Predicting Student Academic Performance with Machine Learning

Background to the Study

Predicting student academic performance has long held a central place in educational research, given its direct implications for student welfare, institutional resource allocation, curriculum design, and the broader socioeconomic outcomes tied to educational attainment. Traditionally, identifying at-risk students and understanding performance determinants relied heavily on descriptive statistics and simple bivariate analysis, approaches that, while informative, struggle to model the complex, multivariate, and often non-linear interactions among the many factors that jointly shape a student's academic trajectory.

The rapid maturation of educational data mining and learning analytics as distinct interdisciplinary fields over the past two decades has transformed this landscape. Institutions now generate and retain unprecedented volumes of structured data, demographic records, attendance logs, continuous assessment scores, learning management system logs, and socioeconomic background information, that together form a rich substrate for statistical and machine learning-based predictive modelling. The International Educational Data Mining Society has organised much of the scholarship in this space, supporting researchers working to turn that data into better understanding of how students learn.

This data abundance has coincided with real advances in machine learning methodology, letting researchers move beyond simple linear regression toward a diverse set of algorithms, including decision trees, random forests, support vector machines, k-nearest neighbours, and naive Bayes classifiers, each carrying distinct statistical assumptions and different trade-offs between predictive accuracy, interpretability, and computational cost. That range of options raises a practically consequential question: given a specific educational dataset and outcome of interest, which modelling approach, or combination of approaches, offers the most statistically defensible and practically useful predictive performance? This cannot be answered in the abstract, since relative algorithm performance depends substantially on the underlying structure of the data, including the degree of linearity in predictor-outcome relationships, the presence of interaction effects, the dimensionality of the feature space, and the sample size available for training and validation. That is what motivates the comparative, rather than singular, modelling approach this study takes.

Nigerian tertiary and secondary institutions, like their counterparts across much of Sub-Saharan Africa, face persistent challenges of student underperformance, high dropout and repetition rates, and constrained resources for individualised academic support. Within that context, the capacity to identify, early and with confidence, the students most likely to underperform, and to understand the modifiable and non-modifiable factors most strongly associated with that risk, represents a genuinely high-value application of applied statistics and data science. This aligns closely with the priorities the UNESCO Institute for Statistics tracks globally around learning outcomes and educational equity. Candidate predictors examined in this study include prior academic achievement, previous GPA, study habits, weekly study hours, attendance behaviour, socioeconomic and parental background factors, access to resources such as internet connectivity, receipt of supplementary tutoring, extracurricular involvement, sleep patterns, and institutional context.

This study addresses the outcome through two complementary lenses: first, as a continuous variable, final examination score, modelled via multiple linear regression, allowing direct, interpretable quantification of each predictor's marginal contribution to score variation; and second, as a binary classification problem, high versus low academic performance, defined through a median split, modelled via binary logistic regression and five additional machine learning classifiers, allowing a rigorous, multi-metric comparison across methodologically distinct approaches. Triangulating findings across continuous regression, binary logistic regression, five machine learning classifiers, and tree-based feature importance measures makes it possible to distinguish genuinely robust, cross-methodologically consistent predictors of academic performance from findings that may simply be artefacts of a single modelling choice or outcome definition, a safeguard that matters given this study's aim of informing practical, resource-allocation-relevant institutional recommendations.

Statement of the Problem

Despite the growing sophistication and accessibility of machine learning tools, many Nigerian educational institutions and researchers continue to rely predominantly on simple descriptive statistics, or at most a single regression model, when analysing student academic performance data, without systematically comparing the relative predictive performance of multiple candidate approaches on the same dataset. This creates a real risk: institutions either underuse the predictive value available in their data by relying on overly simplistic analysis, or invest disproportionate resources in complex machine learning infrastructure without first establishing, through rigorous comparison, that such complexity actually delivers a meaningful predictive advantage over simpler, more interpretable alternatives.

A considerable portion of the existing educational data mining literature also evaluates machine learning classifiers using only a single performance metric, most commonly overall accuracy, without jointly considering precision, recall, F1-score, and area under the ROC curve, despite the well-established statistical principle that accuracy alone can present a misleading picture of classifier performance, particularly when class distributions are imbalanced or when the relative costs of false positives and false negatives differ substantially, as is frequently the case in educational early-warning applications where failing to identify a genuinely at-risk student can carry a far higher institutional and personal cost than incorrectly flagging a student who does not need intervention.

