Demographic Prediction Models for Hidden Bias Detection in Datasets
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Datasets representing network interactions often lack demographic data, limiting the analysis of hidden bias, which can lead to unfair treatment of certain demographics in business interactions.
Innovation Solution
A system and method that formats network interaction data into aggregate datasets, joins it with demographic data, trains models to predict demographics, and confirms the presence or absence of bias, allowing for fair treatment without combining demographic data with interaction data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If demographic data is excluded from interaction datasets, then user privacy is protected and data security is improved, but the ability to detect and analyze hidden bias in datasets deteriorates
Solution Approach 1:
The patent introduces an intermediary approach by training a demographic prediction model separately on labeled data, then using this trained model to predict demographics on unlabeled interaction data. This mediator model enables bias detection without directly exposing or storing sensitive demographic information in the interaction datasets, thus maintaining data security while enabling bias analysis.
Solution Approach 2:
The patent segments the bias detection process into two distinct phases: (1) training phase where demographic data is used to train a prediction model, and (2) detection phase where the trained model predicts demographics on interaction data without requiring demographic labels. This segmentation allows demographic data to be used for model training while being excluded from the actual interaction datasets, resolving the contradiction between security and detection capability.
2Measurement precision
If demographic data is appended to interaction datasets, then bias analysis capability is improved, but data privacy protection and security measures deteriorate
Solution Approach 1:
The patent applies preliminary action by pre-training a demographic prediction model on labeled data before the actual bias detection task. This pre-trained model captures demographic patterns from interaction data without requiring demographic information to be stored or appended to the interaction datasets during normal operations, thus maintaining security while enabling precise bias analysis.
Solution Approach 2:
The patent creates a copy of demographic information through the trained prediction model. Instead of storing actual demographic data in interaction datasets, the system creates a predictive representation that captures demographic characteristics. This copied information enables bias analysis while the original sensitive demographic data remains excluded from the interaction datasets, preserving data security.
3Difficulty of detecting and measuring
If models are trained to predict demographics from interaction data, then hidden bias can be identified, but system complexity increases
Solution Approach 1:
The patent makes the demographic prediction model universal by training it to predict multiple demographic attributes (e.g., gender, race, age) from the same interaction data features. This multi-functional model can be applied to detect various types of bias across different demographic dimensions without requiring separate models for each attribute, thus managing system complexity while enabling comprehensive bias detection.
Data Source
AI summary
Disclosed are example embodiments of systems and methods for use in identify content included in datasets, independent of certain data being included in the datasets. In an example embodiment, a computer-implemented method generally includes accessing interaction data as a dataset, where the interaction data is representative of multiple network interactions and includes a first variable and a second variable, and appending demographic data to the dataset. The method also includes applying an exponential decay function, based on multiple constants, to the first variable of the dataset, where each of the constants is indicative of a different defined interval, and encoding the second variable of the dataset into multiple columns in the dataset, where each of the multiple columns includes a binary value. The method then includes training a classifier model based on the dataset, where the demographic data defines classification of the interaction data.