Оценка «ошибки» классификации нейросети для анализа социально-психологических особенностей учащихся группы риска по наркотизации

Abstract

Background. Presented study shows the potential of using nonlinear analysis algorithms as an opportunity for a comprehensive study of the socio-psychological characteristics of students who are and are not assigned to the risk groups for drug addiction when processing the large data sets. Objectives. The study had its purpose to compare linear and nonlinear algorithms for assigning students to the risk group for narcotic drug use and to determine by longitudinal comparative analysis which of the algorithms has better prognostic potential. Study Participants. The study involved 27,790 high school students (Mage = 16.2; SD = 2.03), from this sample, 11,786 students participated in the longitudinal study in 2020 and 2021. Methods. Organizational method: comprehensive. Empirical method: psychodiagnostic — socio-psychological testing (SPT). The study was carried out using the Unified Methodology of Socio-Psychological Testing (EM SPT-19). Data processing methods: quantitative analysis — neural network method, comparative analysis, contingency analysis, discriminant analysis. Interpretive method: structural. Software: “Statistica” statistical analysis software package, “Loginom” analytical low-code platform. Results. The study allowed to select errors of positive and negative attribution to the risk group made by using the neural network model. The analysis revealed that erroneous assignment to the risk group is associated with homogeneity of the matched groups on all scales of risk factors, while such homogeneity is not identified by the protection factors. Comparative analysis of data from students positively and negatively assigned to the risk group found significant differences only in anxiety and frustration scales. It is interesting fact that anxiety and frustration also turn out to be the scales that, according to discriminant analysis, do not have a distinguishing ability when predicting the assignment of a student to a risk group. Conclusions. Classification implemented by means of the neural network model has an advantage over the linear algorithm, associated with consideration of intra-scale relationships and greater stability over time in the identified risk/norm groups. The correctness of the neural network solution is confirmed by the results of discriminant analysis. The study demonstrated that social desirability of answers, which is the most common reason for recognizing respondents' answers as unreliable, does not play a significant role in classification using a neural network. The problematized categories of “anxiety” and “frustration” require further analysis concerning their role in forming a sample of students at risk.