Personalized Knowledge Graphs

• Personalized Knowledge Graphs (PKGs) are structured representations that capture entities, attributes, and relations specific to an individual to support personalized services.
• They integrate heterogeneous data sources using methods like RDF reification, user-specific embeddings, and dynamic summarization to maintain relevance over time.
• PKGs are applied in domains such as personalized search, healthcare, and education while addressing challenges related to data quality, privacy, and system scalability.

A Personalized Knowledge Graph (PKG) is generally understood as structured knowledge about entities related to an individual, their attributes, and the relations between them, but the literature does not converge on a single definition. Some work treats a PKG as a user-owned knowledge graph whose primary purpose is to support personalized services; some defines it as a user-centric “spiderweb” graph centered on one person; and some adjacent systems achieve personalization through user-specific embeddings over a shared graph rather than by materializing a separate graph for each user (Skjæveland et al., 2023, Shirai et al., 2021, Kasela et al., 18 Jul 2025). The acronym itself is also overloaded: in adjacent literatures, “PKG” can denote “Product Knowledge Graph” or “PubMed Knowledge Graph,” so the intended sense must be read from context (Xu et al., 2019, Xu et al., 2020).

1. Definitions and conceptual boundaries

One influential ownership-centered definition states that a PKG is a knowledge graph for which a single individual has full read and write access and the exclusive right to grant others read and write access, with personalized service delivery as its primary purpose (Skjæveland et al., 2023). This definition distinguishes PKGs from service-controlled personalization artifacts: a provider-maintained user-interest graph can support personalization without satisfying the ownership criterion. The same survey explicitly contrasts this with a “personalized knowledge graph,” defined there as a subset of an existing KG restricted to entities and relationships characterizing a user’s interests (Skjæveland et al., 2023).

A second influential line, associated in the cited literature with Balog and Kenter, defines a PKG as structured knowledge about entities and relations of personal, rather than general, importance, often with a “spiderweb” layout in which all nodes connect directly or indirectly to a central user node (Chakraborty et al., 2022). Healthcare work extends this view into Personal Health Knowledge Graphs (PHKGs), arguing that a graph becomes genuinely personal when it captures patient-specific conditions, constraints, behaviors, and context rather than merely presenting a personalized view over a general biomedical graph (Shirai et al., 2021).

A third boundary issue concerns systems that personalize on top of a shared graph. PARK, for example, builds one global academic knowledge graph over users, papers, venues, and affiliations, then learns user-specific embeddings inside that graph; it explicitly states that this is not an individually materialized per-user PKG, although it is a knowledge-graph-based personalized representation (Kasela et al., 18 Jul 2025). This makes a recurring controversy explicit: whether “PKG” should denote a separate graph per person or can also include shared-graph architectures with personalized projections.

2. Construction patterns and data sources

Across the cited systems, PKGs are built by combining heterogeneous personal, behavioral, and domain knowledge. In a direct personal-data-management setting, “PKG API” represents natural-language user statements as reified RDF rdf:Statement objects, attaches the original statement text, and enriches the result with provenance and statement-level read/write permissions (Bernard et al., 2024). In this pattern, personal knowledge is entered explicitly by the user and can remain only partially grounded until later entity linking resolves it to internal or external identifiers.

A second pattern constructs PKGs from behavior and external knowledge bases. MeKB-Rec builds a per-user Personal Knowledge Graph called MeKB by taking source-domain user behaviors, extracting entities from consumed-item content, linking those entities to a global knowledge base, and aggregating the resulting neighborhood into a user-specific graph for cross-domain recommendation (Su et al., 2023). The graph is therefore personal because different users activate different KB regions, not because the underlying KB is itself personal.

Healthcare variants broaden the input space substantially. The PCKG survey describes structured, semi-structured, and unstructured sources including EHRs, laboratory data, diagnostics, genomics, wearables, clinical notes, and medical literature, with normalization to standards such as SNOMED-CT, RxNorm, and LOINC (Khatib et al., 2024). The clinically oriented PHKG for diet recommendation goes further by combining interview-derived preferences and constraints, five weeks of synthetic food logs, FoodKG, and selected American Diabetes Association dietary guidelines, then instantiating a patient graph that captures temporal dietary patterns and recommendation-relevant context (Seneviratne et al., 2021).

