Knowledge Graph

What is a knowledge graph?

A knowledge graph represents data as interconnected entities rather than isolated records. Instead of just listing techniques, it captures how techniques relate to goals, tags, resources, and each other—enabling richer queries and AI-powered discovery.

Why a Knowledge Graph?

The TEA Techniques dataset contains rich relationships that a flat list can't capture. The knowledge graph makes these connections explicit and queryable:

Rich relationships — SHAP is related to LIME; both address explainability but differ in approach
Semantic queries — "Find all explainability techniques that work with neural networks"
AI-powered discovery — Ask AI Agents (e.g. Claude, ChatGPT, or Gemini) to find and compare techniques via the MCP server
Linked data interoperability — Connect with schema.org, SKOS, and other vocabularies

How It's Structured

The knowledge graph connects four entity types through explicit relationships:

Entity Types

Entity	Count	Description	Examples
Technique	121	AI assurance methods with metadata and ratings	SHAP, LIME, Differential Privacy
Assurance Goal	8	High-level objectives a technique addresses	Explainability, Fairness, Privacy
Tag	171	Hierarchical classification taxonomy	`model-agnostic`, `neural-network`
Resource	464	Academic papers, software, and documentation	Research papers, Python packages

Relationships

Relationship	Description	Example
`addressesGoal`	Which assurance goals a technique supports	SHAP → Explainability, Fairness, Reliability
`hasTag`	Classification from the hierarchical taxonomy	SHAP → `model-agnostic`, `black-box`
`relatedTo`	Explicit links between related techniques	SHAP → LIME, Permutation Importance
`citesResource`	Academic papers and software references	SHAP → Lundberg & Lee (2017)

What You Can Discover

The knowledge graph enables queries that would be difficult with flat data:

Find techniques for a use case

"I need explainability techniques for neural networks"

→ Filter by addressesGoal: explainability + hasTag: neural-network

→ Results: Integrated Gradients, Grad-CAM, Attention Visualisation, SHAP...

Explore related techniques

"I'm using SHAP—what similar techniques should I consider?"

→ Follow relatedTo links from SHAP

→ Results: LIME, Permutation Importance, Sobol Indices, Partial Dependence Plots

Find evidence for assurance claims

"What techniques produce quantitative fairness metrics?"

→ Filter by addressesGoal: fairness + hasTag: evidence-type/quantitative-metric

→ Results: Demographic Parity Assessment, Equalised Odds, Average Odds Difference...

JSON-LD Data Format

Each technique is represented as a JSON-LD document with typed relationships:

JSON

{
  "@context": "https://alan-turing-institute.github.io/tea-techniques/data/ld/context.jsonld",
  "@id": "tea:technique/shapley-additive-explanations",
  "@type": "Technique",
  "name": "SHapley Additive exPlanations",
  "acronym": "SHAP",
  "description": "SHAP explains model predictions by quantifying how much each input feature contributes...",

  "addressesGoal": [
    "tea:goal/explainability",
    "tea:goal/fairness",
    "tea:goal/reliability"
  ],
  "hasTag": [
    "tea:tag/applicable-models/architecture/model-agnostic",
    "tea:tag/evidence-type/quantitative-metric"
  ],
  "relatedTo": [
    "tea:technique/local-interpretable-model-agnostic-explanations",
    "tea:technique/permutation-importance"
  ],
  "citesResource": ["tea:resource/shapShap2016"]
}

Vocabulary Mappings

The graph uses a custom namespace with mappings to established vocabularies:

Prefix	Namespace	Usage
`tea:`	`https://alan-turing-institute.github.io/tea-techniques/ns/`	Entity IRIs and custom predicates
`schema:`	`https://schema.org/`	Standard properties (name, description, url)
`skos:`	`http://www.w3.org/2004/02/skos/core#`	Tag taxonomy hierarchy (broader, narrower)

Entity IRI Patterns

What are IRIs?

IRIs (Internationalized Resource Identifiers) are unique identifiers for entities in the knowledge graph—like URLs but for data. They allow entities to be referenced unambiguously and linked across different systems.

tea:technique/{slug}              e.g. tea:technique/shapley-additive-explanations
tea:goal/{slug}                   e.g. tea:goal/explainability
tea:tag/{category}/{path}         e.g. tea:tag/applicable-models/architecture/model-agnostic
tea:resource/{citationKey}        e.g. tea:resource/shapShap2016

Accessing the Knowledge Graph

Static JSON-LD Files

All JSON-LD files are served from GitHub Pages for programmatic access:

https://alan-turing-institute.github.io/tea-techniques/data/ld/

Endpoint	Description
`graph.jsonld`	Complete knowledge graph (all entities in a single `@graph` array)
`techniques.jsonld`	All techniques collection
`techniques/{slug}.jsonld`	Individual technique documents
`goals.jsonld`	All assurance goals
`goals/{slug}.jsonld`	Individual goal documents
`tags.jsonld`	Full tag taxonomy
`resources.jsonld`	All academic resources
`context.jsonld`	JSON-LD `@context` document
`index.json`	API discovery document with entity counts

MCP Server for AI-Powered Discovery

The TEA Techniques MCP server gives AI assistants like Claude direct access to the knowledge graph through 10 specialised tools. Instead of writing queries, you can ask questions in natural language.

View MCP Server setup guide and tool reference →