Content uploaded by Ali Seyed Shirkhorshidi
Author content
All content in this area was uploaded by Ali Seyed Shirkhorshidi on Oct 28, 2014
Content may be subject to copyright.
Big Data Clustering: A Review
Ali Seyed Shirkhorshidi1, Saeed Aghabozorgi1, Teh Ying Wah1 and Tutut Herawan1,2
1Department of Information Systems
Faculty of Computer Science and Information Technology
University of Malaya
50603 Pantai Valley, Kuala Lumpur, Malaysia
2AMCS Research Center, Yogyakarta, Indonesia
shirkhorshidi_ali@siswa.um.edu.my, saeed@um.edu.my,
tehyw@um.edu.my, tutut@um.edu.my
Abstract. Clustering is an essential data mining and tool for analyzing big data.
There are difficulties for applying clustering techniques to big data duo to new
challenges that are raised with big data. As Big Data is referring to terabytes
and petabytes of data and clustering algorithms are come with high computa-
tional costs, the question is how to cope with this problem and how to deploy
clustering techniques to big data and get the results in a reasonable time. This
study is aimed to review the trend and progress of clustering algorithms to cope
with big data challenges from very first proposed algorithms until today’s novel
solutions. The algorithms and the targeted challenges for producing improved
clustering algorithms are introduced and analyzed, and afterward the possible
future path for more advanced algorithms is illuminated based on today’s avail-
able technologies and frameworks.
Keywords: Big Data; Clustering; MapReduce, Parallel Clustering.
1 Introduction
After an era of dealing with data collection challenges, nowadays the problem is
changed into the question of how to process these huge amounts of data. Scientists
and researchers believe that today one of the most important topics in computing sci-
ence is Big Data. Social networking websites such as Facebook and Twitter have
billions of users and they produce hundreds of gigabytes of contents per minute, retail
stores continuously collect their customers’ data, You Tube has 1 billion unique users
which are producing 100 hours of video each an hour and its content ID service scans
over 400 years of video every day [1], [2]. To deal with this avalanche of data, it is
necessary to use powerful tools for knowledge discovery. Data mining techniques are
well-known knowledge discovery tools for this purpose [3]–[9]. Clustering is one of
them that is defined as a method in which data are divided into groups in a way that
objects in each group share more similarity than with other objects in other groups
[1]. Data clustering is a well-known technique in various areas of computer science
and related domains. Although data mining can be considered as the main origin of
clustering, but it is vastly used in other fields of study such as bio informatics, energy
studies, machine learning, networking, pattern recognition and therefore a lot of re-
search works has been done in this area [10]–[13]. From the very beginning research-
ers were dealing with clustering algorithms in order to handle their complexity and
computational cost and consequently increase scalability and speed. Emersion of big
data in recent years added more challenges to this topic which urges more research for
clustering algorithms improvement. Before focusing on clustering big data the ques-
tion which needs to be clarified is how big the big data is. To address this question
Bezdek and Hathaway represented a categorization of data sizes which is represented
in table 1 [14].
Table 1. Bezdek and Hathaway categorization for big data
Big data
Bytes
106
108
1010
1012
10>12
“Size”
Medium
Large
Huge
Monster
Very large
Challenges of big data have root in its five important characteristics [15]:
Volume: The first one is Volume and an example is the unstructured data stream-
ing in form of social media and it rises question such as how to determine the rele-
vance within large data volumes and how to analyze the relevant data to produce
valuable information.
Velocity: Data is flooding at very high speed and it has to be dealt with in reasona-
ble time. Responding quickly to data velocity is one of the challenges in big data.
Variety: Another challenging issue is to manage, merge and govern data that
comes from different sources with different specifications such as: email, audio,
unstructured data, social data, video and etc.
Variability: Inconsistency in data flow is another challenge. For example in social
media it could be daily or seasonal peak data loads which makes it harder to deal
and manage the data specially when the data is unstructured.
Complexity: Data is coming from different sources and have different structures;
consequently it is necessary to connect and correlate relationships and data linkag-
es or you find your data to be out of control quickly.
Traditional clustering techniques cannot cope with this huge amount of data because
of their high complexity and computational cost. As an instance, the traditional K-
means clustering is NP-hard, even when the number of clusters is k=2. Consequently,
scalability is the main challenge for clustering big data.
