Ponzi Detection Eval

ponzi-detection-eval

Time-aware Metapath Feature Augmentation for Ponzi Detection in Ethereum — Jin et al. (2022) (arXiv:2210.16863, 2022)

What this evaluates

This benchmark evaluates the effectiveness of feature augmentation modules for detecting Ponzi scheme accounts on the Ethereum blockchain. It probes a model's ability to classify account nodes as legitimate or malicious based on transaction graph structures and temporal behavior patterns.

Datasets

• Ethereum Ponzi dataset — total ?; splits: test (-1)

Metrics

• micro-F1 (primary) — range: [0, 1]
  • Micro-averaged F1 score computed across all account nodes by aggregating true positives, false positives, and false negatives globally before calculating precision and recall.

Input / output format

Input: Account node features (derived from manual engineering, graph embeddings like Line/DeepWalk/Node2Vec/Trans2Vec, or GNN initial representations) and the underlying Ethereum transaction graph structure.

Output: Binary classification label per account node (Ponzi vs. Normal).

Scoring recipe

# Pseudo-code for evaluation protocol
all_preds = []
all_labels = []
for fold in range(5):
    for seed in range(5):
        train_data, test_data = split_data(fold, seed)
        model.fit(train_data)
        preds = model.predict(test_data)
        all_preds.extend(preds)
        all_labels.extend(test_data.labels)
micro_f1 = compute_micro_f1(all_labels, all_preds)

Common pitfalls

• The test set contains only 38 Ponzi accounts, which severely limits statistical power and may cause high variance in reported scores.
• Performance is averaged over 25 runs (5 folds × 5 seeds); reporting only the mean can mask instability across different random seeds.
• The augmentation module is applied as a drop-in feature enhancer without modifying the base classifier, so gains are inherently constrained by the base method's representational capacity.

Evidence (verbatim from paper)

For all methods, we repeat 5-fold cross-validation five times with five different random seeds and report the average micro-F1 score over 5 × 5 = 25 experiments.

Citation

@misc{jin2022ponzidetection,
  title={Time-aware Metapath Feature Augmentation for Ponzi Detection in Ethereum},
  author={Jin et al. (2022)},
  year={2022},
  note={arXiv:2210.16863}
}

• arXiv: 2210.16863