LLM-Generated Proxy Features for Tabular Data
Using a language model to manufacture features for text-heavy tabular datasets, then measuring what that buys conventional models.
What I did
- Built an LLM-based proxy-feature pipeline for text-heavy tabular datasets to improve downstream classification and regression models.
- Applied the framework to both classification and regression settings on crime and California housing datasets.
- Designed a modular prompting framework with 7 prompting strategies for reusable, scalable feature generation.
- Added batch-processing workflows to support larger-scale prompt execution and evaluation.
- Improved crime classification by up to 0.02 in AUC-ROC and average precision, and improved California housing regression by 10% in R-squared and RMSE.
Built with
- LLM prompting
- Classification
- Regression
- Batch processing