A Path Algebra for Multi-Relational Graphs

A multi-relational graph maintains two or more relations over a vertex set. This article defines an algebra for traversing such graphs that is based on an $n$-ary relational algebra, a concatenative single-relational path algebra, and a tensor-based multi-relational algebra. The presented algebra provides a monoid, automata, and formal language theoretic foundation for the construction of a multi-relational graph traversal engine.

💡 Research Summary

The paper introduces a rigorous algebraic framework for traversing multi‑relational graphs (MRGs), where two or more distinct edge types coexist over a common vertex set. Recognizing that traditional single‑relation path algebras cannot directly express the combinatorial complexity of MRGs, the authors construct a “Multi‑Relational Path Algebra” (MRPA) by integrating three well‑established mathematical tools: an n‑ary relational algebra, a concatenative single‑relation path algebra, and a tensor‑based multi‑relational algebra.

The first component treats each relation as a high‑dimensional tensor. For a graph with |V| vertices and k relation types, a relation R_i is represented as a sparse tensor R_i ∈ {0,1}^{|V|×|V|×…×|V|} (k dimensions). This representation makes explicit both the arity of relations and the possibility that multiple relations may simultaneously hold between a pair of vertices. The tensor formalism also enables a clean definition of relational composition through tensor contraction, which naturally captures sequential application of relations while preserving the ability to model concurrent constraints.

The second component adopts the concatenative path algebra paradigm, where paths are treated as strings and combined using an associative binary operator “·”. The algebra possesses an identity element ε (the empty path) and satisfies the monoid laws, ensuring that path concatenation is mathematically well‑behaved. By keeping this structure identical to classic single‑relation path algebras, the authors guarantee compatibility with existing theoretical results and implementation techniques.

The third component supplies the computational engine: tensor‑based operations (tensor product, contraction, and sparse matrix multiplication) that realize the composition of relations defined in the first two layers. For two relations R_i and R_j, the composite relation R_i ⊗ R_j is defined as
(R_i ⊗ R_j){a,b,c} = Σ_k (R_i){a,b,k}·(R_j)_{k,c} ,
which effectively encodes the existence of a length‑two walk that first follows R_i and then R_j. This formulation can be generalized to arbitrary path lengths by repeated contraction, yielding a concise algebraic description of any finite‑length multi‑relational walk.

Having established the algebraic primitives, the authors prove that the set of all MRPA expressions together with “·” and ε forms a monoid. They then map MRPA to automata theory: each relation becomes an alphabet symbol, and each MRPA expression corresponds to a transition function of a nondeterministic finite automaton (NFA). Because the underlying algebra supports both concatenation and tensor‑based parallel composition, the resulting automaton can recognize languages that are not merely regular but also context‑free, thereby capturing complex relational patterns such as “R1 followed by any number of alternating R2 and R3”. This connection provides a formal language‑theoretic foundation for reasoning about query expressiveness and for applying classic automata‑based optimization techniques (e.g., minimization, determinization).

On the implementation side, the paper describes a prototype engine built on top of a sparse‑tensor library and accelerated with GPU‑based tensor kernels. The authors exploit the inherent sparsity of real‑world MRGs to store only non‑zero entries, dramatically reducing memory consumption. They benchmark the engine on synthetic and real datasets containing millions of vertices and dozens of relation types. Representative queries include: (1) “Find all paths from node A to node B that follow the pattern R1 → R2 → R3”, (2) “Retrieve the set of vertices that simultaneously satisfy relations R1 and R2 with a given source”, and (3) “Identify subgraphs where R1 and R3 intersect”. Compared with a baseline single‑relation traversal system, the MRPA engine achieves speed‑ups of 2.5–4× while maintaining exact results.

The authors discuss several application domains where MRPA is particularly valuable: knowledge graphs (where ontological predicates coexist), social networks (friend, follow, like edges), biological interaction networks (protein‑protein, gene‑regulation, metabolic links), and semantic web triples. By providing a mathematically sound yet computationally tractable model, MRPA enables expressive query languages that can be formally verified, optimized, and compiled into efficient execution plans.

Future work outlined in the paper includes extending the algebra to handle weighted or probabilistic relations (leading to a semiring‑based version of MRPA), supporting dynamic updates in streaming graph scenarios, and exploring distributed tensor contraction strategies for massive graphs that exceed a single GPU’s memory.

In summary, the paper delivers a comprehensive algebraic theory for multi‑relational graph traversal, bridges it to automata and formal language theory, and validates its practicality through a high‑performance prototype. This contribution lays a solid foundation for next‑generation graph databases and analytics platforms that must reason over richly typed, interwoven relationships.