Monisha S, Edwin Raj A
This paper looks at the interdependence and independency of subject-wise academic scores, as well as identify factors that affect student performance, across Matriculation, CBSE and Government school systems. A descriptive statistical analysis, correlation analysis, graphical visualization, Exploratory Principal Component Analysis and Factor Analysis, regression models, and machine-learning techniques were used to analyze the data of 540 higher secondary students. The analysis revealed two large latent academic dimensions, namely, STEM ability and Language Ability. STEM ability was found to have high loadings in Physics, Chemistry, Biology and Mathematics, whereas Language ability showed high loadings in English and Language. The two combined explained 77.4% of the overall Variance, indicating that student performance can be captured on fewer and more understanding academic dimensions. Models of prediction of latent academic scores included OLS, Ridge, Lasso, Random Forest, and XGBoost. The OLS and Ridge Regression models have the best testing results, with the value of R to 0.974 and 0.973, respectively, and in most cases, the mean absolute error is about 2.10. Permutation-importance analysis showed that science-related subjects strongly contributed to latent academic performance. Propensity-score-weighted analysis indicated that section assignment had no statistically significant effect on latent performance scores. The findings support the use of interpretable linear models for structured educational data and show that latent academic patterns can assist in targeted academic interventions. However, as the study is cross-sectional, future research should include longitudinal data and socioeconomic factors to improve causal interpretation and explanatory strength.