Algebraic Property Graphs (APG)

• Algebraic Property Graphs (APGs) are rigorously defined data models that use type theory and category theory to unify graph representations and conventional schemas.
• APGs employ algebraic data types with sum and product constructions to ensure semantic consistency and facilitate seamless data migration.
• Leveraging categorical sketches, APGs provide a framework for structured schema mappings and predictable transformations across diverse enterprise systems.

Algebraic Property Graphs (APGs) provide a rigorous formalization of the property graphs data model using constructions from type theory and category theory. APGs unify the representation of both graph structures and conventional data schemas by leveraging algebraic data types and categorical sketches. Schemas and instances are characterized via sum and product types, with transformation and integration guaranteed to preserve semantic consistency by construction. This framework is grounded in industry-scale requirements, supporting compositional mappings between data formats and guaranteeing invariance under data migration and transformation operations (Shinavier et al., 2019).

1. Syntactic Foundations and Type Theory

Let $\mathcal{P}$ denote a fixed set of primitive types (such as Integer, String, Boolean), with each $p\in\mathcal{P}$ associated to a set $V_0(p)$ of primitive values. The universe of APG types $\operatorname{Ty}(\mathcal{P})$ is defined recursively via the following grammar:

$$t \in \operatorname{Ty}(\mathcal{P}) ::= 0 \mid 1 \mid t_1 + t_2 \mid t_1 \times t_2 \mid \operatorname{Prim}\, p\, (p \in \mathcal{P}) \mid \operatorname{Lbl}\, \ell\, (\ell \text{ a label})$$

Here, sum ($+$) and product ($\times$) types support union and tuple constructions, respectively, while $0$ and $1$ serve as empty and unit types. $\operatorname{Prim}\,p$ tags primitive types, and $p\in\mathcal{P}$0 tags graph label types.

An APG schema is a pair $p\in\mathcal{P}$1, where:

• $p\in\mathcal{P}$2 is a set of labels,
• $p\in\mathcal{P}$3 assigns to each label an algebraic datatype.

For example:

• $p\in\mathcal{P}$4,
• $p\in\mathcal{P}$5,
• $p\in\mathcal{P}$6.

Given $p\in\mathcal{P}$7, an APG instance $p\in\mathcal{P}$8 over $p\in\mathcal{P}$9 consists of:

1. For each $V_0(p)$0, a set $V_0(p)$1 of $V_0(p)$2-elements,
2. For each $V_0(p)$3, a value function $V_0(p)$4, where $V_0(p)$5 is defined inductively by:
   • $V_0(p)$6,
   • $V_0(p)$7,
   • $V_0(p)$8,
   • $V_0(p)$9,
   • $\operatorname{Ty}(\mathcal{P})$0,
   • $\operatorname{Ty}(\mathcal{P})$1.

Sums encode optionality and unions; products encode tuple and edge structures.

An equivalent, type-theoretic presentation of $\operatorname{Ty}(\mathcal{P})$2 treats the data as a four-sorted system with labeled elements, value terms, and an evaluation map, ensuring for each element $\operatorname{Ty}(\mathcal{P})$3 that $\operatorname{Ty}(\mathcal{P})$4 has type $\operatorname{Ty}(\mathcal{P})$5.

2. Categorical Semantics and Schemas

APG schemas and instances admit a categorical formulation via limit-colimit sketches and algebraic theories. The schema $\operatorname{Ty}(\mathcal{P})$6 is represented as a finite-product, finite-coproduct sketch $\operatorname{Ty}(\mathcal{P})$7, comprising:

• A graph with sort-nodes (the types),
• Cones for products, cocones for coproducts,
• Distinguished objects $\operatorname{Ty}(\mathcal{P})$8 and $\operatorname{Ty}(\mathcal{P})$9.

Nodes correspond to $$t \in \operatorname{Ty}(\mathcal{P}) ::= 0 \mid 1 \mid t_1 + t_2 \mid t_1 \times t_2 \mid \operatorname{Prim}\, p\, (p \in \mathcal{P}) \mid \operatorname{Lbl}\, \ell\, (\ell \text{ a label})$$0; for each labeled sum/product type in $$t \in \operatorname{Ty}(\mathcal{P}) ::= 0 \mid 1 \mid t_1 + t_2 \mid t_1 \times t_2 \mid \operatorname{Prim}\, p\, (p \in \mathcal{P}) \mid \operatorname{Lbl}\, \ell\, (\ell \text{ a label})$$1, a coproduct cocone or product cone is introduced. Each primitive type and label, together with generating arrows $$t \in \operatorname{Ty}(\mathcal{P}) ::= 0 \mid 1 \mid t_1 + t_2 \mid t_1 \times t_2 \mid \operatorname{Prim}\, p\, (p \in \mathcal{P}) \mid \operatorname{Lbl}\, \ell\, (\ell \text{ a label})$$2, generate the full category $$t \in \operatorname{Ty}(\mathcal{P}) ::= 0 \mid 1 \mid t_1 + t_2 \mid t_1 \times t_2 \mid \operatorname{Prim}\, p\, (p \in \mathcal{P}) \mid \operatorname{Lbl}\, \ell\, (\ell \text{ a label})$$3 with finite sums and products per schema.

