基于机器学习构建青少年网络游戏成瘾的预测模型

Building a predictive model for adolescent Internet gaming disorder based on machine learning

目的探索机器学习预测青少年网络游戏成瘾的效果,为制定有效的干预措施提供指导。方法于2023年6—9月,采用分层随机整群抽样方法选取贵州省毕节市、黔西市和金沙县3个地区3所初中和3所高中2100名学生作为研究对象。采用简式网络游戏障碍量表(IGDS9-SF)、父母心理控制与自主支持问卷(PPCASQ)、动机结构问卷、相对剥夺感问卷、越轨同伴交往问卷以及自我控制双系统量表进行数据收集。描述性统计分析确定样本特征,使用χ^(2)检验和Mann-Whitney U检验分析变量的组间差异。以人口学变量和各种影响因素作为自变量,以青少年是否网络游戏成瘾作为因变量,运用随机森林、逻辑回归、支持向量机、梯度提升树、决策树和自适应提升算法多种机器学习算法构建预测模型。结果青少年网络游戏成瘾检出率为4.57%(96名);男生和初中生网络游戏成瘾检出率(5.52%,6.29%)相较女生和高中生(3.32%,3.62%)更高,差异均有统计学意义(χ^(2)值分别为5.71,7.86,P值均<0.01)。网络游戏成瘾组相对剥夺感、越轨同伴交往、父亲心理控制、母亲心理控制、控制动机、冲动系统及其维度(冲动性、易分心、低延迟满足)得分高于非网络游戏成瘾组,而父母自主支持得分低于非网络游戏成瘾组(Z值分别为-2.88,-9.32,-4.13,-4.48,-6.58,-7.50,-7.18,-7.56,-7.43,-2.27,P值均<0.05)。预测模型中,自适应提升算法表现最佳(精确度99%,召回率95%,F1分数97%,AUC值为0.96);其次为随机森林和梯度提升树(精确度均为98%,召回率均为95%,F1分数分别为97%和96%,AUC值均为0.96)。结论相较于其他模型,自适应提升算法对青少年网络游戏成瘾有良好预测效果。应选择适合模型尽早识别存在网络游戏成瘾的个体,制定有效的干预策略,降低青少年网络游戏成瘾风险。

Objective To explore the effectiveness of machine learning in predicting adolescent Internet gaming disorder,so as to provide guidance for formulating effective intervention measures.Methods From June to September,2023,a total of 2100 students from 3 middle schools and 3 high schools in Bijie City,Qianxi City and Jinsha County,Guizhou Province were selected by stratified random cluster sampling as research subjects.Data was collected by using several instruments,including the Nineitem Internet Gaming Disorder Scale-Short From(IGDS9-SF),Parental Psychological Control and Autonomy Support Questionnaire(PPCASQ),Motivation Structure Questionnaire,Relative Deprivation Questionnaire,Deviant Peer Association Questionnaire,and Dual Systems of Selfcontrol Scale.Descriptive statistical analysis was conducted to characterize the sample features,and the distribution differences of categorical variables were analyzed by using Chisquare test and Mann-Whitney U test.Demographic variables and various influencing factors were served as independent variables,and whether adolescents were addicted to Internet gaming was the dependent variable.Various machine learning algorithms,including random forest,Logistic regression,support vector machine,gradient boosting trees,decision trees,and adaptive boosting were employed to construct predictive models.Results The detection rate of Internet gaming disorder among adolescents was 4.57%(96 cases).Males and middle school students had higher Internet gaming disorder detection rates(5.52%,6.29%)than females and high school students(3.32%,3.62%),and the differences were statistically significant(χ^(2)=5.71,7.86,P<0.01).The scores of relative deprivation,deviant peer affiliation,paternal psychological control,maternal psychological control,control motivation,impulsive system and its dimensions(impulsivity,distractibility,low delay of gratification)in Internet gaming disorder group were higher than in nonInternet gaming disorder,while the score of parental autonomy support was lower than that in the nonInternet gaming disorder group(Z=-2.88,-9.32,-4.13,-4.48,-6.58,-7.50,-7.18,-7.56,-7.43,-2.27,P<0.05).The adaptive boosting algorithm performed the best(accuracy=99%,recall=95%,F1 score=97%,AUC=0.96).Random forest and gradient boosting trees also performed excellently(accuracy=98%,recall=95%,F1 score=97%,96%,AUC=0.96).Conclusions Compared to other models,the adaptive boosting algorithm shows a good predictive effectiveness for adolescent Internet gaming disorder.Appropriate models should be selected to identify individuals with Internet gaming disorder as early as possible,to develop effective intervention strategies and reduce the risk of Internet gaming disorder.