Trust, Accountability, and Autonomy in Knowledge Graph-Based AI for Self-Determination

Languages to express policies, including but not limited to data access, can be categorised as either general or specific. In the former, the syntax caters to a diverse range of functional requirements (e.g. access control, query answering, service discovery, negotiation), whereas the latter focuses on just one functional requirement. In the early days of the Semantic Web, research into general policy languages that leverage semantic technologies (e.g. KAoS [108], Rei [48], AIR [51], and Protune [15]) was an active area of research. However, despite the huge potential offered by these general-purpose languages to date, none of them achieved mainstream adoption [55]. More recently, researchers have proposed ontologies that can be used to represent licences, privacy preferences, and regulatory obligations [53]. When it comes to the legal domain specifically, Semantic Web researchers have proposed cross-domain ontologies that can be used to encode legal text in a machine-readable format using LegalRuleML¹²¹²12https://docs.oasis-open.org/legalruleml/legalruleml-core-spec/v1.0/os/legalruleml-core-spec-v1.0-os.html and adaptations thereof (e.g. [3, 80]). Others focus on facilitating legal document indexing and search using the European Law Identifier (ELI)¹³¹³13https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:52012XG1026(01) and the European Case Law Identifier (ECLI)¹⁴¹⁴14https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:52011XG0429(01) (e.g. [78, 22]), or bridging the gap between the EU and member state legal terminology (e.g. [1, 14]). Besides these cross-domain activities, there have also been various domain-specific initiatives. For instance, the ELI ontology has to be extended to facilitate the encoding of the text of the General Data Protection Regulation (GDPR)¹⁵¹⁵15 https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A32016R0679&qid=1681238509224 (e.g. [83]). At the same time, others have focused specifically on modelling privacy policies (e.g. [79, 82]). The Open Digital Rights Language (ODRL)¹⁶¹⁶16https://www.w3.org/TR/odrl-model/, which is a W3C recommendation, has gained a lot of traction in recent years in terms of intellectual property rights management (e.g. [42, 74]). Additionally, the ODRL model and vocabularies have been extended in order to model contracts [39], personal data processing consent [32], and data protection regulatory requirements [110]. There has also been some work on automatically extracting rights and conditions from textual documents (e.g. [21, 20]) or extracting important information from legal cases (e.g., [116, 77]). Although many of the proposed approaches are based on existing standards, there is a lot of overhead involved for systems that need to consider different types of policies that are encoded using different languages. General-purpose policy languages are particularly attractive in such scenarios as they lessen the administrative burden. However, considering the potential complexity of such a language, there is a need for policy profiles with well-defined semantics and complexity classes.

From a policy governance perspective, LegalRuleML researchers have proposed automated compliance approaches based on auditing (e.g. [28, 82]) and business processes (e.g. [81, 9]). While [37] shows how LegalRuleML, together with semantic technologies, is used for business process regulatory compliance checking based on a rule-based logic combining defeasible and deontic logic. One of the advantages of description logic-based approaches, when it comes to consistency and compliance checking, is that they can leverage generic reasoners, such as Pellet¹⁷¹⁷17https://github.com/stardog-union/pellet (e.g. [33]). Although there are presently no ODRL-specific reasoning engines, researchers have demonstrated how ODRL can be translated into rules that can be processed by Answer Set Programming (ASP) [8] solvers such as Clingo [35] (e.g. [42, 110]). Additionally, there have been several custom applications that are designed to support ODRL enforcement or compliance checking, such as a licence-based search engine [74]; generalised contract schema and role-based access control enforcement [39]; and access request matching and authorisation [32]. Despite existing efforts, challenges arise when it comes to ensuring that AI and processing algorithms adhere to the policies. This could potentially be achieved either before or during processing using Trusted Execution Environments (TEEs) [10] or after execution by detecting data misuse via automated compliance checking using system logs [54]. The combination of ex-ante and ex-post compliance checking is particularly appealing for supporting risk-based conformance checking such as that envisaged in the proposed EU AI Act. Nevertheless, the practicality, performance and scalability of these proposals remain to be determined. In order to further support self-determination, data owners and processors should be able to engage in on-demand negotiation over policies, assisted by technology that ensures a safe and fair space and helps assess the compliance of negotiated terms with existing regulations. Negotiation between automated agents has been a topic of interest since the early 2000s, but in the context of self-determination, we must pay attention to the right balance between artificial representation and human involvement [5, 6].

