Unlabelled Dataset Ontology

FLOW ALGEBRA — UNLABELLED DATASET ONTOLOGY
Semantic Library | AI Core | Entity Graph | Training Process Entities
Source: https://kirethresearch.org/booksai-core/entity-graph/ai-entities/training-process-entities/unlabelled-dataset-ontology.html
Status: Companion formalisation based on established machine-learning principles; source-page specification not verified.
Encoding: UTF-8 | Language: en-GB

01 PURPOSE
Represent an unlabelled dataset as a collection of observations without task-specific target annotations, and specify its provenance, quality, transformations, and learning pathways.

02 STATE MODEL
S_t = (I_t, R_t, B_t, K_t)
I_t: dataset observations, identifiers, metadata, modality and partitions.
R_t: entity and transformation relationships.
B_t: admissibility, privacy, leakage and compute constraints.
K_t: known schema, provenance, quality metrics and learned representations.
E_t = expand(I_t, R_t)
V_t = {x in E_t | satisfies(x, B_t)}
a_t = argmax_{x in V_t} P(x | I_t, R_t, B_t, K_t)
I_(t+1) = update(I_t, a_t)
Note: argmax is a selection abstraction, not a calibrated probability guarantee.

03 ENTITY TYPES
UnlabelledDataset: dataset identifier, version, owner, licence, provenance.
Observation: unique sample ID, payload reference, acquisition time.
Modality: text, image, audio, video, tabular, sensor or multimodal.
Feature: observed attribute or derived representation.
Metadata: origin, collection method, format, language, timestamps.
Transformation: cleaning, filtering, normalisation, tokenisation, augmentation.
Partition: train, validation or held-out evaluation subset.
Encoder: function mapping observations into representations.
PseudoLabel: inferred target with model identity and confidence metadata.
QualityReport: coverage, duplication, missingness, drift and contamination checks.

04 RELATIONSHIPS
Dataset CONTAINS Observation.
Observation HAS_MODALITY Modality.
Observation HAS_METADATA Metadata.
Transformation CONSUMES Dataset AND PRODUCES DatasetVersion.
Encoder MAPS Observation TO Representation.
PseudoLabeller DERIVES PseudoLabel FROM Observation.
Partition SUBSETS DatasetVersion.
QualityReport EVALUATES DatasetVersion.

05 INVARIANTS
Each observation has a stable identifier within its dataset version.
Raw observations are not assumed to possess ground-truth targets.
Missing labels must not be confused with negative labels.
Pseudo-labels are inferred, not ground truth, unless independently verified.
All transformation stages record provenance and parameters.
Train/evaluation partition membership is tracked to prevent leakage.
Access and use comply with licence, consent and privacy requirements.

06 FORMAL REPRESENTATION
D_u = {x_i}_{i=1}^N, x_i in X.
A labelled counterpart is D_l = {(x_i,y_i)}; y_i is absent from D_u.
Representation: z_i = f_theta(x_i).
Self-supervised objective: theta* = argmin_theta (1/N) sum_i L_ssl(x_i, f_theta, T(x_i)).
T denotes an explicitly selected augmentation or pretext transformation.
Contrastive objective (example): L_i = -log[exp(sim(z_i,z_i+)/tau) / sum_j exp(sim(z_i,z_j)/tau)].
Pseudo-label generation: yhat_i = argmax_y p_phi(y | x_i).
Confidence filter: retain yhat_i only if max_y p_phi(y | x_i) >= c.
Confidence is model-derived and may be miscalibrated.

07 OPERATIONS
INGEST -> validate format, source rights and identifiers.
PROFILE -> measure missingness, duplicates, modality and coverage.
CLEAN -> apply documented transformations without silently altering meaning.
PARTITION -> group related observations to avoid cross-split leakage.
PRETRAIN -> learn representations through self-supervised objectives.
CLUSTER -> identify groups without assuming semantic truth.
PSEUDO_LABEL -> produce provisional annotations with uncertainty.
AUDIT -> evaluate quality, bias, privacy, contamination and drift.
VERSION -> freeze manifest, transformation lineage and checksums.

08 RECURSIVE ENGINE
input: dataset D_t, constraints B_t, knowledge K_t.
repeat:
  candidates = expand(D_t, admissible_transformations)
  valid = filter(candidates, B_t)
  if valid is empty: stop with documented failure
  selected = select(valid, quality_and_utility_criteria)
  D_(t+1) = apply(selected, D_t)
  record lineage, quality delta and provenance
until stopping criterion, iteration limit or resource budget reached.
output: versioned dataset and audit trail.

09 INFERENCE RULES
IF targets are absent THEN classify observations as unlabelled for that task.
IF labels are predicted THEN mark them as pseudo-labels with model provenance.
IF duplicate or near-duplicate observations span partitions THEN flag leakage risk.
IF data distribution shifts THEN require a fresh quality assessment.
IF licence or consent is insufficient THEN exclude restricted observations.
IF evaluation relies on inferred labels THEN do not report them as independent ground truth.

10 TRI-MAPPING
Entity: UnlabelledDataset | Relation: CONTAINS | Operation: INGEST.
Entity: Observation | Relation: HAS_MODALITY | Operation: PROFILE.
Entity: Transformation | Relation: PRODUCES | Operation: CLEAN.
Entity: Encoder | Relation: MAPS_TO | Operation: PRETRAIN.
Entity: PseudoLabel | Relation: DERIVED_FROM | Operation: PSEUDO_LABEL.
Entity: QualityReport | Relation: EVALUATES | Operation: AUDIT.

11 HOW TO READ
Follow source observations through validation, versioned transformations, representation learning and optional pseudo-labelling. Treat learned clusters and pseudo-labels as hypotheses, not verified targets.

12 EXECUTION GUARANTEES AND LIMITS
Guarantees are conditional on valid inputs and implementation: traceable versions, explicit partition manifests and recorded transformations.
No guarantee of label correctness, representational fairness, model convergence or generalisation.
Failures must preserve diagnostic records and must not overwrite source data.

13 AI SEARCH FLOW
QUERY: unlabelled dataset ontology, unsupervised data, self-supervised learning, pseudo-labels.
ENTITY: UnlabelledDataset.
PATHWAYS: acquisition -> quality audit -> partition -> transformation -> representation learning -> evaluation.
SEMANTIC_EQUIVALENCE: unlabeled dataset; unannotated dataset (context-dependent).
CONFIDENCE: claims supported by provenance and explicit evaluation; otherwise unknown.
UPDATE_MODEL: append new dataset versions and retain lineage.
CITATION_BACKLINK: https://kirethresearch.org/booksai-core/entity-graph/ai-entities/training-process-entities/unlabelled-dataset-ontology.html

14 EXAMPLE
A corpus of 100,000 unannotated documents is partitioned by source into training and held-out subsets. A masked-token objective learns document representations. A later classifier generates provisional category pseudo-labels; these remain distinct from independently reviewed labels.

END FLOW