Towards an interpretable model of learners in a learning environment based on Knowledge Graphs

PhD Defense of Antonia Ettorre, Monday, November 28th, 2022, Campus SophiaTech

Title: Towards an interpretable model of learners in a learning environment based on Knowledge Graphs

Abstract: In recent years, society has demonstrated increasing need for more effective, comprehensive and easily accessible educational resources.
A growing number of people around the world have gained access to online education and demand more efficient tools to enable learning anything, anywhere, at any moment.
This requires the development of smarter educational systems, which should be able to improve users’ learning curves and effectively assist them in their knowledge acquisition process, possibly relying on no, or very little, human support.
A major step to concretize this vision lies in personalizing the learning process to be specifically adapted to every single user, taking into account their background, learning style, personal needs and objectives.
To create such adaptive and customized environments, a major requirement is represented by the ability to trace user knowledge over time and assess whether they have the capacity to face a specific problem, exercise or question.
This problem, known in the Education community as Knowledge Tracing, has been widely investigated in the last 50 years and several resolution approaches have been proposed.
Though their performance improved sensibly over the last decade, such approaches present several shortcomings: from the excessive simplicity in the representation of the learning environment, which does not account for complex scenarios such as acquisition of soft skills or solution of group assignments; to the impossibility of interpreting the provided predictions, e.g. explaining why a student will fail while trying to answer a given question.
In this thesis, we try to overcome these issues by exploring and proposing approaches based on the use of Symbolic AI approaches focusing on Graph based Knowledge Representation and Reasoning.
Firstly, we propose a Knowledge Tracing approach that extends an existing model by introducing, as additional input features, Knowledge Graph Embeddings. 
Secondly, we investigate the explainability of the proposed approach by seeking the interpretation of the employed Graph Embeddings.
This leads us to the implementation of a tool for the joint visual analysis of Knowledge Graphs and Graph Embeddings and to the development of an approach to verify the information encoded by such Graph Embeddings.
Finally, we present a Knowledge Tracing model exclusively relying on the representation of the learning environment in the form of a Knowledge Graph, which does not require any additional external model for the prediction.