According to the differences between Yin and Yang, the five-pattern personality theory of TCM divides individuals into five traits: TYa, SYa, Yy, SYi, and TYi. Table 1 briefly introduces the characteristics of different personalities (Yang and Xue, 2006). To make the model more universal, the participants in this study were individuals without psychological or personality problems. Based on the results of previous studies and the distribution of participants, we divide the true values obtained in the personality inventory into three levels: low, medium, and high. We then discretized these scores to be used as category labels of eye movement signals. This enables us to express the task of personality prediction as a classification problem and use discrete class labels for training and prediction.

The inventory used in this study is a fully researched and verified tool, which has been used in a series of personality testing tasks in personality research (Li L. et al., 2015). This study was approved by the Ethics Committee of Beijing University of Chinese Medicine (2022BZYLL05011). All participants agreed to the experiment and provided signed informed consent.

The eye movement behavior of the participants was recorded using a Tobii 4C eye tracker, and the stimulation was presented by a standard 23.8-inch LCD with a screen resolution of 1,920 × 1,080 pixels. The participants sat in a chair with their heads on a fixed chin rest, keeping their eyes about 70 cm away from the monitor. In addition, to ensure that the eye tracker could accurately track the participants’ fixation points, all participants underwent a standard 9-point calibration and verification test before the experiment officially started. The data were collected in a controlled laboratory environment with fixed illumination and room temperature, and each participant was collected individually.

In this study, static image emotional stimulation was used to arouse the emotional response of the participants. The eye movement behavior of the participants was recorded when they freely observed each emotional face combination image (Figure 1). The Chinese Facial Affective Picture System (Gong et al., 2011) was used in the experiment. Based on the valence and arousal of different emotions, pictures that scored higher for five emotional faces (anger, happy, calm, sad, and fear) were selected, with a total of 20 pictures, 4 for each emotion, maintaining a 1:1 male-to-female ratio. The images were combined in pairs according to the same gender and different emotions, and 10 emotional face combination images were presented to all participants in the same order. The display time of each combined image was set to 5 s, and a 1 s black cross picture appeared between two images to ensure that the original gaze point of each image started from the center.

The features extracted from the recorded eye movement data are related to eye activity and can be categorized into three types: complex eye movement patterns, eye-movement Shannon entropy and fixation duration of specific area of interest. Complex eye movement patterns describe eye movement information through physiological objective indicators (Holland and Komogortsev, 2011), capturing overall patterns and behavioral characteristics of eye movements. Eye-movement Shannon entropy is a key indicator of gaze dispersion, assessing the complexity and uncertainty in eye movement data from the perspective of information entropy (Wang et al., 2020). Fixation on specific parts of the face, such as the eyes, nose, and mouth, can convey different meanings related to emotions, personality, identity, and communication (Schurgin et al., 2014). Fixation duration in these areas is an important metric for evaluating visual attention allocation, reflecting the time spent focusing on specific regions, which provides insights into an individual’s attention strategy in visual tasks.

Complex eye movement patterns include 14 indicators, which are fixation count, mean fixation duration, mean vectorial saccade amplitude, mean horizontal saccade amplitude, mean vertical saccade amplitude, mean vectorial saccade velocity, mean vectorial saccade peak velocity, velocity waveform indicator, scanpath length, scanpath convex hull area, regions of interest, inflection count, slope of the amplitude-duration relationship, and slope of the main sequence relationship. They can be used to objectively evaluate the biological characteristics of eye movement and their ability to accurately and precisely distinguish between unique individuals. The raw eye movement data were processed by MATLAB R2016a, and finally 140 complex eye movement pattern-related features were included.

Eye-movement Shannon entropy is used to measure the statistical randomness or aggregation of the participants’ eye movements. Spatially dispersed gaze leads to higher entropy values, while spatially tightly focused gaze leads to lower entropy values. MATLAB R2016a was used to draw the duration heatmap for each valid trial of each participant, and then Shannon entropy was calculated on the heatmap, and finally 10 related features were included.

For the combination of emotional face pictures, participants can point their visual attention to a face or one of the three face elements (eyes, nose and mouth), thus showing their visual attention preference. First, the 10 combined images were divided into regions of interest using Adobe Photoshop CC 2019 with a single emotional face picture, the eyes, nose, and mouth portions of the different face pictures. Next, the matplotlib and cv2 libraries in Python 3.9 were used to label the fixation points recorded by self-developed eye movement data software, and to calculate the total fixation duration in different interest areas for each image, resulting in 110 related features.

