AI-Assisted Theorem Discovery: Revolutionizing Mathematical Exploration

In more developed countries, this so-called dropout phenomenon has become a critical issue to the sustainability of educational institutions, since it leads to waste of resources (time, effort, money, etc.), squandered potential (of both the student and the entire society) and adverse socio-economic implications for the community. Availability of educational data, as well as the trends on artificial intelligence developments, gives a good enable for machine learning to become a promising technique to foresee and reduce the dropout rate. We conduct a systematic review of the different types of machine learning techniques that have been utilized for student dropout prediction, examining their approaches, datasets, metrics, and practical use cases. The review was performed across peer-reviewed academy papers published in the past decade through the following databases: IEEE Xplore, ScienceDirect, Springer and Google Scholar. A total of 150 studies were screened, and 60 were included based on relevance, methodological clarity and empirical validation. The case studies included K-12 through higher education, and multiple geographic regions. This review presents four categories of machine learning models employed in dropout prediction: (1) supervised learning, (2) unsupervised learning, (3) ensemble learning, and (4) deep learning. This very popular approach is itself a supervised learning (Decision Trees, Support Vector Machines, Random Forest, Naive Bayes and Logistic Regression) because it is easy to implement, it is easy to interpret, and it is easy to understand a clear performance. Only lately, deep learning models (i.e., neural networks and recurrent neural networks) take place for modeling challenging conditions of temporal behavior data of students, while consuming high resources and requiring more explainability. Some significant attributes that can be used for dropout predictive models includes academic performance, demographic feature, behavioral feature and psychological feature. In case of longitudinal and multi-source data, the training of models is improving the quality of prediction of whether a student will be successful or not (Stelmakh, 2020), i.e., predicting that quality by merging the academic records with activity logs from a learning management system (LMS) and survey elaborations; Performance across studies is typically evaluated using metrics such as accuracy, precision, recall, F1-score, and Area Under the Curve.