Sequence to Sequence Models: Knowledge Tracing with Deep Learning
Johannes Nawrath · Open access LMU (Ludwid Maxmilian's Universitat Munchen) · 2021
This thesis consists of two parts.At first, various methods of modeling sequential data with neural networks are presented.The idea is to show the development of sequential deep learning models over time.For this purpose, the problems of early approaches are highlighted in order to better understand the methods in more modern solutions.More precisely, the thesis starts off with simple recurrent neural networks and moves on to gated recurrent units.Furthermore, attention mechanisms are discussed, which ultimately result in the transformer architecture.The second part of the thesis demonstrates that modern sequential deep learning models can be applied successfully to the task of knowledge tracing, the ability to model the knowledge of students over time, as they interact with coursework.Various model configurations are trained on real world data from almost 400K students as part of a competition on Kaggle.The models achieve state-of-the-art results and trump other machine learning algorithms, which cannot exploit the sequential data structure well enough.