Predicting Emerging Topics via Embedding Space Outliers

A new study demonstrates that documents appearing as outliers in embedding space can predict emerging topics with an F1 score above 0.90 in high-consensus subsets. The method distinguishes anticipatory outliers from noise by analyzing their subsequent trajectories in vector space.