Personal Causal Knowledge Graph (PCKG)

• Personal Causal Knowledge Graph (PCKG) is a structured, individualized model that maps causal dependencies among personal variables, habits, and events.
• It combines structural causal modeling with knowledge graph formalism to encode relationships, estimate causal strengths, and support both interventional and counterfactual queries.
• PCKGs enable precise, context-specific recommendations by integrating domain-specific data, graph inference, and LLM-assisted planning, validated through empirical studies.

A Personal Causal Knowledge Graph (PCKG) is a structured, individualized graphical representation of causal dependencies among a person's variables, habits, events, and contexts, constructed to enable personalized causal reasoning, interventional analysis, and counterfactual simulation. PCKGs integrate structural causal modeling (SCM) with knowledge graph formalism, supporting precise, explainable, and context-specific recommendations and predictions in domains ranging from health and lifestyle to behavioral support and explainable artificial intelligence.

1. Formal Definitions and Mathematical Structure

The foundational structure of a PCKG is a directed acyclic graph or multigraph, where nodes correspond to person-specific variables, events, or contexts, and directed edges encode direct causal influence. In the formalism of (Yang et al., 28 Feb 2025), for individual $i$: $G_i = (V_i,\,E_i,\,W_i)$ where:

• $V_i = \{v_1, \ldots, v_m\}$: Person-specific variables. These may include domain-relevant features (nutrient intakes, biomarkers, behavioral habits, etc.).
• $E_i \subseteq V_i \times V_i$: Directed edges $v_j \to v_k$ representing direct causal effects.
• $W_i = \{w_{jk}\}$: Real-valued weights quantifying strength of each edge; these are parameterized in an SCM.

An alternative multigraph/hypergraph schema, used in (Raman et al., 8 Sep 2025) and (Jaimini et al., 2022), defines: $G_p = (V_p,\,E_p,\,\tau_V,\,\tau_E)$ with node sets partitioned as event-nodes ($V_p^{\mathsf{E}}$), habit-nodes ($V_p^{\mathsf{H}}$), and context-variables ($V_p^{\mathsf{C}}$), and edges labeled for causal or temporal semantics: $G_i = (V_i,\,E_i,\,W_i)$0 Each directed edge may carry a weight $G_i = (V_i,\,E_i,\,W_i)$1 and a semantic label (e.g., “causes,” “aggravates”).

The CausalKG framework (Jaimini et al., 2022) generalizes this to hyper-relational graphs—allowing hyper-edges $G_i = (V_i,\,E_i,\,W_i)$2, for with mediators $G_i = (V_i,\,E_i,\,W_i)$3 and causal effect attributes (total, direct, indirect).

All main formalisms specify an SCM on the graph. For a linear model: $G_i = (V_i,\,E_i,\,W_i)$4 where $G_i = (V_i,\,E_i,\,W_i)$5 are parents of $G_i = (V_i,\,E_i,\,W_i)$6, and $G_i = (V_i,\,E_i,\,W_i)$7 is a noise term.

2. Construction and Enrichment of Personal Causal Knowledge Graphs

PCKG synthesis proceeds through several stages:

• Data Acquisition and Preprocessing: In health/lifestyle, this involves collecting timeseries (e.g., CGM, activity sensors), logging macronutrient intake, event journals, user habits, and context variables such as baseline health or external modulating factors (Yang et al., 28 Feb 2025, Raman et al., 8 Sep 2025, Jaimini et al., 2022).
• Entity and Context Extraction: Each data record is mapped to ontology nodes (nutrients, biomarker levels, events like “slept 4 h,” habits like “meditates daily,” static contexts such as “caffeine tolerance”) (Raman et al., 8 Sep 2025, Jaimini et al., 2022).
• Causal Discovery: Algorithms such as the PC algorithm or structural learning with statistical tests construct the graph skeleton, testing conditional independencies among variables; node and edge sets are oriented domain-specifically (e.g., dietary causes precede glycemic outcome nodes) (Yang et al., 28 Feb 2025).
• Edge Weight Estimation: Each causal edge's strength is estimated through regression of the child on its parents (linear SCMs) or by fitting conditional probability tables in causal Bayesian networks (Yang et al., 28 Feb 2025, Jaimini et al., 2022).
• Knowledge Graph Enrichment: External domain graphs (e.g., population health causal networks) are merged or aligned with the PCKG, using semantic similarity and mapping functions (Raman et al., 8 Sep 2025). Mediator nodes, contextual qualifiers (e.g., “StressLevel=High”), and provenance metadata are attached to hyper-edges as needed (Jaimini et al., 2022).
• Edge Pruning and Personalization: Weak or spurious edges (low estimated weight or low causality) may be pruned. The graph is continuously updated as new records or user contexts are ingested (Raman et al., 8 Sep 2025).

3. Causal Reasoning and Inference: Interventional and Counterfactual Queries

PCKGs instantiate an explicit SCM, enabling both interventional (“do-”) and counterfactual (“would-have-been”) queries:

• Interventional Reasoning: The effect of setting variable $G_i = (V_i,\,E_i,\,W_i)$8 to $G_i = (V_i,\,E_i,\,W_i)$9 on $V_i = \{v_1, \ldots, v_m\}$0 is computed via do-calculus,

$V_i = \{v_1, \ldots, v_m\}$1

with $V_i = \{v_1, \ldots, v_m\}$2 an appropriate adjustment set blocking backdoor paths (Raman et al., 8 Sep 2025). In linear SCMs, this reduces to algebraic updates of expected outcomes given interventions.

