For each module, we divided the Moodle logs into chunks of weeks. We then extracted multiple features for each week
- A count of how many times they accessed Moodle
- The ratio of on campus to off campus accesses using IP address
- The average time of day they used Moodle
- How many times times they accessed Moodle during the weekend
These features have been shown to be good indicators of student performance.
Work by Cocea and Weibelzah on feature extraction was an influence on ours. The features used in this study include number of pages accessed, time spent reading pages and student performance on mid-semester tests. Even though they were using a VLE targeted at one particular module, the features used were similar to ours. However their objective was different, with their goal being to estimate learners’ levels of motivations.
We used a Support Vector Machine (SVM) to classify students as a pass or a fail. We trained one SVM for each week of the semester up until the exams. Each training set contained all data available up until that week. For example the ”week 7” SVM was trained with the demographic data, and all weekly Moodle log data up until week 7. This allows us to make predictions on new Moodle log data on a week by week basis.