Enhancing Conformal Prediction via Class Similarity
Researchers have developed a novel method to enhance conformal prediction (CP) by incorporating class similarity. This approach aims to reduce the size of prediction sets while ensuring they contain semantically similar classes, which is particularly useful in high-stakes applications like medical diagnosis. The method theoretically proves advantages for group-related metrics and can even reduce average set sizes, with a variant that leverages model embeddings for further improvement. AI
IMPACT Improves reliability of AI predictions in critical applications by reducing prediction set size and semantic diversity.