There is also comparatively limited Nigerian undergraduate statistics research that jointly models academic performance as both a continuous outcome, via multiple regression, yielding interpretable marginal effect estimates, and a binary classification outcome, via a comprehensive multi-algorithm comparison, despite the complementary insight each framing offers: regression-based analysis clarifies the direction and magnitude of each predictor's association with performance, while classification-based analysis directly addresses the practically important question of how accurately at-risk students can be identified for early intervention. It is against this backdrop that this study statistically models student academic performance using both continuous and binary classification frameworks, applying a comprehensive, multi-metric evaluation protocol, to establish an empirically grounded, comparative basis for academic performance prediction and early-warning system design.

Aim and Objectives of the Study

The aim of this study is to statistically model and predict student academic performance using multiple linear regression and a comparative evaluation of machine learning classification algorithms. The specific objectives are to:

1. Describe the demographic, socioeconomic, and academic-history characteristics of the study sample.

2. Examine bivariate relationships between candidate predictors (gender, tutoring support, parental education, socioeconomic status, school type) and academic performance.

3. Identify the statistically significant predictors of continuous final examination score using multiple linear regression.

4. Identify the statistically significant predictors of high versus low academic performance using binary logistic regression.

5. Comparatively evaluate the classification performance of logistic regression, decision tree, random forest, k-nearest neighbours, naive Bayes, and support vector machine algorithms using accuracy, precision, recall, F1-score, and AUC.

6. Rank the relative importance of predictor variables using random forest and decision tree feature importance measures.

Research Questions

1. What are the demographic, socioeconomic, and academic-history characteristics of students in the study sample?

2. Is there a statistically significant relationship between gender, tutoring support, parental education, socioeconomic status, school type, and academic performance?

3. Which factors are statistically significant predictors of continuous final examination score?

4. Which factors are statistically significant predictors of high versus low academic performance?

5. Which of the six classification algorithms evaluated achieves the best overall predictive performance for identifying high versus low academic performers?

6. Which predictor variables contribute most to academic performance classification according to feature importance measures?

Significance of the Study

This study is significant to educational institutions and academic administrators seeking an empirically grounded basis for designing early-warning systems and targeted academic support interventions, informed by a rigorous comparison of which statistical and machine learning approaches deliver the most reliable predictive performance for their specific institutional data. For educational policymakers, the study's findings on the relative statistical importance of modifiable factors, attendance, study habits, tutoring access, versus largely non-modifiable background factors, parental education, socioeconomic status, offer direct guidance for prioritising intervention resources toward what institutions can actually influence.

For statisticians, data scientists, and machine learning researchers, this study offers a methodologically rigorous, multi-metric evaluation spanning classical regression-based inference and contemporary machine learning classification, directly addressing the practically important and often under-examined question of whether added algorithmic complexity translates into meaningfully improved predictive performance on a real-world educational dataset. Readers working on related applied statistics or computational projects may also find the Computer Science project collection a useful complementary resource.

Academically, the study serves as a comprehensive reference and template for students undertaking educational data mining or comparative statistical and machine learning research, demonstrating a complete analytical workflow from descriptive statistics through advanced multi-algorithm model comparison. It also adds to the relatively limited body of Nigerian and West African educational data mining literature employing this breadth of comparative methodology.

Scope of the Study

This study is restricted to the statistical analysis of a structured, cross-sectional dataset of 800 student records comprising demographic variables (age, gender), socioeconomic variables (parental education, socioeconomic status), behavioural variables (weekly study hours, attendance rate, extracurricular involvement, sleep hours), resource-access variables (internet access, tutoring support), institutional context (school type), and academic-outcome variables (previous GPA, final examination score). The analysis covers descriptive statistics, Pearson correlation, independent samples t-tests, one-way ANOVA, a Chi-square test, multiple linear regression, binary logistic regression, and a six-algorithm machine learning classifier comparison, logistic regression, decision tree, random forest, k-nearest neighbours, naive Bayes, and support vector machine, evaluated via a stratified 70:30 train-test split.

