NPA: Improving Large-scale Graph Neural Networks with Non-parametric Attention
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
Recent works show great interest in designing Graph Neural Networks (GNNs) that scale to large graphs. While previous work focuses on designing advanced sampling techniques for existing GNNs, the design of non-parametric GNNs, an orthogonal direction for scalable performance, has aroused lots of concerns recently. For example, nearly all top solutions in the Open Graph Benchmark leaderboard are non-parametric GNNs. Despite their high predictive performance and scalability, non-parametric GNNs still face two limitations. First, due to the propagation of over-smoothed features, they suffer from severe performance degradation along with the propagation depth. More importantly, they only consider the graph structure and ignore the feature influence during the non-parametric propagation, leading to sub-optimal propagated features. To address these limitations, we present non-parametric attention (NPA), a plug-and-play module that is compatible with non-parametric GNNs, to get scalable and deep GNNs simultaneously. We have deployed NPA in Tencent with the Angel platform, and we further evaluate NPA on both real-world datasets and large-scale industrial datasets. Experimental results on seven homophilic graphs (including the industrial Tencent Video graph) and five heterophilic graphs demonstrate NPA enjoys high performance -- achieves large performance gain over existing non-parametric GNNs, deeper architecture -- improves non-parametric GNNs with large model depth, and high scalability -- can support large-scale graphs with low time costs. Notably, it achieves state-of-the-art performance on the large ogbn-papers100M dataset.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.004 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.001 |
| Open science | 0.001 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it