Overview
FluxNet is inspired by the work CKGConv: General Graph Convolution with Continuous Kernels. Building upon the original CKGConv framework, our architecture incorporates optimized attention mechanisms and adaptive feature modulation to more effectively capture complex and nuanced relationships in graph-structured data.
Key contributions:
- A novel graph convolution framework that leverages pseudo-coordinates and continuous convolution kernels to model graph structures.
- Integration of GATv2-based attention mechanisms for enhanced context awareness and dynamic feature interaction.
- A highly flexible and efficient architecture that achieves state-of-the-art performance on Long-Range Graph Benchmark (LRGB) tasks, offering an excellent balance between accuracy and computational efficiency.
Model Architecture
Benchmark Results
We evaluate our proposed method on five datasets from Benchmarking GNNs (Dwivedi et al., 2022a) and another two datasets from Long-Range Graph Benchmark (Dwivedi et al., 2022c). These benchmarks include diverse node and graph-level learning tasks such as node classification, graph classification, and graph regression. They test an algorithm's ability to focus on graph structure encoding, to perform node clustering, and to learn long-range dependencies.
Graph Benchmark Results
| Dataset | Metric | GCN | GIN | GAT | GRIT | FluxNet |
|---|---|---|---|---|---|---|
| MNIST | Accuracy | 90.70% | 96.48% | 95.53% | 98.10% | 98.42% |
| CIFAR10 | Accuracy | 55.71% | 55.25% | 64.22% | 76.46% | 72.78% |
| PATTERN | W. Accuracy | 71.89% | 85.38% | 78.27% | 87.19% | 88.66% |
| CLUSTER | W. Accuracy | 68.49% | 64.71% | 70.58% | 80.02% | 79.00% |
Long Range Grpah Benchmark Results
| Dataset | Metric | GCN | GINE | GatedGCN | GRIT | FluxNet |
|---|---|---|---|---|---|---|
| Peptides-func | Average Precision | 0.59 | 0.54 | 0.58 | 0.69 | 0.77 |
| Peptides-struct | MAE | 0.34 | 0.35 | 0.34 | 0.24 | 0.24 |
Applications
The CKG model can be applied to a wide range of knowledge graph-based tasks:
Knowledge Graph Reasoning
Predicting missing links in knowledge graphs by effectively modeling node-edge interactions and capturing complex relationships between entities.
Drug Discovery and Molecular Property Prediction
Predict molecular properties and interactions by effectively modeling the structural and chemical relationships within molecular graphs.
Social Network Analysis
Enables advanced analysis of social structures, influence propagation, and relationship prediction within social networks.
Recommendation Systems
Building recommendation engines that can model complex user-item interactions with rich feature representations.
3D Computer Vision
Process 3D point clouds and meshes, enabling tasks like segmentation, classification, and shape analysis.
Installation & Usage
Please navigate to the model documentation (link at the top of this page) :)