International Journal of Geographical Information

Science

ISSN: 1365-8816 (Print) 1362-3087 (Online) Journal homepage: https://www.tandfonline.com/loi/tgis20

An overview of clustering methods for geo-

referenced time series: from one-way clustering to
co- and tri-clustering

Xiaojing Wu, Changxiu Cheng, Raul Zurita-Milla & Changqing Song

To cite this article: Xiaojing Wu, Changxiu Cheng, Raul Zurita-Milla & Changqing Song (2020):
An overview of clustering methods for geo-referenced time series: from one-way clustering
to co- and tri-clustering, International Journal of Geographical Information Science, DOI:
10.1080/13658816.2020.1726922

To link to this article: https://doi.org/10.1080/13658816.2020.1726922

Published online: 16 Feb 2020.

Submit your article to this journal

Article views: 2

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tgis20
INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE
https://doi.org/10.1080/13658816.2020.1726922

REVIEW ARTICLE

An overview of clustering methods for geo-referenced time

series: from one-way clustering to co- and tri-clustering
Xiaojing Wua,b,c,d, Changxiu Chenga,b,c,d, Raul Zurita-Millae and Changqing Songb,c,d
a
Key Laboratory of Environmental Change and Natural Disaster, Beijing Normal University, Beijing, China;
b
State Key Laboratory of Earth Surface Processes and Resource Ecology, Beijing Normal University, Beijing,
China; cFaculty of Geographical Science, Beijing Normal University, Beijing, China; dCenter for Geodata and
Analysis, Beijing Normal University, Beijing, China; eDepartment of Geo-Information Processing, Faculty of
Geo-Information Science and Earth Observation (ITC), University of Twente, Enschede, The Netherlands

ABSTRACT ARTICLE HISTORY

Even though many studies have shown the usefulness of clustering Received 5 June 2019
for the exploration of spatio-temporal patterns, until now there is Accepted 4 February 2020
no systematic description of clustering methods for geo-referenced KEYWORDS
time series (GTS) classified as one-way clustering, co-clustering and Spatio-temporal pattern;
tri-clustering methods. Moreover, the selection of a suitable cluster- classification; method
ing method for a given dataset and task remains to be a challenge. selection; clustering analysis;
Therefore, we present an overview of existing clustering methods data mining
for GTS, using the aforementioned classification, and compare dif-
ferent methods to provide suggestions for the selection of appro-
priate methods. For this purpose, we define a taxonomy of
clustering-related geographical questions and compare the cluster-
ing methods by using representative algorithms and a case study
dataset. Our results indicate that tri-clustering methods are more
powerful in exploring complex patterns at the cost of additional
computational effort, whereas one-way clustering and co-clustering
methods yield less complex patterns and require less running time.
However, the selection of the most suitable method should depend
on the data type, research questions, computational complexity,
and the availability of the methods. Finally, the described classifica-
tion can include novel clustering methods, thereby enabling the
exploration of more complex spatio-temporal patterns.

1. Introduction
Advances in data collection and sharing techniques have resulted in significant increases
in spatio-temporal datasets. Therefore, novel approaches in terms of pattern mining and
knowledge extraction are required for such large datasets (Miller and Han 2009, Cheng
et al. 2014, Shekhar et al. 2015). Geo-referenced time series (GTS), a type of spatio-
temporal data, record time-changing values of one or more observed attributes at fixed
locations and consistent time intervals (Kisilevich et al. 2010). GTS is popular in real
applications, and examples include hourly PM2.5 concentrations observed at a network
of ground monitoring stations. Moreover, sequences of images can also be considered as
GTS, e.g., satellite image time series.

CONTACT Changxiu Cheng [email protected]

© 2020 Informa UK Limited, trading as Taylor & Francis Group
2 X. WU ET AL.

Clustering is a data mining task that identifies similar data elements and groups them
together. As a result, data elements in each group, or cluster, are similar to each other and
dissimilar to those in other groups (Berkhin 2006, Han et al. 2012). This allows an overview
of datasets at the cluster level and also provides insights into details by focusing on
a single cluster (Andrienko et al. 2009). Thus, clustering is useful for extracting patterns
from spatio-temporal datasets.
As mentioned in previous studies (Zhao and Zaki 2005, Henriques and Madeira 2018, Wu
et al. 2018), clustering methods for GTS can be classified as one-way clustering, co-clustering
and tri-clustering methods depending on the dimensions involved in the analysis. In such
a classification, one-way clustering methods, also termed traditional clustering, identify
clusters in one of the dimensions of 2D datasets based on the similarity of data elements
along the other dimension (Dhillon et al. 2003, Zhao and Zaki 2005). Also analyzing 2D
datasets, co-clustering methods identify (co-)clusters along both spatial and temporal
dimensions based on the similarity of data elements along these two dimensions
(Banerjee et al. 2007, Wu et al. 2020a). Tri-clustering methods identify (tri-) clusters based
on the similarity of data elements along spatial, temporal and third, e.g., attribute, dimen-
sions from 3D datasets (Wu et al. 2018, Henriques and Madeira 2018). However, a systematic
description of clustering methods for GTS from this perspective has not yet been reported.
Besides, an important issue concerns selecting appropriate clustering methods for
specific tasks at hand considering various available methods (Grubesic et al. 2014).
Similar issues also exist when choosing clustering methods for GTS and we aim to provide
suggestions for selecting suitable methods. To achieve this, we define a taxonomy of
clustering-related geographical questions, compare clustering methods in the above
classification by answering these questions using representative algorithms and a case
study dataset, and provide suggestions for selecting suitable methods.
Thus, the objective of this study is to provide two important and unique perspectives
on clustering methods for GTS. First, we provide an overview of clustering methods for
GTS using the classification outlined above. Thereafter, we compare different clustering
methods by answering clustering-related geographical questions and provide sugges-
tions on selecting suitable methods.
The structure of this paper is as follows: First, we describe the types of GTS and define
clustering-related questions in Section 2. Thereafter, we systematically describe the
clustering methods for GTS in the classification in Section 3. In Section 4, our case study
dataset and representative algorithms are described. Clustering results are interpreted
and the algorithms are compared in Section 5. Finally, we discuss the results in Section 6
and draw conclusions in Section 7.

2. GTS and questions for clustering GTS

The characteristics of the data to be analyzed heavily influence the choice of clustering
methods (Andrienko and Andrienko 2006, Kisilevich 2010). As a type of spatio-temporal
data, GTS instinctively involves three components: space (S), time (T) and attribute (A) in
a triad framework (Peuquet 1994). Depending on the number of attributes, GTS can be
divided into single attribute GTS (abbreviated to GTS-A, where A indicates the single
attribute) and multiple attributes GTS (abbreviated to GTS-As, where the affixed
s indicates the plural form). Alternatively, if single attribute GTS has one attribute but two
 INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE 3

nested hierarchies in either spatial or temporal dimension (e.g., day and hour in the case
of time), then GTS also includes the single attribute GTS with nested hierarchies in the
spatial dimension (abbreviated to GTS-Ss, where the S after the hyphen indicates the spatial
dimension, and s indicates the plural form) and one with nested hierarchies in the temporal
dimension (abbreviated as GTS-Ts, where T after the hyphen indicates the temporal dimen-
sion, and s indicates the plural form). GTS with more complex structures, e.g., with multiple
attributes and nested spatial and temporal dimensions, are beyond the scope of this paper
and not further discussed – also because they need the development of new clustering
methods.
With two dimensions, GTS-A are 2D GTS and typically organized into a data table where
rows are locations, columns are timestamps in which the attribute is observed, and
elements of the table are values of the attribute (Figure 1(a)); for example, hourly PM2.5
concentrations recorded at monitoring stations. With three dimensions, GTS-As, GTS-Ss
and GTS-Ts are 3D GTS, and any of them can be organized into a data cuboid with rows,
columns and depths as its three dimensions. Take GTS-As for instance, in which rows are
locations, columns are timestamps, depths are attributes, and elements are values of
attributes observed at corresponding locations and timestamps (Figure 1(b)); for example,
hourly PM2.5, PM10, NO2 and CO values recorded at monitoring stations.
In addition to the data characteristics, the other important factor for selecting cluster-
ing methods is the questions researchers are interested to answer (Andrienko and
Andrienko 2006). According to the triad framework developed by Peuquet (left of
Figure 2), three types of questions can be structured for GTS concerning the three
components: (1) where (space) + when (time) → what (attribute); (2) when + what →
where; (3) where + what → when (Peuquet 1994). For these questions, two reading levels

Figure 1. Various formats of GTS under different situations. (a) single attribute GTS (GTS-A); (b)
multiple attributes GTS (GTS-As); (c) single attribute GTS with nested hierarchies in spatial dimension
(GTS-Ss); (d) single attribute GTS with nested hierarchies in temporal dimension (GTS-Ts).

