This repository provides an implementation of LINCOLN as described in the paper: LINCOLN: Learning High-Order Dynamics of Real-World Networks.
- Yunyong Ko (yyko@cau.ac.kr)
- Da Eun Lee (ddanable@hanyang.ac.kr)
- Song Kyung Yu (ssong915@hanyang.ac.kr)
- Sang-Wook Kim (wook@hanyang.ac.kr)
-
Observations
- We observe that high-order relations tend to (O1) have a structural and temporal influence on other relations in a short term and (O2) periodically re-appear in a long term.
-
Method
- We propose a novel dynamic hypergraph learning method, Lincoln, that effectively captures long-term and short term patterns of high-order relations in real-world networks.
-
Evaluation
- Via extensive experiments on seven real-world datasets, we demonstrate that Lincoln outperforms nine state-of-the-art methods in the dynamic hyperedge prediction task.
| Name | # of nodes | # of hyperedges | Range | Snapshot frequency | # of snapshots |
|---|---|---|---|---|---|
| Email-enron | 143 | 10,883 | 3 Years 5 Months | Monthly | 40 |
| Email-eu | 979 | 234,760 | 2 Years 2 Months | Weekly | 78 |
| Tags | 3,021 | 271,233 | 8 Years 6 Months | Monthly | 86 |
| Thread | 90,054 | 192,947 | 7 Years 6 Months | Monthly | 86 |
| Contact-primary | 242 | 106,879 | 32.4 Hours | Quarter of a Hour | 71 |
| Contact-high | 327 | 172,035 | 4.2 Days | Quarter of a Hour | 164 |
| Congress | 1,718 | 260,851 | 33 Years 6 Months | Quarter of a Year | 98 |
We provide 7 real-world datasets for our new benchmark task at Here and sampling negative samples code /data/sampler.py
# File Organization
[ dataset name ]
|__ cns_[dataset_name].pt
|__ mns_[dataset_name].pt
|__ sns_[dataset_name].pt # negative hyperedges samples used for training and evaluation
|__ hyper_[datset_name].csv # original hypergraph dataset
|__ new_hyper_[datset_name].csv # reindexed hypergraph dataset
The source code used in the paper is available at ./LINCOLN/.
python main.py --dataset_name [name of dataset] --snapshot_size [size of snapshot for each datset]
More details about arguments are described in ./LINCOLN/utils.py.
Our code runs on the Intel i7-7700k CPU with 64GB memory and NVIDIA RTX 2080 Ti GPU, installed with CUDA 12.2 and cuDNN
8.2.1., with the following packages installed in environment.yaml
