Optimising student internet navigation: a comparative analysis of machine learning algorithms for action prediction

Web-based learning has been promoted in education and students are required to retrieve online information to complete their assignments and study for exams. Research shows that challenges exist during information retrieval, especially with novice students. In this research, we aim to lessen these challenges by introducing a collaborative framework that gathers students’ searched keyphrases and analyses trends to predict the most effective subsequent keyphrase to search. The proposed solution encourages students to contribute by sharing their information retrieval trends while collectively benefiting from each other’s searching strategies. In addition, novice students will enrich their domain knowledge since the prediction results contain keyphrases searched by students from previous cohorts. Next-word prediction is a well-known area of NLP that is used to forecast the next word given a sentence or predict trends based on time-series data. Word suggestions are popular in mobile devices and studies show that users rely on them while they are typing. The methodology involves the implementation of a framework designed to collect online browsing activity. Undergraduate students studying a BSc in Computer Science were engaged to participate in an experiment wherein they installed a Google Chrome extension capable of collecting data and predicting suitable content related to the researched domain. The collected data consisted of URLs containing keyphrases that students searched during their studies. A feature engineering process was performed to analyse and transform the data into a time-series sequence of actions and to ensure that it is fit for the intended purpose. A grid-search method was employed on various machine learning models to identify the most effective hyper-parameters that can predict the next best keyphrase. The results obtained during an in-class test show that students relying on the predictions generated by the machine learning models outperformed those who depended