[WIP] Stanford CS224W PR for implementing Uncertainty Quantification of Sparse Travel Demand Prediction with Spatial-Temporal Graph Neural Networks (https://arxiv.org/pdf/2208.05908) in PyG.

[sagarkapare]

This pull requested is created as a part of project work for Stanford CS224W. It is WIP and includes following key features:

Graph Data Representation:

A new STZINBGraph class has been added to represent Origin-Destination (O-D) graph data with temporal travel demand features. It enables the construction of graph data, including edges, node features, temporal features, and adjacency matrices.

Models:

BTCN (Bidirectional Temporal Convolutional Network): Implements bidirectional temporal convolutions to model temporal dependencies in graph data.

DiffusionGCN: Introduces a diffusion-based graph convolutional network to capture spatial correlations in the O-D graph structure.

NBNormZeroInflated: Provides a graph-based implementation to predict Zero-Inflated Negative Binomial (ZINB) parameters—n (count), p (success probability), and pi (zero-inflation probability).
ST_NB_ZeroInflated: Combines spatial and temporal models to learn spatiotemporal embeddings and model uncertainty with ZINB.

Utility Enhancements:

A new loss function, nb_zeroinflated_nll_loss, has been introduced to compute the negative log-likelihood for ZINB distributions.

Unit tests for each new model and utility function to ensure correctness and performance.

Work in Progress

This pull request is a work in progress. Key tasks remaining include:
Fine-tuning the models for better performance on real-world datasets.
Adding more documentation for developers to understand the functionality and usage of new classes and methods.
Integrating additional examples for users to easily adopt the new functionality.
Benchmarking the models against existing approaches for spatial-temporal graph problems.