Fig. 1. Progress of developments in clustering algorithms to deal with big data
The main target is to scale up and speed up clustering algorithms with minimum sacri-
fice to the clustering quality. Although scalability and speed of clustering algorithms
were always a target for researchers in this domain, but big data challenges underline
these shortcomings and demand more attention and research on this topic. Reviewing
the literature of clustering techniques shows that the advancement of these techniques
could be classified in five stages as shown in Figure 1.
In rest of this study advantages and drawbacks of algorithms in each stage will be
discussed as they appeared in the figure respectively. In conclusion and future works
we will represent an additional stage which could be the next stage for big data clus-
tering algorithms based on recent and novel methods.
Techniques that are used to empower clustering algorithms to work with bigger da-
tasets trough improving their scalability and speed can be classified into two main
categories:
Single-machine clustering techniques
Multiple-machine clustering techniques
Single-machine clustering algorithms run in one machine and can use resources of
just one single machine while the multiple-machine clustering techniques can run in
several machines and has access to more resources. In the following section algo-
rithms in each of these categories will be reviewed.
2 Big Data Clustering
In general, big data clustering techniques can be classified into two major categories:
single-machine clustering techniques and multiple-machine clustering techniques.
Recently multiple machine clustering techniques has attracted more attention because
they are more flexible in scalability and offer faster response time to the users. As it is
demonstrated in Fig. 2 single-machine and multiple-machine clustering techniques
include different techniques:
Single-machine clustering
o Sample based techniques
o Dimension reduction techniques
Multiple-machine clustering
o Parallel clustering
o MapReduce based clustering
In this section advancements of clustering algorithms for big data analysis in catego-
ries that are mentioned above will be reviewed.
Fig. 2. Big data clustering techniques
2.1 Single-machine clustering techniques
2.1.1 Sampling Based Techniques
These algorithms were the very first attempts to improve speed and scalability of
them and their target was dealing with exponential search space. These algorithms are
called as sampling based algorithms, because instead of performing clustering on the
whole dataset, it performs clustering algorithms on a sample of the datasets and then
generalize it to whole dataset. This will speed up the algorithm because computation
needs to take place for smaller number of instances and consequently complexity and
memory space needed for the process decreases.
Clustering Large Applications based on Randomized Sampling (CLARANS).
Before introducing CLARANS [16] let us take a look at its predecessor Clustering
Large Applications (CLARA) [17]. Comparing to partitioning around medoids
(PAM) [17], CLARA can deal with larger datasets. CLARA decreases the overall
quadratic complexity and time requirements into linear in total number of objects.
PAM calculates the entire pair-wise dissimilarity matrix between objects and store it
in central memory, consequently it consume O (n2) memory space, thus PAM cannot
be used for large number of n. To cope with this problem, CLARA does not calculate
the entire dissimilarity matrix at a time. PAM and CLARA can be regarded conceptu-
ally as graph-searching problems in which each node is a possible clustering solution
and the criteria for two nodes for being linked is that they should be differ in 1 out of
k medoids.
PAM starts with one of the randomly chosen nodes and greedily moves to one of its
neighbors until it cannot find a better neighbor. CLARA reduce the search space by
searching only a sub-graph which is prepared by the sampled O (k) data points.
CLARANS is proposed in order to improve efficiency in comparison to CLARA. like
PAM , CLARANS, aims to find a local optimal solution by searching the entire
graph, but the difference is that in each iteration, it checks only a sample of the neigh-
bors of the current node in the graph. Obviously, CLARA and CLARANS both are
using sampling technique to reduce search space, but their difference is in the way
that they perform the sampling. Sampling in CLARA is done at the beginning stage
and it restricts the whole search process to a particular sub-graph while in
CLARANS; the sampling is conduct dynamically for each iteration of the search pro-
cedure. Observation results shows that dynamic sampling used in CLARANS is more
efficient than the method used in CLARA [18].
BIRCH
If the data size is larger than the memory size, then the I/O cost dominates the compu-
tational time. BIRCH [19] is offering a solution to this problem which is not ad-
dressed by other previously mentioned algorithms. BIRCH uses its own data structure
called clustering feature (CF) and also CF-tree. CF is a concise summary of each clus-
ter. It takes this fact into the consideration that every data point is not equally im-
portant for clustering and all the data points cannot be accommodated in main
memory.
