Analysis of social data for accuracy improvement of collaborative filtering in MOOCs using text mining and deep learning techniques

Analysis of social data for accuracy improvement of collaborative filtering in MOOCs using text mining and deep learning techniques

연구 분야: Databases

학회: Discover Computing

The accuracy and scalability of educational recommender systems are important issues that have been favorably investigated in previous research. However, there is room for improvement in these systems in terms of their prediction accuracy using social data. The accuracy of educational recommender systems is often hindered by the limitations of single-rating approaches, which fail to capture the complex nature of learner preferences. In contrast, multi-criteria recommendation systems can offer a more comprehensive understanding by evaluating multiple aspects of educational resources. Nevertheless, the use of a multi-criteria recommendation approach with the aid of learners’ online reviews is fairly unexplored. This research accordingly puts forward a new approach for educational recommender systems. We rely on a multi-criteria recommendation approach using text mining, clustering with ensemble learning, and deep learning techniques. A Deep Belief Network (DBN) technique is used for the prediction task in the proposed method. Latent Dirichlet Allocation (LDA) is used to construct a multi-criteria dataset from the learners’ online reviews. We also use Self Organizing Map (SOM) clustering to discover similar learners’ preferences in distinct groups. The effectiveness of the proposed method is evaluated using a dataset collected from the Udemy platform which is a comprehensive Massive Open Online Course (MOOC) system. The proposed method is compared with the traditional Collaborative Filtering (CF) recommendation systems for its efficiency in recommending educational resources. The results showed that the method which used LDA, DBN and SOM techniques provides the best recommendation performance (Precision = 0.9434 and F1 = 0.9189) compared with the other methods.