The goal of our research is improvement of mathematics curriculum and popularization of mathematics among students of economics in developing countries. We analyze and compare curricula of pure mathematics courses that are taught to university students of faculties of economics in Japan and in Bosnia and Herzegovina. Data set contains math syllabuses in 2021/22 school year from six public universities in Bosnia and Herzegovina and seven from Japan. The text corpus was pre-processed and then the Term Frequency – Inverse Document Frequency algorithm, and Sentence Transformed Multi QA model were applied to build word vectors, find the similarity among Japanese and Bosnia and Herzegovina mathematics syllabuses using cosine similarity approach, and to find the key competences of these two countries mathematics syllabuses using the word cloud. Our results show the following similarity between the curricula: 60.7 percent using TF-IDF and 80.3 percent using Multi QA model. The key competences in the Japanese mathematics course are narrow and focused, in contrast to Bosnia and Herzegovina’s.

This research aims to reduce dropout rates in higher education by developing a machine learning model to predict churn early, enabling timely interventions. The most important research findings of the thesis are summarized as follow: - The overall dropout rate at the University of Banja Luka, Bosnia and Herzegovina, was nearly half of enrolled students between the 2007/08 and 2018/19 academic years, with rates showing an increase over time. - Half of the student churn occurs within the first year of enrollment, with significantly higher rates in three-year study programs compared to four-year programs. - The results suggest that at least one-third of student churn could potentially be prevented, as it is attributable to institutional factors. - The findings indicate that it is possible to predict student churn at the earliest stages of education, even when using a challenging dataset with missing data and limited pre-academic, academic, and socioeconomic features. - The use of the Histogram-based Gradient Boosting Classifier (HGBC) in the educational data field, for the first time, resulted in an attrition classification accuracy of 75% at the beginning of the first academic year and 83% by the end, outperforming other widely used models. - To validate the effectiveness of HGBC, the model was tested under varying conditions and consistently demonstrated high performance. - The application of pre- and post-hoc interpretability techniques can enhance the credibility and reliability of machine learning models in predicting student discontinuation. - The most significant predictors of churn include gender (female), student cohort, accumulated ECTS credits, scholarship status, age at enrollment, the number of successfully passed courses, program duration, and whether the student attended a Gymnasium. - Certain variables, such as gender (female) and student cohort, consistently maintained a high level of significance across the models over time. - Models with lower performance and quality exhibited a similar ranking of feature importance to those of the best-performing models.