jhkim611/CHASE-datasets

CHASE Dynamic Graph Datasets

Datasets for "Chasing Label Shifters: A Change-Aware Framework for Dynamic Graph Node Classification" (NeurIPS 2026).

Each dataset is a discrete-time dynamic graph where node labels evolve over snapshots. Shifters are nodes whose label changes between consecutive snapshots (label[t] != label[t-1]).

File	Domain	Nodes	Edges	Snapshots (train/val/test)	Classes	Avg Shifter %	Label meaning
DBLP-S.pkl	Academic	34,717	394,385	14 (9/2/3)	2	2.62%	Research area
GossipCop.pkl	Social Media	6,091	2,000,390	37 (26/5/6)	2	4.06%	>50% fake news consumption
Music.pkl	Music Streaming	878	1,009,457	24 (15/4/5)	5	18.77%	Dominant listening genre

Edges are undirected edges summed across snapshots, excluding self-loops. Splits are chronological by snapshot.

• DBLP-S: co-authorship edges, yearly snapshots.
• GossipCop: co-readership edges (shared article consumption), weekly snapshots.
• Music: co-listening edges (shared track in the same month), monthly snapshots, 2007–2008.

Format

Each file is a Python pickle containing a dict:

key	type / shape	description
edge_indices	list of (2, E_t) arrays	per-snapshot edge index
edge_weights	list of (E_t,) arrays	per-snapshot edge weights
feats	(K, N, F) float array	per-snapshot node features
labels	(K, N) array	per-snapshot node labels
mask	(K, N) 0/1 array	per-snapshot node activity mask

Usage

hf download jhkim611/CHASE-datasets --repo-type dataset --local-dir main_files
python run.py --dataset DBLP-S --use-chase

Or in Python:

import pickle
from huggingface_hub import hf_hub_download

path = hf_hub_download("jhkim611/CHASE-datasets", "Music.pkl", repo_type="dataset")
data = pickle.load(open(path, "rb"))

Checksums are in SHA256SUMS.

License and sources

Released under CC BY-NC 4.0 (non-commercial use). The datasets are derived from the sources below; please also respect their original terms and cite them.

• DBLP-S: DBLP-v11.
• GossipCop: FakeNewsNet.
• Music: LastFM-song-listens and the Million Song Dataset.

Code

github.com/jhkim611/CHASE

Citation

TBD