This is a simple side-project showcasing the use of Data Anlysis, Visualization and Machine Learning to predict the adaptability of K-12 students to remote learning. The project is still in development.
Install the required packages using pipenv:
pipenv installThen after choosing the right kernel, run the notebook.
Online Education was suddenly a reality for 2 (or more) years for some students during COVID. It allowed students to connect with teachers and colleagues from the safety of their home. Contrary to some institutions and countries who could adapt pretty quickly to this new reality, students from developing countries were the most affected by this change. Government and Education Insitutions were, for the most part, not prepared to adapt to this new reality.
But it was not just education institutions or governments that were affected. A big share of students had to adapt to learn, most of the times, through a phone or tablet. The following dataset was created to study the adaptability of students to Online Learning. The surveys were conducted in Bangladesh. The dataset can be found in Kaggle HERE and the Conference Paper generated by the authors can be found HERE.
If we know how to predict the adaptability of students to Online Learning, that would allow Governments, schools, universities, etc. to act faster and in a more personalized way to try to answer these limitations of online learning.
It seems we can predict with some accuracy what are the students which can adapt better to online learning. Financial Condition, Age and Class Duration seem to be the most important features for this problem, at least for the final model used.
Looking at the problem and the research done on it, these results seem to be accurate. Financial Condition, Age and Class Duration have all been subject of different studies and researches to see if they impact the academic performance of students, but not many studies have been done about these subjects on K-12 students and in respect to Online Learning.
In terms of technical conclusions, out of all the tested models, the best model for this problem is the Random Forest Classifier. It has the best accuracy and the best F1 Score.