The study does not extend to longitudinal tracking of academic performance across multiple academic terms, deep learning or neural network-based modelling, natural language processing of qualitative student feedback, or learning management system clickstream analysis, all of which are recommended as areas for further research.

Operational Definition of Terms

Academic Performance: Operationalised in two complementary forms in this study: a continuous final examination score on a 0 to 100 scale, and a binary High/Low Performer classification derived from a median split of the continuous score distribution.

Previous GPA: A student's grade point average from the immediately preceding academic term or session, on a standard 5-point scale, used as a proxy for prior academic achievement.

Educational Data Mining (EDM): An interdisciplinary field concerned with developing and applying statistical, machine learning, and data mining methods to understand and improve student learning and educational processes using data generated within educational settings.

Feature Importance: A statistic, typically derived from tree-based machine learning models, quantifying the relative contribution of each predictor variable to a model's predictive accuracy, commonly computed via mean decrease in node impurity, or Gini importance.

Area Under the ROC Curve (AUC): A performance metric summarising a binary classification model's ability to discriminate between two outcome classes across all possible classification thresholds, ranging from 0.5, no better than chance, to 1.0, perfect discrimination.

Train-Test Split: A model evaluation procedure in which a dataset is partitioned into a training subset, used to fit model parameters, and a held-out test subset, used to evaluate model performance on previously unseen data, providing an estimate of out-of-sample predictive performance.

Conclusion

The clearest takeaway from this study is that predictive sophistication and predictive accuracy are not the same thing. A well-specified logistic regression model, built on previous GPA and attendance rate as its strongest predictors, outperformed five more computationally demanding machine learning algorithms in identifying which students needed early support. For institutions weighing where to invest limited analytical resources, that is a genuinely useful finding: start with interpretable models and strong data on attendance and prior achievement before reaching for more complex infrastructure. Students and researchers working on related quantitative projects can find further reference material in the ScholarNest project research library, including comparable studies across statistics and computer science.

Frequently Asked Questions

Which model predicted student academic performance most accurately in this study?

Logistic regression achieved the highest overall classification performance among the six algorithms tested, with 80.83% accuracy and an AUC of 86.65%, outperforming random forest, naive Bayes, support vector machine, decision tree, and k-nearest neighbours.

What are the strongest predictors of student academic performance?

Previous GPA and attendance rate were the two dominant predictors, confirmed consistently across multiple linear regression, random forest variable importance, and decision tree feature importance.

Why did a simpler model outperform more complex machine learning algorithms?

The study found that when predictor-outcome relationships are approximately linear and additive, algorithmic complexity does not guarantee superior predictive accuracy, and logistic regression captured the relevant patterns effectively.

What machine learning algorithms were compared in this study?

The study compared decision tree, random forest, k-nearest neighbours, naive Bayes, and support vector machine classifiers against binary logistic regression, alongside multiple linear regression for the continuous outcome.

How was the dataset structured?

The study used a structured dataset of 800 student records containing demographic, socioeconomic, behavioural, resource-access, institutional, and academic-history variables.

What evaluation metrics were used to compare the classifiers?

The study used accuracy, precision, recall, F1-score, and area under the ROC curve (AUC), evaluated on a stratified 70:30 train-test split, rather than relying on accuracy alone.

What is a train-test split and why does it matter?

A train-test split partitions data into a training subset used to fit a model and a held-out test subset used to evaluate it on unseen data, giving a realistic estimate of out-of-sample predictive performance.

What practical recommendations does the study make?

It recommends that institutions prioritise attendance monitoring and early academic support for students with weak prior academic records, and adopt interpretable models like logistic regression before investing in more complex machine learning infrastructure.

Is the dataset used in this study from a real Nigerian institution?

The dataset is a structured, simulated dataset built to reflect realistic relationships and effect sizes documented in prior educational data mining literature, so figures should be read as illustrative of methodology rather than institution-specific findings.

Where can I find more research like this?

Related studies in applied statistics, data analysis, and computational methods are available in the Statistics and Computer Science sections of the ScholarNest project research library.