Other domains instantiate the same general pattern with different source layers. The context-aware messaging framework builds a person- and context-centric graph from LinkedIn, X, Wikipedia, and news APIs, with nodes for individuals, locations, events, skills, interests, and cultural attributes (Kumar et al., 12 Mar 2025). In educational Graph RAG for MOOCs, the learner-specific PKG is created when a learner marks a concept as “Did Not Understand,” thereby personalizing a shared educational knowledge graph into a learner overlay that stores the DNU concept, the current slide text, and neighboring slide concepts (Abdelmagied et al., 15 May 2025). In academic retrieval, PARK converts a citation graph into an academic KG with four entity types—users/authors, documents/papers, venues, affiliations—and five relation types—wrote, cited, in_venue, affiliated, and symmetric co_author—then learns user embeddings from that shared structure (Kasela et al., 18 Jul 2025).

3. Representation, summarization, and reasoning mechanisms

A generic formalization used in the literature writes a knowledge graph as $$\mathcal{G} = \{\mathcal{E}, \mathcal{R}, \mathcal{F}\}$$, where $\mathcal{E}$ is the entity set, $\mathcal{R}$ the relation set, and $\mathcal{F}$ the set of facts $(h,r,t)$ (Chakraborty et al., 2022). What varies across PKG systems is not the existence of triples but the representation of uncertainty, provenance, temporality, permissions, and user-specific salience.

One representation strategy is statement-centric and symbolic. “PKG API” stores natural-language assertions as rdf:Statement with rdf:subject, rdf:predicate, and rdf:object, preserves the original wording, allows unresolved elements to remain skos:Concepts, and attaches provenance via PAV plus statement-level access control through pkg:readAccessRights and pkg:writeAccessRights (Bernard et al., 2024). This design favors incremental enrichment and explicit governance.

A second strategy is semantic encoding of user-specific graphs. MeKB-Rec represents a user through an encoding of that user’s personal graph,

$\mathbf{h}_u = f_{\text{PKG}}(\mathcal{G}_u),$

and uses pretrained LLMs to align user-graph semantics with target-item semantics for cross-domain recommendation, including zero-shot users with no target-domain behavior (Su et al., 2023). In a related but not identical pattern, PARK first trains MiniLM for dense retrieval, then fixes document-node embeddings to those neural document vectors while learning TransE- or TransH-based embeddings for users, venues, affiliations, and relations in the same space; retrieval-time ranking then combines lexical relevance, dense relevance, and user–author similarity (Kasela et al., 18 Jul 2025).

A third strategy treats PKGs as compact summaries that must adapt to evolving interests. APEX$^2$ defines a KG as $\mathcal{G}=(\mathcal{E},\mathcal{R},\mathcal{T})$ and a PKG as a subgraph $$\mathcal{P}=(\mathcal{E}_p,\mathcal{R}_p,\mathcal{T}_p)$$, then updates user interest through a decayed query-log model and heat diffusion over the graph: $$q_{\rm total}^{(T)} = \sum_{t=0}^{T} \gamma^{T-t} q^{(t)}, \qquad \mathbf{e}^{(T)} = \sum_{l=0}^{d} \alpha^l \mathbf{A}^l q_{\rm total}^{(T)}.$$ Triple preference is then scored from head-entity, relation, and tail-entity interest, and the top-$\mathcal{E}$0 triples form the current summary (Li et al., 2024). This makes PKG maintenance a streaming summarization problem rather than a one-shot extraction problem.

