RangeFC: Temporal Fact Checking with Year-Range Prediction

Predicting coherent validity intervals for knowledge graph assertions.

Accepted at ISWC 2026

RangeFC is an interval-aware temporal fact-checking framework that predicts the start and end years of validity for a given knowledge graph assertion ((s,p,o)). The model extends the TemporalFC codebase with a relation-aware Mixture-of-Experts architecture, a shared calendar representation, coupled boundary prediction, and overlap-aware training.

Highlights

• Predicts complete validity intervals rather than a single timestamp.
• Uses frozen Dihedron entity and relation embeddings.
• Employs relation-aware top-(k) Mixture-of-Experts routing.
• Uses a factorized calendar head for start/end prediction.
• Optimizes interval quality using boundary and overlap-aware objectives.
• Evaluated on five Wikidata temporal property pairs.

Main Results

Model	MAE Start ↓	MAE End ↓	IoU ↑
TemporalFC	12.07	28.74	0.624
RangeFC	18.94	21.37	0.661

RangeFC improves overall interval overlap while substantially reducing end-boundary error compared with the adapted TemporalFC baseline.

Installation

Clone the repository:

git clone https://github.com/dice-group/RangeFC.git
cd RangeFC

Create the Conda environment:

conda env create -f environment.yml
conda activate tfc_clean_gpu

Dataset

The dataset used for the RangeFC experiments is wikidata6_latest2.

Due to the size of the dataset and the pre-trained Dihedron embeddings, the dataset files are distributed separately through the GitHub Releases section.

Download wikidata6_latest2.zip and extract it into the data_TP/ directory.

The expected structure is:

RangeFC/
├── main.py
├── executer_TP.py
├── data_TP.py
├── ...
└── data_TP/
    └── wikidata6_latest2/
        ├── train/
        │   └── train
        ├── valid/
        │   └── valid
        ├── test/
        │   └── test
        ├── embeddings/
        │   └── dihedron/
        │       ├── entity.npy
        │       ├── entity.pkl
        │       ├── relation.npy
        │       └── relation.pkl
        ├── entities
        ├── relations
        └── times

Running RangeFC

Example command:

python main.py \
  --path_dataset_folder "./data_TP/" \
  --eval_dataset "wikidata6_latest2" \
  --task "range-prediction" \
  --model "range-mlp" \
  --emb_type "dihedron" \
  --embedding_dim 100 \
  --batch_size 1024 \
  --val_batch_size 1000 \
  --use_interaction 1 \
  --use_prod 0 \
  --loss_type "huber" \
  --huber_beta 0.7366304701152739 \
  --end_weight 1.0013169655965504 \
  --extra_order_pen 0.06686696713024719 \
  --lr 0.0018887997194523478 \
  --num_workers 4 \
  --max_num_epochs 120 \
  --seed 42 \
  --emb_noise 0.011450934693677826 \
  --hidden_dim 1024 \
  --dropout 0.11699588659730913 \
  --gauss_sigma_idx 1.3507654810031005 \
  --t_dim 128 \
  --num_experts 5 \
  --k_experts 3 \
  --gate_temp_start 1.4307208041278083 \
  --gate_balance 0.023266080281596692 \
  --use_bands 1 \
  --use_prior 0 \
  --band_margin 1.5763792105820356

Authors

This work was developed by the authors of the RangeFC paper:

Abdullah Qamar, Umair Qudus, Michael Röder, Axel-Cyrille Ngonga Ngomo

Acknowledgement

RangeFC builds upon the open-source TemporalFC codebase. We thank the TemporalFC authors for making their implementation publicly available.

Paper and Citation

This repository contains the implementation accompanying the paper:

“RangeFC: Temporal Fact Checking with Year-Range Prediction” Accepted at the 25th International Semantic Web Conference (ISWC 2026).

The official paper link, DOI, and BibTeX citation will be added once the proceedings are available.