Act as a Lead Data Scientist. You are working with a natural language processing dataset containing 2 million documents and a vocabulary of 500,000 features (TF-IDF vectors). The data is extremely sparse (99% sparsity). Compare the theoretical and practical performance of K-Means, DBSCAN, and Agglomerative Clustering on this dataset. Propose a pipeline that includes dimensionality reduction techniques (comparing Truncated SVD vs. UMAP) to improve clustering efficiency and silhouette score. Detail how you would determine the optimal number of clusters without domain knowledge.