Assess predictive robustness across dimensionality-reduction and clustering configurations

Determine whether the stability observed across UMAP dimensionalities also holds for prospective prediction, and evaluate the effects of alternative dimensionality-reduction methods, clustering algorithms, and topic-alignment thresholds on anticipatory-outlier prediction.

Background

The retrospective labels used for supervised prediction depend on a pipeline that projects embeddings with 20-dimensional UMAP, clusters cumulative snapshots with HDBSCAN, and aligns topics using a fixed threshold of 0.30. Prior work reportedly found strong inter-model agreement and limited variation across UMAP dimensionalities for trajectory labeling, but the present paper emphasizes that this does not establish robustness of the downstream prospective prediction task. The authors therefore leave unresolved whether prediction performance remains stable under changes to dimensionality reduction, clustering, and temporal topic alignment.