Traditionally, feature selection algorithms are divided into three main categories: filter, wrapper and embedded methods. Filter methods are independent of specific learning algorithms and rank the importance of features using a particular evaluation function, selecting the optimal feature subset. This method is simple, fast, and computationally efficient, making it very popular in practical applications. Mutual information (MI) measures the degree of information shared between two variables from an information theory perspective (Cover and Thomas, 1991). It is used to select the features with the highest information content for classification tasks. Embedded methods, on the other hand, automatically select the most relevant features during model training based on feature importance. The Least Absolute Shrinkage and Selection Operator (Lasso) is a typical embedded technique introduced by Tibshirani (1996). Lasso reduces the variable set by constructing a first-order penalty function, thereby selecting features with predictive power. This method combines feature selection and model training, making it more efficient and precise during the feature selection process. In this study, we used mutual information and Lasso for feature selection, implemented through the Scikit-Learn library in Python 3.9.

A separate classifier was trained for the prediction of each personality trait, such that we ended up with 5 classifiers. A training data point corresponds to a participant, including the value of the selected feature and the label assigned based on the discrete score of the personality trait. In the realm of machine learning, supervised learning is a crucial paradigm. The core idea behind supervised learning is to train a model using a labeled training dataset, enabling the model to learn the relationship between input features and output labels. This allows the model to accurately predict or classify new, unseen data (Dridi, 2021). Given the significance of supervised learning in machine learning and its excellent predictive and classification capabilities, this study employed five robust supervised algorithms to predict the classification results of the five-pattern personality traits: Decision Tree (DT), K-Nearest Neighbors (KNN), Logistic Regression (LR), Naive Bayes (NB), and Support Vector Machine (SVM). DT is easy to understand and interpret. KNN is suitable for handling small sample data. LR has fewer parameters, making it less prone to overfitting. NB offers high computational efficiency, and SVM excels in high-dimensional spaces. To evaluate the performance of these classifiers, we utilized five-fold cross-validation and adjusted the model parameters to avoid overfitting. The models were then assessed using accuracy, F1 score, and AUC.

A total of 57 people were recruited to participate in the eye movement experiment with the completion of the five-pattern personality inventory. Among them, 4 participants were excluded for scoring less than 5 on the masking scale, 1 was excluded due to problems in completing the inventory, and 1 was excluded due to missing eye movement data. Finally, 51 people were included in the analysis. Table 2 shows the demographic data of the participants and the statistical descriptions of different personality traits.

After performing the chi-square test, it was found that there were no significant differences in gender distribution across the five personality traits (TYa:χ² = 1.097, p = 0.5779; SYa:χ² = 1.048, p = 0.5920; Yy:χ² = 3.527, p = 0.1715; SYi:χ² = 4.656, p = 0.0975; TYi:χ² = 3.398, p = 0.1829). Among the 51 participants included in this study, the minimum age was 19 and the maximum age was 59. The Kruskal-Wallis test also showed no statistically significant differences in age distribution across the five personality traits (TYa: H = 3.189, p = 0.2030; SYa: H = 0.3677, p = 0.8321; Yy: H = 3.341, p = 0.1881; SYi: H = 3.004, p = 0.2227; TYi: H = 1.896, p = 0.3874), indicating that the selected sample has good representativeness.

To evaluate the effect of feature selection method, we first fused the five trait-specific classifiers into one metric by averaging the prediction results over the five personality traits. We also compared it with the baseline setting that uses all 260 features without feature selection. Figure 2 shows the average accuracy and F1 Score of five classifiers and all baselines for each personality trait. As can be seen from the figure, classifiers with feature selection perform significantly better than those without feature selection. This indicates that feature selection is crucial for enhancing the accuracy of personality predictions. Specifically, both mutual information and Lasso methods improve prediction performance.