Figure 3 illustrates how machine-readable policies and norms can be used to support self-determination. Considering our illustrated scenario (Figure 2), individuals may want to establish policies to precisely define the subset of their PKGs to be shared with communities and what forwarding they allow. For example, share with the diabetes community my blood in sugar values measured by my connected device and the output of my AI healthcare assistant, or only share and forward anonymised aggregates to medical research institutions, or contact me for negotiation if the pharma company is interested in using my data for clinical studies. Communities may do the same, e.g. requiring specific data to be shared to join the community, but also requiring agreements in order to ensure that participants will abide by social and behavioural norms needed for self-regulation. Public and private organisations may need to adhere not only to privacy preferences and licences but also to various general regulations, e.g., the GDPR, the proposed AI Act in the EU or the Health Insurance Portability and Accountability Act (HIPAA)¹⁸¹⁸18https://www.hhs.gov/hipaa/index.html in the US, as well as domain-specific regulations (e.g. advanced therapy medicinal products¹⁹¹⁹19https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=celex%3A32007R1394 and rare diseases²⁰²⁰20https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=celex%3A32009H0703%2802%29).

The concept of a Personal Knowledge Graph (PKG) is that an individual can keep their personal or private data in a space belonging to them, rather than with siloed centralised service providers with limited access and control [7]. A Solid pod²¹²¹21https://solidproject.org/ is an example of a PKG platform, and the key to the vision of Solid is that there should be standard interfaces and authorisation models to grant or deny access to the contents of a PKG at a granular level. This is argued in particular²²²²22https://ruben.verborgh.org/blog/2017/12/20/paradigm-shifts-for-the-decentralized-web/ to enable a highly decentralised architecture for Web applications. Rather than a provider aggregating data from all users into a single location controlled by the provider and application code accessing such data there, an individual permits (or does not permit) Web applications of their choice to access whatever subsets of their data they decide from their PKG. As well as autonomy, this enables greater accountability since access to the PKG can be filtered via personal machine-readable policies at source, and activities can be tracked directly (e.g. [29]). Although PKGs offer great potential, they also come with challenges in terms of performance and scalability, as applications will need to interact with multiple distributed data sources as opposed to a single backend server. These challenges, however, may also simultaneously be opportunities for scalability trade-offs, querying over multiple low-powered data sources rather than a high-powered central one.

Distributed Ledger Technology (DLT) [104] promotes trust and empowerment through the replication of data across contributing nodes, which are geographically distributed across many sites, and the use of consensus algorithms which enable collective fair decision-making with no central control. Blockchains are a type of distributed ledger where an ever-growing list of records in blocks is tied together with cryptographic hashes, often, although not necessarily, associated with a securely exchangeable token system or “cryptocurrency”. This technology rose to prominence following the release of Bitcoin [76] in 2008 - a blockchain-based currency that has now been adopted by El Salvador as their legal tender. Ethereum [115], a blockchain platform released in 2015, contains the notion of a “Smart Contract” [19] (originally coined in the 1990s by Nick Szabo [105]), which is a collection of code that executes in a fully decentralised way. Smart Contracts have been used to implement a range of decentralised applications, including Decentralised Autonomous Organisations (DAOs) [64], which are organisations where decisions are made through blockchain consensus mechanisms. The best-known example of a DAO was “The DAO”, which at one point was worth more than $70M; they have been applied to a number of different activities, including scholarly publishing [43]. Despite the fact that immutability and transparency guarantees offered by DLT are very attractive when dealing with personal data, both the ledgers and the smart contracts themselves will need to be protected against unauthorised access and usage. Personal data itself is neither stored in, or derivable from, immutable DLT records. Smart contracts may also introduce scalability issues: the default Ethereum model involves every contributing node executing every run of a smart contract and thus has inherent scale limitations. Relaxing this model may, however, affect trust.

