Can eyes tell the truth? Can the analysis of human eye-movement data reveal psychological activities and uncover hidden information? Lying is a prevalent phenomenon in human society, but research has shown that people's accuracy in identifying deceptive behavior is not significantly higher than chance-level probability. In this paper, simulated crime experiments were carried out to extract the eye-movement features of 83 participants while viewing crime-related pictures using an eye tracker, and the importance of eye-movement features through interpretable machine learning was analyzed. In the experiment, the participants were independently selected into three groups: innocent group, informed group, and crime group. In the test, the eye tracker was used to extract a total of five categories of eye-movement indexes within the area of interest (AOI), including the fixation time, fixation count, pupil diameter, saccade frequency, and blink frequency, and the differences in these indexes were analyzed. Building upon interpretable learning algorithms, further investigation was conducted to assess the contribution of these metrics. As a result, the RF-RFE suspect identification model was constructed, achieving a maximum accuracy rate of 91.7%. The experimental results further support the feasibility of utilizing eye-movement features to reveal inner psychological activities.
Keywords: eye-movement features; identity recognition; random forest; recursive elimination; simulated crime experiment.