CF is a triple <N,LS,SS> which contains the number of the data points in the cluster,
the linear sum of the data points in the cluster and the square sum of the data points in
the cluster, the linear sum of the data points in the cluster and the square sum of the
data points in the cluster. Checking if CF satisfies the additive property is easy, if two
existing clusters need to be merged, the CF for the merged cluster is the sum of the
CFs of the two original clusters. The importance of this feature is for the reason that it
allows merging two existing clusters simply without accessing the original dataset.
There are two key phase for BIRCH algorithm. First, it scans the data points and build
an in memory tree and the second applies clustering algorithm to cluster the leaf
nodes. Experiments conducted in [20] reveal that in terms of time and space, BIRCH
performs better than CLARANS and it is also more tolerable in terms of handling
outliers. Fig. 3 represents a flowchart which demonstrates steps in BIRCH algorithm.
Fig. 3. BIRCH algorithm flowchart
CURE
Single data point is used to represent a cluster in all previously mentioned algorithms
which means that these algorithms are working well if clusters have spherical shape,
while in the real applications clusters could be from different complex shapes. To deal
with this challenge, clustering by using representatives (CURE) [21] uses a set of
well-scattered data points to represent a cluster. In fact CURE is a hierarchical algo-
rithm. Basically it considers each data point as a single cluster and continually merges
two existing clusters until it reaches to precisely k clusters remaining. The process of
selecting two clusters for merging them in each stage is based on calculating mini-
mum distance between all possible pairs of representative points from the two clus-
ters. Two main data structures empower CURE for efficient search. Heap is the first,
which is used to track the distance for each existing cluster to its closest cluster and
the other is k-d tree which is used to store all the representative points for each clus-
ter.
CURE also uses sampling technique to speed up the computation. It draws a sample
of the input dataset and run mentioned procedure on the sample data. To clarify the
necessary sample size Chernoff bound is used in the original study. If the dataset is
very large, even after sampling, the data size is still big and consequently the process
will be time consuming. To solve this issue, CURE uses partitions to accelerate the
algorithm. If we consider 𝑛 as the original dataset and the 𝑛′ as the sampled data, then
CURE will partition the 𝑛′ into 𝑝 partitions and within each partition it runs a partial
hierarchical clustering until either a predefined number of clusters is reached or the
distance between the two clusters to be merged exceeds some threshold. Then another
clustering runs and passes on all partial clusters from all the 𝑝 partitions. At the final
stage all non-sampled data points will assign to the nearest clusters. Results [21] rep-
resent that in comparison to BIRCH, execution time for CURE is lower while it main-
tain the privilege of robustness in handling outliers by shrinking the representative
points to the centroid of the cluster with a constant factor. Fig. 4 demonstrate the
flowchart of CURE algorithm.
Fig. 4. CURE flowchart
2.1.2 Dimension Reduction Techniques
Although the complexity and speed of clustering algorithms is related to the number
of instances in the dataset, but at the other hand dimensionality of the dataset is other
influential aspect. In fact the more dimensions data have, the more is complexity and
it means the longer execution time. Sampling techniques reduce the dataset size but
they do not offer a solution for high dimensional datasets.
Locality-Preserving Projection (Random Projection)
In this method, after projecting the dataset from d-dimensional to a lower dimensional
space, it is desired that the pairwise distance to be roughly preserved. Many clustering
algorithms are distance based, so it is expected that the clustering result in the project-
ed space provide an acceptable approximation of the same clustering in the original
space. Random projection can be accomplished by a linear transformation of the data
matrix 𝐴, which is the original data matrix. If 𝑅 be a 𝑑 × 𝑡 and rotation matrix (𝑡 <<
𝑑) and its elements 𝑅(𝑖, 𝑗) be independent random variables, then 𝐴’ = 𝐴. 𝑅 is the
projection of matrix 𝐴 in 𝑡 dimensional space. It means each row in 𝐴’ has 𝑡 dimen-
sions. Construction of rotation matrix is different in different random projection algo-
rithms. Preliminary methods propose Normal random variables with mean 0 and vari-
ance 1 for 𝑅(𝑖 , 𝑗) although there are studies which represent different methods for
allocating values to 𝑅(𝑖 , 𝑗) [22]. After generating the projected matrix into lower di-
mensional space, clustering can be performed. Samples of algorithm implementation
is illustrated in [23], [24] and recently a method is proposed to speed up the computa-
tion [25].