In the Web space, Self-Sovereign Identity (SSI) is being developed through a combination of Decentralised Identifiers (DIDs) [101] and Verifiable Credentials (VCs) [102], W3C standards for identity and verifiable attestation claims, respectively. DLT is one of the ways in which DIDs can be grounded, although, by design, the DID standard is open in terms of method. A DID is a URL (did:<method>:<...>) which can be resolved in a method-specific manner (e.g. HTTP(S) dereferencing, reading from a smart contract, etc.) to obtain a DID document, a Linked Data set containing information about digital identity in a standard form - for example, how to verify it (e.g. a public key), methods for communicating with the entity controlling it, and so on. DIDs enable SSI; the creation and use of DIDs are open and decentralised, and by using different DIDs with different audiences, individuals can minimise how easily their information can be tracked or correlated across services and can contextually and selectively disclose personal information as desired. This grants individuals significantly greater autonomy than current practices. There is a potential trade-off with trust and accountability of an individual when it comes to information that others need to rely on, which is that effective anonymity of a unique DID can be used to misrepresent oneself (e.g. fake a qualification or entitlement) or pretend to be someone else. VCs are a proposed solution to this. The VC data model is for sharing data alongside information that a recipient can use to verify its integrity or origin, such as a digital signature or DLT record. If a DID is presented to a service that is restricted to legal adults, for example, the DID owner may also present a VC issued by a government body confirming their adulthood; methods for selective disclosure supported by both DID and VC standards allow this to be done verifiably without requiring disclosure of real-world identity. These technologies are relatively new in comparison with standard digital identity models, and while intended and designed to address issues in those models, they may also introduce new difficulties or enable different vulnerabilities to, e.g. identity fraud, than current standards.

In the context of data-driven AI and decentralised infrastructure, there are also techniques for decentralised machine learning. Federated Learning (FL) [118] is the idea that rather than aggregating training data in one location controlled by a model developer (thereby compromising subject privacy), data holders can run learning algorithms to generate model weights for their own data locally and privately, and then send only the weights to the developer to be incorporated into the larger model. An example might be a smartphone text prediction personalisation algorithm, where a user’s own writing is used to generate predictive weights on the device, where periodic selections of these can be aggregated to improve general text prediction models. Refinements of FL approaches include sending not the actual learned model weights but a set of weights with statistically similar properties [112] to further reduce the risk of privacy breaches without affecting model performance. A related approach takes this concept even further, with the idea of embeddings in a larger model, e.g. “Textual Inversion” [34] to personalise large generative image diffusion models. The intuition here is that if someone wants certain personalised types of output from a generative AI, then if a model is sufficiently large, there is a good chance that the desired concept already exists within it. More recently, the idea of federating for preserving privacy has been applied specifically to deep learning, in particular in the context of the Internet of Things. [119] proposes an architecture with a control layer including a distributed ledger, while [117] propose advanced cryptographic mechanisms to reduce the risk of privacy leaks, following more general approaches that apply either differential privacy, homomorphic encryption or secure multi-party computation. Federation also has the positive side effect of potentially speeding up model training when the privacy constraints allow for a helpful distribution of the process [11]. However, when opening the process to multiple parties, there are a number of attack vectors that do not exist in a centralised approach for which we need protection and pay a communication and computation overhead [44].

A decentralised infrastructure supporting self-determination for our illustrative scenario (Figure 2) is depicted in footnote 24. Health data is highly sensitive and private, and individuals may want or need to interact with multiple services where it is relevant, including KG-based AI systems. It thus makes sense to create a personal health knowledge graph (PKG) to be a comprehensive and interconnected representation of an individual’s health information, including their medical history, lifestyle choices, genetic data, and real-time health monitoring data from IoT devices. Data from various sources, such as wearable devices, mobile applications, electronic health records, and even genomic sequencing, can be linked together to form a holistic view of an individual’s health in such a personal health knowledge graph. An early example of a PKG was in [106], where medical, lifestyle, and IoT health monitoring data in a PKG was integrated into a (patient-focused) decision support system built around a public medically-curated KG representing cardiovascular risk factors, giving individuals the autonomy to gain deeper insights into their own health patterns and risks, identify correlations, and make more informed decisions.

