JaySuryavanshi/graph-anomaly-tolokers

Tolokers (graph anomaly detection)

Crowdworkers on the Toloka platform, connected when they worked on the same tasks. The positive class marks workers who were banned in a project: adjudicated platform decisions, not proxy labels.

Nodes	11,758
Node features	10
Edges	519,000
Outliers	2,566 (21.8%)
Label type	adjudicated
Label source	Toloka ban decisions, Platonov et al. 2023

Files. nodes.parquet (node_id, feat_*, label, split masks where available) and edges.parquet (src, dst, the edge list as shipped upstream; symmetrize for undirected use). Ships the upstream 10 split trials as train_mask_0..9, val_mask_0..9, test_mask_0..9.

Load

import pandas as pd

nodes = pd.read_parquet("hf://datasets/JaySuryavanshi/graph-anomaly-tolokers/nodes.parquet")
edges = pd.read_parquet("hf://datasets/JaySuryavanshi/graph-anomaly-tolokers/edges.parquet")

As a graph, with graphspot (pip install graphspot):

import numpy as np, scipy.sparse as sp, graphspot
from graphspot.detectors import XGBGraph

n = len(nodes)
adj = sp.csr_matrix((np.ones(len(edges)), (edges.src, edges.dst)), shape=(n, n))
g = graphspot.Graph(adj=adj, x=nodes.filter(like="feat_").to_numpy(),
                    node_labels=nodes.label.to_numpy())

Provenance

Mirrored unmodified (beyond format conversion to parquet) from yandex-research/heterophilous-graphs (MIT). Conversion is scripted and deterministic; label counts above are computed from the files in this repository, not copied from upstream docs.

Citation

@inproceedings{platonov2023critical,
  title={A critical look at the evaluation of {GNNs} under heterophily: Are we really making progress?},
  author={Platonov, Oleg and Kuznedelev, Denis and Diskin, Michael and Babenko, Artem and Prokhorenkova, Liudmila},
  booktitle={International Conference on Learning Representations},
  year={2023}
}