A model of the schema sketch in $$t \in \operatorname{Ty}(\mathcal{P}) ::= 0 \mid 1 \mid t_1 + t_2 \mid t_1 \times t_2 \mid \operatorname{Prim}\, p\, (p \in \mathcal{P}) \mid \operatorname{Lbl}\, \ell\, (\ell \text{ a label})$$4 is a finite-product and finite-coproduct-preserving functor $$t \in \operatorname{Ty}(\mathcal{P}) ::= 0 \mid 1 \mid t_1 + t_2 \mid t_1 \times t_2 \mid \operatorname{Prim}\, p\, (p \in \mathcal{P}) \mid \operatorname{Lbl}\, \ell\, (\ell \text{ a label})$$5. Such functors are in bijection with APG instances as defined above, yielding a functor category:

$$t \in \operatorname{Ty}(\mathcal{P}) ::= 0 \mid 1 \mid t_1 + t_2 \mid t_1 \times t_2 \mid \operatorname{Prim}\, p\, (p \in \mathcal{P}) \mid \operatorname{Lbl}\, \ell\, (\ell \text{ a label})$$6

where the subscript denotes preservation of products and coproducts.

3. Schema Mappings and Structure-Preserving Functors

Given schemas $$t \in \operatorname{Ty}(\mathcal{P}) ::= 0 \mid 1 \mid t_1 + t_2 \mid t_1 \times t_2 \mid \operatorname{Prim}\, p\, (p \in \mathcal{P}) \mid \operatorname{Lbl}\, \ell\, (\ell \text{ a label})$$7 and $$t \in \operatorname{Ty}(\mathcal{P}) ::= 0 \mid 1 \mid t_1 + t_2 \mid t_1 \times t_2 \mid \operatorname{Prim}\, p\, (p \in \mathcal{P}) \mid \operatorname{Lbl}\, \ell\, (\ell \text{ a label})$$8, a schema morphism is a functor $$t \in \operatorname{Ty}(\mathcal{P}) ::= 0 \mid 1 \mid t_1 + t_2 \mid t_1 \times t_2 \mid \operatorname{Prim}\, p\, (p \in \mathcal{P}) \mid \operatorname{Lbl}\, \ell\, (\ell \text{ a label})$$9 that preserves $+$0, $+$1, finite sums, and products, and carries each primitive in $+$2 to the corresponding primitive in $+$3. This is manifest as terms in the target schema's type language and ensures strict structural preservation.

Concretely, for each $+$4:

• $+$5
• Associated equations mediate compatibility of product and coproduct structure.

This functorial approach enables systematic rewriting of schema types and mediates the transformation of value-terms through the morphism.

4. Graph Transformation and Semantic Consistency

A schema morphism $+$6 induces a pullback functor on APG instances:

$+$7

For $+$8, the instance on $+$9 is given by $\times$0, with corresponding value functions induced by functoriality and universal properties of the cones/cocones.

$\times$1 preserves all typing and consistency constraints—every element in the resulting instance retains the correct data shape per $\times$2. This effect is described as "semantic consistency by construction."

Kan extensions (left and right) $\times$3 implement data migration operations such as projection, join, union, and aggregation, provided the categorical finiteness conditions are met.

5. Enterprise Integration and Worked Examples

Most industry-relevant schema languages, including Apache Avro, Thrift, Protocol Buffers, JSON Schema, RDF with OWL/SHACL schemas, are inherently based on algebraic data types (sums-of-products with recursion). The APG formalism abstracts over these representations, supporting systematic schema alignment and data migration across heterogeneous systems.

To integrate, for example, Avro data into a property-graph-backed knowledge graph:

1. Extract the Avro IDL as an APG schema $\times$4, mapping each record to a label and each field to a product type, with primitives aligned as appropriate.
2. Extract the target property graph schema $\times$5 analogously.
3. Construct a schema morphism $\times$6 mapping Avro structures to graph labels and properties.
4. Given an Avro instance $\times$7, the graph instance is

$\times$8

For a trivial example, mapping user data with fields ("name", "age") as Avro records to graph vertices and property edges (Person, nameProp, ageProp) is handled without auxiliary synchronization logic; consistency and attribute cohesion are a consequence of the model's semantics.

6. Summary, Generalizations, and Outlook

• The property graph model is given formal type-theoretic and categorical semantics by regarding all labels as algebraic datatype constructors.
• APG schemas are precisely finite limit-colimit sketches; APG instances are the corresponding models in $\times$9 preserving specified cones and cocones.
• Schema mappings are sketch morphisms preserving sum and product structure; data migration via pullback preserves all typing and structural invariants by construction.
• The model encompasses not only classic vertices, edges, and multi-valued properties, but also hyperedges, meta-properties, nested records, unions, optionals, and their mixtures.
• As most enterprise schema languages are algebraic, APGs provide a unifying framework for data integration, transformation, and knowledge graph construction, supporting type-safety, referential consistency, and predictable migration semantics.

APGs thus reconcile the flexibility required by engineering teams with mathematical rigor, serving as a lingua franca for enterprise graph-integration pipelines and relying on well-understood algebraic and categorical foundations (Shinavier et al., 2019).