A unified view of density-based methods for
semi-supervised clustering and classification

Gertrudes, Jadson Castro; Zimek, Arthur; Sander, Jörg; Campello, Ricardo J. G. B.

doi:10.1007/s10618-019-00651-1

Cited by 14 publications

(11 citation statements)

References 42 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…These are soft rather than hard constraints and used to guide the selection of clusters. Gertrudes et al [ 19 ] further extended this approach by an alternative HDBSCAN selection method that directly uses labels instead of turning them into pairwise constraints. We will explain this version in more detail in Section 3.2.2 , and also include it in our experiments.…”

Section: Related Workmentioning

confidence: 99%

See 1 more Smart Citation

Constraint-Based Hierarchical Cluster Selection in Automotive Radar Data

Malzer

Baum

2021

Sensors

View full text Add to dashboard Cite

High-resolution automotive radar sensors play an increasing role in detection, classification and tracking of moving objects in traffic scenes. Clustering is frequently used to group detection points in this context. However, this is a particularly challenging task due to variations in number and density of available data points across different scans. Modified versions of the density-based clustering method DBSCAN have mostly been used so far, while hierarchical approaches are rarely considered. In this article, we explore the applicability of HDBSCAN, a hierarchical DBSCAN variant, for clustering radar measurements. To improve results achieved by its unsupervised version, we propose the use of cluster-level constraints based on aggregated background information from cluster candidates. Further, we propose the application of a distance threshold to avoid selection of small clusters at low hierarchy levels. Based on exemplary traffic scenes from nuScenes, a publicly available autonomous driving data set, we test our constraint-based approach along with other methods, including label-based semi-supervised HDBSCAN. Our experiments demonstrate that cluster-level constraints help to adjust HDBSCAN to the given application context and can therefore achieve considerably better results than the unsupervised method. However, the approach requires carefully selected constraint criteria that can be difficult to choose in constantly changing environments.

show abstract

Section: Related Workmentioning

confidence: 99%

“…Gertrudes et al [ 19 ] suggested to directly use labels instead of pairwise constraints. The authors argued that this prevents a bias towards should-not-link relations when adding new constraints, and also requires less effort.…”

Section: The Hdbscan Algorithmmentioning

confidence: 99%

Constraint-Based Hierarchical Cluster Selection in Automotive Radar Data

Malzer

Baum

2021

Sensors

View full text Add to dashboard Cite

show abstract

“…HDBSCAN* has been generalized to perform not only constraint‐based semi‐supervised clustering (Campello et al, ), but also label‐based semi‐supervised clustering and classification (Gertrudes et al, , ). Novel combinations of building‐blocks from different algorithms such as HDBSCAN*, HISSCLU, and SSDBSCAN have also been successfully attempted (Gertrudes et al, , ).…”

Section: Semi‐supervised Density‐based Clusteringmentioning

confidence: 99%

“…The red boxes correspond to the optimal selection of clusters performed by algorithm FOSC, which is an optional postprocessing routine used by HDBSCAN*, with its default stability criterion. This figure was reproduced using the R package "dbscan," with the toy data set "moons," following the codes in the package vignettes HDBSCAN* has been generalized to perform not only constraint-based semi-supervised clustering (Campello et al, 2015), but also label-based semi-supervised clustering and classification (Gertrudes et al, 2018(Gertrudes et al, , 2019. Novel combinations of building-blocks from different algorithms such as HDBSCAN*, HISSCLU, and SSDBSCAN have also been successfully attempted (Gertrudes et al, 2018(Gertrudes et al, , 2019.…”

Section: Semi-supervised Density-based Clusteringmentioning

confidence: 99%

“…The latter category is closely related to the problem of semi‐supervised classification, the difference relies mainly on that: (a) in label‐based semi‐supervised clustering not necessarily all unlabeled objects need to be assigned a label (some objects may be left unclustered as noise), and (b) new labels not present in the original subset of prelabeled objects may be assigned to disclosed clusters not represented in that subset. An in‐depth, comprehensive discussion on the similarities, differences and relationships between the problems of semi‐supervised classification and semi‐supervised clustering under a common umbrella of density‐based approaches has been provided by Gertrudes, Zimek, Sander, and Campello (). In that paper, as well as in a preceding preliminary conference publication (Gertrudes, Zimek, Sander, & Campello, ), Gertrudes et al describe a unified framework that subsumes a number of unsupervised and semi‐supervised density‐based clustering algorithms as particular cases.…”

Section: Semi‐supervised Density‐based Clusteringmentioning

confidence: 99%

See 1 more Smart Citation

Density‐based clustering

Campello

Kröger

Sander

et al. 2019

WIREs Data Min & Knowl

Self Cite

View full text Add to dashboard Cite

Clustering refers to the task of identifying groups or clusters in a data set. In density‐based clustering, a cluster is a set of data objects spread in the data space over a contiguous region of high density of objects. Density‐based clusters are separated from each other by contiguous regions of low density of objects. Data objects located in low‐density regions are typically considered noise or outliers. In this review article we discuss the statistical notion of density‐based clusters, classic algorithms for deriving a flat partitioning of density‐based clusters, methods for hierarchical density‐based clustering, and methods for semi‐supervised clustering. We conclude with some open challenges related to density‐based clustering. This article is categorized under: Technologies > Data Preprocessing Ensemble Methods > Structure Discovery Algorithmic Development > Hierarchies and Trees

show abstract

Non-parametric Semi-supervised Learning by Bayesian Label Distribution Propagation

Gøttcke

Zimek

Campello

2021

Similarity Search and Applications

Self Cite

View full text Add to dashboard Cite

Semi-supervised classification methods are specialized to use a very limited amount of labelled data for training and ultimately for assigning labels to the vast majority of unlabelled data. Label propagation is such a technique that assigns labels to those parts of unlabelled data that are in some sense close to labelled examples and then uses these predicted labels in turn to predict labels of more remote data. Here we propose to not propagate an immediate label decision to neighbors but to propagate the label probability distribution. This way we keep more information and take into account the remaining uncertainty of the classifier. We employ a Bayesian schema that is simpler and more straightforward than existing methods. As a consequence we avoid to propagate errors by decisions taken too early. A crisp decision can be derived from the propagated label distributions at will. We implement and test this strategy with a probabilistic k-nearest neighbor classifier, proving competitive with several state-of-the-art competitors in quality and more efficient in terms of computational resources.

show abstract

A unified view of density-based methods for semi-supervised clustering and classification

Cited by 14 publications

References 42 publications

Constraint-Based Hierarchical Cluster Selection in Automotive Radar Data

Constraint-Based Hierarchical Cluster Selection in Automotive Radar Data

Density‐based clustering

Non-parametric Semi-supervised Learning by Bayesian Label Distribution Propagation

Contact Info

Product

Resources

About