• Counterfactual Reasoning: For counterfactuals, the standard three-step Pearl procedure applies:
  1. Abduction: Infer the posterior over unobserved exogenous variables given observed evidence.
  2. Action: Modify the SCM by replacing the equation for $V_i = \{v_1, \ldots, v_m\}$3 with $V_i = \{v_1, \ldots, v_m\}$4.
  3. Prediction: Propagate changes to obtain the distribution of $V_i = \{v_1, \ldots, v_m\}$5 in the counterfactual world (Jaimini et al., 2022).

• Algorithmic Traversal: Graph-of-thought or $V_i = \{v_1, \ldots, v_m\}$6-hop traversals identify pathways from queried targets back to relevant habits/events, scored by path strength or plausibility (possibly via LLM scoring) (Raman et al., 8 Sep 2025).

• Quantitative Mediation: Hyper-relational representations allow partitioning of effects into total, direct, and indirect (mediated) components via edge attributes $V_i = \{v_1, \ldots, v_m\}$7, facilitating fine-grained reasoning (Jaimini et al., 2022).

4. Integration with LLMs and Agentic Reasoning

PCKGs are designed to interface directly with LLM-based recommendation or planning agents:

• Reasoning Pipeline: User queries are mapped to graph nodes (goals), subgraph traversals aggregate causal influence, external knowledge sources suggest candidate actions or interventions, and counterfactual simulations verify predicted improvements or outcomes (Yang et al., 28 Feb 2025, Raman et al., 8 Sep 2025).
• Plan Schema Instantiation: Schema-based planning modules retrieve action templates parameterized by identified causes, yielding adaptable, context-specific plan steps. Counterfactual checks on the instantiated plan confirm likely efficacy before recommendation (Raman et al., 8 Sep 2025).
• LLM-Orchestrated Flows: LLMs verbalize the outputs of symbolic modules, grounding answers in explicit graph traversals, intervention results, and action schemas. Modular architectures prevent causal hallucinations by confining language generation to post-inference output (Raman et al., 8 Sep 2025).
• Prompt Engineering: Chain-of-thought scaffolds guide the LLM through sequential steps: goal identification, causal path explanation with strength citations, proposal of actionable interventions, counterfactual verification, and structured recommendation generation (Yang et al., 28 Feb 2025).

5. Evaluation, Empirical Results, and Metrics

Evaluation of PCKG-based systems employs both quantitative and qualitative metrics:

• Effectiveness: In dietary recommendation, Mean Glucose Reduction (MGR) is computed as the mean improvement for PCKG-guided food choices relative to historical baselines: $V_i = \{v_1, \ldots, v_m\}$8 where $V_i = \{v_1, \ldots, v_m\}$9 is the counterfactual glucose response under the recommended intervention (Yang et al., 28 Feb 2025).
• Metrics for Personalization and Causal Fidelity: The Personalization Salience Score (PSS) measures the presence of context blocks in the response, while Causal Reasoning Accuracy (CRA) assesses the presence of correct causal factors. Both are defined via similarity-based matching between response content and ground-truth context/factor sets (Raman et al., 8 Sep 2025).
• Sample Results: Over 34 users and multiple queries, PCKG systems significantly outperformed non-causal baselines in lowering postprandial glucose (e.g., MGR for 1h: PCKG 158.2 vs. baseline 120.5, $E_i \subseteq V_i \times V_i$0), and achieved a 98.4% win rate in LLM-as-a-judge personalization ratings (Yang et al., 28 Feb 2025).

6. Limitations, Extensions, and Practical Considerations

PCKGs are subject to several current limitations:

• Domain Specificity and Generalization: Early implementations are tailored to single-domain outcomes (e.g., glycemic management), but real-world applications require multi-objective trade-offs and multitask optimization (e.g., Pareto front strategies for health) (Yang et al., 28 Feb 2025).
• Confounders and Data Limitations: Unmodeled variables (sleep, stress, microbiome) and limited longitudinal data introduce confounding and attenuate causal validity. Multimodal sensor integration and collection of interventional data are suggested for graph strengthening (Yang et al., 28 Feb 2025).
• Scalability and Complexity: Large graphs may incur high computational overhead for traversal and inference. Pruning strategies and adaptive exploration mitigate performance issues. Presenting only top-ranked causes/effects balances explanation with interpretability (Raman et al., 8 Sep 2025, Jaimini et al., 2022).
• Privacy, Security, and Explainability: Storing edge weights and qualifiers demands access control; differential privacy mechanisms can be applied for aggregate release. Explainability requires balancing graph richness against cognitive load (Jaimini et al., 2022).

A plausible implication is that as graph enrichment, data scale, and integration breadth improve, PCKGs could underlie robust, trustworthy agentic reasoning across a wide class of personalized AI systems.

7. Applications and Outlook

PCKGs have demonstrated efficacy in health (personalized dietary guidance for glucose modulation), lifestyle planning, and lifestyle-based AI assistants (Yang et al., 28 Feb 2025, Raman et al., 8 Sep 2025). The hyper-relational and SCM-based constructs in these frameworks afford explainable, verifiable, and user-aligned decision support. Ongoing research aims to extend PCKGs to multi-domain applications, integrate richer data modalities, develop more scalable reasoning algorithms, and achieve clinical or field validation of intervention effectiveness (Yang et al., 28 Feb 2025, Raman et al., 8 Sep 2025, Jaimini et al., 2022).