Figure 2. Triad framework to structure questions of the clustering analysis of GTS.

4 X. WU ET AL.

Table 1. Clustering-related geographical questions.

Components
Reading levels
I. where (SC) + when (TC) → what (C)
Elementary SC What is the value of the cluster observed at spatial-cluster sci and timestamp-cluster tci?
+ elementary TC
Synoptic SC What is the trend of the cluster(s) observed in the whole study area at timestamp-cluster tci?
+ elementary TC
Elementary SC What is the trend of the cluster(s) observed at location-cluster lci over the whole study
+ synoptic TC period?
Synoptic SC What is the trend of the cluster(s) observed in the whole study area over the whole study
+ synoptic TC period?
II. when (TC) + what (C) → where (SC)
Elementary TC + At which location-cluster(s) is the cluster ci observed in timestamp-cluster tci?
elementary C
Synoptic TC At which location-cluster(s) is the cluster ci observed over the whole study period?
+ elementary C
Elementary TC At which location-cluster(s) are all clusters observed in timestamp-cluster tci?
+ synoptic C
Synoptic TC At which location-cluster(s) are all clusters observed over the whole time period?
+ synoptic C
III. where (SC) + what (C) → when (TC)
Elementary SC + In which timestamp-cluster(s) is the cluster ci observed at location-cluster li?
elementary C
Synoptic SC + In which timestamp-cluster(s) is the cluster ci observed in the whole study area?
elementary C
Elementary SC + In which timestamp-cluster(s) are all clusters observed at location-cluster li?
synoptic C
Synoptic SC + In which timestamp-cluster(s) are all clusters observed in the whole study area?
synoptic C

are distinguished as elementary and synoptic, depending on whether the elements of

components are treated individually or not (Bertin 1983, Andrienko and Andrienko 2006);
for instance, questions regarding one location belong to the elementary level whereas
those regarding part of or the whole area belong to the synoptic level. Based on the
aforementioned work, a taxonomy of clustering-related geographical questions is defined
with three new components: spatial-cluster (SC), temporal-cluster (TC) and cluster (C), as
well as two reading levels (right of Figure 2). Correspondingly, questions regarding one
spatial-cluster belong to the elementary level while those regarding all spatial-clusters or
the subsets belong to the synoptic level. According to the taxonomy, 12 questions are
structured (Table 1).

3. Classification of clustering methods for GTS

The classification of clustering methods for GTS into one-way clustering, co- and tri-
clustering methods is systematically described in this section. Each category is further
divided into hierarchical and partitional methods depending on whether nested clusters
are created. For each type of clustering method, we first explain principles of the method
and then provide an overview of the main methods used in previous studies, emphasizing
on the analysis of GTS.
 INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE 5

Figure 3. Partitional clustering methods for GTS: (a&b1&b2) one-way clustering, (a&c) co-clustering
and (d&e) tri-clustering.

3.1. One-way clustering methods

Both traditional partitional and hierarchical clustering methods analyze 2D GTS organized into
the table either from the spatial or the temporal perspective, respectively. For example,
traditional partitional clustering methods regard locations as objects and timestamps as
attributes when analyzing from the spatial perspective (Figure 3(b1)). Then, they partition
locations into location-clusters, based on the similarity of data elements across all timestamps.
As the clustering results are location-clusters, such an analysis is also known as spatial
partitional clustering. When analyzing from the temporal perspective (Figure 3(b2)), tradi-
tional partitional clustering methods regard timestamps as objects and locations as attributes.
Subsequently, they partition timestamps into timestamp-clusters based on the similarity of
elements across all locations. Such an analysis is also known as temporal partitional clustering
because the resulting clusters are timestamp-clusters. Theoretically, any traditional clustering
method can perform spatial and temporal clustering analysis separately.

3.1.1. Overview of traditional clustering methods

Extensive studies have been conducted for the application of clustering methods for spatio-
temporal data, including GTS (Berkhin 2006, Han et al. 2012), most of which are traditional
methods. Regarding traditional partitional clustering algorithms, the most widely used one
is k-means (MacQueen 1967, Lloyd 1982). This algorithm is described in detail as being
representative of one-way clustering methods (Section 4.2). White et al. (2005) and Mills
et al. (2011) employed k-means to locate similar regions in terms of phenology. Using
a partitioning and optimization process similar to that of k-means, iterative self-organizing
data analysis (ISODATA) employs the predefined number of clusters as an initial estimate,
and it is able to delete, split, and merge clusters for further refinement (Tou and Gonzalez
1974). Gu et al. (2010) applied ISODATA, which is widely used in remote sensing, to identify
6 X. WU ET AL.

regions with similar phenological characteristics. Kohonen (1995) developed self-organizing

maps (SOM) to map n-dimensional input data to neurons on a 2D plane. Starting with
a random initialization of values for the neurons, SOM considers each input data as a vector
and aims to determine its best match unit (BMU) in the neurons with the nearest Euclidean
distance. Once chosen as a BMU of a particular vector, the neuron changes the values of its
neighboring neurons in the output space by using a neighborhood function. The above-
mentioned training process ceases when all the input vectors find their corresponding
BMUs, and the output neurons become stable. Thus, SOM groups similar input vectors to
the same or adjacent neurons, thereby proving their feasibility for partitional clustering
analysis. Owing to its effectiveness for dimension reduction, SOM has also been used for
spatial or temporal clustering in many applications such as company location (Guo et al.
2006), crime rate analysis (Andrienko et al. 2010), and weather analysis (Wu et al. 2013).
In terms of traditional hierarchical clustering methods, popular algorithms are
balanced iterative reducing and clustering using hierarchies (BIRCH, Zhang et al.
1996) and hierarchical SOM (Hagenauer and Helbich 2013). Designed for clustering
large datasets, BIRCH first extracts the clustering features (CFs) from data and then
organizes the CFs into a clustering feature tree (CF tree). Then, the next optional step
entails compressing the initial CF tree into a smaller one to remove outliers and group
sub-clusters. Once the smaller CF tree is built, BIRCH uses an existing hierarchical
clustering algorithm to conduct global clustering with the CF tree. The final optional
step is to reassign data elements to the closest existing cluster centroids to refine the
clusters. Inspired by previous work on Kangas Map (KM, Kangas 1992, Bação et al.
2005) and SOM, Hagenauer and Helbich (2013) proposed a hierarchical clustering
algorithm named hierarchical spatio-temporal SOM (HSTSOM), which is designed with
a spatial and temporal KM in the upper layer and a basic SOM in the lower layer. To
separately consider the spatial and temporal dependence of the data, HSTSOM trains
the two KMs in the upper layer independently but in parallel. To identify spatio-
temporal clusters, this algorithm then concatenates the positions of BMUs in the
upper-layer KMs for each input data to create training vectors for the lower-layer
SOM. In their study, HSTSOM was applied to analyze the socio-economic character-
istics of Vienna. Additional traditional clustering methods are discussed in the litera-
ture (Berkhin 2006, Miller and Han 2009, Grubesic et al. 2014).

3.2. Co-clustering methods

Both partitional and hierarchical co-clustering methods treat locations and timestamps
equally and concurrently analyze 2D GTS along the spatial and temporal dimensions. For
example, partitional co-clustering methods (Figure 3(c)) simultaneously partition loca-
tions into location-clusters and timestamps into timestamp-clusters based on the simi-
larity of data elements along both locations and timestamps. In this case, the clustering
results are co-clusters with similar elements along both dimensions, which are intersected
by each of location-clusters and timestamp-clusters.

