Connectivity-Based Clustering for Mixed Discrete and Continuous Data

International Journal on Cybernetics & Informatics (IJCI) Vol.13, No.3, June 2024
Bibhu Dash et al: SCDD -2024
pp. 17-30, 2024. IJCI – 2024 DOI:10.5121/ijci.2024.130303
CONNECTIVITY-BASED CLUSTERING FOR
MIXED DISCRETE AND CONTINUOUS DATA
Mahfuza Khatun1
and Sikandar Siddiqui2
1
Jahangirnagar University, Savar, 1342 Dhaka, Bangladesh
2
Deloitte Audit Analytics GmbH, Europa-Allee 91,
60486 Frankfurt, Germany
ABSTRACT
This paper introduces a density-based clustering procedure for datasets with variables of
mixed type. The proposed procedure, which is closely related to the concept of shared
neighbourhoods, works particularly well in cases where the individual clusters differ
greatly in terms of the average pairwise distance of the associated objects. Using a number
of concrete examples, it is shown that the proposed clustering algorithm succeeds in
allowing the identification of subgroups of objects with statistically significant
distributional characteristics.
KEYWORDS
Cluster analysis, mixed data, distance measures
1. INTRODUCTION
In the field of applied statistics, “clustering” and “cluster analysis” are collective terms for all
procedures by which individual objects can be aggregated into groups of mutually similar
entities.
One set of methods frequently used to perform this task, commonly referred to as centroid-based
approaches, model a cluster as a group of entities scattered around a common central point.
However, this may be counter-intuitive since many observers would also tend to group entities
together if they are scattered along a common, not necessarily linear path or surface, rather than a
common central point.
In response to this challenge, another class of clustering methods, often summarized under the
general term of “connectivity-“ or “density-based” approaches (Kriegel et al. [13]), has been
developed. In this context, clusters are defined as groups of entities located in regions of high
density and separated by near-empty areas within the sample space. Widely used examples of
such connectivity-based clustering procedures are DBSCAN (Ester et al., [5]) and OPTICS
(Ankerst et al., [2]); see also Oduntan [15].
The current paper addresses a key challenge that may occur in this context: It consists of the
possibility that the variables in the relevant datasets may be of mixed type, i.e. some of them may
be continuous, while others may either be ordered and discrete (like in the case of school grades)
or unordered and discrete (like in the case of country identifiers). This raises the question of how
the pairwise distances between the individual entities, which are key inputs for separating low-
and high-density regions, are to be measured. A possible solution to this problem is proposed in

18
Section 2 of this paper. The proposed clustering procedure itself, which is closely related to the
concept of shared neighbourhoods presented in Houle et al. [10], is described in Section 3.
Section 4 summarizes the outcomes of a number of exemplary applications. Section 5 concludes.
2. MEASURING DISTANCES BETWEEN ENTITIES
2.1. Starting Point
As in Khatun and Siddiqui [12], we assume that there is a dataset consisting of N entities i = 1,
…, N, each of which is characterised by a tuple qi = { xi, vi, zi } of features.
 xi is a realisation of a (K1×1) column vector X of numerical variables that either are
continuous or treated as continuous for practical reasons
 vi is a realisation of a (K2×1) vector V of ordered discrete variables, beginning with 1
numbered consecutively in steps of 1, and
 zi is a realisation of a (K3×1) vector Z of unordered, discrete variables.
2.2. Pairwise Distance with Respect to Continuous Variables
The distance between two entities i and j with respect to the values of X is measured by the
Manhattan Distance (see, e.g., Yang [20]) between the standardized xi and xj values as follows:
(1)
where r(Xs) denotes the range of the observed values of Xs, i.e. the difference between the sample
maximum and the sample minimum.
In this context, the Manhattan Distance is preferred to the more commonly used Euclidean
distance because when using the former, the contrast between the distances from different data
points shrinks less rapidly as the dimension K1 of X grows; see Aggarwal, Hinneburg, and Keim
[1]. In addition, the application of the distance measure (1) simplifies the consolidation of
distance measures for the different variable types involved, as will become obvious below.
2.3. Pairwise Distance with Respect to the Ordered Discrete Variables
The distance between i and j with respect to the values taken by the components of V can be
measured by
(2)
where ms is the number of possible realisations of the s-th ordered discrete variable.
This way of proceeding is justified as follows: If Vs is an ordered, discrete variable ranging from
to ms in steps of 1, then the actual value vi,s can be assumed to be dependent on the value taken
by a latent (=unobservable) variable vi,s
*
 ]0; 1] as follows:
vi,s = j if vi,s
*
 ](j-1)/ ms; j/ms ] (3)

