A REAL-TIME PERSONALISED RECOMMENDER SYSTEM FRAMEWORK FOR ONLINE LEARNING PLATFORMS

Long-tail concerns affect traditional recommender systems. They often recommend identical things, limiting the options available to users. Conventional recommender systems also suffer from lack of real-timeliness. In this work, a recommender system framework for online learning platform is proposed using deep reinforcement learning algorithm. The agent takes action by recommending learning materials to the learners based on the interactions of the recommender agent with the learner. Positive reinforcement (positive reward such as likes, longer dwell time, clicks, etc.) and negative reinforcement (punishment such as dislikes, less dwell time, skips, etc.) are used to teach the recommender agent what to recommend. This enables the agent to iteratively refine its policy via interactivities with the environment, using trial-and-error methods, until the model conforms to an ideal policy that produces suggestions that are most suitable for the users’ dynamic preferences. The outcomes of the deep reinforcement learning agent were benchmarked against the performance of a random agent using evaluation metrics such as average episode reward, click through rate, average quality of recommendation and standard deviation of episode reward. The study shows that the average episode reward, click through rate, average quality of recommendation for the DRL agent increased by 2.72, 1.5 and 16.20 percent respectively, while the standard deviation of episode reward for the DRL agent reduced by 20.61 percent. All these are positive indicators of the better performance of the DRL agent