3.2.1. Overview of co-clustering methods

Co-clustering methods have attracted significant attention ever since they were first
proposed in the early 1970s (Hartigan 1972, Padilha and Campello 2017, Shen et al.
 INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE 7

2018, Wu et al. 2020b). However, the majority of previous studies focused on other fields,
especially bioinformatics (Eren et al. 2012), with only a few recent studies focusing on
spatio-temporal data (Wu et al. 2017). To ensure our overview is comprehensive, co-
clustering methods used in other fields are also mentioned here.
Regarding partitional co-clustering methods, Dhillon et al. (2003) proposed the informa-
tion theoretic co-clustering (ITCC) algorithm for simultaneous word-document clustering.
With an initial random mapping from words to word-clusters and document to document-
clusters, ITCC regards the co-clustering issue as the optimization process in information
theory and formulates the objective function as the loss of mutual information between the
original variables (word and document) and the clustered ones (word-clusters and docu-
ment-clusters). Then, it optimizes the objective function by reassigning words and docu-
ments to word-clusters and document-clusters until convergence is achieved. Cho et al.
(2004), who aimed to analyze gene expression data, developed a co-clustering algorithm by
using the minimum sum-squared residual as the similarity/dissimilarity measure. This algo-
rithm organizes data in the form of a 2D matrix and yields the first set of row-clusters and
column-clusters using either random or spectral initialization, and uses residuals to build
the objective function, which it then minimizes to obtain the optimal co-clustering results.
Generalizing these previous studies (Dhillon et al. 2003, Cho et al. 2004), Banerjee et al.
(2007) subsequently proposed the Bregman co-clustering algorithm as a meta co-clustering
algorithm that aims to partition the original data into co-clusters with several distortion
functions such as the Euclidean distance. They also mentioned several applications of co-
clustering such as natural language processing (Rohwer and Freitag 2004) and video
content analysis (Cai et al. 2005). Recently, Wu et al. (2015) applied the Bregman block
average co-clustering algorithm with I-divergence (BBAC_I), a special case of the Bregman
co-clustering algorithm, to analyze temperature series for simultaneous location and time-
stamp clustering. This was the first study that applied co-clustering analysis to spatio-
temporal data. Afterwards, several studies applied BBAC_I for analyzing GTS in a variety
of fields, for example, disease hotspot detection (Ullah et al. 2017) and identification of
favorable conditions for virus outbreaks (Andreo et al. 2018).
Regarding hierarchical co-clustering methods, Hartigan (1972) developed a direct co-
clustering algorithm and applied it to analyze American presidential voting. This algo-
rithm, which is one of the earliest co-clustering algorithms, employs the squared
Euclidean distance to build the objective function and then aims to minimize it by
using a ‘divide and conquer’ direct clustering algorithm in a hierarchical manner.
Another hierarchical co-clustering algorithm proposed by Hosseini and Abolhassani
(2007) aimed to analyze queries and URLs of a search engine log, to mine the query
logs in web information systems. This algorithm uses the queries and URLs to construct
a bipartite graph in which singular value decomposition (SVD) is used to perform dimen-
sion reduction. Subsequently, k-means is used to iteratively cluster queries, and URLs are
used to create the hierarchical categorization. Costa et al. (2008) developed a hierarchical,
model-based co-clustering algorithm and used it to analyze internet advertisements.
Considering the dataset as a joint probability distribution, this algorithm groups tuples
into clusters characterized by different probability distributions. Thereafter, co-clusters are
identified by exploring the conditional distribution of elements over tuples. Inspired by
ITCC, Cheng et al. (2012), and Cheng et al. (2016) proposed a hierarchical co-clustering
algorithm by employing the information divergence as the measure of similarity/
8 X. WU ET AL.

dissimilarity to analyze newsgroups and documents. This algorithm starts with an initial
co-cluster, and then constructs hierarchical structures of rows and columns by iteratively
splitting the rows and columns to achieve convergence. Unlike ITCC, which uses the loss
in mutual information, Ienco et al. (2009) and Pensa et al. (2012) proposed a hierarchical
co-clustering algorithm named Incremental Flat and Hierarchical Co-Clustering (iHiCC).
This algorithm employs Goodman-Kruskal’s τ coefficient to measure the strength of the
link between two variables and uses the result for text categorization. Using the first
hierarchy created by τCoClust (Robardet 2002), this algorithm divides rows and columns
iteratively until only one element remains in all leaves of the hierarchies of both rows and
columns. However, to date, studies that have applied hierarchical co-clustering methods
for the analysis of spatio-temporal data have not been published. Detailed reviews on co-
clustering were presented by Charrad and Ahmed (2011), Eren et al. (2012), and Padilha
and Campello (2017).

3.3. Tri-clustering methods

Both partitional and hierarchical tri-clustering methods concurrently analyze 3D GTS in
the cuboid along the spatial, temporal, and third dimensions. For example, partitional tri-
clustering analysis of GTS-As (Figure 3(e)) simultaneously groups locations into location-
clusters, timestamps into timestamp-clusters and attributes into attribute-clusters based
on the similarity of data elements along all three dimensions. The clustering results are tri-
clusters that contain similar elements along locations, timestamps and attributes, which
are intersected by each of location-clusters, timestamp-clusters, and attribute-clusters.

3.3.1. Overview of tri-clustering methods

Since the proposal of the first tri-clustering algorithm in 2005 (Zhao and Zaki 2005), this
emerging subject has attracted increasing attention (Henriques and Madeira 2018).
Almost all previous studies on tri-clustering methods focused on other fields, with few
on geo-related fields (Wu et al. 2018). Nevertheless, we mention other methods to ensure
our description is complete.
Previous studies focused on partitional tri-clustering methods to a larger extent. Zhao and
Zaki (2005) introduced the first tri-clustering algorithm named TRICLUSTER, which aims to
mine coherent gene expression over time based on graph-based approaches. TRICLUSTER
first identifies co-clusters as the intermediate results by creating multigraphs of ranges and
finding constrained maximal cliques. Subsequently, these candidate co-clusters generate tri-
clusters. Thereafter, Sim et al. (2010) proposed the mining-correlated 3D subspace Cluster
(MIC) to analyze continuous-valued data and stock-financial-ratio-year data as examples.
Initialized by generating pairs of values with highly correlated information as seeds or initial
clusters, MIC greedily refines these clusters to mine correlated 3D subspace clusters by
optimizing the correlation information of those seeds. In addition, Hu and Bhatnagar (2010)
proposed a tri-clustering algorithm to analyze real-valued gene expression data. Their algo-
rithm identifies tri-clusters in two datasets by specifying an upper threshold for the standard
deviations of these tri-clusters. To this end, the algorithm first searches for co-clusters of which
the standard deviation obeys the specified upper bound in each dataset. Then, tri-clusters are
formulated from these candidate co-clusters. Based on their work on co-clustering analysis,
Wu et al. (2018) extended an existing co-clustering algorithm to a tri-clustering algorithm
 INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE 9

named the Bregman cuboid average tri-clustering algorithm with I-divergence (BCAT_I) to
analyze 3D GTS.
Few studies concerned with hierarchical tri-clustering methods have been reported
and these efforts mostly focus on analyzing biological data. Gerber et al. (2007) developed
a tri-clustering algorithm named GeneProgram for gene expression data analysis based
on hierarchical Dirichlet processes. This algorithm first discretizes continuous gene
expression data, and then employs Markov chain Monte Carlo sampling to approximate
the model posterior probability distribution using a three-level hierarchy in the Dirichlet
process, and finally identifies tri-clusters by summarizing the distribution. Amar et al.
(2015) proposed an algorithm known as three-way module inference via Gibbs sampling
(TWIGS) to analyze large 3D biological datasets. TWIGS functions by initially developing
a hierarchical Bayesian generative model for binary data by using the Bernoulli-Beta
assumption and for real-valued data by using the Normal-Gamma assumption.
Subsequently, TWIGS employs a co-clustering solution as the starting point and then
iteratively improves it using the Gibbs sampler. Finally, tri-clusters are inferred from
candidate co-clusters. A detailed overview of tri-clustering algorithms was published by
Henriques and Madeira (2018).

4. Data and representative algorithms of clustering methods

In this section, the dataset we used as a case study is first described. Then, algorithms
representative of each category of clustering methods are briefly described.

4.1. Case study dataset

To illustrate this study, we used the PM2.5 dataset published by Microsoft Research Asia
(MRA, Zheng et al. 2013, 2014), which is freely available. The dataset contains hourly PM2.5
concentrations at 36 monitoring stations in Beijing from 8 February 2013 to
8 February 2014. Because of the incompleteness of the dataset (Li et al. 2016), we selected
18 stations in the central urban areas (Figure 4). A Thiessen polygon map was created
using the coordinates of these stations (also available from MRA) to indicate the area
covered by each station. Furthermore, for the purpose of analysis, 299 days ranging from
1 February 2013 to 31 January 2014 (365 days) were selected as the study period with the
criterion that the days on which PM2.5 concentrations for all stations are zero for 24 h are
removed. The temporal distribution of these 299 days and the number of non-zero days in
each month over the study period are shown in Figure 5. The experiments were imple-
mented in MATLAB 2018a on a laptop running Windows 10 (64-bit) with a 2.20-GHz Intel
core (i7) CPU with 16 GB of RAM. Parallel computing was not implemented in our
experiments although this could be an interesting line for further research.
Patterns of spatial distribution, seasonal and diurnal variation of PM2.5 concentrations
were analyzed in previous studies (Zhang and Cao 2015, Chen et al. 2015). Based on these
existing studies, we constructed example questions for the PM2.5 dataset according to the
clustering-related questions discussed in Section 2. The 22 example questions we con-
structed are listed in Table 2. Clustering methods were then compared by answering
these questions.
10 X. WU ET AL.

