Reducing Gender and Ethnicity Bias In Personality Trait Prediction

Automatic personality trait prediction is the prediction of a person's personality from cues like facial expressions, body language, manner of speaking, etc. There has been a lot of research on implementing machine learning/deep learning models that can infer a person's personality traits, however studies have shown such models to be biased towards certain ethnicities and genders. Our work proposes the use of Adversarial Debiasing to reduce bias in such models. We apply this technique to train a CNN and an LSTM to predict personality traits from images and audio features extracted from the ChaLearn LAP APA 2016 dataset [1]. This method achieves an accuracy between 0.894-0.913 for different traits and reduces bias, as measured by Disparate Impact Ratio.