[Paper Review] Knowledge Graph Embeddings and Explainable AI

Federico Bianchi, Gaetano Rossiello+3 more|arXiv (Cornell University)|Apr 30, 2020

Explainable Artificial Intelligence (XAI)101 references30 citations

TL;DR

This chapter introduces knowledge graph embeddings, explains how they are generated and evaluated, surveys state-of-the-art approaches, and discusses how embeddings can support explainability.

ABSTRACT

Knowledge graph embeddings are now a widely adopted approach to knowledge representation in which entities and relationships are embedded in vector spaces. In this chapter, we introduce the reader to the concept of knowledge graph embeddings by explaining what they are, how they can be generated and how they can be evaluated. We summarize the state-of-the-art in this field by describing the approaches that have been introduced to represent knowledge in the vector space. In relation to knowledge representation, we consider the problem of explainability, and discuss models and methods for explaining predictions obtained via knowledge graph embeddings.

Motivation & Objective

• Define what knowledge graph embeddings are and how they are represented as low-dimensional vectors.
• Explain how KG embeddings are generated and trained, including scoring and loss functions.
• Survey major families of KG embedding models (translational and bilinear) and key examples.
• Discuss how additional information (text, images, logic) can enhance embeddings and their explainability.

Proposed method

• Represent KG as a 3D adjacency tensor and embed entities and relations in a vector space.
• Describe the TransE scoring function and its training via margin-based loss with negative sampling.
• Outline extensions that incorporate extra information (e.g., DKRL combines structure and description).
• Categorize embedding models into translational, bilinear, neural, and recent neural-symbolic approaches with representative examples (e.g., RESCAL, DistMult, ComplEx, RotatE, HAKE).
• Explain the role of negative sampling (e.g., corrupted triples) and adversarial techniques (e.g., KBGAN) in training.
• Discuss evaluation emphasis on link prediction and how datasets and hyperparameters influence results.

Experimental results

Research questions

• RQ1 What are the main categories and representative models of knowledge graph embeddings?
• RQ2 How are KG embeddings trained and evaluated, particularly for link prediction?
• RQ3 What approaches exist to enhance embeddings with external information (text, images, rules) and how do they impact explainability?
• RQ4 What are the challenges and prospects for explainability in KG embeddings?

Key findings

• Translational models (e.g., TransE) and bilinear models (e.g., RESCAL, DistMult, ComplEx) form the core KG embedding families.
• Models like ComplEx and RotatE address asymmetry and relational properties more effectively than simpler diagonal/distance-based variants.
• Enhancements using text, images, and logical rules can improve representations and enable zero-shot or multimodal capabilities.
• Explainability remains challenging since embeddings encode latent factors; recent work explores integrating reasoning and logic with embeddings to aid explanations.
• There is a broad spectrum of advanced approaches (e.g., TKRL, HAKE, TuckER, COKE) and ongoing exploration of neural-symbolic methods for better interpretability.