Improving Model Performance via Private Multi-Party Leaning of Hidden Features – 9a.018.UVA

Project Start Date: Aug 1, 2020
Research Areas: Analytics, Visualization
Funding: Member Funded
Project Tags: ,

Project Summary

  • Many domains would benefit from institutions being able to share sensitive material for improved predictive modeling, but parties may be unwilling or unable to do so for legal, ethical, or competitive considerations, among concerns
  • One potential solution is the generation and sharing of synthetic data, which may also contain variables or features that were previously unrecorded by one party or the other – for example two companies or hospitals may not maintain identical patient or customer records, or equipment logs.
  • We propose to use Generative Adversarial Networks (GANs) including a variant known as conditional GANs or cGANs, to investigate the feasibility of transferring synthetic samples of these unrecorded features between parties, and their resultant impact on predictive model performance


Principal Investigator(s)