More recently, BlockIoT [97, 98] aims to integrate health data seamlessly in a decentralised PKG using blockchain and KG technologies, addressing this trust aspect and using PKG-driven smart contracts to trigger the personalised recommendations for lifestyle modifications, medication adjustments, or even timely interventions by the healthcare providers. Furthermore, the PKG can serve as a powerful tool for healthcare beyond the individual. Communities of patients, providers, researchers, etc., or combinations thereof, can share knowledge about various aspects of, e.g., particular conditions, whether that is clinical evidence and best practice, peer advice and support on living with a condition, or data on novel or rare symptoms and side effects, with this knowledge used for support, care, or medical research across populations. De-identified and aggregated data from multiple individual KGs can be collected in community KGs, with trust securely established using DIDs and VCs, and accessed by community, practitioner, researcher, and service provider stakeholders, allowing for decentralised large-scale analysis and identification of broader health trends from multiple perspectives and intersecting factors. This can lead to advancements in disease prevention, treatment protocols, and the development of personalised medicine in a collaborative manner [99]. KG-based AI systems can be both trained and used across this ecosystem, with FL being applied to train larger models (e.g., the organisation models in Figure 2) and personalised embeddings used by individuals to get the best experience from their therapy bots and healthcare assistants while maintaining privacy and autonomy.

Efficient query processing infrastructures are fundamental for traversing decentralised KGs. There have been notable efforts such as Fedbench [92] in the past. However, these infrastructures should be capable of executing queries against the available KGs while respecting privacy and adhering to norms and policies. With the increasing emphasis on privacy protection with regulations such as GDPR, it is crucial to develop mechanisms that allow users to access and extract knowledge from KGs without compromising sensitive information or violating privacy regulations. Several research directions are worth considering to address the open challenges in decentralised KG management. Firstly, developing the formalisms to describe KG management semantically can provide a common ground for understanding and interoperability across different decentralised KG systems. Such formalisms can enable standardised representations of KGs in the form of ontologies and facilitate seamless integration and collaboration among diverse knowledge sources. Architectures supporting new protocols and standards specific to decentralised KGs are essential for establishing interoperability and seamless communication between knowledge sources and systems. By defining and adopting common protocols and standards, decentralised KGs can collaborate more effectively, share insights, and facilitate cross-domain knowledge discovery.

Note that if we add LLMs to the picture, their current training and execution processes are currently centralised. Decentralised KG management is useful for providing transparency in data used for their training. For approaches involving the interaction between LLM and KGs, the transparency of the LLM itself still depends on the owner.

Furthermore, explainable methods for data integration and curation, as well as KG validation and distribution, such as the Explanation Ontology for user-centric AI, are necessary to ensure the reliability and accuracy of decentralised KGs [23]. By providing transparent and interpretable approaches, users can have better insights into knowledge integration and validation, enhancing trust and accountability of the knowledge contained in the KG and the insights derived. This is especially critical because, in decentralised KGs, data may come from various sources and be represented in different ways. The standardised framework provided in the Explanation Ontology for representing domain-specific explanations of KG entities and relationships helps users and applications understand the meaning and context of the data in the KG. Provenance and traceability also play a vital role in decentralised KG management. Establishing mechanisms to track and validate the origin, history, and lineage of knowledge within KGs is crucial for accountability and the ability to trace back the sources and transformations that contribute to the resulting knowledge. The W3C Provenance Data Management standards [71] provides the basis for encoding provenance attributes in KGs, and subsequent nanopublications specification [38] has gained a lot of traction in the biomedical domains. While these solutions exist, there needs to be a cohesive framework that ties together explanation provenance data management in a decentralised KG context and ensures that users can trace the origins, transformations, and sources of the data, which is crucial for trust, accountability, and data quality assurance. The W3C provenance data management suite of recommendations provides normative interoperable guidance on recording information about data sources, contributors, and how data is collected or transformed, making integrating heterogeneous data into a coherent KG easier. When data quality issues arise, users can trace back to the source of the problem and take corrective actions, ensuring the KG remains accurate and reliable. The W3C recommendations for decentralised provenance management provide a mechanism for attributing data to its sources or contributors. This attribution is essential for accountability, especially when multiple parties contribute to a KG.

