Prediction Sensitivity

Prediction sensitivity assesses how much a model's output changes based on variations in input features, particularly those related to protected attributes like race or gender. Current research focuses on developing metrics to quantify this sensitivity, linking it to established fairness notions, and applying it to diverse domains such as drug efficacy prediction and text classification, often employing machine learning models like variational autoencoders. This work is crucial for ensuring fairness and accountability in AI systems, enabling continuous monitoring and detection of bias in deployed models across various applications.

Papers

June 7, 2022

A generative recommender system with GMM prior for cancer drug generation and sensitivity prediction
Krzysztof Koras, Marcin Możejko, Paulina Szymczak, Eike Staub, Ewa Szczurek
Variational Autoencoder Latent Representation Gaussian Mixture Model Convolutional Autoencoder High Throughput Variational AutoEncoder Drug Candidate Generative Recommendation Prediction Sensitivity

March 16, 2022

Measuring Fairness of Text Classifiers via Prediction Sensitivity
Satyapriya Krishna, Rahul Gupta, Apurv Verma, Jwala Dhamala, Yada Pruksachatkun, Kai-Wei Chang
Group Fairness Text Classifier Counterfactual Fairness Prediction Sensitivity

February 9, 2022

Prediction Sensitivity: Continual Audit of Counterfactual Fairness in Deployed Classifiers
Krystal Maughan, Ivoline C. Ngong, Joseph P. Near
Human Prediction Counterfactual Fairness Prediction Sensitivity

Prediction Sensitivity

Papers

A generative recommender system with GMM prior for cancer drug generation and sensitivity prediction

Measuring Fairness of Text Classifiers via Prediction Sensitivity

Prediction Sensitivity: Continual Audit of Counterfactual Fairness in Deployed Classifiers