Evoluenet Dynamic Graph Transfer Eval

---
name: evoluenet-dynamic-graph-transfer-eval
description: Evaluates dynamic non-IID transfer learning on graphs by measuring how well a model adapts node classification knowledge from a source temporal graph to a target temporal graph with limited labeled samples. It probes the model's ability to handle evolving graph structures, domain discrepancies, and temporal dependencies across heterogeneous datasets. Use when the user wants to benchmark on DBLP-3, DBLP-5, HCP, or asks about evaluating this task. Reports AUC.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2305.00664
  bibtex_key: wang2023evoluenet
  confidence: high
---

# evoluenet-dynamic-graph-transfer-eval

> EvoluNet: Advancing Dynamic Non-IID Transfer Learning on Graphs — Wang et al. (2023) (arXiv:2305.00664, 2023)

## What this evaluates

Evaluates dynamic non-IID transfer learning on graphs by measuring how well a model adapts node classification knowledge from a source temporal graph to a target temporal graph with limited labeled samples. It probes the model's ability to handle evolving graph structures, domain discrepancies, and temporal dependencies across heterogeneous datasets.

## Datasets

- **DBLP-3, DBLP-5, HCP** — total ?; splits: target_last_timestamp (-1); repo https://github.com/wanghh7/EvoluNet

## Metrics

- `AUC` **(primary)** — range: [0, 1]
  - Area Under the Receiver Operating Characteristic Curve. Computed on node classification predictions versus true labels at the final timestamp of the target domain.

## Input / output format

**Input**: Temporal graph data (nodes, edges, node attributes) across multiple timestamps for source and target domains, with 5 labeled samples per class provided for the target domain at the final timestamp.

**Output**: Predicted node labels (or class probabilities) for all nodes in the target domain at the final timestamp.

## Scoring recipe

```python
def compute_auc(predictions, labels):
    # predictions: node classification probabilities or scores
    # labels: ground truth node labels
    fpr, tpr, _ = roc_curve(labels, predictions)
    auc_score = auc(fpr, tpr)
    return auc_score
```

## Common pitfalls

- The evaluation only uses 5 labeled samples per class in the target domain, which is a highly limited-label setting; results do not generalize to fully supervised scenarios.
- Metrics are reported only on the last timestamp of the target domain, ignoring intermediate temporal steps.
- The benchmarks involve cross-domain transfer between structurally heterogeneous graphs (DBLP vs HCP), so performance heavily depends on domain adaptation rather than pure temporal modeling.

## Evidence (verbatim from paper)

> We compare EVOLUNET with eleven baseline methods across three real-world undirected graphs. We report the AUC of different methods on the last timestamp of the target domain in Table 3.

## Citation

```bibtex
@misc{wang2023evoluenet,
  title={EvoluNet: Advancing Dynamic Non-IID Transfer Learning on Graphs},
  author={Wang et al. (2023)},
  year={2023},
  note={arXiv:2305.00664}
}
```

- arXiv: 2305.00664