Early Classification of Information Cascades with Class-Conditional Small Language Models

Abstract

Early detection classifies an information cascade from its first nodes, but the labeled cascades available for training are complete. On four datasets, discriminative models trained on complete cascades are a median of 18 AUC points below their prefix-trained versions at five nodes. We write the structure and timing of a cascade in arrival order, so its first k nodes form a prefix, and fine-tune a small language model (SLM) to generate it per class. Bayes' rule classifies any prefix. Trained and selected without truncated cascades, three SLMs are 7.5 to 32 points above their discriminative versions trained on the same cascades at five nodes. They are a mean of 0.9 AUC points below their prefix-supervised versions over all evaluated k. On complete cascades they are a mean of 3.7 points above a GRU trained from scratch. Preprint. Submitted to the 35th European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN 2027), Bruges, Belgium, 21-23 April 2027.