniques adjust the model’s outputs after training is complete, such as using different decision thresholds for different groups to equalize a chosen fairness metric. Each approach has real trade-offs, and choosing the right stage to intervene at depends on what’s actually causing the bias and how much control the team has over each stage of the pipeline.

The Old Way

Before bias mitigation matured into a set of distinct, well-understood techniques:

Teams facing a biased model often had only one blunt tool available: retrain from scratch and hope a different random seed or slightly different data produced a better result.
There was little shared understanding of the difference between fixing the data, fixing the training process, and fixing the output, so fixes were applied somewhat randomly rather than matched to the actual source of the problem.
Mitigation efforts frequently focused solely on the fairness metric without any tracking of how much accuracy was being traded away to achieve it.

Matching the fix to the actual stage where contamination lives is what distinguishes real mitigation from guesswork.

What’s Changing (and Why AI Is the Reason)

Practitioners increasingly select among pre-processing, in-processing, and post-processing techniques deliberately, based on diagnosis rather than convenience or habit.
This parallels the structured remediation practices covered in this content library’s dedicated responsible AI principles series, which frames mitigation as a deliberate design choice rather than a reactive patch.
As open-source libraries and commercial tools increasingly package these mitigation techniques into accessible, well-documented functions, applying a genuinely appropriate fix has become far more achievable for teams without a dedicated fairness research background.

The Metaphor, Fully Extended

The Lab’s Decontamination Options	Bias Mitigation Concept
Replacing the reagent before preparing any new sample	Pre-processing: fixing the training data before modeling
Treating and salvaging the existing sample directly	In-processing: building fairness constraints into training
Recalibrating the final reading against a known clean standard	Post-processing: adjusting model outputs after training
Matching the fix to where the contamination actually originated	Matching the mitigation technique to the actual source of bias

For Beginners: What to Actually Do

Learn the three broad categories of bias mitigation — pre-processing, in-processing, and post-processing — and roughly what each one changes.
Practice asking “where in the pipeline is this fix actually being applied” whenever you hear about a bias mitigation effort.
Get comfortable with the idea that mitigation almost always involves some trade-off, rather than being a free, purely beneficial fix.

For Practitioners and Leaders: The Deeper Layer

Diagnose the actual source of a bias finding before selecting a mitigation technique, rather than defaulting to whichever is easiest to implement.
Track both the fairness metric and accuracy metrics throughout mitigation, so any trade-off is visible and deliberate rather than accidental.
Document which mitigation technique was applied and why, so the choice can be reviewed and revisited as the model or its context changes.

Quick Recap

Bias mitigation techniques fall into three stages: pre-processing, in-processing, and post-processing.
Each stage intervenes at a different point in the pipeline and carries different trade-offs.
Matching the technique to the actual diagnosed source of bias is essential to a genuine fix.
Mitigation should be tracked and documented, not applied once and forgotten.

Where This Fits in the Series

Article 8 covered how bias gets baked into training data in the first place. This article covered the specific techniques used to decontaminate a model once bias is confirmed. Article 10 turns to a specimen that makes all of this considerably harder: large language models, and the new shape bias takes inside them.