About

KGs represent a convenient data management solution allowing fast and easy data access and sharing. KGs maintain knowledge in a structured form, often organized according to an underlying schema, represented by expressive domain ontologies. They offer the ideal support to knowledge-intensive tasks. This is witnessed by the variety of domains in which KG-based machine/deep learning models have successfully been applied, such as complementing search results and drug repurposing, for citing a few examples.

However, in complex domains, besides providing accurate predictions, these models should also allow one to understand the why of the predictions they provide. This raises the need for explanation tools that can help experts in making more informed decisions.

Expected Results

HypeKG will offer three main contributions:

  • a class of effective semantic-based embedding methods flexible enough to be applied in a variety of learning tasks, including link prediction, node/graph/triple classification;
  • a class of methods, grounded on neural-symbolic integration, providing solutions to tasks like link prediction and node/graph/triple classification;
  • a class of methods, grounded on formal semantics and graph pattern, providing explanations to link prediction and node/graph/triple classification tasks;
  • a proof of concept, targeting the legal domain. Specifically, on the ground of rulings and documents encoded within KGs, the legal operator will be facilitated by our methods in the search for precedents and the predictability of the outcomes of legal actions. This, in any case, in compliance with art. 22 of regulation no. 2016/679 on the protection of individuals with regard to the processing of personal data and the free circulation of such data (GDPR), which affirms the right not to be subjected to a decision based solely on automated processing. To the best of our knowledge, this is the first solution exploiting both semantic technologies and a combination of reasoning paradigms applied to predictive justice.

HypeKG will deliver an open, freely available, and domain-independent general framework to be applied to any existing KG, thus laying the foundations for a wide and easy adoption of HypeKG results. Experimental studies on general, domain-specific, and benchmark KGs will be conducted and made also publicly available. The use case in the legal domain will hint at the societal and economic impact of the developed solutions.

HypeKG’s results will show how empowering ML methods with semantic capabilities can advance the completeness of state-of-the-art KGs in terms of additional and correct knowledge that may be learned/suggested. Improving the overall quality of the existing KGs will have fallout on several domains where KGs are strongly adopted, as well AI solutions, e.g., digital assistants, largely exploiting KGs. Furthermore, HypeKG aims at filling the vital need for explanation tools (required both by domain experts and practitioners) by supplying interpretable white-box solutions complementing the existing black-box KG-related systems.

Objectives

HypeKG sets three main objectives:

  • Obj1: designing semantically-aware Graph Neural Networks (GNNs) to tackle tasks like link prediction, node/graph/triple classification over KGs. At the core of Obj1, there will be the design of scalable hybrid methods that intertwine GNNs machinery, node and edge types, and KG schema information.
  • Obj2: empowering embedding-based solutions for link prediction and triple classification with (formal) semantics provided by domain ontologies. Obj2 focuses on enhancing embedding methods by injecting formal semantics and interleaving inductive and semi-supervised learning with deductive reasoning capabilities.
  • Obj3: explaining the output of our models in link prediction and node/graph/triple classification tasks. Here the focus is on designing novel methods that, building upon the neural-symbolic integration envisioned in Obj1 and Obj2, can provide interpretable explanations that can support users in making more informed decisions.