A fourth strategy emphasizes retrieval planning over graph memory. PersonalAI 2.0 decomposes a question into sub-questions, generates a search plan $\mathcal{E}$1, extracts entities for each step, matches them to graph vertices, generates clue-queries, traverses the graph, filters retrieved triples, summarizes clue answers, tests answer sufficiency, and optionally enhances the search plan before continuing (Menschikov et al., 13 May 2026). This architecture is not a personal-schema proposal, but it is directly relevant to PKGs because it shows how a graph-backed memory can support adaptive, multi-step retrieval rather than one-shot nearest-neighbor access.

Reasoning remains important wherever PKGs are expected to drive action. In the PHKG diet system, OWL class expressions encode guideline conditions and directives, allowing the graph to infer recommendation constraints such as Mediterranean-diet substitution or carbohydrate ranges for fixed insulin dosing from patient-specific graph facts (Seneviratne et al., 2021).

4. Application areas

PKGs appear in at least four broad application families. One is personal data management and memory support. “PKG API” exposes a RESTful API and a user-facing client for entering, querying, deleting, and visualizing personal statements through natural language, while Personal Research Knowledge Graphs (PRKGs) envision researcher-centered graphs over affiliations, interests, publications, tools, datasets, tasks, and local resources to support assistants such as “SciJeeves” (Bernard et al., 2024, Chakraborty et al., 2022).

A second family is personalized search and recommendation. MeKB-Rec uses per-user MeKB graphs for cross-domain recommendation and reports improvements in HR@10 and NDCG@10 of 24%–91% over prior approaches, with a 105% improvement in HR@10 for zero-shot users (Su et al., 2023). PARK uses a global academic KG with personalized user embeddings for academic retrieval and reports improvements in MAP@100 over the second-best model of about 9% in Political Science, 10% in Psychology, and 9% in Physics, while remaining the strongest user-centric model in Computer Science even though citation popularity performs best there (Kasela et al., 18 Jul 2025). Closely related systems that personalize over shared graphs rather than fully materialized PKGs include shopping-oriented entity resolution from customer-product KGs and ontology-backed vehicle recommendation with user and item entities in the same graph (Lin et al., 2021, Le et al., 2023).

A third family is healthcare. The PCKG survey presents patient-centric graphs as a way to integrate diagnoses, medications, procedures, laboratory results, genomics, device data, and literature into a unified patient representation for disease prediction, treatment recommendation, and decision support (Khatib et al., 2024). The diet-focused PHKG instantiates this pattern with Type 2 Diabetes, while the earlier health agenda paper argues that PHKGs remain under-explored and that collection, linkage, and maintenance are still open problems (Seneviratne et al., 2021, Shirai et al., 2021).

A fourth family is education and communication. In MOOCs, PKG-based question generation in CourseMapper recommends learner-specific questions from DNU-marked concepts and local slide context, and instructor evaluation over three MOOCs yielded a weighted average score of 2.862, with $\mathcal{E}$2 and $\mathcal{E}$3 on a 1–3 scale (Abdelmagied et al., 15 May 2025). E-learning proposals also treat PKGs as user-centric semantic layers over larger educational KGs, often linked to DBpedia or other Linked Open Data, to support recommendations, group views, and explainability (Ilkou, 2022). In messaging, graph-conditioned LLM personalization reported message acceptance rates of 42% in healthcare, 53% in education, and 78% in recruitment, though the paper explicitly treats these as acceptance percentages rather than downstream behavioral outcomes (Kumar et al., 12 Mar 2025). A related educational line uses shared contextual KGs of learning objects to enable personalization, even when the graph itself is not learner-specific (Abu-Rasheed et al., 2024).

5. Controversies, constraints, and common misconceptions

The first recurring misconception is that all personalization graphs are PKGs in the same sense. The ownership-centered survey explicitly rejects this: a provider-controlled personalized KG may characterize a user’s interests yet still fail the criterion that the individual owns, governs, and selectively shares the graph (Skjæveland et al., 2023). Conversely, the academic-retrieval literature shows that personalization can also be realized through user-specific embeddings over a shared KG, so “PKG” is not always synonymous with “a separate graph instance per user” (Kasela et al., 18 Jul 2025).

