A Hybrid Recommender System for E-Learning Platforms

Abstract

E-learning has been a common option for students during epidemic scenarios. Modern research relies heavily on software systems because they help scientists solve challeng ing problems, analyse enormous volumes of data, and organise their work processes. In this research, we describe a novel software system called GrepBing. GrepBing is a knowledge repository that strives to offer conceptual clarity in Science, Technology, Engineering and Mathematics (STEM)-related courses. This application’s goal is to give STEM students access to a single site where they can find all the verified, organ ised information about their courses. One of its many crucial product features is the recommender system. The problem of information overload results from the growth in data size over the last ten years. Because they are designed to provide learners with individualised recommendations, recommender systems are commonly employed on the internet to address the issue of information overload. National Language Processing (NLP), Machine Intelligence (MI), and Deep learning-based techniques can all be used to build recommendation methods. We must create an e-learning recommender system to determine the best model to use for recommendations and the best feature to have the best recommendations on in order to deliver learners relevant content at the correct moment and to increase the quality of recommendations. This study suggests a recom mender system for e-Learning courses that is based on content filtering, in which models like the Count Vectorizer, Tf-Idf Vectorizer and KNN are employed, Collaborative fil tering, whuch employs models like KNN, SVD and NCF and Hybrid Recommender sys tem. During the experiment and analysis for content-based, three separate E-learning datasets are used. We build our cutomized dataset for courses through scraping and for users through pseudo labelling using Auxiliary dataset and used it for experimental purposes. Python is used to create the suggested system, which helps students choose online courses depending on their interests. Performance indicators like hit rate (HR), fault rate (FR), Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) are used to assess how well the models perform. Experiments for content based indicate that Tf-Idf on the feature "skills" outperforms recommendations on the basis of title and sub_course, with higher HR of 0.923 and lower fault values of 0.07. Column category is introduced in the dataset to compare the results with literature and it outperforms the existing category based recommendation with the accuracy of 93.65%. In collab orative filtering results, NCF outperforms KNN and SVD and Hybrid RS technique outperforms collaborative filtering technique. At the last, proposod methodology was Attention-based word2vec features with KNN for collaborative filtering and to build Refined Hybrid Recommender System (RHRS). We evaluated the proposed Refined Hy brid Recommender System (RHRS) on the proposed dataset as well as with literature. The Hybrid Recommender System surpasses Collaborative Filtering techniques.