Predicting Gender via Eye Movements

TL;DR

This study demonstrates that eye movement data can predict gender with about 65% accuracy, showing stability across multiple experiments and confirming previous biases related to eye gaze direction.

Contribution

First stable results on gender prediction using eye movements with detailed statistical validation and identification of effective classifiers.

Findings

Random Forests and Logistic Regression perform best

Gender bias towards left eye observed

Prediction accuracy around 65% with low error

Abstract

In this paper, we report the first stable results on gender prediction via eye movements. We use a dataset with images of faces as stimuli and with a large number of 370 participants. Stability has two meanings for us: first that we are able to estimate the standard deviation (SD) of a single prediction experiment (it is around 4.1 %); this is achieved by varying the number of participants. And second, we are able to provide a mean accuracy with a very low standard error (SEM): our accuracy is 65.2 %, and the SEM is 0.80 %; this is achieved through many runs of randomly selecting training and test sets for the prediction. Our study shows that two particular classifiers achieve the best accuracies: Random Forests and Logistic Regression. Our results reconfirm previous findings that females are more biased towards the left eyes of the stimuli.