Investigation of secondary school student performance with MARS model

Multivariate adaptive regression splines (MARS) modeling, one of the non-parametric regression methods, was used in this study. A web application based on machine learning was developed using the MARS model. It is aimed to accurately predict the achievement scores of the 8th-grade students before the LGS (High School Entrance System) exam With the developed application. The study was carried out with 8th-grade students of Mehmet Akif Ersoy Secondary School in the Tosya district of Kastamonu province. The demographic information of the students and all the test results they took in the last year were used before the LGS exam. Models that predict the LGS exam result correctly and have maximum interaction were tried and the most suitable model was determined. The coefficient of determination was estimated as the maximum and the generalized cross-validation coefficient as the minimum in this model provided. Accordingly, the most appropriate model; it has been determined that 1 product degree with 6 basis functions is provided with a value was 0.954, the value was 423.003. The significant variables are the number of siblings, mother's education level, net count of revolution history and Ataturkism lessons more than 1.73, net count of English lessons more than 1.466, net count of the mathematics lesson more than 5.939, and net count of a mathematic lesson under 5.936. The knot point of the mathematic lesson was determined as 5,936. If the average net count of a mathematic lesson is less than 5,936; the parameter estimation coefficient was determined as -11,965. The situation determined here is that the average net number of the mathematics lesson is less than 5.936, which negatively affects the success of the student. However, when the average net count of a mathematic lesson is greater than 5.936, the parameter estimation coefficient becomes 4.060. When the average net count of a mathematics lesson exceeds 5,946, the success of the student is positively affected. A web-based machine learning-based application has been developed to predict students' LGS scores in line with these data. R Shiny program was used in the development of the study. The program is cloud-based and works independently of the operating system and web browsers. The developed web program can be accessed at https://beststat.shinyapps.io/lgspuan/. The developed application helps students prepare for the LGS exam to offer pre-exam advice to guide their studies.