Recent advancements in large language models (LLMs) have shown promise in feature engineering for tabular data, but concerns about their reliability persist, especially due to variability in generated outputs. We introduce a multi-level diagnosis and evaluation framework to assess the robustness of LLMs in feature engineering across diverse domains, focusing on the three main factors: key variables, relationships, and decision boundary values for predicting target classes. We demonstrate that the robustness of LLMs varies significantly over different datasets, and that high-quality LLM-generated features can improve few-shot prediction performance by up to 10.52%. This work opens a new direction for assessing and enhancing the reliability of LLM-driven feature engineering in various domains.

Crafting effective features is a crucial yet labor-intensive and domain-specific task within machine learning pipelines. Fortunately, recent advancements in Large Language Models (LLMs) have shown promise in automating various data science tasks, including feature engineering. But despite this potential, evaluations thus far are primarily based on the end performance of a complete ML pipeline, providing limited insight into precisely how LLMs behave relative to human experts in feature engineering. To address this gap, we propose ELF-Gym, a framework for Evaluating LLM-generated Features. We curated a new dataset from historical Kaggle competitions, including 251 "golden" features used by top-performing teams. ELF-Gym then quantitatively evaluates LLM-generated features by measuring their impact on downstream model performance as well as their alignment with expert-crafted features through semantic and functional similarity assessments. This approach provides a more comprehensive evaluation of disparities between LLMs and human experts, while offering valuable insights into specific areas where LLMs may have room for improvement. For example, using ELF-Gym we empirically demonstrate that, in the best-case scenario, LLMs can semantically capture approximately 56% of the golden features, but at the more demanding implementation level this overlap drops to 13%. Moreover, in other cases LLMs may fail completely, particularly on datasets that require complex features, indicating broad potential pathways for improvement.

Existing text representations such as embeddings and bag-of-words are not suitable for rule learning due to their high dimensionality and absent or questionable feature-level interpretability. This article explores whether large language models (LLMs) could address this by extracting a small number of interpretable features from text. We demonstrate this process on two datasets (CORD-19 and M17+) containing several thousand scientific articles from multiple disciplines and a target being a proxy for research impact. An evaluation based on testing for the statistically significant correlation with research impact has shown that LLama 2-generated features are semantically meaningful. We consequently used these generated features in text classification to predict the binary target variable representing the citation rate for the CORD-19 dataset and the ordinal 5-class target representing an expert-awarded grade in the M17+ dataset. Machine-learning models trained on the LLM-generated features provided similar predictive performance to the state-of-the-art embedding model SciBERT for scientific text. The LLM used only 62 features compared to 768 features in SciBERT embeddings, and these features were directly interpretable, corresponding to notions such as article methodological rigor, novelty, or grammatical correctness. As the final step, we extract a small number of well-interpretable action rules. Consistently competitive results obtained with the same LLM feature set across both thematically diverse datasets show that this approach generalizes across domains.