Global Projection
In global projection the objective is for each data point it is desirable that the project-
ed data point to be as close as possible to original data point while in local preserving
projection the objective was to maintain pairwise distances roughly the same between
two pairs in projected space. In fact if the main dataset matrix is considered to be 𝐴
and 𝐴’ be the approximation of it, in global projection aim is to minimize ||𝐴’ − 𝐴||.
Different approaches are available to create the approximation matrix such as SVD
(singular value decomposition) [26], CX/CUR [27], CMD [28] and Colibri [29].
2.2 Multi-Machine clustering techniques
Although sampling and dimension reduction methods used in single-machine cluster-
ing algorithms represented in previous section improves the scalability and speed of
the algorithms, but nowadays the growth of data size is way much faster than memory
and processor advancements, consequently one machine with a single processor and a
memory cannot handle terabytes and petabytes of data and it underlines the need algo-
rithms that can be run on multiple machines. As it is shown in Fig. 5, this technique
allows to breakdown the huge amount of data into smaller pieces which can be loaded
on different machines and then uses processing power of these machines to solve the
huge problem. Multi machine clustering algorithms are divided into two main catego-
ries:
Un-automated distributing– parallel
Automated distributing– MapReduce
Fig. 5. General concept of multi-machine clustering techniques
In parallel clustering, developers are involved with not just parallel clustering chal-
lenges, but also with details in data distribution process between different machines
available in the network as well, which makes it very complicated and time consum-
ing. Difference between parallel algorithms and the MapReduce framework is in the
comfortless that MapReduce provides for programmers and reveals them form unnec-
essary networking problems and concepts such as load balancing, data distribution,
fault tolerance and etc. by handling them automatically. This feature allows huge
parallelism and easier and faster scalability of the parallel system. Parallel and dis-
tributed clustering algorithms follows a general cycle as represented below:
Fig. 6. General cycle for multi-machine clustering algorithms
In the first stage, data is going to be divided into partitions and they distribute over
machines. Afterward, each machine performs clustering individually on the assigned
partition of data. Two main challenges for parallel and distributed clustering are min-
imizing data traffic and its lower accuracy in comparison with its serial equivalent.
Lower accuracy in distributed algorithms could be caused by two main reasons, first,
it is possible that different clustering algorithms deploy in different machines and
secondly even if the same clustering algorithm is used in all machines, in some cases
the divided data might change the final result of clustering. In the rest of this study,
parallel algorithms and MapReduce algorithms will be discussed subsequently, then
more advanced algorithms proposed recently for big data will be covered.
2.2.1 Parallel clustering
Although parallel algorithms add difficulty of distribution for programmers, but it is
worthfull because of the major improvements in scaling and speed of clustering algo-
rithms. At the following parts some of them will be reviewed.
DBDC
DBDC [30], [31] is a distributed and density based clustering algorithm. Discovery of
clusters of arbitrary shapes is the main objective of density based clustering. The den-
sity of points in each cluster is much higher than outside of the cluster, while the den-
sity of the regions of noise is lower than the density in any of the clusters. DBDC [32]
is an algorithm which obeys the cycle mentioned in Figure 2 . At the individual clus-
tering stage, it uses a defined algorithm for clustering and then for general clustering,
a single machine density algorithm called DBSCAN is used for finalizing the results.
The results show that although DBDC maintain the same clustering quality in com-
parison to its serial interpretation, but it runs 30 times faster than that.
ParMETIS
ParMETIS [33] is the parallel version of METIS [34] and is a multilevel partitioning
algorithm. Graph partitioning is a clustering problem with the goal of finding the good
cluster of vertices. METIS contains three main steps. First step is called as coarsening
phase. In this stage maximal matching on the original graph is done and then the ver-
tices which are matched create a smaller graph and this process is iterated till the
number of vertices become small enough. The second stage is partitioning stage in
which k-way partitioning of the coarsened graph is performed using multilevel recur-
sive bisection algorithm. Finally in third un-coarsening stage, a greedy refinement
algorithm is used to project back the partitioning from second stage to the original
graph.
ParMETIS is a distributed version of METIS. Because of graph based nature of Par-
METIS it is different from clustering operations and it does not follow the general
cycle mentioned earlier for parallel and distributed clustering. An equal number of
vertices are going to distribute initially, then a coloring of a graph will compute in
machines. Afterward, a global graph incrementally matching only vertices of the same
color one at a time will be computed. In partitioning stage this graph broadcast to
machines and recursive bisection by exploring only a single path of the recursive
bisection tree performs in each machine. Finally un-coarsening stage is consisting of
moving vertices of edge-cut. Experiments represent that ParMETIS was 14 to 35
times faster than serial algorithm while maintaining the quality close to the serial
algorithm.
