Hierarchical & Density-Based Clustering

BERTopic, a widely used library for finding topics in collections of documents, runs every document through the same pipeline: embed it, compress the embedding to five dimensions, cluster with an algorithm called HDBSCAN, and describe each cluster by its most distinctive words. Many first runs end with a large share of the documents in a topic numbered −1, and the library's FAQ has a section titled "How do I reduce topic outliers?". Topic −1 is not a bug. It is the clustering algorithm declaring that those documents do not sit in a region dense enough to belong anywhere, a verdict k-means cannot give, because k-means assigns every point to its nearest centroid, however far away that is.