Machine Learning Study Finds Demographics Barely Predict Students’ Love of Mathematics

Who enjoys mathematics? For decades, educators and policymakers have assumed that the answer could be read off a student’s demographic profile: their sex, their family’s income, the type of school they attended, or the academic track they chose in high school. A new study from the Philippines puts that assumption to a rigorous test using one of machine learning’s most versatile tools, and the result is striking. When researcher Arnel S. Travero of the University of Science and Technology of Southern Philippines fed eight demographic variables into a random forest regression model to predict how much first-year college students appreciate mathematics, the model explained essentially none of the variation. The full model produced an R-squared value of exactly 0.000, meaning that age, sex, academic strand, senior high school mathematics grades, school type, socioeconomic status, learning modality preference, and marital status together carried virtually no predictive signal for students’ enjoyment, perceived value, and willingness to engage with the subject.

The study, published in SN Social Sciences, surveyed 260 first-year students enrolled in Mathematics in the Modern World, a general education course at a Philippine public university. The course itself is a product of national higher education reform, mandated by the Commission on Higher Education’s 2013 memorandum to cultivate holistic understanding and intellectual competencies rather than narrow technical skill. That makes the student population unusually diverse: the course draws learners from many senior high school strands, not only those bound for STEM degrees. Travero chose this setting deliberately, since mathematics appreciation, unlike raw achievement, is an affective and motivational construct that shapes whether students remain willing to engage with quantitative reasoning long after their required coursework ends.

Random forest regression, the analytical engine behind the study, is an ensemble learning method introduced by statistician Leo Breiman in 2001. Instead of fitting a single equation to the data, the algorithm builds hundreds of decision trees, each trained on a random bootstrap sample of the observations and considering only a random subset of predictor variables at each split. The trees’ predictions are then averaged, a process that reduces variance and guards against overfitting in ways that classical linear regression cannot. Random forests have become a workhorse of educational data mining precisely because they capture nonlinear relationships and interactions between variables without the analyst having to specify them in advance. Previous work has used the method to predict college engineering major choice from demographic and high school factors, to evaluate pedagogy and inform personalized learning, and to forecast high school mathematics performance.

Travero’s implementation examined eight demographic predictors in detail. Age and sex are the classic background variables. Academic strand refers to the specialized track a Filipino student follows in senior high school, such as science, technology, engineering and mathematics, accountancy and business, or humanities. The senior high school mathematics grade captures prior achievement. School type distinguishes public from private secondary education, a meaningful divide in the Philippine system. Socioeconomic status, learning modality preference, which became salient after the pandemic-driven shift to online and hybrid instruction, and even marital status round out the list. The appreciation outcome itself was measured as a composite of students’ enjoyment of mathematics, their perception of its value, and their willingness to engage with it.

The headline result is the near-total absence of explanatory power. With all eight demographic variables included, the random forest model achieved an R-squared of 0.000, a figure that would be unremarkable if the model were predicting something like lottery numbers but is genuinely surprising for educational attitudes that are so often assumed to track background characteristics. Travero then refined the model, retaining only the five most influential variables in an attempt to strip away noise. The refined model improved only marginally, reaching an R-squared of 0.011. In practical terms, even the best demographic model could account for barely one percent of the differences in mathematics appreciation among these students. The model’s mean absolute error stood at 2.80 points, with a mean absolute percentage error of 8.9 percent, indicating predictions that were systematically no better than simple averages.

Feature importance analysis, a standard output of random forest methods that quantifies how much each predictor contributes to the model’s splits, did reveal a hierarchy among the variables. Academic strand emerged as the most influential, followed by the senior high school mathematics grade and then age. This ordering is intuitively plausible: a student who spent senior high school immersed in the STEM strand and earned strong mathematics grades might reasonably be expected to view the subject more favorably than a humanities-track student who struggled through it. Yet the study’s central finding is that even these top-ranked contributors remained minimal in absolute terms, given the model’s weak overall performance. A variable can rank first in importance within a model that explains almost nothing; the ranking describes relative signal, not predictive strength.

The null result carries real theoretical weight. Travero situates the findings within Expectancy-Value Theory, the influential motivational framework developed by Jacquelynne Eccles and Allan Wigfield, which holds that students’ engagement with a domain depends on their expectancies for success and the value they attach to the task. Those values, in turn, are shaped by perceptions of usefulness, intrinsic interest, and the accumulated weight of learning experiences, not by demographic categories in themselves. The study’s results are broadly consistent with that account: if appreciation flowed directly from demographic background, a flexible nonlinear learner like a random forest should have detected it. Instead, the data suggest that the drivers of mathematical appreciation lie elsewhere, in affective responses, motivational beliefs, instructional quality, and the specific contexts in which students encounter mathematics.

The findings also push back against what the study calls demographic determinism, the tendency in both research and popular discourse to treat background categories as destiny. A substantial body of prior research has linked socioeconomic status to academic achievement, examined gender gaps in STEM participation, and explored how demographic features relate to mathematics performance and beliefs. Those literatures are not invalidated by a single null result, and achievement is a different outcome from appreciation. But the study adds an important caution: patterns observed for performance do not automatically transfer to attitudes, and correlations reported in aggregate can be far too weak to support prediction for individuals. Machine learning, which is often deployed to squeeze predictive signal out of whatever variables are available, here serves as an honest auditor, reporting that the available demographic variables simply do not contain the signal educators might have hoped for.

For mathematics educators, the practical implication is a shift of focus. If demographic profiles cannot identify which first-year students will appreciate mathematics, then interventions cannot be targeted by background alone. Instead, the study argues for context-sensitive, student-centered approaches that address the variables the demographic model left out: students’ perceptions of the value and relevance of mathematics, their motivational beliefs, their instructional experiences, and the affective climate of their classrooms. Future predictive models, Travero suggests, should incorporate these affective, motivational, instructional, and contextual variables if they are to explain meaningful variation in appreciation. In an era when institutions increasingly reach for algorithmic tools to personalize education, the study offers a humbling demonstration that the most important predictors of a student’s relationship with mathematics may not be found in any enrollment record, but in the day-to-day experience of learning itself.

Subject of Research: Predicting first-year college students' mathematics appreciation from demographic factors using random forest regression

Article Title: Demographic factors and mathematics appreciation among first-year college students: a random forest approach

Article References: Travero, A. S. (2026). Demographic factors and mathematics appreciation among first-year college students: a random forest approach. SN Social Sciences, 6(10), Article 476. https://doi.org/10.1007/s43545-026-01730-z

Image Credits: AI Generated

DOI: 10.1007/s43545-026-01730-z

Keywords: mathematics appreciation, random forest regression, demographic factors, Expectancy-Value Theory, mathematics education, machine learning, first-year college students, educational data mining, STEM attitudes, Philippines, socioeconomic status, academic strand