Emission Prediction with Variable Graph Structure

This project was born out of inspiration from the 2021 Climate TRACE dataset. With the goal of predicting various factors about an emissions site, I invented a novel algorithm for the conversion of standard tabular data into a graph-structured form based on geographic proximity with a sliding hyperparameter. When used with modern graph neural networks (GNNs), my algorithms boasts a 25% performance bump from standard multi-layer perceptrons.

Read the whole paper and associated code on Github!

• Skills: Graph data, algorithm design, deep learning, data augmentation
• Tools: PyTorch Geometric, Pandas/GeoPandas, Shapely, NumPy