基于MYA-LSTM的课堂表现预测

Prediction of In-Class Performance Based on MYA-LSTM

摘要: 本文基于长短期记忆(LSTM)模型分析学习过程中生成的行为数据并预测下一阶段的课堂行为表现，使教师能够基于此对学生采取精细的学习干预。考虑到传统的LSTM-Attention模型注意层参数计算方式优化空间不足，导致模型性能低下，本文提出基于改进的飞蛾扑火优化算法寻找注意层参数的长短期记忆(MYA-LSTM)分类预测模型，首先将注意力机制引入到LSTM网络前，其次，针对MFO算法容易陷入局部最优、收敛精度低的缺点，本文提出MYMFO算法，在种群初始化阶段加入混沌策略以及寻优后期引入柯西变异，最后利用MYMFO对注意层参数进行寻优。通过分析12个基准测试函数的仿真结果，MYMFO算法对比MFO算法的寻优精度得到了有效的提升，同时在课堂行为表现的预测实验中，MYA-LSTM对比使用未改进的MFO算法来寻找注意层参数的MA-LSTM模型在“上课积极度”、“课堂参与度”和“知识掌握度”三种课堂行为表现上的F1值分别提升了3.66、3.77和3.14，而MYA-LSTM对比LSTM-Attention在三种课堂行为上的F1值分别提升了4.53、4.46、4.56，充分证明了MYA-LSTM模型的有效性。

Abstract: This study uses the behavior data generated during the Long Short Term Memory (LSTM) model analysis learning process to predict the classroom performance in the next stage, which will enable teachers to take more refined learning interventions for students. Considering that the traditional LSTM-Attention model pays attention to the insufficient optimization space of the calculation method of layer parameters, resulting in the low performance of the model. In this paper, a Long Short Term Memory (MYA-LSTM) classification model based on the improved MFO algorithm to find the parameters of the attention layer is proposed. Firstly, the attention mechanism is introduced before the LSTM network. Secondly, we propose a MYMFO algorithm that adds chaos strategy in the population initialization stage and introduces Cauchy mutation in the later stage of optimization. Finally, MYMFO is used to optimize the parameters of the attention layer. By analyzing the simulation results of 12 benchmark test functions, the optimization accuracy of MYMFO algorithm compared with MFO algorithm has been effectively improved. At the same time, in the prediction experiment of classroom behavior performance, MYA-LSTM compared with the MA-LSTM model, which uses the improved MFO algorithm to find the attention layer parameters, has increased the F1 values of “class enthusiasm”, “class participation” and “knowledge mastery” in three classroom behavior performances by 3.66, 3.77 and 3.14, respec-tively, compared with LSTM-Attention, MYA-LSTM increased the F1 value of three classroom behaviors by 4.53, 4.46 and 4.56 respectively, which fully proves the effectiveness of MYA-LSTM model.