Multi-Task Generative Adversarial Network for Handling Imbalanced Clinical Data

Multi-Task Generative Adversarial Network for Handling Imbalanced Clinical Data

Mina Rezaei, etc.

Machine Learning for Health (ML4H) Workshop

Intro

大致有两类解决方案：data-level和algorithm-level。

第一个是data-level，基本上是通过re-sampling。SMOTE (Synthetic Minority Over-sampling Technique)。但是会忽略一些重要的sample。其他的方法有patient-wise sampling，或者incremental rectification of mini-batches。

第二个是algorithm-level。比如accuracy loss，Dice coefficient loss，和asymmetric similarity loss，也就是cost-sensitive方法。

这里提出的方案就是用一种新的selective weighted loss，来增加minority class impact和减少majority class。