Reproduce “SenseCF: LLM-Prompted Counterfactuals for Intervention and Sensor Data Augmentation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Shovito Barua Soumma, Asiful Arefeen, Stephanie M. Carpenter, Melanie Hingle, Hassan Ghasemzadeh

The SenseCF framework introduces an LLM-prompted approach to generate counterfactual explanations from multimodal sensor data, offering actionable insights for interventions and demonstrating up to 5% average accuracy gains in downstream predictive models through data augmentation. It provides a model-agnostic method to explain predictions and enhance machine learning robustness in digital health applications.

Key insights

  • Large Language Models can effectively generate highly valid and plausible counterfactual explanations from complex, multimodal structured health data, moving beyond traditional NLP applications.
  • Integrating LLM-generated counterfactuals into training datasets consistently and significantly boosts the performance and robustness of various downstream machine learning classifiers, particularly for imbalanced data.
  • The LLM-based approach provides a model-agnostic method for generating explanations, overcoming the limitations of traditional optimization-based techniques that often require access to internal model architectures or gradients.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.