Identification of Emerging Viral News Topics Using HDBSCAN Outlier Projection

Conventional dynamic topic modeling often disregards low-density regions as stochastic noise. This paper challenges that notion by interpreting HDBSCAN outliers as latent semantic precursors to emerging topic narratives. We propose a predictive projection framework applied to a comprehensive dataset of Croatian news articles organized longitudinally and collected during a specific time period. Using a dual-pipeline BERTopic architecture, we generate word embeddings with paraphrase-multilingual-MiniLM-L12-v2 and apply lemmatization for interpretable c-TF-IDF representations using CLASSLA library. By calibrating density models on a reference corpus and projecting future data into the fixed manifold, we isolate a specific distributional pattern for nascent topics. Contrary to the assumption that novel topics appear as orthogonal vectors and cannot be inferred, our results indicate that viral events reported in the news articles manifest within a distinct cosine similarity interval (0.60-0.80) relative to reference outliers. This latent transition interval is significantly distinguishable from uncorrelated noise (similarity < 0.60) and semantic duplicates (similarity > 0.80). These findings demonstrate that viral themes or narratives initially occupy sparse vector space regions before converging into high-density clusters, confirming that outlier projection substantially reduces the search space for event detection. The codebase created for this research is freely available for research purposes at: https://www.kaggle.com/code/jurefranjkovi/emerging-topic-detection