• educational data mining
• moodle
• kdd
• learning analytics
• data pre-processing
• moodle logs
• kdd process
• logs

Learning Analytics research has demonstrated that temporal analysis is relevant since the greater timing flexibility in online environments may affect learning differently. Therefore, capturing time-on-task as a proxy to model learning behaviour, predict performance, and avoid drop-out has been the focus of a number of investigations. However, although data pre-processing decisions influence the outcomes and can lead to inaccurate predictions, the majority of studies do not provide enough information on how their data were prepared for their findings to be replicated. Moodle logging system and in particular its temporal dimension can be difficult to interpret. To illustrate how its knowledge can be used to improve data processing, we first propose an in-depth analysis of Moodle logging system, then, starting from the correct extraction of Moodle logs, we focus on factors to consider when preparing data for temporal analysis. One of the most important aspects of preparing learning data is the detection of anomalous duration values of student activities. The majority of works on this topic fail to account that distinct activities can have vastly different typical execution times.
Thus, we propose a methodology for estimating time-on-task that applies an outlier detection strategy based on a distinct examination of each learning activity and its peculiarities. Prepared data will then be used to uncover the various temporal habits that each student employs when learning online. Typically, when modelling trends, a chosen configuration is set to capture various habits, and a cluster analysis is undertaken. However, the selection of variables to be observed and the algorithm used to conduct the analysis reflect the researcher’s thoughts and ideas. To explore how students behave over time, we present alternative ways of modelling student temporal behaviour. We accompany our theoretical results with implementations and experiments on six courses. Our results reveal that the generated clusters may or may not differ based on the selected profile and unveil different student learning patterns. The temporal learning behaviour of some students is unique, while some others are always similar regardless of the perspective adopted to model the profile.