mdeff/cnn_graph
Cnn Graph
Cnn_graph is a graph convolutional network framework and graph signal processing library designed for machine learning research. It provides computational notebooks and code to process and classify graph-structured data by combining node features with an underlying adjacency matrix representation.
The framework performs spectral graph convolutions through localized filters and accelerates filtering operations using truncated Chebyshev polynomials to avoid explicit graph Laplacian diagonalization. It includes a graph-structured data pipeline and sparse adjacency representations to handle irregular graph topologies and format custom data matrices, target vectors, and adjacency structures for neural network layers.
Features
- Spectral Convolutions - Performs convolution operations on graph structures by applying localized spectral filters defined through graph Laplacian eigen decomposition.
- Graph Neural Networks - Runs research experiments and verifies benchmark results for convolutional neural networks operating on graph-structured data.
- Graph Neural Network Training Routines - Trains graph convolutional models using localized spectral filtering and adjacency matrix representations.
- Chebyshev Polynomial Approximations - Approximates spectral graph filters using truncated Chebyshev polynomials to accelerate kernel computations without explicit Laplacian diagonalization.
- Convolutional Neural Networks - Provides a framework for applying convolutional neural networks to graph-structured data using fast localized spectral filtering.
- Graph Data Structures - Transforms custom data matrices, target vectors, and adjacency structures into batched tensors compatible with neural network layers.
- Research Repositories - Bundles computational notebooks and code to replicate benchmark experiments on graph classification.
- Sparse Representations - Optimizes memory usage and computation speed by storing network connections using sparse matrix formats tailored for irregular topologies.
- Graph Representation Learning - Applies convolutional neural networks to classify graph-structured inputs using fast localized spectral filtering techniques.
- Graph Data Models - Formats data matrices and adjacency structures to model custom graph topologies for deep learning.
- Notebook-Based Experimentation - Executes complete research workflows and verifies published performance benchmarks inside reproducible computational notebooks.
- Digital Signal Processing Libraries - Processes signals on graphs by combining node features with an underlying adjacency matrix representation.
