Interpretation for Variational Autoencoder Used to Generate Financial Synthetic Tabular Data

Official TensorFlow 2.x implementation and gradient interpretability framework for Tabular Variational Autoencoders (TVAE). Featured by RBC Borealis AI.

---

📌 Abstract

Generative models such as Variational Autoencoders (VAEs) are widely adopted for generating high-fidelity synthetic tabular data. However, understanding feature importance and inter-feature interactions within non-linear latent spaces remains a significant challenge. This repository offers an end-to-end interpretability framework for TVAE models, featuring:

1. Local & Global Feature Importance: First-order Jacobian partial derivative aggregations.
2. Feature Interaction Mapping: Second-order partial derivative computations ($\frac{\partial^2 \mu}{\partial x_i \partial x_j}$) capturing latent variable interactions.

---

🚀 Quickstart

Installation

git clone [https://github.com/jhw0010/jinhongwu.github.io.git](https://github.com/jhw0010/jinhongwu.github.io.git)
cd jinhongwu.github.io
pip install -r requirements.txt