校园欺凌是青少年个体或群体之间反复性欺压行为，由于全世界范围内遭受校园欺凌的青少年比例越来越高，逐渐进入了大众视野。校园欺凌并不只是欺凌者与受害者互动，其中牵涉到多种角色参与与多种欺凌方式，以往研究中关于的校园欺凌角色划分并不一致，并且关于校园欺凌角色的预测研究较少。
本文主要包含了两个子研究，被试选取我国中部某省会城市的4所中学抽取2544名7年级学生，进行四次每隔半年一次的调查。研究一对四个时间点上的提名数据建立潜类别分析模型，根据AIC、BIC和aBIC陡坡图，对比各类别条件概率图的可解释性，最终确定4个类别为校园欺凌的最佳角色类型，根据4类别在题项上的条件概率，将校园欺凌角色划分为(1)欺凌与被欺凌概率均较高的欺凌/受害者(2)欺凌概率较高而受欺凌概率较低的欺凌者(3)欺凌概率较低而受欺凌概率较高的受害者(4)欺凌与被欺凌概率均较低的旁观者。
研究二使用了LightGBM算法框架对校园欺凌角色进行预测，根据时序特征工程的思想，将不同时间点未测量的变量进行缺失处理。在进行过采样以及调参后得到了最佳模型，其预测准确率为0.83。其中欺凌/受害者、欺凌者、受害者、旁观者预测准确率为0.89、0.76、0.83和0.81。成绩、受欢迎、抑郁、自尊和幸福感成为影响校园欺凌角色类型预测的主要因素，其中成绩、受欢迎和幸福感与被试的受欺凌水平成负相关，自尊与抑郁与被试的受欺凌水平成正相关。成绩、受欢迎、自尊、外貌、抑郁成为影响欺凌/受欺凌预测的主要因素；受欢迎、成绩和道德推脱成为影响欺凌者预测的主要因素；父母心理控制、问题行为、成绩、受欢迎成为影响旁观者预测的主要因素；受欢迎、成绩、抑郁和道德推脱成为影响受害者预测的主要因素。

School bullying is a kind of repeated bullying behavior among individuals or groups of adolescents. Due to the increasing proportion of adolescents suffering from school bullying worldwide, it has gradually entered the public view. Campus bullying is not just the interaction between the bully and the victim, which involves a variety of role participation and a variety of bullying methods. The division of campus bullying roles in previous studies is not stable, and there are few studies on the prediction of campus bullying roles. 
This paper mainly includes two sub-studies. In Study 1, 2544 seventh-grade students were selected from four middle schools in a provincial capital city in central China, and four surveys were conducted every six months. A latent class analysis model was established for the nomination data at four time points. According to AIC, BIC and aBIC steep aBIC, the interpretability of conditional probability graphs of various categories is compared, and four categories are finally determined as the best role types of campus bullying. According to the conditional probability of the four categories on the item, the role of campus bullying is divided into ( 1 ) bullies / bullies with higher probability of bullying and being bullied, ( 2 ) bullies with higher probability of bullying and lower probability of being bullied, ( 3 ) bullies with lower probability of bullying and higher probability of being bullied, ( 4 ) non-participants with lower probability of bullying and being bullied.
Study 2 used the LightGBM algorithm framework to predict the role of bullying on campus. According to the idea of temporal feature engineering, the unmeasured variables at different time points were missing. The best model was obtained after sampling and parameter adjustment, and the prediction accuracy was 0.83. The prediction accuracy of bullies / victims, bullies, victims and non-participants was 0.89, 0.76, 0.83 and 0.81. Achievement, popularity, depression, self-esteem and well-being have become the main factors affecting the prediction of bullying role types on campus. Among them, achievement, popularity and well-being are negatively correlated with the bullying level of participants, while self-esteem and depression are positively correlated with the bullying level of participants. Achievement, popularity, self-esteem, appearance and depression are the main factors affecting bullying / bullying prediction ; popularity, achievement and moral disengagement are the main factors affecting bullies ' predictions ; parents ’ psychological control, problem behaviors, performance and popularity are the main factors affecting the prediction of non-participants ; popularity, achievement, depression and moral disengagement are the main factors affecting the prediction of victims.