Predictive Model Training With Missing Sensitive Attributes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing predictive modeling techniques struggle to accurately account for sensitive attributes when they are missing or uncertain, leading to biased models and violations of sensitive attribute inclusion constraints.
Innovation Solution
A computer-implemented method that learns a predictive model with missing or crippled data, formulates a problem based on sensitive attribute criteria, and reduces tasks to a quadratically constrained quadratic problem (QCQP) to characterize loss function values, ensuring strict constraint adherence without performance loss.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If predictive models are trained using uncertain or missing sensitive attribute data, then model training can proceed with available data, but the model fails to accurately capture target sensitive attribute inclusion and may violate fairness constraints
Solution Approach 1:
The patent introduces an intermediary relaxation mechanism that mediates between the hard fairness constraint and uncertain data. Instead of directly enforcing the sensitive attribute inclusion constraint with uncertain labels, the method uses a relaxed constraint that allows some violation probability, enabling the model to learn from uncertain data while still progressing toward the target constraint through iterative optimization
Solution Approach 2:
The patent changes the parameter of constraint strictness by introducing a relaxation parameter that transforms the hard constraint problem into a softer optimization problem. This parameter allows the model to handle uncertain sensitive attribute data by adjusting the degree of constraint enforcement, thereby enabling training to proceed while maintaining progress toward accurate constraint satisfaction
2Reliability
If strict sensitive attribute inclusion constraints are enforced, then fairness is improved, but model performance may be compromised when data is uncertain or missing
Solution Approach 1:
The patent applies partial action by enforcing the fairness constraint only to the extent possible given the data quality. When sensitive attribute labels are uncertain or missing, the method partially enforces the constraint by working with the available reliable data while acknowledging limitations, rather than attempting full enforcement that would fail or require imputation
Solution Approach 2:
The patent implements feedback through iterative optimization where the model repeatedly adjusts its parameters based on feedback from the relaxed constraint satisfaction. The optimization process uses gradients from the relaxed objective function to progressively improve constraint satisfaction, allowing the model to learn from uncertain data while gradually achieving better fairness performance
3Quantity of substance
If sensitive attribute labels are added to existing datasets, then more training data becomes available, but additional annotation effort and potential noise are introduced
Solution Approach 1:
The patent treats uncertain or potentially noisy sensitive attribute labels as disposable data that can be used without requiring high confidence. The relaxed constraint formulation allows the model to learn from these potentially imperfect labels, discarding the need for perfect annotations while still making progress toward accurate constraint satisfaction
Solution Approach 2:
The patent performs preliminary action by using the relaxed constraint to guide initial model training on uncertain data before refining to the target constraint. This preliminary training phase allows the model to learn patterns from available data even with potential label noise, building a foundation that can be optimized further without requiring perfect labels from the start
Data Source
AI summary
A computer-implemented method and system for generating a predictive model include a computation engine learning a predictive model having missing or crippled data. A processor applies a formulated problem of missing or crippled data based learning to the predictive model. The computation engine reduces one or more tasks associated with the predictive model to a quadratically constrained quadratic problem (QCQP). The computation engine characterizes one or more solutions associated with the QCQP, where each of the one or more solutions is a loss function value.