一种LSTM优化算法在高校学生学业预警中的应用  

Application of an LSTM optimization algorithm in early warning of college students′ academic performance

作　　者：成浩 欧阳宁[1] 林乐平[1] CHENG Hao;OUYANG Ning;LIN Leping(Guilin University of Electronic Technology,Guilin 541004,China)

机构地区：[1]桂林电子科技大学,广西桂林541004

出　　处：《现代电子技术》2022年第10期142-147,共6页Modern Electronics Technique

基　　金：国家自然科学基金项目(62001133);国家自然科学基金项目(61661017);国家自然科学基金项目(61967005);国家自然科学基金项目(U1501252);广西自然科学基金项目(2017GXNSFBA198212);认知无线电教育部重点实验室资助项目(CRKL150103);桂林电子科技大学研究生创新项目(2019YCXS020)。

摘　　要：高校现有学业预警系统对学生行为数据要求齐整、预测准确率低、可推广性差。针对上述问题,文中提出一种基于LSTM优化神经网络的预警模型。该模型主要由数据处理、特征提取、优化训练三部分组成,其中数据处理是通过RBF核函数对学生信息数据进行从低维到高维的空间映射,从而保留完整的学生数据信息,降低数据缺失对预警系统的干扰;特征提取是指通过多维正态分布的前馈特征提取对学生信息数据进行分类,然后进行归一化处理,根据不同类别进行学习运算,提高预测正确率;优化训练则是将数据输入至自适应激励函数优化后的LSTM神经网络训练,通过概率拟合得到毕业概率及弱项特征,从而输出学业预测报告。将收集的G大学2017—2019年三届毕业生共15211人的第一学年必修课成绩、消费信息、图书借阅频率等数据作为数据训练集,将2020届5301名毕业生数据作为测试数据进行试验。结果表明,文中模型预测准确率稳定在94.21%,最高可达到98.17%,平均准确率较现有预警模型提升2个百分点,尤其是负召回率有明显提升,数据依赖性和结果稳定性也有显著改善。The existing academic early warning system in colleges and universities requires neat student behavior data,low prediction accuracy and poor popularization. On this basis,an early-warning model based on LSTM(long short term memory) optimized neural network is proposed. The model is mainly composed of the three parts:data processing,feature extraction and optimization training. For the data processing,the RBF kernel function is used to map the student information data from low dimension to high dimension,so as to retain the complete student data information and reduce the interference of data loss to the early warning system. For the feature extraction,the feedforward feature extraction of multi-dimensional normal distribution is used to classify and normalize student information data,and perform learning operations according to different categories to improve the prediction accuracy. For the optimization training,data is inputted to the LSTM neural network optimized by the adaptive excitation function to training,the graduation probability and weakness characteristics are obtained by means of probability fitting,so as to output the academic prediction report. The collected data of 15 211 graduates of G university from 2017 to 2019,including the results of compulsory course scores in the first academic year,consumption information and book borrowing frequency,are used as the data training set,and the data of 5 301 graduates of 2020 are used as the test data. The results show that the prediction accuracy of the model in this paper is stable at 94.21%,up to 98.17%. The average accuracy is two percentage points higher than the existing early warning model,especially the negative recall rate is significantly improved,and the data dependence and result stability are also significantly improved.

关 键 词：学业预警 LSTM优化算法 预警建模 特征提取 数据分类 优化训练 学业预测 

分 类 号：TN919-34[电子电信—通信与信息系统]