Latent deep sequential learning of behavioural sequences

The growing use of asynchronous online education (MOOCs and e-courses) in recent years has resulted in increased economic and scientific productivity, which has worsened during the coronavirus epidemic. The widespread usage of OLEs has increased enrolment, including previously excluded students, resulting in a far higher dropout rate than in conventional classrooms. Dropouts are a significant problem, especially considering the rising proliferation of online courses, from individual MOOCs to whole academic programmes due to the pandemic. Increased efficiency in dropout prevention techniques is vital for institutions, students, and faculty members and must be prioritised. In response to the resurgence of interest in the student dropout prediction (SDP) issue, there has been a significant rise in contributions to the literature on this topic. An in-depth review of the current state of the art literature on SDP is provided, with a special emphasis on Machine Learning prediction approaches; however, this is not the only focus of the thesis. We propose a complete hierarchical categorisation of the current literature that correlates to the process of design decisions in the SDP, and we demonstrate how it may be implemented. In order to enable comparative analysis, we develop a formal notation for universally defining the multiple dropout models examined by scholars in the area, including online degrees and their attributes. We look at several other important factors that have received less attention in the literature, such as evaluation metrics, acquired data, and privacy concerns. We emphasise deep sequential machine learning approaches and are considered to be one of the most successful solutions available in this field of study. Most importantly, we present a novel technique - namely GRU-AE - for tackling the SDP problem using hidden spatial information and time-related data from student trajectories. Our method is capable of dealing with data imbalances and time-series sparsity challenges. The proposed technique outperforms current methods in various situations, including the complex scenario of full-length courses (such as online degrees). This situation was thought to be less common before the outbreak, but it is now deemed important. Finally, we extend our findings to different contexts with a similar characterisation (temporal sequences of behavioural labels). Specifically, we show that our technique can be used in real-world circumstances where the unbalanced nature of the data can be mitigated by using class balancement technique (i.e. ADASYN), e.g., survival prediction in critical care telehealth systems where balancement technique alleviates the problem of inter-activity reliance and sparsity, resulting in an overall improvement in performance.