Clustering feedback data with Logistic Network Lasso

Feedback gathered from university courses' are the most effective tool given to students to shape lectures according to their needs and difficulties. Nonetheless, feedback is often neglected by both students and teachers, even if studies show how feedback can improve the quality of the student's experience as well as being an essential tool for the teacher's development. With the improvement in Natural Language Processing and the Machine Learning field, it is now possible to analyze text and effortlessly gather insights. These insights ease the professor's work when faced with the task of reading feedback which, depending on the course, can be thousands. The Logistic Network Lasso is a novel semi-supervised machine learning algorithm that has been applied in the binary classification scenario and compared to algorithms such as Maxflow and Belief Propagation, resulting in overall increased accuracy. This thesis aims to answer two research questions: Firstly, the performance of various standard clustering algorithms such as K-Means, Affinity Propagation, Spectral Clustering and DBSCAN, are compared using different Natural Language Processing techniques to encode university courses' feedback, showing which embedding techniques are better in terms of clustering feedback data. Secondly, the standard clustering algorithms are compared with the Logistic Network Lasso incorporating the best embedding techniques discovered during the first research question. Applying the Logistic Network Lasso as a clustering algorithm resulted in up to a 30% increase in the external indexes clustering performance metrics that were analyzed.