University of Crete (UOC)
Taming knowledge graphs heterogeneity and bias in entity alignment
Abstract
dc:descriptionEntity alignment (EA), also known as entity resolution (ER), aims to identify pairs of descriptions from different Knowledge Graphs (KGs) that refer to the same real-world entity. In recent years, there has been increasing interest in leveraging representation learning techniques in order to find matching pairs of entities. The idea is to learn a continuous representation of KGs in a low-dimensional vector space that preserves the similarity of entities (nodes) based on the structural and factual information of their descriptions in the two KGs. Then, EA is solved as a similarity-based matching problem of the entity embeddings produced by various Neural Network architectures. In this thesis we show that both structural heterogeneity (in terms of dissimilar and poor neighborhoods) and factual heterogeneity (in terms of dissimilar literal values, names) of entities severely impact the effectiveness embedding-based methods for EA, especially when the two KGs exhibit simultaneously both forms of heterogeneity at different degrees. It should be stressed that structural diversity may additionally lead to unfair matching decisions due to missing facts or relations of entities belonging to sensitive populations. To our knowledge, no previous work has studied how the factual and structural heterogeneity of KGs affects both EA performance and fairness. More precisely, we propose a hybrid system called HybEA, that exploits both the entity relations and facts. The HybEA factual component is capable of weighting the contribution of each property in the prediction of matching entities using a new attention-based model based on Transformers and contextualized attribute value embeddings of Sentence-BERT. Then, several hybrids can be formed by plugging different structural models that exploit the relations connecting neighbor entities in KGs at the level of triples (e.g., Knowformer or graphs (e.g., RREA). As each component model predicts its own entity alignments, we propose a new semi-supervised framework, where at every iteration, some high-confidence matching pairs detected by each model are added to the training set used for the subsequent co-training of both models. Intuitively, one component model can compensate for lost matches of the other, leading to a high adaptability of HybEA to different types and levels of heterogeneity exhibited by KGs in real applications. HybEA outperforms the state-of-the-art EA systems, achieving a 16% average relative improvement of Hits@1, ranging from 3.6% up to 40% in 5 monolingual datasets, with some datasets that can now be considered as solved. We also show that HybEA outperforms state-of-the-art methods in 3 multi-lingual datasets, as well as on 2 datasets that drop the unrealistic, yet widely adopted, one-to-one assumption. Overall, HybEA outperforms all (11) baseline methods in all (3) measures and in all (10) datasets evaluated, with a statistically significant difference. To assess the robustness of EA systems to structural bias in KGs, we propose an exploration-based sampling algorithm, named SUSIE, which samples subgraphs of varying structural settings based on the connectivity of KGs. We experimentally show that HybEA is the most robust method, exhibiting high tolerance in structural variations (e.g., H@1 drops only 8.79% in the most extreme case). On the basis of our previous contributions, we have developed a fairness-aware end-to-end system that allows experimentation, both on the adaptability of the entity alignments to different aspects of heterogeneity and their robustness to bias related to different connectivity measures of the two KGs. The system incorporates both HybEA and SUSIE, while statistical fairness measures are used, with the protected groups being defined on the basis of the structural characteristics of the two aligned KGs.
Degree
thesis:*- Grantor dc:publisher
- University of Crete (UOC)
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Fanourakis, Nikolaos
- Φανουράκης, Νικόλαος
Subjects
dc:subject × 8Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/58986