Figure 4. Thiessen polygon map indicating the area covered by each station and the location of the
study area in Beijing (inset).

Figure 5. Temporal distribution of PM2.5 dataset collected in Beijing with non-zero days indicated in
green (outer circle) and the number of non-zero days in each month over the study period (inner
histogram).
 INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE 11

Table 2. Example questions of the clustering PM2.5 dataset in Beijing.

Components
Reading levels Number
I. where (SC) + when (TC) → what (C)
elementary SC What is/are the pollution level(s) of PM2.5 at station-cluster1 and in day-cluster1? 1
+ elementary TC What is/are the pollution level(s) of PM2.5 at station-cluster1, in day-cluster1 and 2
in hour-cluster1?
synoptic SC What is the pattern of pollution in the study area in day-cluster2? 3
+ elementary TC What is the pattern of pollution in the study area in day-cluster1 and hour-cluster1? 4
Elementary SC What is the pattern of pollution at station-cluster1 over the study period? 5
+ synoptic TC
synoptic SC What is the seasonal distribution of the pollution in the study area over the study 6
+ synoptic TC period?
What is the spatial distribution and seasonal variation of the pollution in the study 7
area over the study period?
What is the spatial distribution, seasonal and diurnal variation of the pollution in the 8
study area over the study period?
II. when (TC) + what (C) → where (SC)
elementary TC + At which station-cluster(s) is the PM2.5 pollution level of Good observed in day- 9
elementary C cluster2?
At which station-cluster(s) is the PM2.5 pollution level of Good observed in day- 10
cluster2 and hour-cluster1?
Synoptic TC At which station-cluster(s) is the PM2.5 pollution level of Good observed over the 11
+ elementary C study period?
elementary TC At which station-cluster(s) is the PM2.5 pollution level becoming worse in day-cluster1? 12
+ synoptic C At which station-cluster(s) is the PM2.5 pollution level becoming worse in day- 13
cluster2 and hour-cluster1?
Synoptic TC At which station-cluster(s) is the PM2.5 pollution level becoming worse over the time 14
+ synoptic C period?
III. where (SC) + what (C) → when (TC)
elementary SC In which day-cluster(s) is the PM2.5 pollution level of Good observed at station- 15
+elementary C cluster1?
In which day-cluster(s) and hour-cluster(s) is the PM2.5 pollution level of Good 16
observed at station-cluster1?
synoptic SC + In which day-cluster(s) is the PM2.5 pollution level of Good observed in the study 17
elementary C area?
In which day-cluster(s) and hour-cluster(s) is the PM2.5 pollution level of Good 18
observed in the study area?
elementary SC + In which day-cluster(s) does the PM2.5 pollution level worsen at station-cluster1? 19
synoptic C In which day-cluster(s) and hour-clusters does the PM2.5 pollution level worsen at 20
station-cluster1?
synoptic SC + In which day-cluster(s) does the PM2.5 pollution level worsen in the study area? 21
synoptic C In which day-cluster(s) and hour-cluster(s) does the PM2.5 pollution level worsen in 22
the study area?

4.2. K-means
Since the case study dataset is an hourly PM2.5 dataset, i.e., single attribute GTS with day
and hour as nested temporal dimensions (GTS-Ts), it was averaged to daily PM2.5 dataset,
i.e., single attribute GTS (GTS-A) when subjected to the traditional clustering method. The
dataset is organized into a table where rows are stations, columns are days and elements
are daily PM2.5 concentrations. Such a data table could also be seen as the co-occurrence
matrix OSD between a spatial and temporal variable, the former taking values in m (18)
stations and the latter in n (299) days.
Because of its wide use in many applications (Berkhin 2006), k-means was selected as
the algorithm representative of traditional clustering methods and used in this study to
perform temporal clustering. It is noteworthy that k-means can also be used to perform
spatial clustering.
12 X. WU ET AL.

Figure 6. Pseudocode of the k-means algorithm.

Suppose the days are clustered into l day-clusters. The pseudocode of the process of
iteratively optimizing day-clusters by k-means is depicted in Figure 6. With a random
initialization, l days are first selected as the cluster centers (step 1). Then, the iterative
process starts by assigning each of n days to the most similar cluster center measured by
the Euclidean distance indicated by Deuc ð; Þ(step 2.1). Next, for each of l day-clusters, the
cluster center is updated as the mean of all days assigned to this cluster (step 2.2). The
objective function of k-means is typically formulated as the sum of squared errors
between the days and corresponding day-clusters. The iterative process continues until
the objective function converges (i.e. reaches a value below a predefined threshold) and
the optimized l day-clusters are yielded.

4.3. Bregman block average co-clustering algorithm with I-divergence (BBAC_I)

The co-clustering method was also used to analyze the daily PM2.5 dataset in the table and
considers the table as the co-occurrence matrix, OSD. BBAC_I was chosen as the repre-
sentative algorithm because of its effectiveness in analyzing GTS (Wu et al. 2015, 2016).
Suppose the stations are clustered into k station-clusters, and the days are clustered
into l day-clusters for the analysis of the daily PM2.5 data matrix, OSD. The pseudocode of
BBAC_I (shown in Figure 7) demonstrates the process of optimizing the station-clusters
and day-clusters iteratively. With a random initial mapping as the starting point (step 1),
stations are partitioned into k station-clusters and days into l day-clusters concurrently,
resulting in a co-clustered 2D data matrix (O^ SD ). The objective function is then formulated
as the distortion between the original and the co-clustered matrices measured using the
 INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE 13

Figure 7. Pseudocode of BBAC_I.

information divergence (step 2), where DI ðjjÞ indicates the information divergence of
two matrices. Thereafter, the iterative process starts by re-assigning stations to station-
clusters and days to day-clusters, to optimize the objective function (step 3). This process
has been proven to monotonically decrease the objective function after each reassign-
ment (Banerjee et al. 2007). The iterative process terminates when the objective function
achieves convergence (i.e., gets below a predefined threshold) and k × l optimized sta-
tion-day co-clusters are yielded.

4.4. Bregman cuboid average tri-clustering algorithm with I-divergence (BCAT_I)

The tri-clustering method was used to analyze the hourly PM2.5 dataset, which is orga-
nized into a data cuboid where rows represent stations, columns represent days, depths
are 24 hours, and elements are hourly PM2.5 concentrations. Such a data cuboid can be
regarded as the 3D co-occurrence matrix, OSDH, among one spatial variable taking values
in m (18) stations, and two temporal variables, taking values in n (299) days and p (24)
hours, respectively.
BCAT_I, which was developed and proven to be effective for analyzing GTS (Wu et al.
2018), was selected as the representative algorithm. Suppose the stations, days and
24 hours in OSDH are clustered into k station-clusters, l day-clusters and z hour-clusters in
the tri-clustering analysis. The pseudocode of BCAT_I (Figure 8) demonstrates the
14 X. WU ET AL.

Figure 8. Pseudocode of BCAT_I.

optimization process of partitioning the hourly PM2.5 matrix into tri-clusters in an iterative
manner. Starting with a random initialization by mapping stations to k station-clusters, days
to l day-clusters and hours to z hour-clusters (step1), the algorithm first generates a tri-
clustered 3D data matrix (O^ SDH ). In the next step, BCAT_I measures the distortion between
the original and the tri-clustered matrices using the information divergence to build its
objective function (step 2). Thereafter, it aims to minimize the objective function by
iteratively updating mappings from stations to station-clusters, days to day-clusters and
hours to hour-clusters (step 3). The iterative process ceases when the objective function is
below a preset threshold, which yields the optimized k × l × z station-day-hour tri-clusters.