A second difficulty is data quality and sparsity. PARK states that the effectiveness of its KG user model depends on citation completeness and quality, and that sparse or incomplete academic metadata degrade both embeddings and retrieval (Kasela et al., 18 Jul 2025). The context-aware messaging system likewise depends on public profile and event data; it acknowledges unresolved cold-start and sparsity issues for users with little public information (Kumar et al., 12 Mar 2025). In healthcare, missingness, heterogeneity, and evolving patient data are pervasive constraints rather than edge cases (Khatib et al., 2024).

A third constraint is governance. PKG API makes access rights and provenance first-class at the statement level, while the ecosystem survey argues that fine-grained access control, user-friendly management interfaces, and trusted hosting are central to real PKG adoption (Bernard et al., 2024, Skjæveland et al., 2023). Healthcare surveys add privacy, authorization, and interoperability pressures from clinical data, where patient-centric graphs must remain semantically useful without undermining confidentiality (Khatib et al., 2024).

A fourth issue is temporality and maintenance. Health-oriented PKG work emphasizes that personal relations may be short-lived and that updates must respond both to changing personal states and to changes in linked external knowledge (Shirai et al., 2021). PRKG work proposes explicit validity intervals $\mathcal{E}$4 for time-bounded relations such as affiliations and projects (Chakraborty et al., 2022). APEX$\mathcal{E}$5 shows that even after a PKG has been built, it may need continual re-summarization to follow interest drift under severe storage limits (Li et al., 2024).

A final controversy concerns explainability. Many PKG papers motivate graphs partly because they are more interpretable than opaque latent vectors, but several systems stop short of explicit explanation interfaces. PARK notes that its schema is more interpretable than latent-only user vectors yet does not implement a dedicated explanation mechanism (Kasela et al., 18 Jul 2025). The messaging framework similarly observes that graph-derived personalization could in principle support explanations, but does not implement them (Kumar et al., 12 Mar 2025).

6. Current research directions

Recent work pushes PKGs toward continuous adaptation, extreme compression, and LLM-native use. APEX$\mathcal{E}$6 reframes PKG maintenance as adaptive summarization under evolving query streams and compression ratios at or below $\mathcal{E}$7, showing that PKG utility can be treated as a dynamic budget-allocation problem rather than a static subgraph-extraction problem (Li et al., 2024). PersonalAI 2.0, in turn, treats graph memory as an external substrate for iterative LLM planning, and reports that plan enhancement yields an 18% boost over the disabled version while traversal-based retrieval gains about 6% over flat retrieval on average (Menschikov et al., 13 May 2026).

Another direction is hybrid KG+LLM systems. In context-aware messaging, the graph is used as a retrieval and conditioning layer for prompt construction rather than as a direct symbolic reasoner (Kumar et al., 12 Mar 2025). In educational Graph RAG, the PKG determines what the learner should ask, while the shared EduKG determines what evidence should answer it (Abdelmagied et al., 15 May 2025). This suggests a broader architectural shift: PKGs increasingly act as structured memory and retrieval controllers for generation systems rather than only as downstream analytics substrates.

A further direction is ecosystem building. The PKG ecosystem survey argues that progress depends not only on graph construction but also on standardized vocabularies, interoperable interfaces, synchronization with data sources, and usable administration by non-experts (Skjæveland et al., 2023). “PKG API” operationalizes part of that agenda by combining a natural-language client, a RESTful API, RDF reification, provenance, and statement-level permissions (Bernard et al., 2024). Healthcare surveys add that richer temporal modeling, better data harmonization, and stronger privacy-preserving integration remain necessary before patient-centric graphs can support routine precision medicine at scale (Khatib et al., 2024).

Taken together, the literature presents PKGs not as a single graph formalism but as a family of user-centered knowledge representations that range from user-owned personal data stores to domain-specific per-user graphs to shared knowledge graphs with personalized projections. The field’s unifying problem is stable: to encode knowledge of personal rather than merely general importance in a form that supports retrieval, reasoning, explanation, and adaptation under real constraints of privacy, sparsity, temporality, and scale.