Clustering basic benchmark

Cite as:
P. Fänti and S. Sieranoja
K-means properties on six clustering benchmark datasets
Applied Intelligence, 48 (12), 4743-4759, December 2018
https://doi.org/10.1007/s10489-018-1238-7
BibTex


Machine Learning
School of Computing
University of Eastern Finland
P.O.Box 111
FIN-80101 Joensuu
Finland

 
S-sets
S1
S1
S3
S3
 S2
S2
S4
S4
 Synthetic 2-d data with N=5000 vectors and k=15 Gaussian clusters with different degree of cluster overlap

P. Fränti and O. Virmajoki, "Iterative shrinking method for clustering problems", Pattern Recognition, 39 (5), 761-765, May 2006. (Bibtex)

S1:  ts  txt
S2:  ts  txt
S3:  ts  txt
S4:  ts  txt

Ground truth centroids and partitions:  zip
s3 and s4 updated 4.2.2015
Tabs converted to spaces 25.9.2024
 
A-sets
A1
A1
N=3000, k=20
 A2
A2
N=5250, k=35
   Synthetic 2-d data with increasing number of clusters (k). There are 150 vectors per cluster.

I. Kärkkäinen and P. Fränti, "Dynamic local search algorithm for the clustering problem", Research Report A-2002-6 (pdf)(Bibtex)

A1:  ts  txt
A2:  ts  txt
A3:  ts  txt
A3
A3
N=7500, k=50
     Ground truth centroids:   cb and txt
Ground truth partitions:   pa
 
Birch-sets

Birch1
Birch2
 Synthetic 2-d data with N=100,000 vectors and k=100 clusters

Zhang et al., "BIRCH: A new data clustering algorithm and its applications", Data Mining and Knowledge Discovery, 1 (2), 141-182, 1997. (Bibtex)

Data sets (TS and TXT), ground truth centroids (CB and TXT) and partitions (PA):

Birch3

 Birch1: Clusters in regular grid structure  ts  txt  cb  gt  pa 
Birch2: Clusters at a sine curve  ts  txt cb  gt  pa 
Birch3: Random sized clusters in random locations  ts  txt  cb  gt 

Birch2 subsets: Varying N=1,000-1,000,000  ts  txt   Varying k=1-100  ts  txt
G2 sets
g2-2-30
G2 datasets


 g2-2-50
N=2048, k=2
D=2-1024
var=10-100
 Gaussian clusters datasets with varying cluster overlap (var) and dimensions (D).
txt (17 MB)   ts (50 MB)  

P. Fränti R. Mariescu-Istodor and C. Zhong, "XNN graph"
IAPR Joint Int. Workshop on Structural, Syntactic, and Statistical Pattern Recognition
Merida, Mexico, LNCS 10029, 207-217, November 2016. (Bibtex)

Ground truth centroids: cb and txt
Ground truth partitions: pa
DIM-sets (high)
dim032
dim032
D=32
 dim064
dim064
D=64
 High-dimensional data sets N=1024 and k=16 Gaussian clusters.
Clusters are well separated even in the higher dimensional cases.

P. Fränti, O. Virmajoki and V. Hautamäki, "Fast agglomerative clustering using a k-nearest neighbor graph",
IEEE Trans. on Pattern Analysis and Machine Intelligence,
28 (11), 1875-1881, November 2006. (Bibtex)

Ground truth centroids: cb and txt
dim128
dim128
D=128
 dim256
dim256
D=256
 Data sets in TS and TXT, ground truth partitions in PA format:
dim032:  ts  txt  pa
dim064:  ts  txt  pa
dim128:  ts  txt  pa
dim256:  ts  txt  pa
dim512:  ts  txt  pa
dim1024:  ts  txt  pa
dim512
dim512
D=512
 dim1024
dim1024
D=1024
  
 
DIM-sets (low)
Dim2
Dim2
 Synthetic data with Gaussian clusters.
N=1351-10126 vectors in k=9 clusters in 2-15 dimensional space

I. Kärkkäinen and P. Fränti, "Gradual model generator for single-pass clustering", Pattern Recognition, 40 (3), 784-795, March 2007. (Bibtex)

ts  txt
Unbalance
Unbalance
Unbalance
N=6500, k=8
         Synthetic 2-d data with N=6500 vectors and k=8 Gaussian clusters
ts  txt

M. Rezaei and P. Fränti, "Set-matching measures for external cluster validity",
IEEE Trans. on Knowledge and Data Engineering, 28 (8), 2173-2186, August 2016. (Bibtex)

Ground truth centroids:   cb and txt
Ground truth partitions:   pa 		 
 



Other clustering datasets

To cite the datasets please use the original articles.