5. Results
In our analysis, the number of station-clusters was chosen to be three in accordance with
previous studies (Zhao et al. 2014, Wang et al. 2015) and the number of day-clusters is set
as four, with the expectation that days would fall into four ‘real’ seasons to enable us to
explore patterns of seasonal variations. Additionally, the number of hour-clusters was set
 INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE 15

to six because the air pollution index (AQI) for PM2.5 is categorized into six levels: Excellent
(0–50), Good (51–100), Lightly polluted (101–150), Moderately polluted (151–200), Heavily
polluted (201–300) and Severely polluted (>300) (according to the Technical Regulation
on Ambient Air Quality Index (on Trial) (China 2012)). Clustering results are interpreted by
answering example questions for clustering the PM2.5 dataset and then the three cluster-
ing algorithms are compared in terms of several aspects.

5.1. K-means clustering results

After the temporal clustering analysis, the 299 days are grouped into four day-clusters.
The ringmap in Figure 9 displays the temporal distribution of the four clusters of days. The
innermost circle in the ringmap shows the distribution of zero values in 365 days and the
other four circles are four clusters of days with increasing concentrations from inside

Figure 9. Ringmap displaying the results of k-means clustering. The innermost circle indicating days
with zero values. Other four circles from inside outward indicating day-cluster 1 to day-cluster 4 and
days in each day-cluster colored the same using average value.
16 X. WU ET AL.

outward. Each circle indicates 365 days, which is divided into 12 months from
February 2013 to January 2014 in a clockwise direction, and days falling into each cluster
are colored using the average value of that cluster.
With the ringmap, four out of the 22 questions (numbers: 3, 6, 17, 21) can be answered.
In response to question number 3, the ringmap shows that day-cluster2 has the averaged
value of ‘Lightly polluted’ according to (China 2012) and days therein mainly occur in
Spring (April and May), Summer and early Autumn (July, August, September and October).
As for question number 6, it can be seen that ‘Good’ days in day-cluster1 are sparsely
spread in April, July, August, December and January while ‘Lightly polluted’ days occupy
most of the study area. ‘Heavily polluted’ days are scattered throughout Winter and also
October. In response to question number 16, days in day-cluster4 are ‘Heavily polluted,’
and the fewest days are scattered throughout January 2014, February, early March and
October. For question number 20, because day-clusters are arranged from 1 to 4 with
increasing values of PM2.5 concentrations, the pollution level becomes worse from day-
cluster1 to day-cluster4.

5.2. BBAC_I co-clustering results

After the co-clustering analysis, 18 stations were grouped into three station-clusters and
299 days were grouped into four day-clusters, resulting in 12 (3 × 4) co-clusters.
The heatmap (Figure 10) straightforwardly shows all co-clusters: by arranging day-
clusters and station-clusters with increasing values from left to right along the x-axis and
from bottom to top along the y-axis, respectively. Consequently, values of co-clusters
increase from the bottom left to the top right. Each geographical map in the small multiples
(top of Figure 11) displays the spatial distribution for each of three station-clusters with
PM2.5 concentrations increasing from left to right. For each map, the region covered by each
station-cluster is colored with an average value in that cluster. The ringmap (bottom of

Figure 10. Heatmap displaying BBAC_I co-clustering results. The color of each co-cluster intersected
by each station- and day-cluster indicating the average value of that co-cluster.
 INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE 17

Figure 11. Small multiples (top) and ringmap (bottom) displaying BBAC_I co-clustering results. In the
small multiples, stations falling into each station-cluster colored using the average value of that
station-cluster. In the ringmap, the innermost circle indicating days with zero values. Other four circles,
from inside outward, representing day-cluster1 to day-cluster4 and days in each day-cluster colored
the same as the average value.

Figure 11) shows the temporal distribution of four day-clusters using four circles inside
outward with increasing values. For each circle, days in corresponding day-clusters are
displayed in the same color as an average value.
With these visualizations, more than half of the example questions (13) can be
answered (numbers: 1, 3, 5, 6, 7, 9, 11, 12, 14, 15, 17, 19, 21). Questions answered by
k-means clustering results are not repeated. For question number 1, the heatmap shows
18 X. WU ET AL.

that the co-cluster intersected by station-cluster1 and day-cluster4 is ‘Heavily polluted’.

For question number 3, days in day-cluster2 are mostly spread from July to October.
During these days, the pollution level worsens from ‘Good’ at stations in the east (station-
cluster1&2) to ‘Lightly polluted’ at stations in the west (station-cluster3). In response to
question number 5, the pollution level of the Haidianbeijingzhiwuyuan station (1002, 海
淀北京植物园) in station-cluster1 changes from ‘Excellent’ in day-cluster1 to ‘Heavily
polluted’ in day-cluster4. For question number 7, the pollution level worsens from the
west to the east of the study area and from Summer to Winter in the study period.
Moreover, the highest fluctuations of PM2.5 values occur during Winter, with the highest
and lowest levels of the entire year, whereas the fluctuations in Spring and Summer are
much reduced with medium-level concentrations (Li et al. 2015, 2016). For question
number 9, the heatmap shows that in station-cluster1, the pollution level of ‘Good’ is
observed in day-cluster2. The result also shows that in station-cluster1 and station-
cluster2, the pollution level is observed to be ‘Good’ over the study period as the response
to question number 11. For question number 12, in day-cluster1, the pollution level
worsens from station-cluster1 in the west to station-cluster3 in the east of the study
area. The same trend can also be observed over the time period for question number 14.
The answer to question number 15 is the same as that to question number 9 and the
answer to the last question is the same as that to question number 5.

5.3. BCAT_I tri-clustering results

After the tri-clustering analysis, 18 stations, 299 days and 24 hours were grouped into
three station-clusters, four day-clusters and six hour-clusters, respectively, resulting in
72 (3 × 4 × 6) tri-clusters. The quasi-3D heatmap in Figure 12 provides a direct view of all
tri-clusters arranged according to station-clusters, day-clusters, and hour-clusters with
values increasing from bottom to top of rows, from left to right of columns, and from
front to back of the depths, respectively. The overall view is that the values of tri-clusters
increase from the bottom left front to the top right back. The spatial distribution for
each station-cluster is displayed in the small multiples with PM2.5 values increasing from
left to right (top of Figure 13). Four circles in the ringmap (middle of Figure 13) use
colors to display the temporal distribution of days in four day-cluster with values
increasing from the innermost to the outermost ring. The set of six bar timelines
(bottom of Figure 13) displays the temporal distribution of six hour-cluster over
24 hours with concentrations increasing from the bottom to the top. Each bar timeline
represents 24 hours and hours in each hour-cluster are colored using the average value
of that hour-cluster to show the distribution.
With the above visualizations, all example questions can be answered. Questions that
were answered by k-means and co-clustering results are not repeated. For question
number 2, the quasi-3D heatmap shows that the pollution level of PM2.5 is ‘Excellent’ at
station-cluster1, day-cluster1 and hour-cluster1. For question number 4, the overall
pollution level in day-cluster1 and hour-cluster1 is ‘Excellent’ with the pollution slightly
worsening from the west to the east of the study area. For question number 8, the
combination of visualizations shows that most stations in the west and other stations
mostly in the southern and eastern areas have the highest value. These results are
consistent with those of previous studies, i.e., low PM2.5 values exist in the north (west),
 INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE 19

Figure 12. Quasi-3D heatmap displaying the tri-clustering results. The color of each tri-cluster
intersected by each station-, day- and hour-cluster indicating the average value of that tri-cluster.

whereas high values exist in the south (east) (Zhao et al. 2014, Wang et al. 2015). With
respect to the seasonal variation, it is shown that high fluctuations of PM2.5 concentrations
occur in the Autumn and especially in the Winter, whereas a more stable pattern of
middle-valued concentrations appear in the Spring and Summer (Li et al. 2015, 2016).
Furthermore, hours from 7:00 to 14:00 are characterized by low concentrations, whereas
hours from 21:00 to 24:00 and 1:00 to 3:00 occur the highest PM2.5 concentrations. These
results are supported by previous studies on diurnal variations (Zhao et al. 2014, Chen
et al. 2015). For question number 10, the heatmap shows that in day-cluster2 and hour-
cluster1, station-cluster2 and station-cluster3 are observed to have a ‘Good’ pollution
level. This includes all stations except Haidianbeijingzhiwuyuan (1002, 海淀北京植物园)
as shown in the small multiples. The response to question number 13 is that, in day-
cluster2 and hour-cluster1, the pollution level worsens from ‘Excellent’ to ‘Good’ from
station-cluster1 to station-clusters2&3. For question number 16, the heatmap shows that,
in station-cluster1, the pollution level is observed to be ‘Good’ at several intersections
of day-clusters and hour-clusters, e.g., the intersections of day-cluster1 and hour-cluster6,
day-cluster2 and hour-cluster2. It also shows that the pollution level of ‘Good’ is observed
at additional intersections of day-clusters and hour-clusters in the study area for question
number 18 (e.g., that of day-cluster2 and hour-clusters1-6 at station-cluster3). For ques-
tion number 20, at station-cluster1, the pollution level worsens from day-cluster1
and hour-cluster1 to day-cluster4 and hour-cluster6, i.e., from hours 7:00–9:00 on days
scattered throughout April and November to hours 21:00–24:00 on days spread sparsely
across September, October and January 2014. Moreover, it shows that the pollution level
20 X. WU ET AL.

