According to Gardner and Brooks (2018) [2], regarding the statistical models used to map features to predictions, supervised learning techniques are extensively used in predictive student modeling in MOOCs, compared to unsupervised approaches, as student dropout/stopout is easily observable. In supervised learning, machines are trained using well “labelled” training data, and on the basis of that data, machines predict the output. Authors chose to present indicatively popular techniques for MOOC learner modeling, with very good empirical performance in large-scale MOOC modeling studies, (e.g., Dass et al., 2021 [18]). LR and SVM are among the most used, while NB, kNN, and DT are less frequently used in surveys ([2]). There is no pattern found for an algorithm to be distinguished compared to others. According to Herrmannova et al., 2015 [19], each model captures different properties of input data and the results are complementary.

Self-regulated learning (SRL) is a complex multidimensional phenomenon often described by a set of individual cognitive, social, metacognitive, and behavioral processes embedded in a cyclical model. Students with limited application of SRL strategies do not perform well [1,20,21,22]. Zimmerman [23] proposed a cyclical model consisting of three interrelated phases of the learning process. In the forethought phase, self-regulated learners set learning goals and design the strategy for their learning. This is followed by the performance and control phase, during which self-regulated learners employ strategies to process the learning material. They seek help when needed, manage their time, structure their environment, and monitor their learning processes. In the third phase of self-reflection, self-regulated learners evaluate their performance and adjust their strategies to achieve their learning goals [1,18,22,24,25].

Research using questionnaires has shown positive correlations between the mentioned SRL activity and the completion of MOOCs [24]. Despite the limitations of self-reported data [26], the large sample size enhances their usefulness. On the other hand, the use of trace data in measuring SRL has increased, but their interpretation remains challenging [26]. Jansen et al. [24] propose the combined use of SRL data from traces and questionnaires. Timely SRL support interventions in MOOCs should be considered significant pedagogical tools that contribute to achieving positive outcomes for students [20,21].

The features of MOOCs, such as internet-based massiveness, openness, and flexible learning, create a unique blend of a large number of learners, making the prediction of learner success (as well as providing support based on these predictions) particularly challenging. Several researchers have developed prediction models by employing machine learning (ML) algorithms [27] and adopting supervised, unsupervised, and semi-supervised architectures [28]. Deep learning methods are also utilized for predicting dropout. For instance, Moreno-Marcos et al. [29] applied a combination of random forest (RF), generalized linear model (GLM), support vector machines (SVM), and decision trees (DT). Feng et al. [10] utilized logistic regression (LR), support vector machine with a linear kernel (SVM), random forest (RF), gradient boosting decision tree (GBDT), and a three-layer deep neural network (DNN) for their analysis.

Diverse features, even in limited quantities, provide a more comprehensive, multidimensional view of learners and can improve the quality of predictive models. Collecting additional data, especially during the initial weeks of a course, enhances prediction performance [2,28]. For successful timely interventions, predictive models need to be transferable, meaning they perform well on new course iterations by utilizing historical data [30,31,32]. Some researchers have examined specific aspects of SRL and observed its impact on predicting success, including goal setting and strategic planning [6], student-programmed plans [33], the combination of self-reported SRL strategies and patterns of interaction sequences, demographic features, and intentions [34]. For example, Kizilcec et al. (2017) [6] investigated which specific SRL strategies predict the attainment of personal course goals and how they are mapped in specific interactions with MOOCs’ online content. Yielding to form a longitudinal account of SRL, authors combined learners’ self-reported SRL strategies and characteristics, achievement data, and records of individual’s engagement with the course’s content. Using multiple linear and logistic regression modeling, Maldonado-Mahauad et al. (2018) [34] concluded that specific self-reported SRL strategies (“goal setting”, “strategic planning”, “elaboration” and “help seeking”), complex behavioral data on MOOC’s platform like meaningful online activity sequence patterns, self-reported prior experience and level of interest in MOOC’s assessments as well as total time spent online belong to factors that contribute to the prediction of MOOC learners’ success.