Publications

Detailed Information

Fast Incremental Density-Based Clustering over Sliding Windows : 슬라이딩 윈도우상의 빠른 점진적 밀도 기반 클러스터링

DC Field Value Language
dc.contributor.advisor문봉기-
dc.contributor.author김보경-
dc.date.accessioned2022-12-29T07:44:29Z-
dc.date.available2022-12-29T07:44:29Z-
dc.date.issued2022-
dc.identifier.other000000172160-
dc.identifier.urihttps://hdl.handle.net/10371/187775-
dc.identifier.urihttps://dcollection.snu.ac.kr/common/orgView/000000172160ko_KR
dc.description학위논문(박사) -- 서울대학교대학원 : 공과대학 컴퓨터공학부, 2022. 8. 문봉기.-
dc.description.abstractGiven the prevalence of mobile and IoT devices, continuous clustering against streaming data has become an essential tool of increasing importance for data analytics. Among many clustering approaches, density-based clustering has garnered much attention due to its unique advantage that it can detect clusters of an arbitrary shape when noise exists. However, when the clusters need to be updated continuously along with an evolving input dataset, a relatively high computational cost is required. Particularly, deleting data points from the clusters causes severe performance degradation.

In this dissertation, the performance limits of the incremental density-based clustering over sliding windows are addressed. Ultimately, two algorithms, DISC and DenForest, are proposed. The first algorithm DISC is an incremental density-based clustering algorithm that efficiently produces the same clustering results as DBSCAN over sliding windows. It focuses on redundancy issues that occur when updating clusters. When multiple data points are inserted or deleted individually, surrounding data points are explored and retrieved redundantly. DISC addresses these issues and improves the performance by updating multiple points in a batch. It also presents several optimization techniques. The second algorithm DenForest is an incremental density-based clustering algorithm that primarily focuses on the deletion process. Unlike previous methods that manage clusters as a graph, DenForest manages clusters as a group of spanning trees, which contributes to very efficient deletion performance. Moreover, it provides a batch-optimized technique to improve the insertion performance. To prove the effectiveness of the two algorithms, extensive evaluations were conducted, and it is demonstrated that DISC and DenForest outperform the state-of-the-art density-based clustering algorithms significantly.
-
dc.description.abstract모바일 및 IoT 장치가 널리 보급됨에 따라 스트리밍 데이터상에서 지속적으로 클러스터링 작업을 수행하는 것은 데이터 분석에서 점점 더 중요해지는 필수 도구가 되었습니다. 많은 클러스터링 방법 중에서 밀도 기반 클러스터링은 노이즈가 존재할 때 임의의 모양의 클러스터를 감지할 수 있다는 고유한 장점을 가지고 있으며 이에 따라 많은 관심을 받았습니다. 그러나 밀도 기반 클러스터링은 변화하는 입력 데이터 셋에 따라 지속적으로 클러스터를 업데이트해야 하는 경우 비교적 높은 계산 비용이 필요합니다. 특히, 클러스터에서의 데이터 점들의 삭제는 심각한 성능 저하를 초래합니다.

