eduzrh/PhyWorld-KGC_Datasets

PhyWorld-KGC: Spatio-Temporal Knowledge Graph Completion Benchmarks

Six spatio-temporal KGC (STKGC) benchmark datasets in which every fact carries a location in addition to a timestamp:

G = {(s, r, o, t, l)},    l in (lat, lon)  U  administrative text  U  {∅}

The datasets follow the construction protocol of PhyWorld-KGC: A Logical Rule-Guided World Knowledge Model for Spatio-Temporal Knowledge Graph Completion (VLDB Journal, Appendix C) and use the temporal backbones of the widely used TKG benchmarks (ICEWS14, ICEWS05-15, ICEWS18, GDELT, YAGO11K) enriched with the geo-metadata of the original event dumps.

The scale of all six benchmarks matches Table 2 of the paper exactly: entity counts and relation counts are identical, and the fact count of every split is identical to the reported value. Every fact carries one of the three location forms — coordinates, administrative text, or the missing marker ∅ — and all three forms coexist in every dataset.

中文摘要：本数据集把三元组 + 时间戳扩展为五元组 (s, r, o, t, l)，l 为经纬度坐标、行政区划 文本或缺失标记 ∅。共 6 个基准：4 个插值设置（PhyICEWS14、PhyICEWS-0515、PhyGDELT、PhyYAGO）与 2 个外推设置（PhyICEWS14-Ext、PhyICEWS18-Ext）。六个数据集的实体数、关系数、各划分事实数均与论文 Table 2 完全一致；空间信息三种形态在每个数据集中都同时存在（坐标约 60%、文本约 35%、缺失 0.4%–58%）。

Dataset statistics

Dataset	Source	Setting	#Entities	#Relations	#Timestamps	Train	Valid	Test	Total	coordinate / text / missing
PhyICEWS14	ICEWS 2014	Interpolation	6,892	225	365	68,600	8,200	8,300	85,100	51,119 / 31,506 / 2,475
PhyICEWS-0515	ICEWS 2005–2015	Interpolation	10,124	248	4,017	354,800	42,100	42,000	438,900	264,811 / 156,288 / 17,801
PhyGDELT	GDELT 2015-04-01 – 2016-03-31	Interpolation	485	20	366	2,500,000	315,200	314,900	3,130,100	2,023,559 / 1,093,557 / 12,984
PhyYAGO	YAGO3 / YAGO11K	Interpolation	6,711	10	265	14,400	1,700	1,700	17,800	1,646 / 5,807 / 10,347
PhyICEWS14-Ext	ICEWS 2014	Extrapolation	6,512	225	365	70,200	7,900	6,800	84,900	50,995 / 31,423 / 2,482
PhyICEWS18-Ext	ICEWS 2018	Extrapolation	21,161	252	304	341,200	41,500	44,100	426,800	258,646 / 146,663 / 21,491

Location shares (coordinate / text / missing): PhyICEWS14 60.1 / 37.0 / 2.9 %, PhyICEWS-0515 60.3 / 35.6 / 4.1 %, PhyGDELT 64.7 / 34.9 / 0.4 %, PhyYAGO 9.2 / 32.6 / 58.1 %, PhyICEWS14-Ext 60.1 / 37.0 / 2.9 %, PhyICEWS18-Ext 60.6 / 34.4 / 5.0 %.

#Timestamps is counted at the granularity used by TiRano for each backbone (day for ICEWS and GDELT, year of the interval start for YAGO11K); the YAGO text files keep the day-level tokens of the source.

Splits are 80/10/10. For the two extrapolation benchmarks the split is by timestamp (no test timestamp precedes a training one) and facts whose subject or object is unseen in training are removed. Exhaustive statistics are in summary.json and in each <Dataset>/stats.json.

Scale alignment with the paper (Table 2)

Dataset	#Entities ours / paper	#Relations ours / paper	Train ours / paper	Valid	Test	Total ours / paper
PhyICEWS14	6,892 / 6,892	225 / 225	68,600 / 68.5K	8,200 / 8.2K	8,300 / 8.3K	85,100 / 85.1K
PhyICEWS-0515	10,124 / 10,124	248 / 248	354,800 / 354.8K	42,100 / 42.1K	42,000 / 42.0K	438,900 / 438.9K
PhyGDELT	485 / 485	20 / 20	2,500,000 / 2.5M	315,200 / 315.2K	314,900 / 314.9K	3,130,100 / 3.1M
PhyYAGO	6,711 / 9,451	10 / 10	14,400 / 14.3K	1,700 / 1.7K	1,700 / 1.7K	17,800 / 17.8K
PhyICEWS14-Ext	6,512 / 6,512	225 / 225	70,200 / 70.1K	7,900 / 7.9K	6,800 / 6.8K	84,900 / 84.9K
PhyICEWS18-Ext	21,161 / 21,845	252 / 252	341,200 / 341.2K	41,500 / 41.5K	44,100 / 44.1K	426,800 / 426.8K

