You are here

  1. Home
  2. eSTEeM Projects
  3. Supporting students
  4. An investigation into the use of Artificial Neural Networks to predict student failure, and the efficacy of sustainable additional support for those students

An investigation into the use of Artificial Neural Networks to predict student failure, and the efficacy of sustainable additional support for those students

  • Project leader(s): John Woodthorpe
  • Theme: Supporting students
  • Faculty: STEM
  • Status: Archived
  • Dates: December 2013 to March 2016

The project analysed VLE data with Artificial Neural Networks (ANNs), in order to identify patterns of behaviour that correlate with the likelihood of a student failing the EMA. The ANN models were capable of flagging such students early enough in the module for action to be taken to improve their chances of passing and progressing.

The early stages of the investigation involved collating, synthesizing and processing large datasets from previous TU100 cohorts with the additional aim of significantly increasing the knowledge of patterns of student behaviour. Two crucial steps in this were the selection and pre-processing of the student data, as different parts of the University hold different data in different formats, and derived from differing criteria. After obtaining and feeding the data into the neural networks, the models were refined to determine which of the criteria chosen were most important. This iterative process of training and testing continued until the predictions and observed results from a previous known presentation matched each other. The model was then used as a predictive tool on a current presentation, where it was shown to work well. Doing the same for other modules should be possible, but would require data on those modules and the time to train the models and their users. It is hoped to do some work on this later in 2016/17.

Once the predictions were made, the project showed that personalised telephone guidance from their tutor improves the chances of ‘at risk’ students passing their module. So if they can be identified, something can be done to try to help them. It has also alerted tutors to the merits of focussing on those students, as that contact has the potential to help them pass the module. 

It is important to note that the tutor contact can be triggered by any predictive model, such as those used by OU Analyse, the Information Office, from the Student Support Tool categories, or from a manual selection of criteria applicable to a particular module. So other modules may like to develop that aspect of the project as it involves potentially relevant extensions to the role of the tutor. However, it is important that the tutors are funded for this work, either by an extra payment or (as was done here) by removing an existing activity included within the contact time. Merely adding this to the ever-growing tutor workload is not an option.

The tutor contact also has the extra benefit of encouraging tutors to do something that uses their knowledge of the module and of their students, and which many tutors enjoyed. Consequently, there is potential to extend this so that tutors and Learner Support work more closely in supporting ‘at risk’, ‘fail’ and ‘cannot pass’ students.

Some of these aspects are being continued within TU100, with further development being funded by the STEM Faculty.

Related Resources: 
PDF icon John Woodthorpe, Artificial Neural Networks. eSTEeM Final Report.pdf318.12 KB

eSTEeM final report.

PDF icon SST Enhancement Digest_Ed2_WEB.pdf1.4 MB

SST Enhancement Digest article.