Researchers Design Model to Reduce AI Bias

A new model from Waterloo researchers is looking for a method to make AI more fair and trustworthy.

One of the main issues with AI is how to ensure that systems are fair, unbiased, and transparent in their decision-making and knowledge organization processes.

Bias in machine learning is a phenomenon that occurs when a model produces results that are systematically skewed or inaccurate due to the assumptions, methods, or data used in the training process. Bias can have negative impacts on the performance, fairness, and reliability of machine learning applications, especially when they affect human lives or decisions. Therefore, it is important to identify, measure, and mitigate bias in machine learning as much as possible.

The Solution

Researchers from the University of Waterloo have developed a new explainable AI (XAI) model that aims to address this problem. The model, called Pattern Discovery and Disentanglement (PDD), can reduce bias and enhance trust and accuracy in machine learning-generated outcomes.

PDD, or Pattern Discovery and Disentanglement, is designed to handle imbalanced groups and anomalies in relational datasets. Here’s how PDD addresses these challenges:

  • Imbalanced groups: In many real-world datasets, the distribution of samples across different groups or classes is not balanced. This can lead to biased models that perform poorly on minority groups. PDD addresses this issue by employing techniques such as oversampling, undersampling, or class weighting to ensure that all groups have sufficient representation during the pattern discovery process. By balancing the groups, PDD can improve the accuracy and fairness of predictions for all classes.
  • Anomalies: Anomalies are observations that deviate significantly from the normal patterns or behaviors in the dataset. They can be noise or outliers that can negatively impact the pattern discovery process. PDD handles anomalies by employing robust statistical methods or outlier detection algorithms to identify and filter out these abnormal data points. By removing or downweighting anomalies, PDD can focus on discovering meaningful patterns that are representative of the majority of the data and avoid being influenced by outliers.

By addressing imbalanced groups and anomalies, PDD aims to ensure that the discovered patterns are more accurate, reliable, and generalizable across the dataset, leading to improved performance and interpretability of the machine learning models built using PDD.

PDD works by untangling complex patterns from data and relating them to specific underlying causes that are unaffected by anomalies and mislabeled instances. This way, PDD can reveal the deep knowledge that is hidden or mixed at the data level due to the entanglement of multiple factors.

The researchers applied PDD to various domains, such as protein binding analysis, medical diagnosis, and image recognition. They showed that PDD can improve the performance and interpretability of machine learning models, as well as provide scientific evidence for the discovered patterns.

The researchers hope that PDD can bridge the gap between AI technology and human understanding and help enable trustworthy and reliable XAI applications in various fields.


Photo of author

Robert Brown

When you ask Robert to describe himself, he always says - what you see is what you get.