Multiplier Postprocessing for ML Bias Mitigation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning models face challenges in achieving high levels of fairness and accuracy simultaneously, with existing methods being non-deterministic, requiring retraining, limited to binary classification, and lacking configurability, which complicates fairness and accuracy optimization in industrial applications.
Innovation Solution
A bi-objective optimization approach for postprocessing class probabilities using multipliers that are specifically optimized for fairness and accuracy, allowing for flexible metric optimization and applicability to multiclass classification, while avoiding randomness and retraining, and ensuring deterministic and interpretable predictions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If preprocessing of training corpus or adjustment inside ML model is used to mitigate bias, then fairness can be improved, but retraining is required which is slow and discourages exploration
Solution Approach 1:
The patent applies preliminary action by pre-computing and storing fairness adjustment multipliers for different sensitive feature values before actual classification. These multipliers are calculated in advance based on fairness metrics, so that when new data comes in, the system can immediately apply the pre-computed multipliers without retraining the model, thus resolving the contradiction between improving fairness and avoiding retraining time loss
Solution Approach 2:
The patent extracts the fairness adjustment mechanism from the core ML model by separating it into a postprocessing step. Instead of modifying the model internals or retraining, the system extracts the original model predictions and applies fairness multipliers as a separate postprocessing layer, allowing fairness improvement without affecting model training speed or complexity
2Reliability
If random selection or random generation is used to adjust inference, then fairness can be improved, but the inference becomes non-deterministic and difficult to explain
Solution Approach 1:
The patent applies parameter changes by using deterministic mathematical operations (multiplication by pre-computed multipliers) instead of random processes. The fairness adjustment is achieved by changing the probability parameters through controlled scalar multiplication, which maintains determinism while improving fairness metrics, thus resolving the contradiction between fairness and determinism
3Reliability
If adjustment of inferred probability is applied to improve fairness, then fairness can be improved, but the probability may become outside supported range breaking classification boundaries
Solution Approach 1:
The patent applies dynamics by making the multiplier values adaptive and context-dependent. Different multipliers are selected based on the sensitive feature values and the specific class probabilities, allowing the system to dynamically adjust probabilities while maintaining them within valid ranges. This dynamic selection prevents probability overflow while achieving fairness improvements
4Device complexity
If only binary classification is supported, then model complexity is reduced, but versatility and adaptability to different classification problems is limited
Solution Approach 1:
The patent applies universality by designing a postprocessing framework that works universally across different classification problems. The same multiplier-based approach can be applied to binary, multiclass, and even regression problems by adjusting which probabilities are modified and which are kept as reference, thus achieving multi-functionality without increasing core model complexity
Data Source
AI summary
In an embodiment, a computer infers, from an input (e.g. that represents a person) that contains a value of a sensitive feature that has a plurality of multipliers, a probability of a majority class (i.e. an outcome). Based on the value of the sensitive feature in the input, from the multipliers of the sensitive feature, a multiplier is selected that is specific to both of the sensitive feature and the value of the sensitive feature. The input is classified based on a multiplicative product of the probability of the majority class and the multiplier that is specific to both of the sensitive feature and the value of the sensitive feature. In an embodiment, a black-box bi-objective optimizer generates multipliers on a Pareto frontier from which a user may interactively select a combination of multipliers that provide a best tradeoff between fairness and accuracy.