ShapGPT
Having reviewed different machine learning methods to identify fraud, I started wondering about what factors influenced the models in making specific predictions. I began to ask myself if there's a way to really understand what's behind these predictions. After some research I came across SHAP values which do a great job of interpretting models outputs, but I still thought there could be a more user-friendly, slightly less technical, medium to output the response. Which is when I thought about combining the SHAP value outputs with ChatGPT to provide a text description, extending my original credit fraud personal project.
​
Project Repo:
Non-Fraudulent
Using Shap values in python the model's decisions were able to be decomposed and expressed in terms of how each feature impacted final choice. Providing a numeric and graphic view of how these features each contributed.
Fraudulent
Additionally, uncover the root causes behind fraudulent predictions from the AI model. Understanding the top levers influencing the model, allows comparisons to be made and possible trends in model predictions.
Text Descriptions
Producing text outputs explaining the models decision in plain English. This opens up opportunities to provide outputs to operators in business setting or automating communications involving AI outputs.