dsgiitr/graph_nets
Graph Nets
Graph Nets is a graph neural network library and educational toolkit implemented in PyTorch, providing implementations of popular graph representation learning algorithms and research papers. The project covers core graph machine learning tasks including semi-supervised node classification, inductive and unsupervised node embedding generation, and neighborhood feature aggregation.
The library supports diverse algorithmic approaches for processing network structures, ranging from shared-parameter graph convolutions and attention-weighted neighborhood aggregation to spectral Chebyshev filtering and random-walk skip-gram objectives.
Features
- Graph Neural Network Implementations - Implements popular graph neural network algorithms and research papers using PyTorch for representation learning.
- Graph Representation Learning - Generate continuous vector representations for nodes in a network using unsupervised learning approaches inspired by natural language modeling techniques.
- Dynamic Attention - Assigns adaptive attention weights to neighboring nodes during feature aggregation without requiring full graph visibility.
- PyTorch-Based Frameworks - Provides a collection of deep learning models and representation algorithms built on top of PyTorch.
- Graph Neighborhood Sampling - Computes inductive node embeddings by sampling fixed-size local neighborhoods and aggregating neighbor features for unseen graph structures.
- Graph Convolutions - Applies uniform filter parameters across non-euclidean graph neighborhoods to perform semi-supervised node classification using localized message passing.
- Skip-Gram Model Architectures - Learns continuous vector representations of nodes by generating random walk sequences and optimising a language-modeling objective function.
- Chebyshev Polynomial Approximations - Formulates graph convolutions using truncated Chebyshev polynomial expansions to enable fast localized spectral filtering without explicit eigen-decomposition.
- Node Classification - Apply shared filter parameters across graph locations to perform semi-supervised classification on non-euclidean graph data structures.
- Inductive Learning - Leverage node feature information through sampling and aggregation strategies to efficiently compute embeddings for previously unseen graph nodes.
- Graph Node Filtering - Apply convolutional operations across non-euclidean graph structures using shared filter parameters to perform semi-supervised classification on nodes.
- Unsupervised Node Embeddings - Generates continuous vector representations of nodes in an unsupervised manner using random walks and language modeling objectives.
- Attention-Weighted Feature Pairs - Assign learned attention weights dynamically to prioritize relevant connections during feature aggregation without requiring prior knowledge of the full graph structure.
- Spectral Graph Convolution Tools - Formulate convolutional neural networks on graphs using fast localized spectral filtering based on Chebyshev polynomials.
- Node Embeddings - Use inductive frameworks with neighborhood feature sampling to efficiently compute vector representations for newly added graph nodes.
- PyTorch Model Components - Organises deep learning operations into reusable neural network layers using tensor manipulation primitives provided by a tensor computing framework.
- Graph Traversal Embeddings - Generates continuous vector representations by simulating random paths across network structures and optimising co-occurrence objectives using language modeling.
- Educational Toolkits - Serves as an educational toolkit for exploring graph convolutions, attention mechanisms, and random walks.
