RESUMO
Introduction: In recent years, research has used psycholinguistic features in public discourse, networking behaviors on social media and profile information to train models for depression detection. However, the most widely adopted approach for the extraction of psycholinguistic features is to use the Linguistic Inquiry Word Count (LIWC) dictionary and various affective lexicons. Other features related to cultural factors and suicide risk have not been explored. Moreover, the use of social networking behavioral features and profile features would limit the generalizability of the model. Therefore, our study aimed at building a prediction model of depression for text-only social media data through a wider range of possible linguistic features related to depression, and illuminate the relationship between linguistic expression and depression. Methods: We collected 789 users' depression scores as well as their past posts on Weibo, and extracted a total of 117 lexical features via Simplified Chinese Linguistic Inquiry Word Count, Chinese Suicide Dictionary, Chinese Version of Moral Foundations Dictionary, Chinese Version of Moral Motivation Dictionary, and Chinese Individualism/Collectivism Dictionary. Results: Results showed that all the dictionaries contributed to the prediction. The best performing model occurred with linear regression, with the Pearson correlation coefficient between predicted values and self-reported values was 0.33, the R-squared was 0.10, and the split-half reliability was 0.75. Discussion: This study did not only develop a predictive model applicable to text-only social media data, but also demonstrated the importance taking cultural psychological factors and suicide related expressions into consideration in the calculation of word frequency. Our research provided a more comprehensive understanding of how lexicons related to cultural psychology and suicide risk were associated with depression, and could contribute to the recognition of depression.
RESUMO
As traditional methods such as questionnaires for measuring risk propensity are not applicable in some scenarios, a nonintrusive method that could automatically identify individuals' risk propensity could be valuable. This study utilized Defense of the Ancients 2 (DOTA 2) single match data and historical statistics to train predictive models to identify risk propensity by machine learning methods. Self-reported risk propensity scores from 218 DOTA 2 players were paired with their behavioral metrics. The best-performing model occurred with Gaussian process regression. The root mean square error of this model was 1.10, the correlation between predicted scores and self-reported questionnaire scores was 0.44, the R-squared was 0.17, and the test-retest reliability was 0.67. We discussed how selected behavioral features could contribute to predicting risk propensity and how the approach could be of potential value in the application of perceiving individuals' risk propensities. Moreover, the limitations of our study were discussed, and recommendations were made for future studies in this field.
