Research On Learning Status And Performance Prediction Method Based On Deep Learnin

To a certain extent,the learning behavior data generated by students’ classroom learning or participation in learning activities can reflect their recent learning status.By analyzing the potential information of behavioral data with the help of data mining technology,teaching managers can quickly grasp students’ individual learning status and future performance trends,provide more directional and personalized teaching interventions for learners,urge and cultivate students to develop good learning habits,and achieve the purpose of improving teaching quality.Therefore,in this thesis,we use students’ learning behavior data to predict learning status and performance respectively.(1)Firstly,this thesis combines classroom learning behavior data into three typical classroom performances,transforms learning status prediction into classroom performance prediction task,and then explores the classroom performance category definition using clustering algorithm to complete the dataset production.Secondly,based on the time-series characteristics of the dataset,while the traditional attention layer parameter calculation method is easy to fall into local optimum,MYA-LSTM classroom performance classification prediction model is proposed.By incorporating the attention mechanism into the LSTM neural network model and optimizing the attention layer parameters using the MFO algorithm,the model is able to more accurately identify the feature information that is helpful for the classroom performance classification prediction task.In addition,this thesis further explores the impact of improving the MFO algorithm on the performance of the classification prediction model.Finally,the experimental results show that the improvement of the MFO algorithm can enhance its optimization-seeking ability,which in turn can quickly find the appropriate attention layer parameters in the MYA-LSTM model and improve the training efficiency and accuracy of the model.(2)Considering the impact of high feature dimensionality of the achievement prediction dataset on the performance of the achievement prediction model,this thesis explores the performance of the combination of cascade and parallel approaches between three feature selection algorithms,namely,principal component analysis,genetic algorithm and XGboost,on SVM.Finally,it is concluded that the combination of different types of feature selection algorithms can obtain feature subsets with fewer redundant features and higher feature information content compared to single feature selection algorithms,among which the combined way of PCA+XGboost(C-PXS),which takes parallel sets in parallel,has the best performance.(3)To address the problem of class imbalance and temporal characteristics of the achievement prediction dataset,as well as the limitations of a single model.In this thesis,we fully consider model diversity,synthesize a few classes using BorderlineSMOTE algorithm,use logistic regression,Light GBM,MYA-LSTM and C-PXS as base classifiers,and SVM as a meta-classifier,and propose a multi-model stacked student performance prediction model(BS-Stacking)based on the integrated learning Stacking method.The experimental results show that the Borderline-SMOTE algorithm can effectively solve the category imbalance problem,while the BS-Stacking grade prediction model proposed in this thesis can more ideally reduce the gap between the prediction effects of each category and improve the accuracy,recall and F1 value of the model.