In recent years, the integration of blockchain technologies and tokenomics has gained attention in the context of decentralised KG management. Projects such as OriginTrail²⁵²⁵25https://origintrail.io have contributed to the development of ownable DKGs, which leverage blockchain’s inherent properties to enhance trust, provenance, and accountability. By utilising blockchain, KG management systems can ensure the integrity and traceability of data and metadata across various nodes in the network. The OriginTrail protocol aims to create a trustless environment where data providers, consumers, and verifiers can interact and validate the authenticity and reliability of data stored within the knowledge graph. Their protocol issues tokens as incentives for data contributors, validators, and curators within the KG ecosystem. The integration of blockchain technologies and tokenomics in decentralised KG management addresses several critical aspects. Firstly, blockchain’s immutability and transparency enable the traceability and provenance of data and metadata, ensuring accountability throughout the KG management pipeline. Secondly, the decentralised nature of blockchain mitigates single points of failure and promotes the distribution of knowledge and decision-making power among participants. This decentralised approach aligns with the principles of self-determination, empowering individuals to have control over their data and knowledge. By rewarding contributors, validators, and curators with tokens, these systems encourage continuous improvement, data quality assurance, and community engagement. Token-based economies can drive the development of sustainable KG management pipelines, enabling the growth and evolution of DKGs over time. However, the tokenomics have to be carefully designed and monitored to avoid the possibility contributors have a motivation (possibly extrinsic) to misbehave. There is also the risk that a sudden churn in blockchain participants impacts performance and availability. There is also the question of the performance of the consensus algorithm of a specific blockchain itself.

An approach to decentralised knowledge graph management in the context of healthcare in our illustrative scenario (Figure 2), where users retain control over their personal information while benefiting from enhanced privacy measures and seamless collaboration in a community, is illustrated in Figure 5. At the heart of this framework lies the concept of PKGs, such as Solid, which empowers individuals to store and manage their personal health data securely. Central to the architecture are specific components aimed at safeguarding user privacy and ensuring data transparency. The process begins with knowledge sanitisation, which anonymises sensitive information and filters the data according to the user’s preferences and data policies. These policies encompass not only globally recognised regulations like GDPR and HIPAA but also individual data policies, enabling users to set granular restrictions on how their data is used, such as opting out of genetic data usage for medical research. To ensure interoperability and standardisation, the creation of knowledge graphs leverages community-defined ontologies and vocabularies. These shared frameworks facilitate seamless integration and alignment of personal knowledge graphs within the broader ecosystem, promoting data exchange and collaboration. Users are incentivised to aggregate their knowledge graphs, contributing to the construction of community-based knowledge graphs focused on specific diseases. Through community-based verification, validation, and knowledge aggregation processes, these disease-based knowledge graphs are created, providing valuable insights and fostering collaborative efforts among healthcare professionals, researchers, and the wider community. Blockchain-based incentives drive user participation, rewarding both community users and healthcare experts for their verification, validation, and aggregation activities. The utilisation of an immutable ledger and verifiable credentials ensures the integrity and trustworthiness of the verification process. The validation process, powered by RDF SHACL and Shape descriptions, further enhances data quality and consistency, instilling confidence in the aggregated knowledge. The integrated knowledge graphs, encompassing personal, community-based, and healthcare expert knowledge, can be queried using federated querying mechanisms powered by SPARQL. This allows various institutions, including insurers, pharmaceutical companies, and medical research organisations, to access and leverage the rich insights stored within the knowledge graphs, enabling evidence-based decision-making and advancing medical research and healthcare practices. By combining decentralised knowledge graph management, user-centric privacy controls, and collaborative data sharing, this innovative framework represents a significant step forward in transforming decentralised KG management, fostering a secure, privacy-enhanced environment that empowers users, facilitates collaboration, and drives advancements in domains such as medical knowledge and patient care.

Despite the unquestionable reasoning features of symbolic systems and the studies reporting limitations of LLMs in human-like tasks (e.g., explanations, memories, and reasoning over factual statements) [40], there is an ongoing debate about LLM’s reasoning their causal inference capabilities [52]. Although LLMs excel at certain reasoning tasks, they do poorly in others, raising the question of whether they genuinely engage in causal reasoning or merely function as unreliable mimics, generating memorised responses (e.g. [45]). Methods to reason can be roughly divided into methods using only the LLM itself (e.g. with prompt engineering) and methods combining the LLM with an external reasoner and/or external source of knowledge (e.g. a Knowledge Graph) [86]. Our vision posits that external help will always be needed, especially for concrete use cases. There are discussions about the need for knowledge graphs in the era of LLMs. Sun et al. [103] and Dong [31] report on an empirical assessment of ChatGPT [93] with respect to DBpedia, illustrating the need of symbolic systems that over-fit for the truth whenever factual statements are collected from KGs. In addition, symbolic approaches can support sanity checking and be easily auditable and traceable. These features position the combination of both approaches in neuro-symbolic AI as a feasible option to provide KG-based AI. Neuro-symbolic AI delivers the basis to integrate the discrete methods implemented by symbolic AI with high-dimensional vector spaces managed by LLMs. They must decide when and how to combine both systems, e.g., following a principled integration (combining neural and symbolic while maintaining a clear separation between their roles and representations) or integrated (e.g. a symbolic reasoner integrated into the tuning process of an LLM).