본 박사 학위 논문에서는 슬라이딩 윈도우상의 밀도 기반 클러스터링의 성능 한계를 다루며 궁극적으로 두 가지 알고리즘을 제안합니다. 첫 번째 알고리즘인 DISC는 슬라이딩 윈도우상에서 DBSCAN과 동일한 클러스터링 결과를 찾는 점진적 밀도 기반 클러스터링 알고리즘입니다. 해당 알고리즘은 클러스터 업데이트 시에 발생하는 중복 문제들에 초점을 둡니다. 밀도 기반 클러스터링에서는 여러 데이터 점들을 개별적으로 삽입 혹은 삭제할 때 주변 점들을 불필요하게 중복적으로 탐색하고 회수합니다. DISC 는 배치 업데이트로 이 문제를 해결하여 성능을 향상시키며 여러 최적화 방법들을 제안합니다. 두 번째 알고리즘인 DenForest 는 삭제 과정에 초점을 둔 점진적 밀도 기반 클러스터링 알고리즘입니다. 클러스터를 그래프로 관리하는 이전 방법들과 달리 DenForest 는 클러스터를 신장 트리의 그룹으로 관리함으로써 효율적인 삭제 성능에 기여합니다. 나아가 배치 최적화 기법을 통해 삽입 성능 향상에도 기여합니다. 두 알고리즘의 효율성을 입증하기 위해 광범위한 평가를 수행하였으며 DISC 및 DenForest 는 최신의 밀도 기반 클러스터링 알고리즘들보다 뛰어난 성능을 보여주었습니다.
-
dc.description.tableofcontents1 Introduction 1
1.1 Overview of Dissertation 3
2 Related Works 7
2.1 Clustering 7
2.2 Density-Based Clustering for Static Datasets 8
2.2.1 Extension of DBSCAN 8
2.2.2 Approximation of Density-Based Clustering 9
2.2.3 Parallelization of Density-Based Clustering 10
2.3 Incremental Density-Based Clustering 10
2.3.1 Approximated Density-Based Clustering for Dynamic Datasets 11
2.4 Density-Based Clustering for Data Streams 11
2.4.1 Micro-clusters 12
2.4.2 Density-Based Clustering in Damped Window Model 12
2.4.3 Density-Based Clustering in Sliding Window Model 13
2.5 Non-Density-Based Clustering 14
2.5.1 Partitional Clustering and Hierarchical Clustering 14
2.5.2 Distribution-Based Clustering 15
2.5.3 High-Dimensional Data Clustering 15
2.5.4 Spectral Clustering 16
3 Background 17
3.1 DBSCAN 17
3.1.1 Reformulation of Density-Based Clustering 19
3.2 Incremental DBSCAN 20
3.3 Sliding Windows 22
3.3.1 Density-Based Clustering over Sliding Windows 23
3.3.2 Slow Deletion Problem 24
4 Avoiding Redundant Searches in Updating Clusters 26
4.1 The DISC Algorithm 27
4.1.1 Overview of DISC 27
4.1.2 COLLECT 29
4.1.3 CLUSTER 30
4.1.3.1 Splitting a Cluster 32
4.1.3.2 Merging Clusters 37
4.1.4 Horizontal Manner vs. Vertical Manner 38
4.2 Checking Reachability 39
4.2.1 Multi-Starter BFS 40
4.2.2 Epoch-Based Probing of R-tree Index 41
4.3 Updating Labels 43
5 Avoiding Graph Traversals in Updating Clusters 45
5.1 The DenForest Algorithm 46
5.1.1 Overview of DenForest 47
5.1.1.1 Supported Types of the Sliding Window Model 48
5.1.2 Nostalgic Core and Density-based Clusters 49
5.1.2.1 Cluster Membership of Border 51
5.1.3 DenTree 51
5.2 Operations of DenForest 54
5.2.1 Insertion 54
5.2.1.1 MST based on Link-Cut Tree 57
5.2.1.2 Time Complexity of Insert Operation 58
5.2.2 Deletion 59
5.2.2.1 Time Complexity of Delete Operation 61
5.2.3 Insertion/Deletion Examples 64
5.2.4 Cluster Membership 65
5.2.5 Batch-Optimized Update 65
5.3 Clustering Quality of DenForest 68
5.3.1 Clustering Quality for Static Data 68
5.3.2 Discussion 70
5.3.3 Replaceability 70
5.3.3.1 Nostalgic Cores and Density 71
5.3.3.2 Nostalgic Cores and Quality 72
5.3.4 1D Example 74
6 Evaluation 76
6.1 Real-World Datasets 76
6.2 Competing Methods 77
6.2.1 Exact Methods 77
6.2.2 Non-Exact Methods 77
6.3 Experimental Settings 78
6.4 Evaluation of DISC 78
6.4.1 Parameters 79
6.4.2 Baseline Evaluation 79
6.4.3 Drilled-Down Evaluation 82
6.4.3.1 Effects of Threshold Values 82
6.4.3.2 Insertions vs. Deletions 83
6.4.3.3 Range Searches 84
6.4.3.4 MS-BFS and Epoch-Based Probing 85
6.4.4 Comparison with Summarization/Approximation-Based Methods 86
6.5 Evaluation of DenForest 90
6.5.1 Parameters 90
6.5.2 Baseline Evaluation 91
6.5.3 Drilled-Down Evaluation 94
6.5.3.1 Varying Size of Window/Stride 94
6.5.3.2 Effect of Density and Distance Thresholds 95
6.5.3.3 Memory Usage 98
6.5.3.4 Clustering Quality over Sliding Windows 98
6.5.3.5 Clustering Quality under Various Density and Distance Thresholds 101
6.5.3.6 Relaxed Parameter Settings 102
6.5.4 Comparison with Summarization-Based Methods 102
7 Future Work: Extension to Varying/Relative Densities 105
8 Conclusion 107
Abstract (In Korean) 120
-
dc.format.extentix, 121-
dc.language.isoeng-
dc.publisher서울대학교 대학원-
dc.subjectDensity-BasedClustering-
dc.subjectDataStreams-
dc.subjectSlidingWindow-
dc.subjectDeletion-
dc.subjectIncrementalClustering-
dc.subjectDISC-
dc.subjectDenForest-
dc.subject.ddc621.39-
dc.titleFast Incremental Density-Based Clustering over Sliding Windows-
dc.title.alternative슬라이딩 윈도우상의 빠른 점진적 밀도 기반 클러스터링-
dc.typeThesis-
dc.typeDissertation-
dc.contributor.AlternativeAuthorBogyeong Kim-
dc.contributor.department공과대학 컴퓨터공학부-
dc.description.degree박사-
dc.date.awarded2022-08-
dc.identifier.uciI804:11032-000000172160-
dc.identifier.holdings000000000048▲000000000055▲000000172160▲-
Appears in Collections:
Files in This Item:

Altmetrics

Item View & Download Count

  • mendeley

Items in S-Space are protected by copyright, with all rights reserved, unless otherwise indicated.

Share