GPU based parallel clustering
A new issue is opened recently in parallel computing to use processing power of GPU
instead of CPU to speed up the computation. G-DBSCAN [35] is a GPU accelerated
parallel algorithm for density-based clustering algorithm, DBSCAN. It is one of the
recently proposed algorithms in this category. Authors distinguished their method by
using a graph based data indexing to add flexibility to their algorithm to allow more
parallelization opportunities. G-DBSCAN is a two-step algorithm and both of these
steps have been parallelized. The first step constructs a graph. Each object represents
a node and an edge is created between two objects if their distance is lower than or
equal to a predefined threshold. When this graph is ready, the second step is to identi-
fy the clusters. It uses breath first search (BFS) to traverse the graph created in the
first step. Results show that in comparison to its serial implementation, G-DBSCAN
is 112 times faster.
2.2.2 MapReduce
Although parallel clustering algorithms improved the scalability and speed of cluster-
ing algorithms still the complexity of dealing with memory and processor distribution
was a quiet important challenge. MapReduce is a framework which is illustrated in
Fig. 7 initially represented by Google and Hadoop is an open source version of it [36].
In this section algorithms which are implemented based on this framework are re-
viewed and their improvements are discussed in terms of three features:
Speed up: means the ratio of running time while the dataset remains constant and
the number of machines in the system is increased.
Scale up: measures if x time larger system can perform x time larger job with the
same run time
Size up: keeping the number of machines unchanged, running time grows linearly
with the data size
Fig. 7. MapReduce Framework
MapReduce based K-means (PK-means)
PKMeans [37] is distributed version of well-known clustering algorithm K-means
[38], [39]. The aim of k-means algorithm is to cluster the desire dataset into k clusters
in the way that instances in one cluster share more similarity than the instances of
other clusters. K-means clustering randomly choose k instance of the dataset in initial
step and performs two phases repeatedly: first it assigns each instance to the nearest
cluster and after finishing the assignment for all of the instances in the second phase it
updates the centers for each cluster with the mean of the instances.
PKMeans distributes the computation between multiple machines using MapReduce
framework to speed up and scale up the process. Individual clustering which contain
the first phase happens in the mapper and then general clustering perform second
phase in the reducer.
PKMeans has almost linear speed up and also a linear size up. It also has a good scale
up. For 4 machines it represented a scale up of 0.75. At the other hand, PKMeans is
an exact algorithm, it means that it offer the same clustering quality as its serial coun-
terpart k-means.
MR-DBSCAN
Big Data
Big Data
A very recent proposed algorithm is MR-DBSCAN [40] which is a scalable MapRe-
duce-based DBSCAN algorithm. Three major draw backs are existed in parallel
DBSCAN algorithms which MR-DBSCAN is fulfilling: first they are not successful
to balance the load between the parallel nodes, secondly these algorithms are limited
in scalability because all critical sub procedures are not parallelized, and finally their
architecture and design limit them to less portability to emerging parallel processing
paradigms.
MR-DBSCAN proposes a novel data partitioning method based on computation
cost emission as well as a scalable DBSCAN algorithm in which all critical sub-
procedures are fully parallelized. Experiments on large datasets confirm the scalabil-
ity and efficiency of MR-DBSCAN.
MapReduce based on GPU
As it discussed in G-DBSCAN section, GPUs are much more efficient than CPUs.
While CPUs have several processing cores GPUs are consisted of thousands of cores
which make them much more powerful and faster than CPUs. Although MapReduce
with CPUs represents very efficient framework for distributed computing, but if
GPUs is used instead, the framework can improve the speed up and scale up for dis-
tributed applications. GPMR is a MapReduce framework to use multiple GPUs. Alt-
hough clustering applications are not still implemented in this framework, but the
growth of data size urge researcher to represent faster and more scalable algorithms so
maybe this framework could be the appropriate solution to fulfill those needs.