Figure 13. Small multiples (top), ringmap (middle) and bar timelines (bottom) displaying the tri-clustering
results. In the small multiples, stations in each station-cluster colored the same using average value. In the
ringmap, the innermost circle indicating days with zero values. Other four circles (from inside outward)
indicating day-cluster1 to day-cluster4 and days in each day-cluster colored the same using average value.
In the bar timelines, hours in each hour-cluster colored the same using average value.
 INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE 21

is worsening from day-cluster1, hour-cluster1 and station-cluster1 to day-cluster4, hour-

cluster6 and station-cluster3, respectively, in response to the last question.

5.4. Comparisons of clustering algorithms

The k-means, BBAC_I and BCAT_I algorithms are compared using the case study dataset
and the results in terms of the input data, size of the data matrix, the number of
parameters needed, the number of iterations & initializations, computational efficiency
represented by average running time and also the number of example questions
answered (Table 3).
The results in Table 3 indicate that BCAT_I analyzes the dataset with finer resolution
and larger size than k-means and BBAC_I, whereas BCAT_I requires a larger number of
input parameters. Both k-means and BBAC_I analyzed the daily PM2.5 dataset with the size
18 × 299, whereas BCAT_I analyzed the hourly dataset with the size 18 × 299 × 24. As such,
the tri-clustering algorithm allows the inclusion of more information in the clustering
process and consequently in the results. In terms of the number of input parameters for
the case study, k-means requires the least, namely three, i.e., the number of day-clusters,
iterations and initializations. In comparison, BBAC_I needs an additional parameter as the
number of station-clusters and BCAT_I also needs the number of hour-clusters.
In terms of computational efficiency, k-means requires the shortest average running time
for analysis in the case study, followed by BBAC_I, whereas BCAT_I needs the longest time.
Using the same number of iterations and initializations for each algorithm, the results in
Table 3 indicate that k-means is 100 times faster than BBAC_I and thousands of times faster
than BCAT_I. Compared with BCAT_I, the average running time of BBAC_I is 60 times faster.
In terms of answering example questions, BCAT_I is the most capable method because
it allows us to answer all questions. This method is followed by BBAC_I (answers more
than half of all questions), and then k-means (answers less than one-fifth of all questions).
By performing temporal clustering, k-means can answer any question on the spatial-
cluster at the synoptic level (synoptic SC). Because traditional clustering methods can
perform spatial clustering separately, theoretically k-means can also answer three exam-
ple questions: 5, 11 and 14 (which are questions on the temporal-cluster at the synoptic
level (synoptic TC)). As such, it can reveal spatial or temporal patterns, e.g., the seasonal
variation in the PM2.5 dataset. BBAC_I concurrently performed spatial and temporal
clustering with the clustering results allowing us to answer all questions except those
with two nested temporal dimensions. In view of this, BBAC_I can reveal more complex
patterns, e.g., the spatial distribution and seasonal variation in the case study dataset. The
analysis of the hourly dataset using BCAT_I enabled us to answer all questions and explore
more patterns in the dataset, e.g., the spatial distribution, seasonal and diurnal variations.

Table 3. Comparisons of the three clustering algorithms.

Number of Number of Average Number of exam-
Clustering Size of the parameters iterations & running ple questions
algorithm Input data data matrix needed initializations time answered
k-means Daily PM2.5 dataset 18 × 299 3 100 & 20 0.01 second 4 (out of 22)
BBAC_I Daily PM2.5 dataset 18 × 299 4 100 & 20 1 second 13 (out of 22)
BCAT_I Hourly PM2.5 dataset 18 × 299 × 24 5 100 & 20 60 seconds 22 (out of 22)
22 X. WU ET AL.

6. Discussion
6.1. Suggestions for selecting clustering methods
As mentioned above, tri-clustering methods represented by BCAT_I are more powerful in
analyzing GTS with fine resolutions and exploring complex patterns but are less compu-
tationally efficient than other methods. In comparison, traditional clustering methods
represented by k-means and co-clustering methods represented by BBAC_I are capable of
exploring less complex patterns but require less running time. Then, given one-way
clustering, co- and tri-clustering methods for GTS, is there one type as the best and
most suitable for any task and dataset? Or is it possible to select a single method as being
superior? There is no clear cut answer to such a question, as stated by Grubesic et al.
(2014). Selection of the most suitable method should consider the data type to be
analyzed, the research questions with which researchers are concerned, the computa-
tional effort, and the availability of the methods (Table 4).
If the data at hand are 2D GTS and research questions relate to the whole study area or
period, traditional clustering methods instead of co-clustering methods are recommended,
especially for large datasets. That is because the computational complexity of co-clustering
methods is generally higher than that of traditional clustering methods. As shown in
Table 4, the computational complexity of k-means is O(mnki) (where m is the number of
rows in GTS, n is the number of columns, k is the number of row-clusters and i is the number
of iterations needed to reach convergence). In comparison, the complexity of BBAC_I is
higher, i.e., O(mni(k + l)) (where l is the number of columns in GTS). Nevertheless, if research

Table 4. Comparison of one-way clustering, co- and tri-clustering methods.

Clustering-related Typical Computational
Methods Data questions algorithms complexity Availability
Traditional clustering 2D-GTS (GTS-A) Synoptic SC + k-means O(mnki)a Codes available in
elementary TC; different
synoptic SC + languages, e.g.,
synoptic TC; Python
synoptic SC + BIRCH O(m) Codes available in
elementary C; different
synoptic SC + languages, e.g.,
synoptic C; Python
elementary SC +
synoptic TC;
synoptic TC +
elementary C;
synoptic TC +
synoptic C
Co-clustering 2D-GTS (GTS-A) All BBAC_I O(mni(k + l)) Available onlineb
iHiCC O(i(k + l)2) Available upon
request
Tri-clustering 3D-GTS (GTS-Ss; All TRICLUSTER O(mn2p) Available onlinec
GTS-Ts; BCAT_I O(mnpi Available onlined
GTS-As) (k + l + z))
a
where m is the number of rows, n is the number of columns, p is the number of depths, i is the number of iterations
needed until convergence, k is the number of row-clusters, l is the number of column-clusters and z the number of
depth-clusters.
b
https://figshare.com/s/48324046400cac9489f8.
c
http://www.cs.rpi.edu/~zaki/software/TriCluster.tar.gz.
https://figshare.com/s/48324046400cac9489f8.
 INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE 23

questions also relate to individual spatial-clusters or timestamp-clusters, then co-clustering

methods are suggested even though they are more time-consuming.
Tri-clustering methods are suggested if researchers are interested in analyzing 3D
GTS and answering any clustering-related research questions, even at the expense of
considerable computational effort. As shown in Table 4, the computational complexity
of TRICLUSTER, the first tri-clustering algorithm, is O(mn2 p) (where p is the number of
depths in the 3D data cuboid that GTS is organized into) and that of BCAT_I is O(mnpi
(k + l + z)) (where z is the number of depth-clusters). Compared with that of
traditional clustering and co-clustering methods, the complexity of tri-clustering
methods is much higher because they generally need to search all three dimensions
of the data cuboid for potential tri-clusters. Moreover, the computational complexity
is directly linked to the size of the dataset and that could be challenging when the
size increases.

