Blog_gnn_naics

A project exploring categorical feature encodings/embeddings, especially with graph neural networks techniques.

Towards Data Science "Exploring Hierarchical Blending in Target Encoding"

Table data is in the top level in the "tables.xlsx" document.

Code is at the top level; notebooks would run in order. Metrics are collected and summarized in 80_perf_summary.ipynb.

Running Code

First, download the SBA Loans Dataset from Kaggle.

Then, change setup.py

• Make input_path point to the SBA Loans dataset on your system
• temp_path should point to a writeable directory on your system

Towards Data Science "No Label Left Behind: Alternative Encodings for Hierarchical Categoricals"

See the subdirectory A_target_count_encoding

Hardware and GPUs

Everything runs on my home hardware (Mac Book Air). GNN methods use Stellargraph, which requires certain Python versions; I use 3.8.18.

Stellargraph uses keras/tensorflow. Here are some relevant package versions

• keras 2.15.0
• networkx 3.1
• tensorflow 2.15.0
• tensorflow-macos 2.15.0 (mac users only)
• tensorflow-metal 1.1.0 (mac users only)
• xgboost 1.7.5

The full pip list output is also in the top level of this repository.

Because I use a Mac, I had to install tensorflow metal for neural networks and graph neural networks. Here is a helpful article: https://medium.com/@angelgaspar/how-to-install-tensorflow-on-a-m1-m2-macbook-with-gpu-acceleration-acfeb988d27e