From classical to quantum machine learning for analyzing and predicting alumni outcomes

Abstract

This thesis is submitted to the graduate program at the School of Engineering and Sciences as part of the requirements for obtaining the degree of Doctor of Philosophy in Computer Science. This study aims to generate models using both classical and quantum machine learning (ML) methodologies to accurately predict three key outcomes for alumni: job level, career satisfaction, and first employment. The data analyzed comes from Tec de Monterrey university alumni surveys. The study’s objectives also include the identification of important and actionable features for alumni outcome predictions. Among the challenges in finding models to predict and explain alumni outcomes, we encountered issues such as handling imbalanced classification, hyperparameter tuning, model prediction interpretation, and long training times. To address the latter, we proposed a method that reduces execution time when working with large datasets, particularly in methodologies like support vector machines. This proposal effectively resolves time and memory limitations in high-dimensional classification problems without compromising performance accuracy. The results show that classical machine learning models accurately predicted alumni outcomes. For instance, gradient boosting was most accurate in predicting job level and career satisfaction, while support vector machines outperformed in employment prediction. Significant features identified included current salary and number of people supervised for job level, with higher salaries and more supervisory responsibilities correlating with higher job positions. For career satisfaction, life and income satisfaction were important indicators, as higher satisfaction levels in these areas predicted greater career satisfaction. In the case of employment, networking support resulted as the most important feature, with stronger professional connections significantly increasing the likelihood of securing employment shortly after graduation. Additionally, the research identified actionable features that can impact both educational institutions and students. For job level, soft skills, particularly communication and teamwork, were found to be crucial in advancing to higher positions. Institutions can focus on enhancing these skills through their programs, while students are encouraged to develop them actively. For career satisfaction, the effective use of skills and technological tools acquired during education was a strong predictor, indicating the importance of aligning academic training with the demands of the job market. Facilitating robust professional networks proved essential for employment, emphasizing the need for institutions to create networking opportunities and for students to build social connections proactively. Many more interesting trends and findings related to alumni outcomes are highlighted in the thesis. Regarding quantum machine learning (QML) models, this research demonstrates the v feasibility of predicting alumni outcomes. A hybrid quantum-classical approach was particularly effective in predicting the three alumni outcomes in reduced datasets without substantially affecting accuracy. For example, quantum support vector classifiers (QSVC) showed comparable performance to classical support vector classifiers (SVC) while utilizing a reduced dataset versus SVC with complete datasets. Although QML is still in its early stages, this research suggests that QML could become a viable alternative in educational data mining as the field expands.