Image data
[bridge.pgm]
Bridge
(256x256)
N=4096, D=16
      4x4 pixel blocks  ts  txt
4x4 binarized pixel blocks  ts  txt
4x4 pixel blocks: 25% randomly sampled (for training)  ts  txt
4x4 pixel blocks: 75% randomly sampled (for testing)  ts  txt
[house.ppm]
House
(256x256)
N=34112, D=3
 RGB-values, quantized to 5 bits per color  ts  txt
RGB-values, 8 bits per color  ts  txt
[missa001.pgm]
Miss America
(360x288)
N=6480, D=16
 4x4 pixel blocks from the difference image of frame 1 and 2  ts  txt
4x4 pixel blocks from the difference image of frame 2 and 3  ts  txt
europe
Europe
(vector)		 europe differentials
Europe
N=169308, D=2
Differential coordinates of Europe map ts  txt  original 

P. Fränti, M. Rezaei and Q. Zhao, "Centroid index: cluster level similarity measure",
Pattern Recognition, 47 (9), 3034-3045, September 2014, 2014. (Bibtex)

Nested datasets
dim032
N3
k=3
 N6
N6
k=6
          Nested Gaussian clusters N3 (N=2250) and N6 (N=5500).


P. Fränti et al., "Article to be written". zip
Worms
Worms
Worms
N=105,600, k=35, D=2
N=105,000, k=25, D=64
                  Synthetic 2-d and 64-d data with worm like shapes.
Dataset and MATLAB generation scripts: worms.zip 

S. Sieranoja and P. Fränti, "Fast and general density peaks clustering",
Pattern Recognition Letters, 128, 551-558, December 2019. (pdf) 		 
Variations
Unbalance2
Unbalance2
N=6500, k=8
ts  txt gt 		Asymmetric
Asymmetric
N=1000, k=5
ts  txt gt 		     Synthetic 2-d Gaussian clusters to test variations in cluster size unbalanace, symmetry, overlap and skewness
M. Rezaei and P. Fränti, "Can the number of clusters be determined by external indices?",
IEEE Access, 8 (1), 89239-89257, December 2020 (pdf).

Overlap
Overlap
N=1000, k=6
ts  txt gt 		Skewed
Skewed
N=1000, k=6
ts  txt gt
Graph datasets

          varDeg: Artificial graphs, varying average degree
varMu: Artificial graphs, varying mixing parameter mu (cluster overlap)
varN: Artificial graphs, varying number of nodes
icd10: Disease co-occurence networks

Dataset: gclu_data.zip (437 MB)

S. Sieranoja and P. Fränti, "Adapting k-means for graph clustering"
Knowledge and Information Systems (KAIS), 4:1-28, December 2021. (pdf)

More information here 		          
K-Sets data
sets_data
Sets data
N=1200
k=4,8,16,32
D=100,200,400,800
Overlap=0,5%,10%,20%,40%
Imbalance types=1,2,3,4,5

imbalance_types
          15 synthetic datasets of sets with N=1200 vectors and diverse number of clusters, dimensionality, overlap, and imbalance types
Items of sets are codes for classification of diseases (ICD-10) introduced by World Health Organization (WHO).

Data   Ground truth   Data generator

M. Rezaei and P. Fränti, "K-sets and k-swaps algorithms for clustering sets", Pattern Recognition, 139, 109454, July 2023.
(pdf)
  
KDDCUP04Bio set
KDDCUP04Bio
KDDCUP04Bio
N=145751, k=2000, D=74
   KDDCUP04Bio biology dataset.

