Forecasting The Adoption Rate of Students to the E-Learning Platforms Using Multilayer Recurrent Neural Network with Long Short-Term Memory

Abstract

E learning platform has increased learning outcomes of the students in the education environment. Recent advances in technologies have increased potential of e learning system on offering student centric materials through recommendation and forecasting approach to increase their user adoption rate. Despite of several advantages of machine learning and deep learning approaches in e- learning platform, those approaches fails to accurately forecast the adoption rate of the student to their learning different modules due to evolving dynamic user behaviors and perception. To meet above objective, a deep learning model from artificial intelligence has to be incorporated. In this paper, long short term memory mechanism integrated multilayer recurrent neural network has been employed as deep learning model to forecast adoption rate of the students to different modules in the e learning platforms. Initially data is preprocessed using stop word removal, stemming, tokenization and token weighting mechanism. Weighted Tokens of the user feedback in form of vector is applied to recurrent neural network. Recurrent Neural Network processes each weighted token in hidden layer. Hidden layer uses activation function to identify the relationship between the sequences of token and organizes as dependency map in association of the long short term memory model. Long short term memory model uses gating mechanism to store the different state of the hidden layer of RNN in different states as hidden state and forget state. In particular, LSTM model compute the long dependency between the tokenized vectors effectively. Finally those dependency maps is processed in the dense layer of the model using softmax function to predict user opinion or feedback as positive (High Adoption Rate) or negative (Low Adoption Rate) effectively. Experimental analysis of the model is performed using Contextual e learning learner interaction dataset extracted from kaggle repository in the Google colab environment incorporating tensorflow to obtain GPU capabilities. Performance analysis of model using cross fold validation of the test data proves that proposed model attains 97.4% accuracy which found to be better compared to conventional fraud detection approaches.