19
The distance between the mid-points of two neighbouring sub-intervals given in (3) equals 1/ms.
2.4. Pairwise Distance based on Unordered Discrete Characteristics
With regard to Z, the distance between two entities i and j can be quantified by their separateness
or lack of overlap (see, e.g., Stanfill and Waltz [16]), based on the Hamming Distance (Hamming
[9]):
(4)
In the above equation, I(.
) is the indicator function that equals 1 if the condition in brackets is
fulfilled and 0 if not. The scalar ms stands for the number of distinct possible realisations of the s-
th unordered discrete variable.
The distance prevailing between two specific realisations zi,s and zj,s of the s-th unordered discrete
variable Zs with respect to the values thus equals 0 if zi,s equals zj,s, and 1/ms if not. The distance
between pairs of observations with different values of Zs, as measured by (3), thus shrinks as the
number of possible realisations of the relevant variable increases. This normalisation rule is
motivated by the idea that the finer the classification scheme according to which individual
entities are grouped, the smaller the average number of entities per group, and the less certain we
can be that differences in group membership reflect actual disparities between the entities, rather
than merely random “noise”.
2.5. Overall Pairwise Distance Between Two Entities
The overall distance score between two entities i and j can then be calculated by summing up the
distance measures for the different types of variables given in (1), (2), and (4).
(5)
3. CLUSTERING PROCEDURE
As mentioned in the introduction, the clustering procedure proposed in this section is based on
the concept of shared neighbourhood, as presented in the seminal paper by Houle et al. [10]. This
approach is adopted here because, according to the authors, it is less affected by the “curse of
dimensionality” (Bellman [3]). What is meant by this term is that as the number of variables
under consideration grows, the size differences between the pairwise distances of the individual
data points decrease, which makes it increasingly impossible to form meaningful clusters based
on such distances.
In line with the above specifications, it is possible to calculate, for each entity i in the sample, the
distance between itself and each of the remaining (N-1) entities, and to sort the results of these
calculations in ascending order. Let i(1) ≤ i(2) ≤ … ≤ i(N-1) denote the sorted distances from i,
and let g > 0 be a user-specified integer number. Then, the adjacency set Si of entity i, is defined
as the sets of all entities j i for which the inequality d(i,j) ≤ i(g) holds:
Si := { j i | d(i,j) ≤ i(g)} (6)

20
Two entities j and i are considered mutually interlinked if their adjacency sets Si and Sj have at
least one element in common:
Si ⚭ Sj  Si  Sj  {} (7)
The proposed clustering procedure can then be summarized as follows:
Listing 1: Clustering Procedure
Step 1: Gather all entities in the sample in the subset of hitherto unassigned entities
Step 2: Initialize the cluster index c as 0.
Step 3: Increase the cluster index c by 1.
Step 4: Set the number of elements in cluster c, denoted by nc, to 0
Step 5: Set i to 1.
Step 6: If
 entity number i has not yet been assigned to a cluster and
 nc exceeds 0 and
 entity number i and at least one element of cluster c are mutually
interlinked
then
 assign entity number i to cluster c,
 remove entity number i from the subset of hitherto unassigned entities, and
 increase nc by 1
Step 7: If
 entity number i has not yet been assigned to a cluster, and
 nc equals 0,
then
 assign entity number i to cluster c,
 remove entity number i from the subset of hitherto unassigned entities, and
 increase nc by 1
Step 8: Increase i by 1
Step 9 If i ≤ N, continue with Step 6
Step 10 If i > N, and if there is at least one object in the subset of hitherto unassigned
entities that is mutually interlinked with at least one element of cluster c, continue
with Step 5.
Step 11 If i > N and if there is not a single element in the subset of hitherto unassigned
entities is mutually interlinked with at least one element of cluster c, continue with
Step 3.
Step 12 If i > N and the set of hitherto unassigned entities is empty, terminate.
An implementation of this procedure in the matrix language GAUSS, together with the datasets
for the examples from the following section, is available on the WWW via
https://drive.google.com/drive/folders/1putjdHHMJjg2TafWenFl9lMBr3ShxC93?usp=drive_link
.
The above procedure will cause each of the entities in the sample to be unequivocally assigned to
a single cluster. However, particularly when the number g of neighbouring entities from which
the adjacency set of each entity i is derived is small (say, e.g. 1 or 2), some or even many objects
will be assigned to “degenerate” clusters comprising only a single object. On the other hand,
raising the value of g above a certain threshold (the level of which depends on the distributional
characteristics of the underlying dataset) will cause all objects to be gathered in a single,
maximally heterogeneous group. It thus becomes obvious that the choice of g implies a trade-off
between the potentially conflicting objectives of within-cluster homogeneity on one hand and