KDDCUP04Bio:  ts  txt
Shape sets
from_literature
Third column is the label.
Aggregation
N=788, k=7, D=2
  Aggregation:  txt
A. Gionis, H. Mannila, and P. Tsaparas, Clustering aggregation. ACM Transactions on Knowledge Discovery from Data (TKDD), 2007. 1(1): p. 1-30.
Compound
N=399, k=6, D=2
  Compound:  txt
C.T. Zahn, Graph-theoretical methods for detecting and describing gestalt clusters. IEEE Transactions on Computers, 1971. 100(1): p. 68-86.
Pathbased
N=300, k=3, D=2
  Pathbased:  txt
H. Chang and D.Y. Yeung, Robust path-based spectral clustering. Pattern Recognition, 2008. 41(1): p. 191-203.
Spiral
N=312, k=3, D=2
  Spiral:  txt
H. Chang and D.Y. Yeung, Robust path-based spectral clustering. Pattern Recognition, 2008. 41(1): p. 191-203.
D31
N=3100, k=31, D=2
  D31:  txt
C.J. Veenman, M.J.T. Reinders, and E. Backer, A maximum variance cluster algorithm. IEEE Trans. Pattern Analysis and Machine Intelligence 2002. 24(9): p. 1273-1280.
R15
N=600, k=15, D=2
  R15:  txt
C.J. Veenman, M.J.T. Reinders, and E. Backer, A maximum variance cluster algorithm. IEEE Trans. Pattern Analysis and Machine Intelligence, 2002. 24(9): p. 1273-1280.
Jain
N=373, k=2, D=2
  Jain:  txt
A. Jain and M. Law, Data clustering: A user's dilemma. Lecture Notes in Computer Science, 2005. 3776: p. 1-10.
Flame
N=240, k=2, D=2
  Flame:  txt
L. Fu and E. Medico, FLAME, a novel fuzzy clustering method for the analysis of DNA microarray data. BMC bioinformatics, 2007. 8(1): p. 3.
UCI datasets
Thyroid
Thyroid
N=215, k=2, D=5
ts  txt
 Wine
Wine
N=178, k=3, D=13
ts  txt
 UCI datasets original source is http://archive.ics.uci.edu/ml/

Breast-Cancer-Wisconsin: We have removed features 1 (sample id) and 11 (class label). All missing values are given value 1.
Yeast
Yeast
N=1484, k=10, D=8
txt
ts  integer
 Breast
Breast
N=699, k=2, D=9
ts  txt
Iris
Iris
N=150, C=3, D=4
ts  txt  labels
 Glass
Glass
N=214, k=7, D=9,
ts  txt  labels
Wdbc
Wdbc
N=569, k=2, D=32
ts  full  numeric (D=31)
leaves
leaves
N=1600, k=100, D=64
zip
letter
Letter
N=20000, k=26, D=16
zip
Categorical
Census
Census
N=1000-512000, D=68
zip  		  Categorical attributes from Public Use Microdata Samples (PUMS) person records. Includes subsets of size 1000, 2000, 4000, ..., 512000. Source
Mopsi locations

users_locations_Finland
User locations
(Finland)
N=13467, D=2


MopsiLocations2012-Joensuu
User locations
(Joensuu)
N=6014, D=2

User locations until 2012 (FINLAND)
User locations:  cb  txt 

User locations until 2012 (JOENSUU)
User locations Joensuu:  ts  txt 

Mopsi datasets
Miscellaneous
t4.8k
t4.8k
N=8000, k=6, D=2
t4.8k.txt 
ConfLongDemo
N=164,860, k=11, D=3
txt 
 t4.8k: G. Karypis, E.H. Han, V. Kumar, CHAMELEON: A hierarchical 765 clustering algorithm using dynamic modeling, IEEE Trans. on Computers, 32 (8), 68-75, 1999.

ConfLongdemo has eight attributes, of which only three numerical attributes are included here.

MNIST
N=10000, k=10, D=748
txt 
MiniBooNE
N=130,065, D=50
txt 		MNIST includes 10 handwriting digits and contains 60,000 477 training patterns and 10,000 test patterns of 784 dimensions.

MiniBooNE

Related links