Skip to content
HyperMesh

What is a Temporal Hypergraph?

Beyond Nodes and Edges

Section titled “Beyond Nodes and Edges”

Traditional graph databases model relationships as pairwise edges — one node connects to another. This works well for social networks, dependency trees, and simple knowledge graphs.

But many real-world events involve more than two participants simultaneously:

  • A cybersecurity event links a machine + process + account + IP at a specific time
  • A drug interaction involves 3–5 substances co-administered to a patient
  • A financial transaction connects a sender + receiver + intermediary + jurisdiction

A hypergraph generalizes the graph by allowing each edge (called a hyperedge) to connect any number of nodes — 2, 5, 20, or more — in a single atomic record.

Adding Time

Section titled “Adding Time”

A temporal hypergraph extends this further: every hyperedge carries a timestamp (or time range), making the structure a natural fit for event-driven data where the when matters as much as the who.

Hyperedge #1: {Machine-A, Process-svchost, Account-SYSTEM, IP-10.0.0.5} @ t=1000
Hyperedge #2: {Machine-A, Process-cmd, Account-Admin, IP-10.0.0.5}    @ t=1005
Hyperedge #3: {Machine-B, Process-svchost, Account-SYSTEM, IP-10.0.0.5} @ t=1010

Why Not a Regular Graph?

Section titled “Why Not a Regular Graph?”
ApproachLimitation
Relational DBMulti-party events require JOIN-heavy star schemas; temporal range queries are slow without custom indexing
Property Graph (Neo4j, etc.)Must decompose n-way events into multiple binary edges + intermediate “event” nodes, losing atomicity
RDF / Triple StoreEach fact is a triple (subject-predicate-object); n-way events require reification, which explodes storage and query complexity
Temporal HypergraphNative representation: one record per event, variable membership, indexed by time

HyperMesh’s Approach

Section titled “HyperMesh’s Approach”

HyperMesh stores temporal hyperedges in a purpose-built binary index (TPI — Temporal Property Index) that partitions data into time buckets. This enables:

  1. O(1) bucket seek — jump directly to the relevant time window
  2. O(k) scan — read only the hyperedges within that window
  3. O(log N) node lookup — find all events involving a specific node via the FMI (Forward Member Index)

The result: sub-millisecond range queries on datasets with millions of hyperedges, without the overhead of B-tree traversal or graph pattern matching.

Key Terminology

Section titled “Key Terminology”
TermDefinition
HyperedgeA single record linking 2–64 nodes at a specific timestamp
MemberA node participating in a hyperedge
FormationThe semantic type/label of a hyperedge (e.g., “CoProximity”, “NetworkFlow”)
TPITemporal Property Index — the primary time-bucketed storage layer
FMIForward Member Index — maps node IDs to their hyperedge participation
PSIProperty Secondary Index — optional indexes on numeric properties

Next Steps

Section titled “Next Steps”