3 Conclusion and Future Works
Clustering is one of the essential tasks in data mining and they need improvement
nowadays more than before to assist data analysts to extract knowledge from terabytes
and petabytes of data. In this study the improvement trend of data clustering algo-
rithms were discussed. to sum up, while traditional sampling and dimension reduction
algorithms still are useful, but they don’t have enough power to deal with huge
amount of data because even after sampling a petabyte of data, it is still very big and
it cannot be clustered by clustering algorithms, consequently the future of clustering is
tied with distributed computing. Although parallel clustering is potentially very useful
for clustering, but the complexity of implementing such algorithms is a challenge. At
the other hand, MapReduce framework provides a very satisfying base for implement-
ing clustering algorithms. As results shows, MapReduce based algorithms offer im-
pressive scalability and speed in comparison to serial counterparts while they are
maintaining same quality. Regarding to the fact that GPUs are much powerful than
CPUs as a future work, it is considerable to deploy clustering algorithms on GPU
based MapReduce frameworks in order to achieve even better scalability and speed.
Acknowledgments. This work is supported by University of Malaya High Impact
Research Grant no vote UM.C/625/HIR/MOHE/SC/13/2 from Ministry of Education
Malaysia.
References
1. T. C. Havens, J. C. Bezdek, and M. Palaniswami, “Scalable single linkage hierarchical
clustering for big data,” in Intelligent Sensors, Sensor Networks and Information
Processing, 2013 IEEE Eighth International Conference on. IEEE, 2013, pp. 396–401.
2. “YouTube Statistic,” 2014. [Online]. Available:
http://www.youtube.com/yt/press/statistics.html.
3. P. Williams, C. Soares, and and J. E. Gilbert, “A Clustering Rule Based Approach for
Classifi-cation Problems,” Int. J. Data Warehous. Min., vol. 8, no. 1, pp. 1–23, 2012.
4. R. V. Priya and A. Vadivel, “User Behaviour Pattern Mining from Weblog,” Int. J. Data
Warehous. Min., vol. 8, no. 2, pp. 1–22, 2012.
5. T. Kwok, K. A. Smith, S. Lozano, and D. Taniar, “No Title,” in Parallel Fuzzy c-Means
Clustering for Large Data Sets, 2002, pp. 365–374.
6. H. Kalia, S. Dehuri, and A. Ghosh, “A Survey on Fuzzy Association Rule Mining,” Int. J.
Data Warehous. Min., vol. 9, no. 1, pp. 1–27, 2013.
7. O. Daly and D. Taniar, “Exception Rules Mining Based on Negative Association Rules,”
in Proceedings of the International Conference on Computational Science and Its
Applications (ICCSA 2004), 2004, pp. 543–552.
8. M. Z. Ashrafi, D. Taniar, and K. A. Smith, “Redundant association rules reduction
techniques,” Int. J. Bus. Intell. Data Min., vol. 2, no. 1, pp. 29–63, 2007.
9. D. Taniar, W. Rahayu, V. C. S. Lee, and O. Daly, “Exception rules in association rule
mining,” Appl. Math. Comput., vol. 205, no. 2, pp. 735–750, 2008.
10. Meyer, F. G., and J. Chinrungrueng., “Spatiotemporal clustering of fMRI time series in the
spectral domain,” Med. Image Anal., vol. 9, no. 1, pp. 51–68, 2004.
11. J. Ernst, G. J. Nau, and Z. Bar-Joseph, “Clustering short time series gene expression data.,”
Bioinforma. 21, vol. 21, no. suppl 1, pp. i159 – i168, Jun. 2005.
12. F. Iglesias and W. Kastner, “Analysis of Similarity Measures in Times Series Clustering
for the Discovery of Building Energy Patterns,” Energies, vol. 6, no. 2, pp. 579–597, Jan.
2013.
13. Y. Zhao and G. Karypis, “Empirical and theoretical comparisons of selected criterion
functions for document clustering,” Mach. Learn., vol. 55, no. 3, pp. 311–331, 2004.
14. R. Hathaway and J. Bezdek, “Extending fuzzy and probabilistic clustering to very large
data sets,” Comput. Stat. Data Anal., vol. 51, no. 1, pp. 215–234, 2006.
15. “Big Data, What is it and why it is important.” [Online]. Available:
http://www.sas.com/en_us/insights/big-data/what-is-big-data.html.
16. R. T. Ng and J. Han, “CLARANS: A method for clustering objects for spatial data
mining,” IEEE Trans. Knowl. Data Eng., vol. 14, no. 5, pp. 1003–1016, 2002.
17. L. Kaufman and peter J. Rousseeuw, Finding Groups in Data: An Introduction on Cluster
Analysis. John Wiley and Sons, 1990.
