Research On Learning Performance Prediction In Online Learning Environment

In the learning environment that is emerged based on the integration of information and education technologies,it is critical to understand the learning rules of online learners and predict their learning performance,which is of great significance to improve the learning effect efficiency,and experience.This study first explores online learners' spatial-temporal behavioral patterns and its relationship with learning performance;explores related factors of learning cognitive participation in online learning,and constructs an evaluation model for automatically identifying learning cognitive participation;then,proposes an image processing method for learning process sequence data and a learning performance prediction model based on deep learning algorithm;furthermore,in order to improve the prediction outcomes of learning performance prediction model,especially for at-risk students,a two-level enhanced integration algorithm based on the stacking strategy is developed to solve the problem of poor prediction of single model,and an integrated prediction framework based on the latent variational autoencoder(LVAE)is proposed to mitigate the impact of imbalanced educational dataset on learning performance prediction.The main achievements of this study are as follows:Firstly,in order to depict the spatial-temporal distributions of online learning behaviors,this paper proposes an analysis method based on learning time entropy and learning location entropy to explore the spatial-temporal distribution characteristics of different types of online learners,and to reveal the relationship between different spatial-temporal behavioral patterns and learning performance.Experimental results confirm that entropy analysis is an appropriate method to study the spatial-temporal behavioral distributions of online learning behaviors,and it is found that frequent changes of learning locations have a negative impact on learning performance.Secondly,in order to understand the depth of online learners' cognitive participation,this paper tries to understand online learners' cognitive participation depth and rules based on online discussions.Considering that the cognitive participation depth is not only related to the semantics of the discussion contents,but also related to the way of thinking and expression,this paper proposes an automatic evaluation algorithm based on multi,feature fusion for automatically evaluating online learners'cognitive participation depth and understanding their thinking process.Experimental results show that due to the combination of multiple features,the proposed method can effectively identify the cognitive participation depth that was involved in the discussion posts,and different types of textual features have different evaluation advantages.Furthermore,considering that the online learning process data contains rich information such as online learners' learning rule and state differences in the learning process,this paper proposes an image processing method for learning process sequence data and a learning performance prediction model based on deep learning algorithm.Specificly,it transforms online learners' learning process sequence data into single channel or multi-channel learning images,and then identifies the differences between learning rules and states based on convolutional neural network(CNN)architecture,and then predicts the learning performance.Experimental results show that the proposed prediction method can accurately identify different sub-types of at-risk students,achieve good prediction outcomes,and provide teachers with intuitive and accurate information about the problems in the learning process in order to support the design of personalized interventions.Finally,in order to further improve models' prediction performance,especially for at-risk students,this study proposes novel algorithms and models from the algorithm level and data level based on the consideration of imbalanced education dataset,for the purpose of improving and optimizing prediction performance,which includes(1)a two-level enhanced integration algorithm based on the stacking strategy,that is,the secondary model combines the prediction advantages of different models in the primary stage to achieve the final prediction of at-risk students.Experimental results show that the proposed two-level enhanced integration prediction model has high prediction accuracy,high recall rate,and low false positive recognition rate;(2)an integrated prediction framework based on the latent variational autoencoder(LVAEPre),that is,LVAE generates samples that are similar to real at-risk students by learning the potential characteristic distributions of at-risk students in order to improve the imbalanced distributions of training data,and then the final prediction is performed by deep neural network.Experimental results show that LVAE outperforms the traditional sampling methods in the educational environment,and LVAEPre achieves high prediction accuracy,high recall rate,as well as good robustness.