21
inclusiveness (i.e. the assignment of as many entities as possible to valid, or “non-degenerate”,
clusters) on the other.
The proposed solution to this problem is to compare different possible outcomes of the clustering
procedure with a measure of separation accuracy that can be calculated as follows: Let min
denote a user-defined minimum cluster size, c(i) the index number of the cluster to which object i
has been assigned, and nc(i) the number of objects in that cluster. Then, the quantity
(8)
equals the distance between object i and its closest neighbour within its cluster, provided that
at least reaches the specified minimum size, whereas
(9)
equals the distance between object i and its closest neighbour outside its cluster whenever c(i)
does not fall short of min. Then, for any given value min, the particular value g*
of g that
maximizes the quantity
(10)
can be looked upon as the one that maximizes the separation accuracy associated with the chosen
value of min.
4. EXEMPLARY APPLICATIONS
4.1. Two-Dimensional Datasets with Continuous Variables Only
Although the main purpose of the proposed method is to allow the clustering of mixed data, some
of its key characteristics are probably best understood when applying it to a small set of
deliberately simple cases. The first important feature is that, unlike the centroid-based
procedures, the procedure from Sections 2 and 3 can identify clusters of arbitrary shape rather
than being limited to ones with round or oval profiles. In an exemplary manner, this is shown for
the “Aggregation” dataset examined by Gionis, Mannila and Tsaparas (2007). Figure 1 displays
the outcome of an application of the proposed procedure in this case:

22
Figure 1
In the proposed procedure, the measure of neighbourliness that is used to decide whether two or
more data points belong together or are considered separate is not defined by a fixed distance
threshold; rather it is based on the presence or absence of at least one common neighbour. In a
setup where the individual clusters differ greatly in terms of the average pairwise distance of the
associated objects, this feature enables the procedure from Section 3 to identify such
accumulations of objects nevertheless, as is shown by its application to the “toy dataset” dealt
with in Jain and Law [11]:
Figure 2
However, the application of the proposed method to the R15” dataset studied in Veenman,
Reinders and Backer [18] also points to the flip side of the apparent advantages of our algorithm:
Because of its emphasis on common close neighbours, it tends to merge groups of objects
surrounding two or more different central points into a single group whenever there is some
degree of overlap between them. This can sometimes lead to counterintuitive results, as Figure 3
indicates: Here, our algorithm forms a single cluster out of eight centrally located point clouds
near the centre of the diagram, although most human observers would probably have perceived
them as separate groups.
The outcome for the dataset “Unbalance” examined by Rezaei and Fränti [14] (see Figure 4)
further underscores this point. Here, most human observers would probably have split Cluster 1
in four and Cluster 2 into three separate “sub-clusters”.
The above findings show that the proposed approach is far from a universal, objective solution to
clustering problems. It should rather be considered one out of several related approaches which,
in view of the variety of possible data and research objectives, can produce results with very
different degrees of plausibility from case to case.

