A course hybrid recommender system for limited information scenarios

Abstract

Recommender systems in educational contexts have proven to be effective in identifying learning<br> resources that fit the interests and needs of learners. Their usage has been of special interest in online<br> self-learning scenarios to increase student retention and improve the learning experience. In this article,<br> we present the design of a hybrid course recommendation system for an online learning platform. The<br> proposed hybrid system articulates the recommendation carried out by collaborative and content-based<br> filter strategies. For the collaborative filtering recommender, we address the challenge of recommending<br> meaningful content with limited information from users by using rating estimation strategies from a log<br> system (Google Analytics). Our approach posits strategies to mine logs and generates effective ratings<br> through the counting and temporal analysis of sessions. We evaluate different rating penalty strategies<br> and compare the use of per-user metrics for rating estimation. For the content-based recommender, we<br> compare different text embeddings that range from well-known topic models (LSA and LDA) to more<br> recent multilingual contextual embeddings pre-trained on large-scale unlabelled corpora. The results<br> show that the best model in terms of P@5 was the Collaborative filtering recommendation model with<br> a value of 0:4, i.e., two out of five courses recommended could be of the user’s interest. This result is<br> satisfactory considering that our models were trained from ratings inferred from implicit user data. The<br> content-based strategies did not yield significant results, however, these strategies help to mitigate the<br> cold start problem and validate the use of a combined hybrid strategy.