According to the results in Figure 2, Lasso demonstrated higher accuracy and F1 score across the KNN, LR, SVM, and NB, while the mutual information performed better with the DT. Friedman test revealed that there were statistically significant differences in prediction accuracy between the different feature selection methods (χ² = 8.400, p = 0.0085). Further analysis using Dunn’s multiple comparisons test revealed significant differences in accuracy between the baseline method and the Lasso method (p = 0.0133). Additionally, one-way repeated measures ANOVA showed statistically significant differences in F1 scores between the different feature selection methods (F = 58.43, p < 0.0001). The post-hoc analysis with Bonferroni adjustment further identified significant differences between the baseline method and two of the feature selection methods (MI: p = 0.0003; Lasso: p < 0.0001). Overall, Lasso has indeed shown performance improvements across most classifiers, with the Lasso-LR combination achieving the best results on average across five traits.

Table 3 shows the results of five classifiers using the Lasso feature selection method to predict five-pattern personality traits. LR achieved the best accuracy in the TYa, SYa, SYi, and TYi. For the Yy, NB and SVM outperformed LR. When considering individual traits, we observe that all traits were predicted with accuracy greater than 0.7. Averaging performance across all the classifiers, TYa had the highest accuracy and F1 Score. The statistical analysis, including a one-way repeated measures ANOVA with Geisser-Greenhouse (F = 9.814, p = 0.0058) and the Friedman Test (χ² = 7.200, p = 0.1257), revealed significant differences in prediction accuracy and mean F1 scores between five classifiers. The post hoc analysis with a Bonferroni adjustment revealed that the accuracy of LR is significantly different from NB (DT: p = 0.0501; KNN: p = 1.000; NB: p = 0.0375; SVM: p = 1.000). Among the five traits, the average accuracy of LR was 2.62–13.78% higher than other, significantly surpassing the accuracy of other classifiers. Furthermore, the highest accuracy in Figure 2 was achieved through the Lasso feature selection method. Overall, we conclude that the combination of LR with Lasso feature selection predicts traits more accurately than other combinations, and we will focus on this combination in future experiments.

In further analysis, we will determine which eye movement feature type—complex eye movement patterns, eye-movement Shannon entropy, or fixation duration of specific area of interest—is more informative for predicting personality traits. For this purpose, we divided the data into three groups and used the Lasso-LR model to train and predict different types of eye movement features. As shown in Figure 3, each type of eye movement features has its own strengths and weaknesses in predicting different traits. The average accuracy across all personality traits was 0.6898, 0.4698, and 0.6716, respectively, with average F1 scores of 0.6558, 0.4264, and 0.6385. The complex eye movement patterns outperformed the other two types in predicting TYa, Yy, and TYi, while the fixation duration of specific area of interest was more effective in evaluating SYa and SYi. On average, using complex eye movement patterns and the fixation duration of specific area of interest resulted in a prediction accuracy improvement of 22.00% and 20.18%, respectively. This difference may be due to more informative features are extracted from complex eye movement patterns and fixation duration of specific area of interest. Moreover, the entropy value solely employs Shannon entropy, which accounts for the distribution of gaze points, or the uncertainty of eyes resting on different observation targets, without introducing other metrics such as approximate entropy and fuzzy entropy (Harezlak and Kasprowski, 2020; Lee et al., 2022).

There are statistically significant differences in prediction accuracy and F1 scores among different feature types (Accuracy: F = 23.02, p = 0.0039; F1 Score: F = 34.92, p = 0.0015). The pairwise multiple comparisons revealed that the accuracy of eye-movement Shannon entropy is significantly different from complex eye movement patterns (p = 0.0286) and fixation duration of specific area of interest (p = 0.0111). Regarding F1 scores, statistically significant differences were observed between eye movement Shannon entropy and both complex eye movement patterns (p = 0.0135) and fixation duration of specific area of interest (p = 0.0055). While complex eye movement patterns and fixation duration of specific area of interest each demonstrated different strengths in predicting various personality dimensions, the differences between these two feature types were not statistically significant after Bonferroni correction (Accuracy: p > 0.9999; F1 Score: p > 0.9999). Earlier studies on personality have indicated that different eye movement features reflect various aspects of personality. For example, there is a significant correlation between BAS score and blink response (Taib et al., 2020). Individuals with high Neuroticism score show a significantly prolonged average gaze time and dwell time (Rauthmann et al., 2012), and Openness has a significant impact on eye movement parameters (Matsumoto et al., 2010; Wu et al., 2014). These findings emphasize the importance of carefully selecting eye movement features for predicting personality, as different feature combinations offer a more comprehensive understanding of individuals.