Trust in AI systems stems from various factors, including transparency, reproducibility, predictability, and explainability. Neuro-symbolic systems play a vital role in enhancing trustworthiness by enabling communication between modules and facilitating tracing. Modularity enables the specification, verification, and validation of each component and its interactions. As a result, a system’s behaviour can be traced and validated. Specifically, within the domain of KG-based AI for self-determination, the seamless integration of KGs and symbolic semantic reasoning offers a comprehensive and unified perspective on curated knowledge. This integration holds immense value in addressing critical tasks such as validating, refuting, and explaining incorrect, biased, or misleading information that may potentially be generated by LLMs. By combining symbolic reasoning over KGs with LLMs, the propagation of misinformation can be mitigated while simultaneously enhancing the transparency and trustworthiness of AI-generated outputs. Consequently, KG-based AI systems can effectively emulate human behaviour by subjecting mistakes arising from false or incomplete information to a process of validation and enrichment using curated and potentially peer-reviewed sources of knowledge [109].

A notable application of KGs in neuro-symbolic AI is as a source of informative prior knowledge to increase the quality of machine learning models. An example is the work by Rivas et al. [89], where a deductive database, expressed in Datalog, establishes an axiomatic system of the pharmacokinetic behaviour of a treatment’s drugs and enables the deduction of new drug-drug interactions in cancer treatments. This prior knowledge plays a crucial role in elucidating the characteristics of a therapy and justifying its efficacy by considering all the interactions and the dynamic movement of drugs within the body. It encompasses factors such as the absorption, bioavailability, metabolism, and excretion of drugs over time. A KG embedding model improves its prediction of the effectiveness of a treatment based on prior knowledge, which encodes statements about a treatment’s characteristics; these statements are inferred by a deductive system which comprises the symbolic component of the hybrid approach. An approach for explaining link prediction (e.g. [90]) allows the justification of why this added prior knowledge affects the model’s decisions, potentially improving trust in the model’s results.

Grounding on the example presented in Figure 2, when individuals and professionals engage in communities with bots and assistants powered by AI models, it is critical to ensure the transparency of their decision-making process. However, despite the increasing focus on LLMs in healthcare and their continual improvement in terms of precision and accuracy [100], their outcomes can still be susceptible to hidden biases and a lack of traceability [62]. To tackle these challenges, the utilisation of a neuro-symbolic system can enhance LLMs by incorporating reasoning capabilities. This system operates as a deductive system on a user’s Knowledge Graph (KG). These hybrid AI systems can be effectively modelled using patterns proposed by [109]. Figure 6 depicts a pattern describing a hybrid AI system that enhances the explainability of the LLMs described in our running example. At the community level, symbolic reasoning applied to the ontology of shared PKGs can generate prior knowledge, enabling precise and concrete questioning of an LLM and providing additional contextual information. Moreover, a symbolic system facilitates the linking of shared PKGs with corresponding entities in KGs related to insurance, pharmaceuticals, and medical research. By incorporating this prior knowledge, the LLM’s answers are improved and validated with the assistance of the symbolic system. The systems operating at the community level and involving heterogeneous sources can be described using the explainable system with prior knowledge pattern; data alignments comprising prior knowledge enhance contextual knowledge provided to the therapy bot, facilitating thoughtful health recommendations.

DI1: Comprehensive recording.

A DLT can provide an immutable ledger, but work remains on how best to connect KG-based AI activities, e.g. to a possible federated query engine.

DI2: Personalised tracing.

Providing individual and community owners of PKGs with personalised traces of how acquired data was processed and used will involve dis-aggregating KG-processing and inferencing according to different user data and ensuring that privacy is not violated when individual results are returned.