23
Figure 3
Figure 4
The four examples in this subsection relate to synthetically generated data points from a bivariate
distribution involving continuous variables only. The purpose of the two following subsections is
to demonstrate that the proposed algorithm can also produce plausible results in real-world
situations involving more variables, and variables of different types.
4.2. Country Grouping by Macroeconomic Indicators
The above procedure can be applied to form groups of countries based on similarity comparisons
with regard to their geographic location as well as a number of macroeconomic indicators. In the
particular case examined here, we use the seven World Bank [19] regions as geographical
assignment indicators and a set of four macroeconomic variables, which include (i) GDP per
capita, as well as (ii) government debt, (iii) the current account balance, and (iv) the government
budget balance, the last three of which are being expressed as a percentage of GDP. This choice
is motivated because the four variables just mentioned are among the most commonly used
indicators used to assess the resilience of countries to adverse economic shocks; see, e.g.,
Briguglio et al. [4] for a more comprehensive treatment of this issue. The common source for all
the data in use is Trading Economics [17], and the reference date is generally the year-end of
2022. If no data is available for this date, the most recent earlier key date has been chosen. After
removing sovereign states with missing data, we end up with a sample of 163 countries.
With the minimum cluster size min set to 2, application of the above procedure to the dataset
compiled accordingly yields an optimum number of close neighbours (g*
) of 2 and leads to two
large clusters, one more small cluster of only two countries (Cambodia and the Maldives), and a
total of six “outliers” (Afghanistan, Bhutan, Cyprus, Guinea, Indonesia, and Lebanon) that cannot

24
be assigned to a cluster that reaches or exceeds the above size threshold. Table 4.1.1. enumerates
the countries assigned to Cluster 1 and 2 by region. Descriptive statistics of the four continuous
variables in use are given in Tables 4.1.2 to 4.1.5.
Table 1: Distribution of countries across Clusters 1 and 2
Cluster 1 Cluster 2
East Asia &
Pacific
Fiji, Japan Australia, Brunei, China, Laos, Malaysia,
Mongolia
Myanmar, New Zealand, Papua New
Guinea,
Philippines, Singapore, South Korea,
Thailand,
Vietnam
Europe & Central
Asia
Albania, Armenia, Belgium,
Bosnia and Herzegovina, France,
Georgia, Greece, Italy,
Macedonia, Moldova,
Montenegro, Portugal, Serbia,
Spain, Ukraine, United Kingdom,
Uzbekistan
Austria, Azerbaijan, Belarus, Bulgaria,
Czech Republic, Croatia, Denmark,
Estonia, Finland, Germany, Hungary,
Kazakhstan, Kosovo, Kyrgyzstan,
Iceland, Latvia, Lithuania, Luxembourg,
Netherlands, Norway, Poland, Romania,
Russia, Slovakia, Slovenia, Sweden,
Switzerland, Tajikistan, Turkey,
Turkmenistan
Latin America &
Caribbean
Argentina, Bahamas, Belize,
Bolivia, Brazil, Chile, Colombia,
Costa Rica, Dominican Republic,
Ecuador, El Salvador, Guatemala,
Haiti, Honduras, Jamaica,
Mexico, Nicaragua, Panama,
Paraguay, Peru, Suriname,
Trinidad and Tobago, Uruguay,
Cayman Islands
Guyana
Middle East &
North Africa
Algeria, Bahrain, Djibouti, Egypt,
Iran, Iraq, Jordan, Libya,
Palestine, Tunisia
Israel, Kuwait, Malta, Saudi Arabia,
United Arab Emirates, Oman, Qatar
North America Canada, United States
South Asia Bangladesh, Nepal, India,
Pakistan, Sri Lanka
Sub-Saharan
Africa
Angola, Benin, Botswana,
Burkina Faso, Burundi,
Cameroon, Cape Verde, Central
African Republic,
Chad, Comoros, Congo, Ethiopia,
Equatorial 1Guinea, Gabon,
Gambia, Ghana, Guinea, Guinea
Bissau, Ivory Coast, Kenya,
Lesotho, Liberia, Madagascar,
Malawi, Mauritania, Mauritius,
Mozambique, Namibia, Niger,
Nigeria, Republic of the Congo,
Senegal, Sierra Leone, Rwanda,
Seychelles, South Africa, Sudan,
Swaziland, Tanzania, Togo,
Uganda, Zambia, Zimbabwe
Central African Republic, Liberia,
Mauritania, Niger, Senegal, Seychelles

