Eye Movement Classification and Prediction via Attention-Based Distribution Distance

Eye-tracking classification often relies on the training of deep neural networks. Despite good performance, the interpretability of the hidden output of the classification networks is unclear, making the representation learning and regularization challenging. In this study, we propose a simple but effective eye movement classification model to identify children with autism spectrum disorder (ASD) based on Attention-Based Distribution Distance (ABDD). We first explore a Reference Attention-Based Distribution in each image from the actual eye movement data. Then, we utilize the difference in eye movement distribution between the ASD group and the typically developing (TD) group to assign higher weights for the images. Finally, we estimate the classification threshold by fitting a Weibull distribution and achieve subject-wise classification. In eye movement classification tasks, we can take the statistic as a teacher model to pre-train neural networks as a student model, enabling them to converge quickly. Furthermore, we know that the hidden layer output of the neural network is related to the attention distribution on the image. The experiment confirmed that the proposed method can achieve results comparable to deep models and assist in pretraining of Eye-tracking classification tasks.