Abstract: Eye-tracking has long been a classic and popular research method in psychological studies. Traditional analysis of eye-movement data mainly focuses on the spatial distribution and the duration of the eye fixations. In the current review, we use eye movement in face processing as an example to introduce the new methods of data analysis that have been developed in recent years.
In the first part of the review, we briefly introduce the traditional methods of data analysis and discuss their limitations. The main traditional approach is to gather the fixations during face processing and to plot the fixation distribution in the form of a heatmap to show the critical facial regions for information processing. However, in the spatial dimension, the boundaries of the regions of interest (ROI) are often poorly defined, limiting the power to obtain highly quantitative results and to make conclusive statistical inferences; in the temporal domain, the dependencies between the sequential fixations are often not quantified.
In the second and major part of the review, we discuss how the application of machine learning and computational modeling to the analysis of eye movement data can advance the understanding of the cognitive mechanism of visual processing. Based on recently published work, we introduce three new methods for eye movement data analysis in face processing. We elucidate the technical implementation and the open-source toolkits supporting these methods. We also cover the scientific questions and the statistical inferences related to these methods. The first method concerns how to combine machine-learning approaches to quantify the clustering of fixations and to define accurate boundaries of the face ROIs. In contrast to the intuitive fixation densities shown by the traditional heatmaps, the machine-learning approaches render quantitatively separated fixation clusters and the specific landmarks of the face ROIs. The second method is to take into account the multidimensional features of the eye movement data to reveal structural patterns of visual processing. The model trained with the multivariate eye movement features can be used to recognize and predict specific patterns of eye movements. Importantly, the model has the advantage of recognizing and predicting the pattern of microsaccades, which is not covered by the traditional method. The representational similarity analysis can provide quantitative distinctions between the different patterns of eye movement data. The third method concerns the modeling of the dependencies between the fixation sequences. The hidden Markov model quantifies the transitional probabilities between the fixation clusters to show the statistical dependencies between the fixation sequences. Based on the model, individual-level eye movement strategies can be distinguished, and the order of the eye movement pattern can be quantified in terms of entropy. The most recent model developed by the recent artificial intelligence (AI) technique is to use the top-notch Transformer model to train and predict the fixation sequences.
In the third part of the review, we summarize how the advances in eye movement data analysis benefit both basic research and clinical applications. Specifically, we highlighted that the traditional methods, which are largely theory-driven, and the newly developed methods, which are largely data-driven, should not be treated as exclusive. Instead, the two types of methods are mutually complementary to advance the understanding of face processing, and the good practice of combining the two kinds of methods will benefit future studies. Although the current work focused on face processing, the introduced methods reflect how visual information is obtained and processed in general. The basic principle can be generalized and the methods can be applied to other areas such as memory, text reading, and the identification of various mental disorders. In combination with the booming AI techniques, the ongoing and further development of eye movement data analysis would advance these investigations. Although the current work focuses on the relation between face images and eye movements, the methods can also help to understand the neural mechanism of face processing by modeling the function between the eye movements and the neural activities during face processing. In summary, the current work provides new perspectives and methodological foundations for both basic research and applications of eye tracking.

Key words: eye tracking, machine learning, spatiotemporal characteristics, face processing.

摘要： 眼动追踪是面孔加工的经典和热门研究方法。传统的数据分析方法多聚焦于注视点在面孔的空间分布和持续时间,近年来多个研究结合机器学习和计算建模开发了一系列新的数据分析方法：采用机器学习对注视点聚类和精准界定感兴趣区,提高了空间解析度和统计推论的明确性;基于多变量模式分析和表征相似性分析量化眼动模式的空间结构性;采用隐马尔科夫模型和结合新兴人工智能模型架构,建立眼动数据的时空序列量化信息采样与整合等。这些方法在空间和时间两个维度建立高度量化的指标,推动面孔认知机制的实证研究和理论进步。论文介绍眼动数据分析新方法所回答的科学问题、基本原理以及所涉及的统计推断知识,为眼动研究提供新视角和方法论依据。

关键词: 眼动追踪, 机器学习, 时空特性, 面孔加工