25
Table 2: Descriptive Statistics by Cluster, GDP per capita
Cluster # Mean Std. dev. Min. Max # obs.
1 13829.40 13347.69 705.00 63670.00 95
2 36308.82 26802.29 838.00 115683.00 60
3 11560.00 10189.41 4355.00 18765.00 2
Sample 22067,75 22162,15 705.00 115683.00 163
Table 3: Descriptive Statistics by Cluster, Government debt as a percentage of GDP
1 71.38 40.39 14.60 264.00 95
2 46.80 24.67 1.90 160.00 60
3 47.35 14.92 36.8 57.9 2
Sample 62,32 37.89 1,9 264 163
Table 4: Descriptive Statistics by Cluster, Current account balance as a percentage of GDP
1 -3.15 6.56 -23.8 19.2 95
2 0.42 12.36 -26.8 30.5 60
3 -21.75 7.28 -26.9 -16.6 2
Sample -2.21 10.17 -33.80 30.50 163
Table 5: Government budget balance as a percentage of GDP
1 -4.28 5.07 -35.00 6.50 95
2 -2.41 4.35 -19.30 11.60 60
3 -10.80 4.95 -14.30 -7.30 2
Sample -3.62 4.88 -35.00 11.60 163
It turns out that the entities assigned to Cluster 1, on average, have a lower GDP per capita and
higher ratios of government debt and government budget deficits to GDP than those gathered in
Cluster 2. In both of these cases, a standard two-sample t-test leads to a rejection of the null
hypothesis of equal means on a confidence level exceeding 99%. Moreover, the average
government budget balance and the average current account balance, both expressed as a
percentage of GDP, are lower in Cluster 1 than in Cluster 2. In both of these cases, the absolute
values of the associated t-statistics exceed the 95% critical value for a two-sided test by far.
Clusters 1 and 2 also exhibit considerable differences between the correlation patterns prevailing
between the continuous variables involved, only the most striking of which are mentioned in the
following:
 In Cluster 1, the sample correlation coefficient between the GDP per capita and the
government budget balance per unit of GDP exceeds 0.5 and is statistically significant on
a 95% level. In Cluster 2, the same coefficient is below 0.06 and statistically
insignificant.

26
Figure 4:
 In Cluster 1, the correlation coefficient between GDP per capita and the government’s
budget balance takes a negative value (-0.20), whereas in Cluster 2, the corresponding
coefficient has the opposite sign (0.297). A t-test derived a corresponding bivariate linear
regression in which the slope parameters were allowed to differ in Cluster 1 and Cluster 2
resulted in the null hypothesis of identical parameter values being rejected on a 99%
confidence level.
Figure 5:
The above results clearly indicate that, in this particular case, the proposed clustering procedure
succeeds in allowing the identification of subgroups of entities with statistically significant
distributional characteristics that might not have been detectable with other tools of exploratory
data analysis, such as histograms and scatterplots.
4.3. Clusters of Credit Card Applicants
Another dataset to which the procedure from sections 2 and 3 can be applied is the sample of
applicants for a specific type of credit card provided in the online complements of Greene’s [8]
econometrics textbook. The dataset consists of 1,319 observations on 12 variables. The variables
used in the current example are listed in Table 4.2.1.

27
Table 6: Variables Used in the Credit Card Example
Name Type Description
card Discrete, unordered = 1 if application was accepted, = 0 otherwise
reports Treated as continuous Number of major derogatory reports
age Continuous Age of the applicants in years
income Continuous Annual income in USD 10 000
owner Discrete, unordered = 1 if applicants own their home, = 0 otherwise
selfemp Discrete, unordered = 1 if applicants are self-employed, = 0 otherwise
dependents Treated as continuous Number of dependents
In this case, the application of the proposed procedure with a minimum cluster size of 2 leads to
the formation of two large clusters without any outliers. Cluster 1, the larger of these two, number
1, comprises little more than two-thirds of the sample. Cluster-specific descriptive statistics for
those that either are continuous or treated as such are given in Tables 7 to 10 below.
Table 7: Descriptive Statistics by Cluster, Variable “reports”
1 0.6212 1.5917 0.0000 14.0000 887
2 0.1180 0.3942 0.0000 3.0000 432
Sample 0.4564 1.3453 0.0000 14.0000 1319
Table 8: Descriptive Statistics by Cluster, Variable “age”
1 35.4493 10.1895 0.1667 83.5000 887
2 28.6215 8.3509 0.5000 67.1667 432
Sample 33.2131 10.1428 0.1667 83.5000 1319
Table 9: Descriptive Statistics by Cluster, Variable “income”
1 3.6686 1.8618 0.2100 67.1667 887
2 2.7427 1.0347 1.3200 10.9999 432
Sample 3.3654 1.6939 0.2100 13.5000 1319
Table 10: Descriptive Statistics by Cluster, Variable “dependents”
1 1.3766 1.3372 0.0000 6.0000 887
2 0.2083 0.4066 0.0000 1.0000 432
Sample 0.9939 1.2477 0.0000 6.0000 1319
The outcome of a two-sample t-test indicates that, for each of the four variables named above, the
cluster-specific differences in the means are statistically significant at a 99% confidence level.
Below is information on cluster-specific frequency distributions for the three binary indicator
variables in use. Here, too, the absolute values of the t-statistics relating to the differences
between the two clusters exceed the critical values for the 99% confidence interval by far.

