Can you predict your students’ final grade at the start of the course? Yes, you can with Artificial Intelligence

ia.jpg

One essential need for almost every teacher at the beginning of the course is to get to know their students as soon as possible in order to channel pedagogical strategies that will lead to better learning. To address this challenge, Tecnológico de Monterrey professors developed a predictive model based on machine learning that, with the help of artificial intelligence, predicts students’ grades for homework, quizzes, and even partial evaluations. 

By Omar Olmos López (oolmos@itesm.mx) y Miguel Ángel Hernández (mihernan@itesm.mx)

Imagine this scene… The long-awaited first day of school arrives and you, as a teacher, are given a sheet with your students’ performance forecast, their grades and even who might drop out of the group… Before starting any educational process whatsoever. This is the holy grail for teachers!

One essential need for almost every teacher at the beginning of the course is to get to know their students as soon as possible in order to channel pedagogical strategies that will lead to a better learning. The task of identifying the profiles is generally developed through dynamics, exams and/or diagnostic tests. In most cases, these studies of correlating profiles with academic performance are usually qualitative and the information obtained is limited. Therefore, characterizing each student’s behavior improvement areas is often uncertain and inaccurate.

Using appropriate techniques and tools, we can detect learning problems of each student and also group with a possibility of very high certainty.

As teachers, we know that educational processes are not only teaching-learning processes that affect learning, but they also involve behaviors, values and skills previously developed by students, which is something we must consider in our analysis. This could be why this task is so complex and diverse when establishing reliable prediction models. However, it can be achieved if the appropriate techniques and tools are applied. We could improve the way in which learning methods and actions are channeled, to detect students with deficiencies, address each student’s and group’s particular learning problems, with a very high possibility of attention and certainty.

Derived from collaborative work with other science teachers, we developed a mathematical model supported by Artificial Intelligence that makes it possible to identify academic performance at the beginning of the course. As part of the preliminary work of the last five years, the teachers conducted student performance profile studies, which are diagnostic tests to determine learning performance profiles and/or multiple intelligences, as well as a bio-response study of a group of students under a period of observation. The predictive model is based on a Machine Learning algorithm, which can be trained, with the help of artificial intelligence, to predict students’ grades for homework, quizzes and even partial evaluations. 

Over four years, the teachers involved in the project examined 106 students who were studying bachelor’s degrees in Engineering and Business at Tecnológico de Monterrey. Based on the study sample population, they analyzed data such as evaluations of the academic performance logs for each of their courses, habit and behavior tendencies, the teacher’s evaluation model, his/her experience as a teacher, preferences for taking courses related to sports or cultural abilities, and biometric aspects during a school period, including sleep quality, heartrate, hydration, nutrition, facial morphology, stress levels and neuronal response through certain standardized stimuli.

The predictive model is based on a Machine Learning algorithm, which can be trained, with the help of artificial intelligence, to predict students’ grades for homework, quizzes and even partial evaluations. 

Based on 70 academic and nonacademic variables, a prediction model was developed supported by Artificial Intelligence with unassisted random tree networks. The purpose of this Learning Machine algorithm is to find matching patterns among the 70 variables defined to generate a correlation model of students’ performance. It is important to mention that the similarity criteria are not selected in advance; the algorithm, learns the possible coincidences with each analyzed student and will generate patterns with the students being studied.

Therefore, a regression model was obtained, which accurately determines the evaluations that the students achieved at the end of the course, since they had 100% of the students’ data. In this type of studies, the Machine Learning algorithm establishes patterns through the population’s coincidences for certain types of performance, and is not defined at any time by the researchers.

The question raised by the research professors was: Is it possible to apply this model to a population where the academic and nonacademic information is unknown? Since they wanted to use this model in a new population, they decided to select only one nonacademic variable from the 70 study variables. They used facial biometry as a user identifier to be able to implement the performance prediction study.

In this type of studies, the Machine Learning algorithm establishes patterns through the population’s coincidences for certain types of performance, and is not defined at any time by the researchers.

They applied the calibrated model among the study population of almost 350 students, divided into 12 groups, from diverse courses and with different teachers. The results at the end of the school period in comparison with the prediction model were amazing. Of the 12 groups that were studied, the accuracy of the prediction ranged from 96%-98% so that, from this perspective, it can be affirmed that the model makes it possible to predict students’ academic performance.

This solution is highly accurate with a deterministic and non-probabilistic model. Currently it is still being studied, validating the process and prediction model, but we are encouraged by the results obtained. Above all, many positive implications are envisioned in the near future. This valuable information can be used to evaluate educational processes, take preventive academic engagement actions and design efficient adaptive courses, among others.

This study confirms that we can improve attention to individual learning needs, which is why we invite all teachers to learn more about the use of these new technologies and to take advantage of what predictive information, through intelligent and dynamic algorithms, can offer us as improvement tools in our courses.  


About the authors
Omar Olmos López holds a Doctorate in Physics and is a full-time professor at Tecnológico de Monterrey, Toluca Campus. His research focuses on new evaluation models, adaptive learning, artificial intelligence applied to education and teaching models through active learning supported by technology. He won the institution’s educational innovation prize every year from 2010 through 2013.

Miguel Ángel Hernández is a professor at Tecnológico de Monterrey, Toluca Campus. He is also a specialist and consultant in areas of cryptography, mathematical modeling, complex systems, data scientist, Big Data and adaptive models supported by Artificial Intelligence. He has developed educational innovation in the areas of personalized and adaptive learning, and predictive models for academic performance. He has created diverse tools and applications for national and international institutions in the Machine Learning area.