Harnessing Knowledge Graphs for Explainable AI

Korte inhoud:Artificial intelligence (AI) has become an integral part of our daily lives, influencing decisions in diverse domains such as healthcare, finance, and autonomous systems. Despite its widespread adoption, a major limitation of modern machine learning (ML) models—especially deep neural networks—is their lack of transparency. These models often function as black boxes, making it difficult to understand the reasoning behind their decisions. This opacity raises concerns about trust, bias, and accountability, particularly in high-stakes applications. Furthermore, neural networks are known to learn spurious correlations from training data, leading to unreliable and biased predictions. Addressing these challenges requires explainability methods that offer meaningful insights into the model's decision-making process without relying on labour-intensive manual annotations or subjective interpretations. To overcome these limitations, this thesis investigates the use of knowledge graphs as structured, information-rich resources that can mitigate the need for manual annotations and facilitate structured explanations. Specifically, it introduces novel methods integrating knowledge graphs to enable unsupervised explainability in machine learning models, focusing particularly on image classifiers and knowledge graph embedding models. First we introduced a framework to explain neuron activations in convolutional neural networks by mapping them to semantic attributes extracted from commonsense knowledge graphs. This method enables a more intuitive understanding of what features influence a model's predictions. Next we tackled the interpretability of Knowledge graph embedding models by introducing a feature selection-based framework, FeaBI, to interpret knowledge graph embeddings. By identifying key features that influence learned embeddings, FeaBI provides insights into how knowledge graph representations are formed, improving interpretability for downstream tasks. To make these explanations more accessible, an interactive tool is developed, allowing users to compute entity similarities and visualize their explanations dynamically. Finally, we combined the structure of knowledge graphs with the power of foundation models to introduce a novel framework for discovering learned patterns in ML models. This framework extracts key patterns that influence classifier decisions. The extracted patterns are used for error analysis and bias detection, providing deeper insights into dataset composition and model behavior. In the process of refining explainability methods, a new approach is also developed for generating comprehensive image descriptions, capturing contextual elements such as background, object relationships, and environmental factors. This enhances interpretability beyond simple object detection, revealing subtle patterns that might otherwise go unnoticed. By integrating structured knowledge representations with advanced foundation models, this thesis offers an automated approach to explainability, addressing critical gaps in AI transparency. The proposed methods enhance model interpretability, facilitate bias detection, and improve debugging, contributing to the development of more trustworthy and accountable AI systems.

Jaar van publicatie:2025

Toegankelijkheid:Embargoed