Short-Term Gaze Prediction: Analysis of Individual Differences, Typical and Extreme-Case Errors

Gaze prediction is a diverse field of study with multiple research focuses and practical applications. This article investigates how recurrent neural networks and transformers perform short-term gaze prediction. We used three models: a three-layer long-short-term memory (LSTM) network, a simple transformer-encoder model (TF), and a classification-predictor network (ClPr), which simultaneously classifies the signal into eye movement events and predicts the positions of gaze. The performance of the models was evaluated for ocular fixations and saccades of various amplitudes and as a function of individual differences in both typical and extreme cases. On average, LSTM performed better on fixations and saccades, whereas TF and ClPr demonstrated more precise results for post-saccadic periods. In extreme cases, the best-performing models vary depending on the type of eye movement. We reviewed the difference between the median $P_{50}$ and high-percentile $P_{95}$ error profiles across subjects. The subjects for which the models perform the best overall do not necessarily exhibit the lowest $P_{95}$ values, which supports the idea of analyzing extreme cases separately in future work. We explore the trade-offs between the proposed solutions and provide practical insights into model selection for gaze prediction.