Explainable and Interactive Link Prediction in Knowledge Graphs
TL;DR
This thesis addresses the challenge of making link prediction models correct for the correct reasons with the help of human feedback by proposing a technical and conceptual pipeline in which model explanations serve as interfaces for collecting and incorporating user feedback to iteratively align model behaviour with human intent.
Abstract
Machine learning models move from the safe walls of research labs to high-risk real-world applications. In these environments, users need to understand why a machine learning model behaves as it does and correct it if necessary. One critical application domain is Root Cause Analysis in manufacturing, where data-driven models identify fault causes in time-critical environments. Link prediction models in Knowledge Graphs can be used to tackle this task. However, Knowledge Graph Completion models, particularly embedding-based link predictors, often operate as black-boxes. This limits human oversight, complicates validation, and undermines trust. This thesis addresses the challenge of making link prediction models correct for the correct reasons with the help of human feedback. Specifically, it proposes a technical and conceptual pipeline in which model explanations serve as interfaces for collecting and incorporating user feedback to iteratively align model behaviour with human intent. Three research questions are addressed in this work: (1) How can the decision-making process of a link prediction model be explained in ways that support downstream alignment? (2) What types of user feedback on explanations are effective for guiding model updates? (3) How can feedback on model explanations be integrated into the training and inference processes of link prediction models? In response, three major contributions are made in this thesis: First, KGEPrisma is introduced, a post-hoc explanation method for embedding-based link prediction models. KGEPrisma translates latent embeddings into instance-based, analogy-based, and rule-based explanations. This method provides a multi-perspective view on model decisions and establishes a foundation for the subsequent exploration of feedback modalities. Second, RootFinder is presented, a system that treats Root Cause Analysis as a link prediction task in a manufacturing Knowledge Graph. RootFinder enables domain experts to assess model explanations and to provide simple, customised feedback for real-world settings. This application motivates an in-depth discussion of feedback modalities. Finally, LiEr is introduced in the thesis. LiEr is a reinforcement learning approach that iteratively aligns link prediction with preference-based feedback over model explanations. LiEr reformulates Knowledge Graph Completion as a Markov Decision Process and incorporates preference-based user feedback into the reward function. This enables the model to learn policies that converge toward human-plausible reasoning. LiEr is applied to Root Cause Analysis in electric vehicle manufacturing. The case study confirms that preference-based feedback over explanations enables root cause prediction alignment with domain knowledge. Together, these contributions provide a methodological and empirical foundation for using explanations to collect human feedback and align the behaviour of link prediction models with it. Explanations enable an interactive learning loop. This thesis presents the first work on interactive and explainable Knowledge Graph Completion.