VelvetFlow: An engineering pipeline for robust multi-density clustering

Eyvazi, Hossein; Badzohreh, Mohammad; Shahrokhi, Seyed Ali

doi:10.22061/jdma.2025.12039.1131

VelvetFlow: An engineering pipeline for robust multi-density clustering

Document Type : Full Length Article

Authors

Department of Computer Science, University of Tarbiat Modares, Tehran, I. R. Iran

10.22061/jdma.2025.12039.1131

Abstract

Problem. Real-world datasets seldom respect a single density scale: tight blobs, elongated ribbons, and isolated points often coexist. Classical algorithms such as DBSCAN or \textit{k}-means require domain-specific parameter tuning and provide only ad-hoc support for anomaly detection.
Solution. We introduce VelvetFlow, an engineering pipeline that turns a set of well-understood building blocks into a cohesive, end-to-end workflow for multi-density clustering \emph{and} principled outlier detection. The pipeline is composed of three reusable stages:
(i) \emph{Contextual-density splitting} assigns every point to a high- or low-density partition using a single neighbourhood size $k$.
(ii) \emph{Density-aware clustering} applies a Jaccard-guided \textit{FusedNeighbor}+DBSCAN routine to the sparse partition and HDBSCAN to the dense partition-without introducing new hyper-parameters.
(iii) \emph{Scaled-MST verification} re-examines the complete $k$-NN graph, flags weakly connected components, and validates them with a $k$-NN gate; this step recovers small remote clusters while filtering genuine anomalies.

Graphical Abstract