Abstract

Knowledge graphs are structured representations of real-world information, where entities are connected by edges with labels known as predicates. These graphs contain logical patterns, known as inference patterns, such as symmetry, transitivity and composition, to name a few. To leverage the information contained within a knowledge graph, embedding methods have been developed to encode entities and predicates as numerical vectors in a multi-dimensional space. These embeddings aim to capture the semantic and structural information of the graph, including the inference patterns. However, the sub-symbolic nature of these embeddings present significant challenges in terms of interpretability. Additionally, each embedding model utilizes different embedding representations and scoring functions, adding to the complexity of interpretations. Therefore, it is important to develop interpretation methods that do not solely rely on the embeddings of the model, and can be applied to any knowledge graph embedding model. To address these challenges, this dissertation proposes model-agnostic methods for interpreting knowledge graph embeddings. By focusing on approaches that are independent of specific embedding architectures, we aim to provide a unified framework for analyzing and comparing model interpretability. Our methods leverage the link prediction evaluation task, a common protocol for assessing knowledge graph embedding models. During this task, we collect edges that are deemed plausible by the embedding model at hand into a separate knowledge graph, which serves as the basis for our interpretability analyses. The main contribution of this dissertation is twofold. First, we extract Horn rules from the knowledge graph of deemed-plausible edges to provide dataset-level (global) interpretations of a model's behavior. We introduce the concept of interpretation accuracy, which quantifies how well these extracted rules represent the edges deemed plausible. Interpretation accuracy enables quantitative comparisons among diverse knowledge graph embedding models. Second, we propose a method to analyze the ability of embedding models in capturing inference patterns. We propose an overlap metric that assesses how effectively an embedding model captures inference patterns present in the original graph. Both interpretability methods allow us to gain insights into the strengths, weaknesses of different embedding models. By providing a clear understanding of the logical relationships captured by the embeddings, these methods help identify potential biases or gaps in the models under evaluation.

Library of Congress Subject Headings

Graph theory; Embeddings (Mathematics); Knowledge representation (Information theory); Knowledge management; Big data; Information visualization

Publication Date

2-2025

Document Type

Dissertation

Student Type

Graduate

Degree Name

Computing and Information Sciences (Ph.D.)

Department, Program, or Center

Computing and Information Sciences Ph.D, Department of

College

Golisano College of Computing and Information Sciences

Advisor

Carlos R. Rivero

Advisor/Committee Member

Qi Yu

Advisor/Committee Member

Zack Butler

Campus

RIT – Main Campus

Plan Codes

COMPIS-PHD

Download

Share

COinS