Development of analytics tools for E-Learning (A)

Abstract

Nowadays many universities around world have enhanced their educational system with the so-called e-learning system, and a considerable amount of educational data generated from such system every second. Analytics of such educational data could be used as a tool to improve the education quality of academic institutes, which motivated the proposal of this project. Some similar works had been done since early 2000's and had been applied for e-learning platforms partially, nonetheless, there is still potential for improvements and inspirations for new conceptions. This project aims to develop an analytic tool for students to investigate and make inference on their academic performance based on data collected from E-learning platform, and providing with a possible implementation strategy of such analytic system. The system is expected to be progressively responsive, in the sense that prediction results alter with time, which is a hidden but dominant input, proceeds. At the beginning of a predefined academic period, e.g. beginning of one semester, or the start of a four-year undergraduate study, the system would be fed with background information of students, and providing a rough and inaccurate prediction of the student's final grade. As time in the real-world progress, more and more data generated from the e-learning platform should be added into the system to tune the analytic model to produce prediction results of higher accuracy and smaller expectation bias interval. A prototype analytic system model was developed in this project as a demonstration of the possible implementation of the algorithms of analytic models. The prototype system had been trained with real-case student data, with all confidential personal details been converted to symbolic notations, avoiding actual personal information of students being revealed. Several prediction models were built and tested to evaluate their performance, including but not limited to: Support Vector Machine, Decision Tree, Random Forests, k'th Nearest Neighbour. The accuracy of the system to predict student grades was ranging from 30%-75%, as time progresses, for numbered scores; and 40%-85% for a pseudo letter grade. The variance in the accuracy was introduced by the change of amount of information provided as input to the system. Lower accuracy results are yielded with a limited scale of input, typically with only the background information, and as the virtual timeline proceeds, the prediction results converge to an appreciably fine-grained interval.