6.2. Comparisons of the classifications of clustering methods

To date, there has been no uniform classification of clustering methods for spatio-
temporal data. For instance, Han et al. (2009) provided an overview of clustering methods
for spatial point data by classifying them as partitional, hierarchical, density-based and
grid-based methods. Han et al. (2009) and Kisilevich (2010) proposed two different
classifications for trajectory data. In the work of Han et al. (2009), clustering methods
were first categorized depending on whether they cluster entire or partial trajectories.
Thereafter, entire trajectory clustering methods were further divided into probabilistic
and density-based methods. Kisilevich (2010) broadly divided clustering methods into
two types: descriptive & generative model-based clustering methods and density-based
methods. Recently, Grubesic et al. (2014) provided a classification of clustering methods
for hotspot analysis, by dividing the methods into partitional, hierarchical, scan-based and
autocorrelation-based methods.
Compared with the aforementioned classifications, the one presented in this study is
straightforward and reveals new insights. One-way clustering methods analyze GTS along
a single dimension and result in spatial or temporal patterns. Co-clustering methods focus
on the analysis of two dimensions of GTS and concurrent spatial and temporal patterns
can be explored. In comparison, tri-clustering methods focus on the analysis of three
dimensions and result in spatio-temporal patterns in 3D GTS.
Furthermore, the classification described in this study is necessary because it allows to
include novel clustering methods. In the era of big data, various clustering methods for
patterns exploration are needed with increasing amounts of GTS. However, most cluster-
ing methods are categorized as one-way clustering methods and only explore spatial or
temporal patterns in 2D GTS. Co-clustering methods are needed to explore complex
patterns, e.g., concurrent spatio-temporal patterns. Moreover, the emergence of higher
dimensional GTS, e.g., 3D GTS, requires the use of clustering methods that can analyze
data along more dimensions. Although few of these methods have been applied to GTS
(Wu et al. 2015, 2018, Ullah et al. 2017, Andreo et al. 2018), the classification presented in
this study shows the potential of including other co- and tri-clustering methods, which
enable the exploration of more complex spatio-temporal patterns.
24 X. WU ET AL.

7. Conclusions
In this paper, we systematically described the classification of clustering methods for GTS
categorized into one-way clustering, co- and tri-clustering methods. Furthermore, we
compared different categories to offer suggestions for selecting appropriate methods. To
achieve this, we defined a taxonomy of clustering-related questions with three compo-
nents (spatial-cluster, temporal-cluster and cluster) and two reading levels (elementary,
synoptic). Different methods were then compared by answering these questions using
representative algorithms and a case study dataset.
Our results show that tri-clustering methods are more powerful in exploring complex
patterns from GTS with fine resolutions at the cost of considerably extended running time.
In relative terms, one-way clustering and co-clustering methods require less running time
but are less capable of exploring complex patterns. However, the selection of the most
appropriate method should consider the data type, research questions, computational
complexity, and also the availability of methods. Traditional clustering methods are
recommended for analyzing large 2D datasets when research questions focus on the
whole study area or period; otherwise, co-clustering methods are recommended for 2D
GTS. Tri-clustering methods are recommended for analyzing 3D GTS for complex patterns,
albeit at the expense of additional computational effort. Finally, the classification
described in this study is necessary because it can include more co- and tri-clustering
methods for GTS and thus explore more complex spatio-temporal patterns.

Acknowledgments
We thank the reviewers for their constructive comments.

Disclosure statement
No potential conflict of interest was reported by the authors.

Data availability statement

The data and codes that support the findings of this study are available in figshare.com with the
identifier(s) at the link https://figshare.com/s/48324046400cac9489f8.

Funding
This work was supported by the National Natural Science Foundation of China [41771537,
41901317]; China Postdoctoral Science Foundation Grant [2018M641246]; National Key Research
and Development Plan of China [2017YFB0504102];Fundamental Research Funds for the Central
Universities.

References
Amar, D., et al., 2015. A hierarchical Bayesian model for flexible module discovery in three-way
time-series data. Bioinformatics, 31 (12), i17–i26. doi:10.1093/bioinformatics/btv228
 INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE 25

Andreo, V., et al., 2018. Identifying favorable spatio-temporal conditions for west nile virus outbreaks
by co-clustering of modis LST indices time series. In: IGARSS 2018-2018 IEEE International
Geoscience and Remote Sensing Symposium. Valencia, Spain, 4670–4673.
Andrienko, G., et al., 2009. Interactive visual clustering of large collections of trajectories. In: 2009
IEEE Symposium on Visual Analytics Science and Technology (VAST) 12-13 Oct. Atlantic City, New
Jersey, 3–10.
Andrienko, G., et al., 2010. Space-in-time and time-in-space self-organizing maps for exploring spatio-
temporal patterns. Computer Graphics Forum, 29 (3), 913–922. doi:10.1111/cgf.2010.29.issue-3
Andrienko, N. and Andrienko, G., 2006. Exploratory analysis of spatial and temporal data -
a systematic approach. Berlin: Springer-Verlag.
Bação, F., Lobo, V., and Painho, M., 2005. The self-organizing map, the Geo-SOM, and relevant variants
for geosciences. Computers & Geosciences, 31 (2), 155–163. doi:10.1016/j.cageo.2004.06.013
Banerjee, A., et al., 2007. A generalized maximum entropy approach to Bregman co-clustering and
matrix approximation. Journal of Machine Learning Research, 8, 1919–1986.
Berkhin, P., 2006. A survey of clustering data mining techniques. Grouping Multidimensional Data:
Recent Advances in Clustering, 25–71.
Bertin, J., 1983. Semiology of graphics: diagrams, networks, maps. London: University of Wisconsin
Press.
Cai, R., Lu, L., and Cai, L.-H. 2005. Unsupervised auditory scene categorization via key audio effects
and information-theoretic co-clustering. Proceedings. (ICASSP’05). IEEE International Conference on
Acoustics, Speech, and Signal Processing, ii/1073-ii/1076 Vol. 1072. Philadelphia, Pennsylvania.
Charrad, M. and Ahmed, M.B., 2011. Simultaneous clustering: A survey. In: International Conference
on Pattern Recognition and Machine Intelligence. Moscow, Russia, 370–375.
Chen, W., Tang, H., and Zhao, H., 2015. Diurnal, weekly and monthly spatial variations of air
pollutants and air quality of Beijing. Atmospheric Environment, 119, 21–34. doi:10.1016/j.
atmosenv.2015.08.040
Cheng, T., et al., 2014. Spatiotemporal data mining. Handbook of regional science. Heidelberg,
Germany: Springer, 1173–1193.
Cheng, W., et al., 2012. Hierarchical co-clustering based on entropy splitting. Proceedings of the 21st
ACM international conference on Information and knowledge management. Maui, Hawaii,
1472–1476.
Cheng, W., et al., 2016. HICC: an entropy splitting-based framework for hierarchical co-clustering.
Knowledge and Information Systems, 46 (2), 343–367. doi:10.1007/s10115-015-0823-x
China, 2012. Technical regulation on ambient air quality index (on trial). China: China Environmental
Science Press Beijing.
Cho, H., et al., 2004. Minimum sum-squared residue co-clustering of gene expression data. Fourth
SIAM Int’l Conf. Data Mining. Florida, USA.
Costa, G., Manco, G., and Ortale, R., 2008. A hierarchical model-based approach to co-clustering
high-dimensional data. Proceedings of the 2008 ACM symposium on Applied computing. Maui,
Hawaii, 886–890.
Dhillon, I.S., Mallela, S., and Modha, D.S., 2003. Information-theoretic co-clustering. In: The 9th
International Conference on Knowledge Discovery and Data Mining (KDD). Washington, DC,
89–98. doi:10.1159/000071010
Eren, K., et al., 2012. A comparative analysis of biclustering algorithms for gene expression data.
Briefings in Bioinformatics, 14 (3), 279–292.
Gerber, G.K., et al., 2007. Automated discovery of functional generality of human gene expression
programs. PLoS Computational Biology, 3 (8), e148. doi:10.1371/journal.pcbi.0030148
Grubesic, T.H., Wei, R., and Murray, A.T., 2014. Spatial clustering overview and comparison: accuracy,
sensitivity, and computational expense. Annals of the Association of American Geographers, 104
(6), 1134–1156. doi:10.1080/00045608.2014.958389
Gu, Y., et al., 2010. Phenological classification of the United States: A geographic framework for
extending multi-sensor time-series data. Remote Sensing, 2, 526–544. doi:10.3390/rs2020526
26 X. WU ET AL.