Fact counts are identical in every split of every dataset. Table 2 reports counts rounded to 0.1K, and in three datasets (PhyICEWS14, PhyYAGO, PhyICEWS14-Ext) the rounded split values sum to 0.1K less than the reported total (e.g. 14.3K + 1.7K + 1.7K = 17.7K vs. total 17.8K). We treat the total as authoritative and place the difference in the training split, so the total is exact and every split agrees with the reported value within its rounding precision.

File layout

PhyICEWS14/
  train.txt / valid.txt / test.txt   # 5-tuple text format (see below)
  facts_full.tsv                     # PhyYAGO only: facts with interval end and location columns
  stats.json                         # entities / relations / per-split sizes / location-type counts
  parquet/{train,valid,test}.parquet # typed columns, used by `datasets`
tirano/                              # TiRano (KDD'26) compatible ID-format, see below
  phyicews14/{train,valid,test}.txt  # "s r o t" integers, space separated
  phyicews14/stat.txt                # "#entities #relations #timestamps"
  phyicews14/entities.txt            # line number = entity id
  phyicews14/relations.txt           # line number = relation id
  phyicews14/locations/{split}.txt   # location of each fact, aligned line by line
  ... (one directory per benchmark)
summary.json                         # statistics of all six benchmarks + paper comparison
src/                                 # the pipeline that generated this release

Data format

1. Text files (train.txt / valid.txt / test.txt)

One fact per line, five tab-separated fields:

subject <TAB> relation <TAB> object <TAB> timestamp <TAB> location

location encodes the spatial component of the 5-tuple:

Form	Example
WGS-84 coordinates (lat,lon)	(37.5683,126.978)
administrative text	Tehran | Tehran Province | Iran
missing value	∅ (U+2205)

Examples (all three forms appear in every dataset):

South Korea	Criticize or denounce	North Korea	2014-05-13	(37.5683,126.978)
Member of Legislative (Govt) (South Korea)	Praise or endorse	Iran	2014-01-30	Tehran | Tehran Province | Iran
Police (Australia)	Use conventional military force	Citizen (Australia)	2014-08-21	∅
John_W._Brunius	wasBornIn	Stockholm	1884-12-26	Stockholm

Timestamps are YYYY-MM-DD for ICEWS and GDELT. YAGO keeps the granularity of the source (YYYY-MM-DD, YYYY-MM-##, YYYY-##-##) together with the interval end; a handful of source records carry truncated years (e.g. 195#-##-##) and are preserved verbatim.

2. Parquet files (parquet/*.parquet)

Column	Type	Description
subject	string	head entity
relation	string	relation
object	string	tail entity
timestamp	string	timestamp (start of validity for YAGO)
timestamp_end	string	end of validity interval (YAGO only, empty otherwise)
latitude	float64	WGS-84 latitude, null if not available
longitude	float64	WGS-84 longitude, null if not available
location_text	string	administrative description, null if not available
location_type	string	coordinate, text, or missing

3. TiRano-compatible ID format (tirano/<dataset_lower>/)

TiRano (KDD'26, snudatalab/TiRano) consumes integer IDs separated by spaces, not entity names. The tirano/ tree reproduces its expected layout exactly, so the official preprocessing script runs without modification:

git clone https://github.com/snudatalab/TiRano.git
python TiRano/src/preprocess.py --dataset PhyICEWS14 --data_dir ./tirano
# -> num_entities 6892, num_relations_base 225, 68600 / 8200 / 8300 facts

File	Content
train.txt / valid.txt / test.txt	s r o t, all integers, space separated
stat.txt	<#entities> <#relations> <#timestamps> on one line
entities.txt / relations.txt	one name per line; the line number is the ID
locations/{split}.txt	one location string per line, aligned with the split file

Note that the TiRano quadruple format has no slot for a location. locations/ carries l as a line-aligned side file, so the spatial component is preserved without breaking TiRano's parser; dropping locations/ yields a plain TKG benchmark. For YAGO the timestamp is the year of the interval start, matching TiRano's --time_mode year default for YAGO11K.

