As a result of the COVID19 pandemic, more higher-level education courses have moved to online channels, raising challenges among educators in monitoring students’ learning progress. Thanks to the development of learning technologies, learning behaviours can be recorded at a more fine-grain level of detail, which can then be further analysed. Inspired by the premise of approaching education as a complex system, this research aims to develop a novel approach to analyse students’ learning behavioural data in programming education, utilising complexity methods. First, essential learning behavioural features are extracted. Second, a novel method based on Random Matrix Theory is developed to remove the noise and trend effect in the data in order to better highlight the differences in students’ learning behaviours. Third, Community Detection is applied to cluster the students into groups with similar learning behavioural characteristics. In the thesis also, motivated by a need to determine likely outcomes of students, a range of machine learning classification techniques have also been applied to predict the student learning outcomes based on behavioural data which has been cleaned of the noise and trend. The proposed approaches have been applied to datasets collected from a bespoke online learning platform in an Irish University. The datasets contain information from 566 students in different programming-related modules over a range of years encompassing pre and during the COVID19 pandemic. This gives us a unique opportunity to test our methods for the effects of the pandemic on learning. Results indicate the similarities and deviation in learning behaviours between student cohorts. Overall, we found that students interacted similarly with all course resources during the semester. However, while higher-performing students seem to be more active in practical tasks (e.g. programming exercises on labs), lower-performing students have been found to focus overmuch on lecture notes and lose their focus at the later phase of the semester. Additionally, students’ learning behaviours in a conventional university setting tend to differ significantly to those students in a fully online setting during the pandemic. We have also attempted to reduce the noise component in the data and the experimental results further demonstrate the better prediction performance of models which are trained based on the cleaned dataset, in comparison with the original dataset. Recommendations for current educational practice are made, including the continuous analysis of learning behaviours by the proposed methods and suggestions for the prompt interventions to in order to max- imise student supports.