Tree-Based Inference Models Latent Bias Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Latent bias in tree-based inference models can lead to undesirable impacts on computer-implemented services due to unintended correlations in training data, resulting in discriminatory outcomes.
Innovation Solution
A training procedure that utilizes a splitting rule to incentivize predictive power for desired labels while disincentivizing predictive power for bias features, thereby reducing the likelihood of latent bias in tree-based models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional tree-based inference models are trained using standard training procedures, then the models achieve good predictive performance, but latent bias is introduced into the models due to unintended correlations in training data
Solution Approach 1:
The patent modifies the training objective parameters by introducing a bias penalty term to the loss function. This changes the optimization criteria from purely maximizing predictive accuracy to simultaneously penalizing bias in predictions based on protected attributes, thereby resolving the contradiction between predictive performance and bias reduction
Solution Approach 2:
The patent implements a feedback mechanism where the model's predictions are evaluated not only for accuracy but also for bias relative to protected attributes. This feedback loop allows the model to learn from its biased predictions and adjust its parameters to reduce latent bias while maintaining predictive performance
2Object-affected harmful factors
If the training procedure is modified to reduce latent bias, then the model becomes more equitable, but the complexity of the training procedure increases
Solution Approach 1:
The patent merges the bias reduction objective with the existing training framework by incorporating the bias penalty term into the standard loss function. This integration allows bias reduction to be achieved through a unified training process rather than requiring separate complex procedures, thereby limiting the increase in training complexity
Solution Approach 2:
By changing the loss function parameters to include bias penalty terms, the patent simplifies the training procedure modification. Instead of requiring completely new training algorithms, the approach uses parameter adjustments within the existing framework, thereby limiting the complexity increase while achieving bias reduction
Data Source
AI summary
Methods, systems, and devices for providing computer implemented services are disclosed. To provide the computer implemented services, inference models may generate and provide inferences used in the computer implemented services. The inference models may be obtained through training using training data. Training processes used to train the inference models may proactively attempt to reduce the likelihood of the trained inference models exhibiting latent bias. The training process may disincentivize predictive power with respect to bias features and incentivize predictive power for labels.