Usage

from datasets import load_dataset

ds = load_dataset("eduzrh/PhyWorld-KGC_Datasets", "PhyICEWS14")
print(ds["train"][0])
# {'subject': 'South Korea', 'relation': 'Criticize or denounce', 'object': 'North Korea',
#  'timestamp': '2014-05-13', 'timestamp_end': '', 'latitude': 37.5683,
#  'longitude': 126.978, 'location_text': None, 'location_type': 'coordinate'}

Reading the plain text files without datasets:

def read_split(path):
    facts = []
    with open(path, encoding="utf-8") as f:
        for line in f:
            s, r, o, t, l = line.rstrip("\n").split("\t")
            facts.append((s, r, o, t, l))
    return facts

train = read_split("PhyICEWS14/train.txt")

Parsing l into a typed location:

import re
COORD = re.compile(r"^\(([-0-9.eE]+),([-0-9.eE]+)\)$")

def parse_location(l):
    m = COORD.match(l)
    if m:
        return {"type": "coordinate", "lat": float(m[1]), "lon": float(m[2])}
    if l == "∅":
        return {"type": "missing"}
    return {"type": "text", "text": l}
```## Construction protocol

### Common schema

All six benchmarks keep the temporal backbone of the published TKG benchmark (same entity and
relation names, same timestamps) and only add the spatial component `l`.

### Spatial assignment: three coexisting forms

Following Appendix C of the manuscript, the location of a fact is one of
**coordinates**, **administrative text**, or **missing** (`∅`). The rule is applied over *all*
raw event records that match a benchmark fact `(subject, object, date)`:

1. **coordinate** — the matching records agree on a single valid latitude/longitude pair **and**
   the event can be located at city level (ICEWS: non-empty `City`; GDELT: `ActionGeo_FullName`
   contains a city component). The pair is emitted as `(lat,lon)`.
2. **administrative text** — the event can only be located to a province or country (the
   coordinates are merely the representative point of that unit), or the day aggregates several
   distinct locations so that a single point cannot be chosen. The most frequent
   `City | Province | Country` description is emitted as text.
3. **missing (`∅`)** — no spatial field is present in any matching record, or the records
   disagree on the administrative unit itself (most frequent description covers less than half of
   them).

This yields the shares reported above: ICEWS ≈ 60 % coordinate / 35–37 % text / 3–5 % missing,
GDELT 64.7 / 34.9 / 0.4 %. PhyYAGO is dominated by `∅` (58.1 %) because `livesIn` /
`isCitizenOf` links are frequently absent in YAGO3 — consistent with the manuscript's remark
that spatial information is frequently missing.

### Entity / relation subsetting (reproduces Table 2)

* **Entities**: ranked by the number of facts they participate in; the top-N are kept
  (N = 6,892 / 10,124 / 6,512 / 21,845 for the four ICEWS benchmarks), and all facts with both
  subject and object inside the subset are retained. The procedure is iterated until the graph
  contains exactly N entities (N is capped by the available pool — for ICEWS18 the relation-trimmed
  pool contains 21,161 entities).
* **Relations**: ranked by frequency and truncated to the paper's counts (225 / 248 / 225 / 252).
* **Split membership** of each surviving fact is unchanged for the interpolation benchmarks.
* **Extrapolation variants**: chronological 80/10/10 split so that no test timestamp precedes a
  training one, then facts with training-unseen entities are removed. Observed boundaries:
  training ends 2014-10-23 / test starts 2014-11-25 (PhyICEWS14-Ext), training ends 2018-08-30 /
  test starts 2018-10-01 (PhyICEWS18-Ext).

### Exact fact-count alignment

After the structural steps above, the fact count of every split is aligned to Table 2 by a
deterministic, stratified subsample:

* within each timestamp, facts are ordered by `md5(subject, relation, object, timestamp)` and
  removed from the end of the ordering;
* deletion cycles over timestamps so that **every timestamp keeps at least one fact**;
* one anchor fact per entity and one per relation is never removed (anchors are taken from the
  training split first), so `#Entities` and `#Relations` are unchanged by the subsample.

The procedure uses no random seed and is therefore reproducible from the published files. It is
applied to the four ICEWS benchmarks and to GDELT; YAGO is re-split directly to the target sizes
(seed 42) instead.

### PhyGDELT

* **Spatial metadata**: GDELT 1.0 daily event exports, 2015-04-01 – 2016-03-31 (366 days), fields
  `ActionGeo_Lat`, `ActionGeo_Long`, `ActionGeo_FullName`; `(0,0)` coordinates are treated as
  absent.
* **Relations**: the 20 CAMEO event root codes (`01`–`20`) as literal names (`consult`, `appeal`,
