Periodic Reporting for period 2 - ENEXA (Efficient Explainable Learning on Knowledge Graphs)
Reporting period: 2024-04-01 to 2025-09-30
Knowledge graphs are a family of data structures that are used in an increasing number of user-facing applications. They are now widely regarded as a key enabler for explainable machine learning for the masses. However, the knowledge graphs used in real applications are often large, inconsistent and incomplete. Hence, they come with three main challenges. First, they are challenging to store and query, especially when analytical queries are required to gather high-value data. Second, standard inference engines cannot be used on them, as they simply terminate their computation when faced with inconsistencies. Third, their incompleteness often leads to downstream applications remaining unaware of highly relevant assertions. Finally, their non-vectorial representation makes it challenging to deploy classical machine learning techniques for analysis or exploitation.
The goal of ENEXA was to address these challenges. To achieve this goal, the project specified, implemented, and evaluated novel extraction, storage, inference, and explainable machine learning techniques for knowledge graphs at large scale. All approaches in the project were evaluated in knowledge graphs with sizes of 1 billion triples or more. Our results clearly show that our approaches are now ready to empower small and large European enterprises to use knowledge graphs reliably as cornerstone for their data strategy.
The goal of ENEXA was to address these challenges. To achieve this goal, the project specified, implemented, and evaluated novel extraction, storage, inference, and explainable machine learning techniques for knowledge graphs at large scale. All approaches in the project were evaluated in knowledge graphs with sizes of 1 billion triples or more. Our results clearly show that our approaches are now ready to empower small and large European enterprises to use knowledge graphs reliably as cornerstone for their data strategy.
ENEXA delivered the following core innovations:
- We implemented Tentris, one of the world‘s most efficient graph databases. Our solution is up to 1000x faster than the state of the art on large datasets such as WikiData with real and synthetic query loads.
- We developed a unification theory for graph embedding techniques based on Clifford algebras. This theory show that are now able to learn embedding spaces as well as embeddings and achieve significantly higher embedding accuracy.
- We implemented a modular reasoner for large-scale knowledge graph that can detect and alleviate inconsistencies in large knowledge graphs with over 1 billion triples.
- We developed several concept learning algorithms that exploit reinforcement learning and supervised to achieve runtime improvements of an order of magnitude or more while retaining the accuracy and explainability of the computed models.
- We created the first retrieval-augmented generated explainer for concept learning that implements co-constructive explanations of machine learning results.
- We implemented Tentris, one of the world‘s most efficient graph databases. Our solution is up to 1000x faster than the state of the art on large datasets such as WikiData with real and synthetic query loads.
- We developed a unification theory for graph embedding techniques based on Clifford algebras. This theory show that are now able to learn embedding spaces as well as embeddings and achieve significantly higher embedding accuracy.
- We implemented a modular reasoner for large-scale knowledge graph that can detect and alleviate inconsistencies in large knowledge graphs with over 1 billion triples.
- We developed several concept learning algorithms that exploit reinforcement learning and supervised to achieve runtime improvements of an order of magnitude or more while retaining the accuracy and explainability of the computed models.
- We created the first retrieval-augmented generated explainer for concept learning that implements co-constructive explanations of machine learning results.
See section 2. The corresponding publications are at https://enexa.eu/publications/(opens in new window). Key needs to ensure further uptake will be elaborated upon during the second phase of the project.