Guo, D., et al., 2006. A visualization system for space-time and multivariate patterns (VIS-STAMP).
IEEE Transactions on Visualization and Computer Graphics, 12 (6), 1461–1474. doi:10.1109/
TVCG.2006.84
Hagenauer, J. and Helbich, M., 2013. Hierarchical self-organizing maps for clustering spatiotemporal
data. International Journal of Geographical Information Science, 27 (10), 2026–2042. doi:10.1080/
13658816.2013.788249
Han, J., Kamber, M., and Pei, J., 2012. Data mining concepts and techniques. 3rd ed. Burlington, MA:
Morgan Kaufman MIT press.
Han, J., Lee, J.-G., and Kamber, M., 2009. An overview of clustering methods in geographic data
analysis. In: H.J. Miller and J. Han, eds. Geographic data mining and knowledge discovery. 2nd ed.
New York: Taylor & Francis Group, 150–187.
Hartigan, J.A., 1972. Direct clustering of a data matrix. Journal of American Statistical Association, 67
(337), 123–129. doi:10.1080/01621459.1972.10481214
Henriques, R. and Madeira, S.C., 2018. Triclustering algorithms for three-dimensional data analysis:
A comprehensive survey. ACM Computing Surveys (CSUR), 51 (5), 95. doi:10.1145/3271482
Hosseini, M. and Abolhassani, H., 2007. Hierarchical co-clustering for web queries and selected urls.
In: International Conference on Web Information Systems Engineering. Nancy, France, 653–662.
Hu, Z. and Bhatnagar, R., 2010. Algorithm for discovering low-variance 3-clusters from real-valued
datasets. In: 2010 IEEE 10th International Conference on Data Mining (ICDM). Sydney, Australia,
236–245.
Ienco, D., Pensa, R.G., and Meo, R., 2009. Parameter-free hierarchical co-clustering by n-ary splits.
In: Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Bled,
Slovenia, 580–595.
Kangas, J., 1992. Temporal knowledge in locations of activations in a self-organizing map. In: I.
Aleksander and J. Taylor, eds. Artificial neural networks, 2. Vol. 1. Amsterdam, Netherlands:
North-Holland, 117–120.
Kisilevich, S., et al., 2010. Spatio-temporal clustering. In: O. Maimon, et al., eds. Data mining and
knowledge discovery handbook. Springer US, 855–874.
Kohonen, T., 1995. Self-organizing maps. Berlin: Springer-Verlag.
Li, H., Fan, H., and Mao, F., 2016. A visualization approach to air pollution data exploration—a case
study of air quality index (PM2. 5) in Beijing, China. Atmosphere, 7 (3), 35. doi:10.3390/atmos7030035
Li, R., et al., 2015. Diurnal, seasonal, and spatial variation of PM2. 5 in Beijing. Science Bulletin, 60 (3),
387–395. doi:10.1007/s11434-014-0607-9
Lloyd, S., 1982. Least squares quantization in PCM. IEEE Transactions on Information Theory, 28 (2),
129–137. doi:10.1109/TIT.1982.1056489
MacQueen, J., 1967. Some methods for classification and analysis of multivariate observations. the
Fifth Berkeley Symposium on Mathematical Statistics and Probability. Berkeley, California, 281–297.
Miller, H.J. and Han, J., 2009. Geographic data mining and knowledge discovery: an overview. In: H.
J. Miller and J. Han, eds. Geographic data mining and knowledge discovery - 2nd edition. London:
Taylor & Francis Group, 1–26.
Mills, R.T., et al., 2011. Cluster analysis-based approaches for geospatiotemporal data mining of
massive data sets for identification of forest threats. Procedia Computer Science, 4, 1612–1621.
doi:10.1016/j.procs.2011.04.174
Padilha, V.A. and Campello, R.J., 2017. A systematic comparative evaluation of biclustering
techniques. BMC Bioinformatics, 18 (1), 55. doi:10.1186/s12859-017-1487-1
Pensa, R.G., Ienco, D., and Meo, R., 2012. Hierarchical co-clustering: off-line and incremental
approaches. Data Mining and Knowledge Discovery, 28 (1), 31–64. doi:10.1007/s10618-012-0292-8
Peuquet, D.J., 1994. It’s about time: a conceptual framework for the representation of temporal
dynamics in geographic information systems. Annals of the Association of American Geographers,
84 (3), 441–461. doi:10.1111/j.1467-8306.1994.tb01869.x
Robardet, C., 2002. Contribution à la classification non supervisée: proposition d’une méthode de bi-
partitionnement. Doctoral dissertation, Lyon, 1.
Rohwer, R. and Freitag, D., 2004. Towards full automation of lexicon construction. Proceedings of the
HLT-NAACL Workshop on Computational Lexical Semantics. Boston, MA, 9–16.
 INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE 27

Shekhar, S., et al., 2015. Spatiotemporal data mining: a computational perspective. ISPRS
International Journal of Geo-Information, 4 (4), 2306–2338. doi:10.3390/ijgi4042306
Shen, S., et al., 2018. Spatial distribution patterns of global natural disasters based on biclustering.
Natural Hazards, 92 (3), 1809–1820. doi:10.1007/s11069-018-3279-y
Sim, K., Aung, Z., and Gopalkrishnan, V., 2010. Discovering correlated subspace clusters in 3D
continuous-valued data. 2010 IEEE 10th International Conference on Data Mining (ICDM),
471–480. doi:10.1016/j.nano.2009.09.005
Tou, J.T. and Gonzalez, R.C., 1974. Pattern recognition principles. Boston, MA: Addison-Wesley
Publishing Company.
Ullah, S., et al., 2017. Detecting space-time disease clusters with arbitrary shapes and sizes using a
co-clustering approach. Geospatial Health, 12 (2), 567.
Wang, Z., et al., 2015. Spatial-temporal characteristics of PM2.5 in Beijing in 2013. Acta Geographica
Sinica, 70 (1), 110–120.
White, M.A., et al., 2005. A global framework for monitoring phenological responses to climate
change. Geophysical Research Letters, 32 (4), L04705. doi:10.1029/2004GL021961
Wu, X., et al., 2020a. Spatio-temporal differentiation of spring phenology in China driven by
temperatures and photoperiod from 1979 to 2018. Science China-Earth Sciences. doi:10.1360/
SSTe-2019-0212
Wu, X., et al., 2020b. An interactive web-based geovisual analytics platform for co-clustering
analysis. Computers & Geosciences, 104420. doi:10.1016/j.cageo.2020.10442
Wu, X., et al., 2018. Triclustering georeferenced time series for analyzing patterns of intra-annual
variability in temperature. Annals of the American Association of Geographers, 108 (1), 71–87.
doi:10.1080/24694452.2017.1325725
Wu, X., Zurita-Milla, R., and Kraak, M.J., 2015. Co-clustering geo-referenced time series: exploring
spatio-temporal patterns in Dutch temperature data. International Journal of Geographical
Information Science, 29 (4), 624–642. doi:10.1080/13658816.2014.994520
Wu, X., Zurita-Milla, R., and Kraak, M.-J., 2013. Visual discovery of synchronization in weather data at
multiple temporal resolutions. The Cartographic Journal, 50 (3), 247–256. doi:10.1179/
1743277413Y.0000000067
Wu, X., Zurita-Milla, R., and Kraak, M.-J., 2016. A novel analysis of spring phenological patterns over
Europe based on co-clustering. Journal of Geophysical Research: Biogeosciences, 121, 1434–1448.
Wu, X., et al., 2017. Clustering-based approaches to the exploration of spatio-temporal data.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
(ISPRS’17). Wuhan, China, 1387–1391.
Zhang, T., Ramakrishnan, R., and Livny, M., 1996. BIRCH: an efficient data clustering method for very
large databases. ACM SIGMOD Record, 25 (2), 103–114.
Zhang, Y.L. and Cao, F., 2015. Fine particulate matter (PM 2.5) in China at a city level. Scientific
Reports, 5, 14884. doi:10.1038/srep14884
Zhao, C., et al., 2014. Temporal and spatial distribution of PM2.5 and PM10 pollution status and the
correlation of particulate matters and meteorological factors during winter and spring in Beijing.
Environmental Science, 35 (2), 418–427.
Zhao, L. and Zaki, M.J., 2005. TRICLUSTER: an effective algorithm for mining coherent clusters in 3D
microarray data. Proc. of the 2005 ACM SIGMOD International Conference on Management of Data.
Baltimore, Maryland, 694–705.
Zheng, Y., et al., 2014. A cloud-based knowledge discovery system for monitoring fine-grained air
quality. Preparation, Microsoft Tech Report, http://research.microsoft.com/apps/pubs/default. aspx
Zheng, Y., Liu, F., and Hsieh, H.-P., 2013. U-Air: when urban air quality inference meets big data.
Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data
mining. Chicago, IL, 1436–1444.