Frontiers in Education (Jul 2020)
Identifying Reliable Predictors of Educational Outcomes Through Machine-Learning Predictive Modeling
Abstract
Results-based financing has guided the development of policies with measurable results improving learning outcomes at micro/macro levels. However, it is then necessary to identify factors which predict early and accurately favorable or challenging conditions for learning. Learning outcomes depend on complex interactions between multiple variables, many of which are not fully understood. The objective was to develop valid and accurate models predicting low and high levels of math performance and Vietnamese language, using machine-learning algorithms, as part of an international large-scale project in primary education in Vietnam. The models achieved very high accuracy (95–100%). A strong common pattern has been found for both Math and Vietnamese language, for the low and high levels of performance: the individual cognitive characteristics, physical factors and daily routines/ activities of the child are very important predictive factors of academic performance, as measured by student performance in the final Grade 5 test in math and Vietnamese, respectively. Parental expectations, pre-school attendance and school trajectory of students have added relative importance in the classification. In order to accurately identify an expected low or high academic performance outcome, it is the full pattern of variables contained in the vector of information from each case that should be considered. Because, although each variable in a particular vector has a small contribution to the total predictive weight, it is the overall pattern containing the interactions between these variables that carries the necessary information for the accurate predictions. In addition, the identification of specific patterns for extreme groups of performance provides the necessary guidance for more focused educational interventions/investment and sound educational policies.
Keywords