DI3: “Decency” check.

There is a need for easy-to-use services that allow users and communities to check if an organisation has behaved in a “decent” way when it processes acquired data. Research here will examine how “decency” can be defined and validated by comparing PKG declarations of use (e.g. policies) with generated traces of use.

DI4: Interoperability.

Develop mechanisms that facilitate comprehensive, interoperable identification of human and machine participants in KG-based AI processes. For example, users and communities will wish to know and be able to validate claims that a data request comes from a particular organisation, unit and even individual KG processor. This will provide a foundation for accountability at all levels of granularity.

DI5: Self-sovereignty.

True self-sovereign KG-based AI needs to be: (i) based upon easy-to-use self-sovereign identities and data management; and (ii) capable of supporting the continuous monitoring of organisational behaviours in a transparent fashion.

DKG1: Knowledge Sanitisation.

Develop robust techniques for knowledge sanitisation that ensure user privacy by anonymising and filtering sensitive information based on data policies. These policies can be regulations such as GDPR and HIPAA, as well as individual-level data policies enforced at their personal data store, empowering users to specify their sharing preferences and control the aspects of data they disclose.

DKG2: Knowledge Graph Aggregation.

Design and implement mechanisms to encourage users to contribute their PKGs towards aggregated knowledge graphs, such as a concerted effort towards developing specific disease KGs. Blockchain-based incentive models that reward users for contributing to constructing such knowledge graphs, fostering collaborative efforts, and enriching the overall quality of shared knowledge are components of this goal.

DKG3: Knowledge Verification.

Develop community-based and expert processes to verify the knowledge available in the global KGs. On the community front, it is critical to ensure that a knowledge item that was previously contributed through an individual has not been altered (either through error or with malicious intent), for instance, via blockchain primitives, as explained in the previous section.

DKG4: Knowledge Validation.

Validation of knowledge is paramount to ensure KG interoperability and the consumption of knowledge in target applications. By employing RDF and SHACL technologies, we ensure that the DKGs across different data stores conform to a specific template, thus enabling their integration with community-supported KGs.

DKG5: Federated Querying.

Explore and implement federated querying mechanisms, specifically utilising SPARQL, to enable efficient querying across integrated KGs. This process includes developing techniques to support various institutions, such as insurers, pharmaceutical companies, and medical research organisations, accessing and extracting insights from the knowledge graphs to enhance decision-making and advance their respective domains.

XNS1: User-dependent Recommendations.

Neuro-symbolic systems need to be empowered to present results transparently to the users according to their interests. For example, in our illustrative scenario (Figure 2), an individual may not expect the same level of detail in a health recommendation as a medical doctor or a community representative.

XNS2: Adaptive Hybrid AI.

Define models that can adaptively combine predictive models with logical reasoning, encompassing abilities such as generalisation and causal inference. For accountability, the neuro-symbolic system should explain when the combination of logical reasoning with a therapy bot or healthcare assistant will be beneficial. For autonomy, the neuro-symbolic system should include the user in the loop and consider their opinion in this decision. Finally, trust requires verifying and validating these decisions.

XNS3: Contextual-based Hybrid AI.

Equip neuro-symbolic systems with contextual knowledge, reasoning capabilities, and causal inference to effectively evaluate the strengths and limitations of machine learning components. This goal empowers the system to identify optimal combinations of statistical and symbolic AI methods, requiring the definition of causal models on top of KGs capable of combining reasoning over KGs with causal inference.

XNS4: Symbolic Reasoning.

Employ inference processes, both inductive and deductive, on knowledge graphs to enable ML models, and LLMs in particular, to adjust hyper-parameters and a model’s configuration to new environments (i.e., Personal, community-based, and integrated healthcare KGs) and provide explanations for their decisions. Despite the advances of Automated Machine Learning (AutoML) systems (e.g., AutoML²⁶²⁶26https://www.automl.org/ and AutoWeka [58], to the best of our knowledge, there are no developments for AutoML over KGs or for neuro-symbolic systems, which will enhance accountability, autonomy, and trust.

XNS5: Learning Transparency.

Investigate if existing XAI mechanisms can be tailored for learning transparency, such that it is possible to explain what action was taken, how the decision making was performed, and why this was perceived as the outcome offering the greatest expected satisfaction.