18. R. T. Ng and J. Han, “CLARANS: A method for clustering objects for spatial data
mining,” IEEE Trans. Knowl. Data Eng., vol. 14, no. 5, pp. 1003–1016, 2002.
19. T. Zhang, R. Ramakrishnan, and M. Livny, “BIRCH: An efficient data clustering method
for very large database,” in SIGMOD Conference, 1996, pp. 103–114.
20. T. Zhang, R. Ramakrishnan, and M. Livny, “BIRCH: An efficient data clustering method
for very large database,” in SIGMOD Conference, 1996, pp. 103–114.
21. S. Guha and R. Rastogi, “CURE: An efficient clustering algorithm for large database,” Inf.
Syst., vol. 26, no. 1, pp. 35–58, 2001.
22. D. Achlioptas and F. McSherry, “Fast computation of low rank matrix approximations,” J.
C=ACM, vol. 54, no. 2, p. 9, 2007.
23. X. Z. Fern and C. E. Brodley, “Random projection for high dimensional data clustering: A
cluster ensemble approach,” in ICML, 2003, pp. 186–193.
24. S. Dasgupta, “Experiments with random projection,” UAI. pp. 143–151, 2000.
25. C. Boutsidis, C. Chekuri, T. Feder, and R. Motwani, “Random projections for k-means
clustering,” in NIPS, 2010, pp. 298–306.
26. G. H. Golub and C. F. Van-Loan, Matrix computations, 2nd ed. The Johns Hopkins
University Press, 1989.
27. P. Drineas, R. Kannan, and M. W. Mahony, “Fast Monte Carlo algorithms for matrices III:
Computing a compressed approximate matrix decomposition,” SIAM J. Comput., vol. 36,
no. 1, pp. 132–157, 2006.
28. J. Sun, Y. Xie, H. Zhang, and C. Faloutsos, “Less is More: Compact Matrix
Decomposition for Large Sparse Graphs.,” in SDM, 2007.
29. H. Tong, S. Papadimitriou, J. Sun, P. S. Yu, and C. Faloutsos, “Colibri: Fast mining of
large static and dynamic graphs,” in Proceedings of the 14th ACM SIGKDD international
conference on Knowledge discovery and data mining, 2008, pp. 686–694.
30. E. Januzaj, hans P. Kriegel, and M. Pfeifle, “DBDC: Density based distributed clustering,”
in Advances in Database Technology-EDBT 2004., 2004, pp. 88–105.
31. Aggarwal, C. C., C. K. Reddy, and Eds, Data Clustering: Algorithms and Applications.
2013.
32. M. Ester, hans P. Kriegel, J. Sander, and X. Xui, “A density-based algorithm for
discovering clusters in large spatial database with noise,” in KDD, 1996, pp. 226–231.
33. G. Karypis and V. Kumar, “Parallel multilevel k-way partitioning for irregular graphs,”
SIAM Rev., vol. 41, no. 2, pp. 278–300, 1999.
34. G. Karypis and V. Kumar, “Multilevel k-way partitining scheme for irregular graphs,” J.
Parallel Disteributed Comput., vol. 48, no. 1, pp. 96–129, 1998.
35. G. Andrade, G. Ramos, D. Madeira, R. Sachetto, R. Ferreira, and L. Rocha, “G-DBSCAN:
A GPU Accelerated Algorithm for Density-based Clustering,” Procedia Comput. Sci., vol.
18, pp. 369–378, 2013.
36. P. P. Anchalia, A. K. Koundinya, and S. NK., “MapReduce Design of K-Means Clustering
Algorithm,” in Information Science and Applications (ICISA), 2013 International
Conference on, 2013, pp. 1–5.
37. W. Zhao, H. Ma, and Q. He, “parallel k-means clustering based on MapReduce,” in Cloud
Computing, 2009, pp. 674–679.
38. J. Han, M. Kamber, and J. Pei, Data mining: concepts and techniques. Morgan Kaufmann,
2006.
39. B. Mirkin, Clustering for data mining a data recovery approach. CRC Press, 2012.
40. Y. He, H. Tan, W. Luo, S. Feng, and J. Fan, “MR-DBSCAN: a scalable MapReduce-based
DBSCAN algorithm for heavily skewed data,” Front. Comput. Sci., vol. 8, no. 1, pp. 83–
99, 2014.
