Predictive Modeling for Career and University Pathways Using Small-Scale

Many students struggle to make informed decisions regarding their future university or career paths. Traditional counseling often provides general guidance and cannot account for individual preferences, academic performance, or extracurricular activities. This study addresses this issue by developing a predictive modeling system for career and university pathway recommendation using small-scale datasets collected from Myanmar students. The system, referred to as NovisAI, serves as a demonstration platform for evaluating predictive models, however, this research focuses exclusively on prediction accuracy and model performance rather than recommendation strategies or system development.
The primary objective of this dissertation was to predict students’ suitable career and university pathways and provide actionable recommendations. A survey was conducted to collect data from current and former high school students in Myanmar. Although responses were collected from a wider pool of participants, only 209 complete and valid responses were retained for analysis. The dataset included academic, behavioural, personality, and extracurricular attributes. Small dataset size posed challenges in model training, including poor generalization and potential overfitting.
To address these challenges, a prototype system, was developed to implement a machine learning pipeline. Data preprocessing included cleaning, encoding, and augmentation using Gaussian noise and class balancing with SMOTE. Feature selection was performed using ANOVA and mutual information to identify the most informative predictors. Several machine learning models were trained, including Random Forest, XGBoost, Multi-Layer Perceptron, Logistic Regression, Support Vector Machine, and an ensemble of Random Forest and XGBoost. Model performances were evaluated using accuracy, F1-score, and cross-validation to enhance robustness and reduce the risk of performance bias arising from the limited dataset size.
Experimental results show that the Multi-Layer Perceptron trained with mutual-information features achieved the highest mean cross-validation accuracy of 0.956 and a final test accuracy of 0.968. Random Forest, XGBoost, and the ensemble model also demonstrated strong predictive performance. Following prediction, a gap analysis module compared student profiles with successful peers in the predicted career path, highlighting areas for academic and extracurricular improvement. Rule-based university recommendations were provided based on student profiles, academic scores, and career alignment.
Overall, this study shows that small-scale datasets can still provide accurate and actionable career guidance with careful preprocessing, balanced data, and appropriate feature selection. Based on the trained machine learning predictive models, a prototype system named NovisAI was implemented to demonstrate how these predictions can be applied in helping students better align their skills and interests with future education and career choices. As future work, this system will extend to larger datasets for predictions and recommendations and explore advanced deep learning techniques for improved prediction models.