Machine Learning Domain Ontology

FLOW ALGEBRA
TITLE: Machine Learning Domain Ontology
DOMAIN: Domain Ontologies / Computing Domains / Machine Learning
SOURCE: https://kirethresearch.org/booksai-core/ontologies/domain-ontologies/computing-domains/machine-learning-domain-ontology.html
VERSION: 1.0
AUTHORS: Mark Anthony Higgins — Sky (Recursive AI)
STATUS: Companion formalisation; source-page specification not independently verified.

flow "MachineLearningDomainOntology" v1.0 {
  ontology:
    id: "machine-learning-domain-ontology"
    parent: "computing-domains"
    purpose: "Represent machine learning tasks, datasets, models, training, evaluation, deployment and evidence-backed relationships."

  types:
    LearningTask: {id: string, kind: enum[classification, regression, clustering, dimensionality_reduction, forecasting, generation, reinforcement_learning, other], objective: string}
    Dataset: {id: string, version: string, provenance: string, split: enum[train, validation, test, production, other], access_policy: string?}
    Feature: {id: string, datatype: string, unit: string?, derivation: string?}
    Label: {id: string, datatype: string, annotation_source: string?}
    Model: {id: string, family: string, version: string, parameters_ref: string?, training_run: string?}
    TrainingRun: {id: string, algorithm: string, hyperparameters: map, seed: integer?, timestamp: datetime, provenance: string}
    Evaluation: {id: string, dataset_id: string, metric: string, value: number, uncertainty: number?, protocol: string, timestamp: datetime}
    Deployment: {id: string, model_id: string, environment: string, version: string, monitoring_policy: string?}
    Assertion: {id: string, subject: string, predicate: string, object: string, confidence: number, evidence: string[], context: string?}

  predicates:
    addresses(model: Model, task: LearningTask)
    trained_on(model: Model, dataset: Dataset)
    uses_feature(model: Model, feature: Feature)
    predicts(model: Model, label: Label)
    produced_by(model: Model, run: TrainingRun)
    evaluated_by(model: Model, evaluation: Evaluation)
    deployed_as(model: Model, deployment: Deployment)
    derived_from(dataset_a: Dataset, dataset_b: Dataset)
    supersedes(model_new: Model, model_old: Model)

  constraints:
    - "Every entity and assertion must have a stable identifier."
    - "Confidence values must lie within [0, 1]."
    - "Dataset splits and lineage must be recorded to assess leakage."
    - "Evaluation metrics require a defined task, dataset, protocol and interpretation."
    - "Do not equate high test performance with causal validity, fairness or production reliability."
    - "Training, validation and test datasets must not be assumed independent without verification."
    - "Model versions and data provenance must remain traceable."
    - "Predictions must not be represented as verified ground truth."

  inference:
    rule training_lineage:
      if produced_by(m,r) and run_dataset(r,d)
      then trained_on(m,d)
      confidence: "preserve training-run evidence"
    rule evaluation_trace:
      if evaluated_by(m,e) and evaluation_dataset(e,d)
      then evaluated_on(m,d)
      confidence: "retain evaluation protocol and uncertainty"
    rule model_version_lineage:
      if supersedes(m3,m2) and supersedes(m2,m1)
      then version_descendant_of(m3,m1)
      confidence: "preserve version provenance; do not infer performance improvement"

  SEMANTIC_EQUIVALENCE:
    - "Resolve model, dataset, metric and task aliases using versioned identifiers."
    - "Distinguish identical models from related architectures and compatible interfaces."
    - "Treat equivalent metrics as context-dependent on definition and evaluation protocol."

  CONFIDENCE:
    range: [0,1]
    factors: [data_quality, representativeness, evaluation_uncertainty, reproducibility, distribution_shift, provenance]
    policy: "Calibrate empirical confidence where possible; never manufacture certainty by chaining inferences."

  VALIDATION:
    - "Check type compatibility, unique IDs, dataset provenance and version consistency."
    - "Verify metric direction, units, sample sizes and uncertainty where available."
    - "Flag leakage, incompatible labels, unsupported performance claims and distribution shift."
    - "Preserve unresolved conflicts and alternative explanations."

  FLOW_HASH:
    algorithm: "SHA-256"
    input: "canonicalised ontology version, typed assertions, rules and provenance"
    note: "Compute at execution; no hash value is claimed here."

  UPDATE_MODEL:
    strategy: "versioned, provenance-preserving incremental update"
    steps: [ingest, normalise, validate, compare, reconcile, recompute_inferences, publish]
    rollback: true

  AI_CONSUMPTION:
    representation: "typed machine-learning knowledge graph with evidence-backed edges"
    outputs: [entity_links, model_lineage, dataset_lineage, evaluation_records, uncertainty_annotations, provenance_records]
    safeguards: [no_unsupported_performance_claims, prevent_data_leakage, preserve_dataset_versions, distinguish_prediction_from_fact]
}

END FLOW ALGEBRA