tcomplex-icews14

TComplEx ICEWS14 (dim 256)

Temporal knowledge-graph embeddings for ICEWS14: TComplEx (Lacroix, Obozinski, Usunier, ICLR 2020) trained with the tranz CLI (v0.7.1) on Burn's wgpu/Metal backend, with the paper's weighted nuclear-3 (Ω³) and temporal smoothness (Λ₃) regularizers.

Time-aware filtered link prediction, test split: MRR 0.520, Hits@1 0.424, Hits@3 0.580, Hits@10 0.695.

Data

ICEWS14 event quads (head, relation, tail, YYYY-MM-DD), the Garcia-Duran, Dumancic & Niepert (EMNLP 2018) splits, fetched from the mmkb release (TemporalKGs/icews14/icews_2014_{train,valid,test}.txt): 72,826 / 8,941 / 8,963 train/valid/test quads over 7,128 entities, 230 relations, and 365 daily timestamps (all of 2014). Entity and relation vocabularies are interned in first-appearance order over the splits; timestamps are interned in sorted (chronological) order.

Training command

# data fetched by heyting's scripts/fetch_icews14.sh (mmkb, see above)
tranz train-temporal --data data/icews14 \
  --dim 256 --epochs 100 --batch-size 1024 --lr 0.01 --init-scale 0.01 \
  --label-smoothing 0.1 --n3-reg 0.0025 --time-smooth 1.0 \
  --output data/icews14-tcomplex --eval

Build: cargo install tranz --features "burn-ndarray,burn-wgpu". About 190 s on an Apple M-series GPU; the --eval flag prints the metrics above (1-N cross-entropy both directions, AdamW, no reciprocals). Hyperparameters were selected on the validation split. Keep --init-scale 0.01 when the regularizers are on: at the 1e-3 default the origin is a fixed point of the trilinear score and the N3 pull wins (loss converges to exactly ln|E|, MRR ~0).

Files and format

• entities.tsv — 7,128 entity embeddings
• relations.tsv — 230 relation embeddings
• times.tsv — 365 timestamp embeddings, rows in chronological order (row 0 = 2014-01-01)

Each file is word2vec-style text. The first line is <count> <dim> (space-separated); every following line is a name and dim floats, all TAB-separated (names contain spaces):

7128 512
South Korea	-0.19331053	-0.6447717	0.63432753	...

dim is 512 because the complex dimension is 256: columns 0..256 are the real parts, columns 256..512 the imaginary parts. The score of a quad is Re(<h, r ∘ w_τ, conj(t)>) (higher = more likely; tranz's TemporalScorer negates it into a lower-is-better energy).

Loading

Rust (tranz):

let (names, vecs) = tranz::io::import_embeddings(Path::new("entities.tsv"))?;
// ... same for relations.tsv and times.tsv, then:
let model = tranz::temporal::TComplEx::from_vecs(ent, rel, time, 256);

Python:

import numpy as np

def load_w2v_tsv(path):
    names, rows = [], []
    with open(path) as f:
        n, d = map(int, f.readline().split())
        for line in f:
            parts = line.rstrip("\n").split("\t")
            names.append(parts[0])
            rows.append(np.array(parts[1:], dtype=np.float32))
    return names, np.stack(rows)   # (n, d); re = [:, :d//2], im = [:, d//2:]

Temporal complex-query answering over this checkpoint (windowed hops, witnesses, conformal answer sets): the heyting crate's icews14_temporal_clqa example.