28
Table 11:
Cluster # application accepted % self-employed % homeowners
1 66.63% 10.26% 65.50%
2 100.00% 0.00% 0.00%
Sample 77.56% 6.90% 44.04%
In most of the above cases, the cluster-specific pairwise correlation patterns between the four
numerical variables involved (“reports”, “age”, “income”, and “dependents”) do not show any
pronounced differences. An exception, however, is the relationship between the income variable
and the number of dependents, which is positive and statistically significant on a 95% confidence
level in the case of Cluster 1 but close to zero in Cluster 2. Hence, the supposition that the
proposed clustering method can be applied to identifying subgroups of entities with significantly
different statistical characteristics is also confirmed by the results obtained in this case.
4.4. Precautionary Remarks
Given the obvious suitability of the proposed approach in the context of the above examples, it
appears necessary to mention a number of problems that cannot be resolved by its application:
 The performance of the algorithm presented and the characteristics of the outcomes
obtained are very sensitive to the choice of the minimum cluster size and the number of
neighbouring entities from which the adjacency set of each entity is derived. If the latter
is chosen too small, the proposed method may result in the formation of "degenerate"
clusters with only very few elements.
 Many datasets to which this method can, in principle, be applied may contain one or
more “irrelevant” variables that do not contain any information on the basis of which
objects can be meaningfully divided into groups of interconnected elements.
 Especially in data sets with many variables, strong dependency relationships between
individual attributes, or subsets thereof, can prevent the application of distance-based
grouping procedures of the kind described here.
 The problem persists that notions like “distance” or “neighbourhood” become less
significant as the dimension of a dataset increases. The method proposed in this paper
may help mitigate this under favourable conditions, but it is by no means a complete
solution.
 The performance of the proposed algorithm is very sensitive to the structure and
distribution of the data to which it is applied. Hence, it remains an open question to what
extent the proposed algorithm can be generalized across different datasets.
5. CONCLUSIONS
In this paper, a distance-based clustering procedure for data of mixed type has been proposed.
The feasibility of the proposed method and its ability to produce empirically plausible results
were demonstrated using some application examples of different nature and complexity.
However, the curse of dimensionality and the possible presence of irrelevant or strongly
interrelated (groups of) variables remain issues that can lead to great difficulties for such
applications. Hence, augmenting the proposed technique with feature selection and/or

