Module temporal 

Temporal knowledge graphs: quads (head, relation, tail, timestamp).

Facts in an event KG (ICEWS-style) carry a discrete timestamp. This module loads quad datasets, defines the TemporalScorer trait (the temporal counterpart of Scorer, same lower-is-better energy convention), provides the TComplEx CPU scorer (Lacroix, Obozinski, Usunier, ICLR 2020: ComplEx with a complex timestamp embedding multiplied into the relation), and evaluates temporal link prediction in the time-aware filtered setting.

Training lives in burn_temporal (feature burn-ndarray or burn-wgpu); the CLI subcommand is tranz train-temporal.

Structs

QuadIds

One interned quad: (head, relation, tail, time) as dense ids.

TComplEx

TComplEx (Lacroix et al., ICLR 2020): -Re(<h, r ∘ w_τ, conj(t)>).

TemporalDataset

An interned temporal dataset: vocabularies plus train/valid/test quads.

TemporalFilterIndex

Time-aware filter index: known tails of (h, r, τ) and known heads of (r, t, τ), over all splits.

Traits

TemporalScorer

Trait for scoring temporal quads. Same convention as Scorer: lower values indicate more likely quads.

Functions

evaluate_temporal_link_prediction

Evaluate temporal link prediction in the time-aware filtered setting.

load_temporal_dataset

Load a temporal dataset from a directory with train.txt, valid.txt, test.txt, each line head \t relation \t tail \t timestamp.