Predicting academic performance - A practical study using Moodle log data and sociodemographic traits

Project Work presented as the partial requirement for obtaining a Master's degree in Data Science and Advanced AnalyticsWith the increase of computational power, usage of IT systems and tools in several industries also increased the amounts of data generated and stored. Education is one of these fields. The opportunity to utilize analytical and data related techniques to the data generated and stored by computer-based educational systems is more significant than ever. Performance prediction is one of the most popular uses for all the data generated by educational systems. In this line of thought, the main objective of this paper is to build a predictive model capable of classifying a student´s grade based on its Moodle system activity and several sociodemographic variables taken from the Netpa System. All the data used belongs to student´s that attended the first semester of 2019 at Nova Information Management School. To achieve the objective, SEMMA Methodology was implemented. Python Language was used, with particular emphasis on the Scikit-Learn, pandas and Seaborn packages. Raw Moodle logs were processed and transformed into variables that represented the number of times a student navigated to a specific page in the platform. This information was then joined with Netpa variables, and a dataset was built. Exploratory data analysis was performed, and several model configurations were tested. The main differences that separate the models are outlier treatment, sampling techniques, feature scalers, feature engineering and type of algorithm – Logistic Regression, K-Neighbours Classifier, Random Forest Classifier and Multi-Layer Perceptron. Using a K-Neighbours Classifier and the SMOTE sampling technique an F1-Score of 0.624 and a ROC AUC of 0.828 was obtained