29
dimensionality reduction techniques that may help mitigate this problem is a promising area for
future research.
ACKNOWLEDGEMENTS
The authors are grateful for several helpful comments and suggestions by two anonymous
referees.
The project underlying this publication was funded by the German Federal Ministry of Economic
Affairs and Climate Action under project funding reference number 01MK21002G.
REFERENCES
[1] Aggarwal, C.C., Hinneburg, A., & Keim, D. A. (2001). On the Surprising Behavior of Distance
Metrics in High Dimensional Space. In J. van den Bussche, & V. Vianu (Eds.), Database Theory —
ICDT 2001. Berlin (Springer).
[2] Ankerst, M., Breunig, M. M., Kriegel, H.-P., & Sander, J. (1999). OPTICS: Ordering Points To
Identify the Clustering Structure. In: ACM SIGMOD Record, 28(2), 49-60.
https://doi.org/10.1145/304181.304187
[3] Bellman, R. E. (1961). Adaptive Control Processes: a Guided Tour. Princeton (University Press).
[4] Briguglio, L., Cordina, G., Farrugia, N. & Vella, S. (2008). Economic vulnerability and resilience
concepts and measurements. Helsinki (United Nations University).
[5] Ester, M., Kriegel, H.P., Sander, J., & Xu, X. (1996). A density-based algorithm for discovering
clusters in large spatial databases with noise. In E. Simoudis, J. Han & U.M. Fayyad (Eds.),
Proceedings of the Second International Conference on Knowledge Discovery and Data Mining
(KDD-96), Washington, D.C. (AAAI Press).
[6] Gionis, A., Mannila, H., & Tsaparas, P. (2007). Clustering aggregation. ACM Transactions on
Knowledge Discovery from Data, 1(1), 1-30. https://doi.org/10.1145/1217299.1217303
[7] Gower, J. C. (1971). A general coefficient of similarity and some of its properties. Biometrics, 27,
623-637. https://doi.org/10.2307/2528823
[8] Greene, W.H. (2003). Econometric Analysis. Upper Saddle River, NJ (Prentice Hall).
[9] Hamming, R.W. (1950). Error detecting and error correcting codes. Bell System Technical Journal,
29 (2), 147–160. https://doi.org/10.1002/j.1538-7305.1950.tb00463.x
[10] Houle, M. E., Kriegel, H. P., Kroger, P., Schubert, E., and Zimek, A. (2010). Can shared-neighbor
distances defeat the curse of dimensionality? In: M. Gertz, & B. Ludäscher (Eds.), Scientific and
Statistical Database Management (SSDBM 2010). Lecture Notes in Computer Science, vol 6187.
Berlin and Heidelberg (Springer).
[11] Jain, A. and Law, M. (2005). Data Clustering: A User’s Dilemma. In S.K. Pal, S. Bandyopadhyay &
S. Biswas (Eds.), Pattern Recognition and Machine Intelligence. (Lecture Notes in Computer
Science, vol 3776), Berlin (Springer). https://doi.org/10.1007/11590316_1
[12] Khatun, M., & Siddiqui, S. (2023). Estimating Conditional Event Probabilities with Mixed
Regressors: a Weighted Nearest Neighbour Approach. Statistika 103(2), 226-234.
[13] Kriegel, H.-P., Kröger, P. Sander, J., & Zimek, A. (2011). Density-based Clustering. WIREs Data
Mining and Knowledge Discovery, 1 (3), 231–240. https://doi.org/10.1002/widm.30
[14] Rezaei, M., & Fränti, P. (2016). Set-matching measures for external cluster validity. IEEE Trans. on
Knowledge and Data Engineering 28 (8), 2173-2186. https://doi.org/10.1109/TKDE.2016.2551240
[15] Oduntan, O. I. (2020). Blending Multiple Algorithmic Granular Components: A Recipe for
Clustering. Department of Computer Science, The University of Manitoba, Winnipeg, Canada.
[16] Stanfill, C., & Waltz, D. (1986). Toward memory-based reasoning. Commun. ACM, 29(12),1213–
1228. https://doi.org/10.1145/7902.7906
[17] Trading Economics (2023): Indicators. Retrieved July 1st, 2023 from
https://tradingeconomics.com/indicators .
[18] Veenman, C.J., Reinders, M.J.T., & Backer, E. (2002). A maximum variance cluster algorithm.
IEEE Trans. Pattern Analysis and Machine Intelligence, 24(9), 1273-1280.
https://doi.org/10.1109/TPAMI.2002.1033218

30
[19] World Bank (2023). World Bank Units. Retrieved July 1, 2023 from
https://www.worldbank.org/en/about/unit
[20] Yang, X.-S. (2019). Introduction to Algorithms for Data Mining and Machine Learning. Amsterdam
(Elsevier).

Connectivity-Based Clustering for Mixed Discrete and Continuous Data

Recommended

More Related Content

Similar to Connectivity-Based Clustering for Mixed Discrete and Continuous Data (20)

Recently uploaded (20)

Connectivity-Based Clustering for Mixed Discrete and Continuous Data