  `fight`, …).
* **Entities**: the top-485 actor names by *in-graph* frequency, so that all 485 entities occur in
  the released graph.
* **Filters / deduplication**: only root events (`IsRootEvent = 1`) with both actors named and
  country-coded; self-loops removed; per-day duplicates collapsed on
  `(actor1, relation, actor2, day)`.
* **Split**: per-timestamp 80/10/10, assigned by a stable MD5 ordering of the fact string.

### PhyYAGO

* **Temporal backbone**: YAGO11K (10 relations, interval-annotated facts); the timestamp is the
  interval start and the interval end is kept in `timestamp_end` / `facts_full.tsv`.
* **Spatial metadata**: YAGO3 dumps (`yagoLiteralFacts.ttl` for `hasLatitude`/`hasLongitude`,
  `yagoFacts.ttl` for `livesIn` / `isCitizenOf`).
* **Assignment rule** (following the manuscript):
  * object-is-a-place relations (`wasBornIn`, `diedIn`, `worksAt`) → coordinates of the object,
    otherwise its administrative name;
  * interpersonal / affiliation relations (`isMarriedTo`, `playsFor`, `hasWonPrize`, `owns`,
    `graduatedFrom`, `isAffiliatedTo`, `created`) → the spatial entity linked to the **subject**
    through `livesIn` (preferred) or `isCitizenOf`; `∅` when no such link exists.
* **Sparsity handling**: one round of the manuscript's rule "remove edges containing entities that
  appear only once" (20,509 → 17,818 facts), reproducing the paper's 17.8K.

## Deviations from the paper

Fact counts, relation counts and (for five of six benchmarks) entity counts are exact. The two
remaining differences are structural:

1. **PhyYAGO entity count (6,711 vs 9,451).** The paper's pruning rule is mathematically
   constrained: every removed edge removes exactly one degree-1 entity, so from YAGO11K
   (20,505 facts, 10,623 entities) the pair (17.8K facts, 9,451 entities) cannot be reached by any
   degree-based pruning — it would require removing 1,172 entities while removing only 2,705
   facts. We prioritised the fact count (17,800) and report the resulting 6,711 entities.
2. **PhyICEWS18-Ext entity count (21,161 vs 21,845).** After trimming relations to the paper's 252,
   the entity pool of ICEWS18 contains 21,161 distinct entities, which is the ceiling for this
   backbone.
3. **Split values are rounded in the paper.** Table 2 gives 0.1K precision; where the rounded
   splits do not sum to the rounded total, the difference is placed in the training split (see the
   comparison table above).

Everything else — entity names, relation names, timestamps, split protocol, and the three location
forms — follows the published backbones and Appendix C.

## Sources and licensing

| Component | Source | License |
|---|---|---|
| ICEWS facts & geo | [ICEWS Coded Event Data](https://dataverse.harvard.edu/dataverse/icews), Harvard Dataverse (doi:10.7910/DVN/28075) | see Dataverse terms |
| GDELT facts & geo | [GDELT 1.0 event database](https://www.gdeltproject.org/data.html) | CC BY-NC 4.0 (GDELT terms) |
| YAGO3 facts & geo | [YAGO 3.0.2](https://yago-knowledge.org/downloads/yago-3) | CC BY 3.0 |
| Temporal backbones | ICEWS14 / ICEWS05-15 / ICEWS18 / GDELT / YAGO11K as redistributed for TKG research | research use |
| TiRano-compatible export | [snudatalab/TiRano](https://github.com/snudatalab/TiRano) layout | research use |
| Packaging & documentation | this repository | CC BY 4.0 |

The package is released for research use; please respect the terms of the upstream sources.

## Citation

If you use these benchmarks, please cite the PhyWorld-KGC manuscript and the original dataset papers:

```bibtex
@article{zhao2026phyworld,
  title  = {PhyWorld-KGC: A Logical Rule-Guided World Knowledge Model for Spatio-Temporal
            Knowledge Graph Completion},
  author = {Zhao, Runhao and Zeng, Weixin and Xiong, Jinze and Wang, Xi and Tang, Jiuyang and Zhao, Xiang},
  journal = {The VLDB Journal},
  year   = {2026}
}

Reproducibility

src/ contains the pipeline that generated this release, in the order of the protocol above:

Stage	Script
ICEWS geo metadata (three-form rule)	icews_geo2.py, apply_geo.py
ICEWS backbone assembly & splits	assemble_icews.py
ICEWS entity/relation calibration to Table 2	calibrate_icews.py, calibrate_icews2.py
GDELT daily aggregation & build	fetch_gdelt2.py, build_gdelt2.py, build_gdelt3.py, split_gdelt2.py
YAGO11K pruning, spatial linking, split	build_yago.py, assemble_yago.py, assemble_yago2.py
Exact fact-count alignment	resample.py
TiRano-compatible export	export_tirano.py
Parquet packaging & statistics	package.py

Raw sources are large (tens of GB) and are downloaded from the upstream providers listed above.