Learning Analytics Integrating Student Attendance Data

Electronic versions


    Research areas

  • Leaning analytics, Student Attendance, Student Engagement, Student Retention, Data Science, Machine Learning, PhD, School of Computer Science and Electronic Engineering


With UK Higher Education (HE) ever more scrutinised with the Teaching Excellence Framework, changes to finance and media attention student welfare and retention is becoming ever more important. Institutions are turning to technological methods to assist wherever possible. Use of Learning Analytics (LA) systems, as a result, is booming. Institutions are deploying these systems to exploit the wealth of information that they hold on their students. LA systems have historically focused on interactions between a student and their course in the form of results or participation within a Virtual Learning Environment. However, their focus tends to be on analysing outcomes based on performance, with other factors occasionally mixed in. If early identification is to the be goal of an analytics system, new sources of data must be added to the models. These new data sources will need to be timely and robust. Almost all institutions, as they have a visa monitoring requirement, monitor student attendance at timetabled and other events. This attendance data is rarely included in LA models.

This thesis shows how a simple synthetic metric, the Bangor Engagement Metric (BEM), can be used in conjunction with standard Data Science and ML techniques can produce a powerful predictive model as early as Week 4 of Semester 1 (\textasciitilde end of October each year). It proves, through a series of experiments, that a single algorithm is suited to this unbalanced class classification problem. The model is then proven, against both past and future unseen data, to produce prediction accuracies in excess of 93\% at Bangor University - reaching as high as 97.33\%. These trials go on to show that the only improvement to this level of accuracy is to utilise the attendance data from all of Semester 1. In addition, the results provide a recommendation that models need to be re-trained every two years to avoid any disproportional effects from a single cohort.

As part of this project a new tool to visualise and communicate student achievement was developed. This tool uses contemporary Information Visualisation techniques to provide both a macro view of the entire cohort, down to the micro view of a single module for one student. One of the views provides an overview of the student's entire academic career by semester. This work has produced a set of sixteen descriptions for the unique (after accounting for time and scale effects) possibilities in this view. The author has termed them `Degree Pictures'. While the form of the intervention is outside of the scope of this work, these standardised forms can allow educators to develop best practice responses to any pattern that starts towards an undesirable outcome.

During testing of this model, an anomaly was observed where students were highlighted for poor attendance/engagement but were not identified as potentially failing in the model. The attendance data for these students showed a common feature, the disengagement began at week 5. Using statistics tools, this work was able to demonstrate that a significant shift occurred within the student body between week 5 and 8. This timing coincides with the timing of Reading Weeks for that semester. The work goes on to infer a structure to student cohorts which suggest that intervention with certain groups would be more effective than others. This insight shows that educators need to be mindful about introducing any disruption into the student experience, as it can be a trigger for disengagement.


Original languageEnglish
Awarding Institution
Award date25 Nov 2019

Research outputs (5)

View all