See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/278403191

Numerical Coding of Nominal Data

Article · January 2015

DOI: 10.26348/znwwsi.12.53

CITATIONS READS

2 8,932

2 authors, including:

Zenon Gniazdowski
Warsaw School of Computer Science
34 PUBLICATIONS   81 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Clustering of variables in the method of PCA View project

Extension of the 1D Stoney algorithm to a 2D case View project

All content following this page was uploaded by Zenon Gniazdowski on 28 August 2016.

The user has requested enhancement of the downloaded file.

 Zeszyty Naukowe WWSI, No 12, Vol. 9, 2015, pp. 53-61

Numerical Coding of Nominal Data

Zenon Gniazdowski∗1 and Michał Grabowski1

1 Warsaw School of Computer Science

Abstract
In this paper, a novel approach for coding nominal data is proposed. For the given
nominal data, a rank in a form of complex number is assigned. The proposed
method does not lose any information about the attribute and brings other proper-
ties previously unknown. The approach based on these knew properties can been
used for classification. The analyzed example shows that classification with the
use of coded nominal data or both numerical as well as coded nominal data is
more effective than the classification, which uses only numerical data.

Keywords — nominal data, numerical data, classification

1 Introduction
Different types of data are used in data analysis. Generally, they can be numerical data or
nominal data. Numerical data are linearly ordered, which leads to the conclusion that two
elements are equal or one element precedes the second one. Nominal data cannot be naturally
ordered. In the set of nominal data, the identity equivalence relation can be defined, at most. It
means that two elements may be equal or different.
For both types of data specific methods of analysis are developed. Particular difficulties
arise when continuous and nominal data are analyzed simultaneously. Usually, by discretization
continuous data are treated as nominal data. In this way, there avoids the opportunity of setting
the data in order. On the other hand, the procedure can be reversed. In this case, nominal data
are coded with the use of numbers [1]. Unfortunately, numerically coded nominal data cannot
be naturally ordered.
In this paper, a novel approach for coding nominal data with the use of complex numbers
will be presented [2]. For the given nominal data, it will be assigned a rank in a form of number.
Proposed approach can be employed for classification and clustering.
∗
E-mail: [email protected]

Manuscript received April 11, 2015 53

 Zenon Gniazdowski, Michał Grabowski

Table 1: Ranks of numerical data

No. Sorted data Assigned rank
1 21 1
2 28 2
3 33 3
4 44 4
5 45 5
6 54 6
7 55 7
8 60 8
9 63 9
10 76 10

2 Preliminaries – ranks of numerical data

In statistics, in some cases, the numerical results of observations are replaced by their ranks. In
the first stage of the ranking procedure, the data set is sorted in ascending order. Next, a rank
equal to the item in the sorted set is assigned for the ordinal elements [3]. An example of this
procedure is shown in Table 1.
On the other hand, it may happen that in the sorted set there are different elements with
the same values. In this case, the ranks assigned to identical values should be the same. Such
elements receive rank that is equal to their average position in the sorted set [3]. These are
so–called tied rank. Table 2 shows an example of tied ranks.

3 Coding of nominal data

As it is noted above, attaching ranks to numerical data is a function of the value of numerical
data as well as the cardinality of its occurrences. In the case of nominal data, there is no dif-
ferentiation of values. Possible method of ranking of nominal data has no chance to utilize the
values. On the other hand, intuition suggests that in random community, both for numerical
and nominal data, more numerous elements are more important than elements of less cardinal-
ity. Therefore, there is a proposal to rank of nominal data using only cardinality of identical
elements.

3.1 Different frequencies of different nominal values

Nominal data cannot be sorted according to their values, but it can be grouped according to
identical values. In the n–element subset consisting of identical elements, these elements may
be numbered from 1 to n. For each of them can be assigned a rank that is equal to the average

54
 Numerical Coding of Nominal Data

Table 2: Tied ranks of numerical data

No. Sorted data Assigned rank
1 21 1
2 28 2
3 44 4
4 44 4
5 44 4
6 54 6
7 55 8.5
8 55 8.5
9 55 8.5
10 55 8.5

Table 3: Rank of nominal data – differences in the cardinality of elements

First value Second value Third value
Value Position Rank Value Position Rank Value Position Rank
a 1 3.5 b 1 3 c 1 2.5
a 2 3.5 b 2 3 c 2 2.5
a 3 3.5 b 3 3 c 3 2.5
a 4 3.5 b 4 3 c 4 2.5
a 5 3.5 b 5 3
a 6 3.5

value of these numbers:

n+1
R= (1)
2
More numerous elements will have higher rank than less numerous elements. As an exam-
ple, a set consisting of 15 elements with nominal values {a, a, a, a, a, a, b, b, b, b, b, c, c, c, c} is
considered. The set can be divided into three subsets. Each subset is a class of equivalence,
which contains identical elements. To each nominal value, respective rank was assigned, ac-
cording to the formula (1). The result of the ranking is shown in Table 3.

3.2 The same frequencies of different nominal values

For equinumerous subsets, method (1) gives the same rank. This would lead to the situation
in which equinumerous elements will be indistinguishable. Therefore, method (1) should be
modified. If the variable has several equinumerous subsets, the nominal elements belonging to

55
 Zenon Gniazdowski, Michał Grabowski

Table 4: Ranks of nominal data - possible equinumerosity of elements in subsets

Nominal Module Phase Exponential Algebraic
No. variable |R| φ[rad] form Reiφ form a + bi
1 a 2 0 2 2
2 a 2 0 2 2
3 a 2 0 2 2
4 b 2 2π/3 2ei2π/3 −1 + 1.73i
5 b 2 2π/3 2ei2π/3 −1 + 1.73i
6 b 2 2π/3 2ei2π/3 −1 + 1.73i
7 c 2 4π/3 2ei4π/3 −1 − 1.73i
8 c 2 4π/3 2ei4π/3 −1 − 1.73i
9 c 2 4π/3 2ei4π/3 −1 − 1.73i
10 d 3.5 0 3.5 3.5
11 d 3.5 0 3.5 3.5
12 d 3.5 0 3.5 3.5
13 d 3.5 0 3.5 3.5
14 d 3.5 0 3.5 3.5
15 d 3.5 0 3.5 3.5

the j − th subset (j = 0, 1, . . . , k − 1) can be coded with the use of k successive roots of unity:
√
Rj = R · k −1 = R · eiφ = R · (cos φ + i sin φ) (2)
√
In the above expression i = −1, φ = 2πj/k (j = 0, 1, . . . , k − 1) and R is the rank
calculated by the formula (1). Value of φ is the phase assigned to the successive (j − th)
nominal value. In the presented concept R is a module of complex rank, depending on the car-
dinality of the subset that contains given nominal value. This approach gives the same modules
R for equinumerous subsets contained identical nominal elements, and distinguishes ranks of
different nominal values via different phases.
Table 4 shows an example of the ranking for the case when the cardinality of elements a, b
and c are equal to three, and the cardinality of the element d is equal to six. For nominal values
of a, b and c assigned phases are respectively equal to 0, 2π/3 and 4π/3. Hence, the rank
assigned to the value of a is real, and ranks assigned to nominal values of b and c are complex.
Real rank is assigned to the nominal value of d.

4 Properties of complex coding

In nominal data set, an equivalence relation can be defined for a data set in column correspond-
ing to the given attribute. This relation divides this column into classes of equivalence. Each

56
 Numerical Coding of Nominal Data

class will contain identical elements. The cardinality of each class is the only attribute informa-
tion that is important from our analysis point of view. Coding with the use of complex numbers
is unambiguous, i. e. after coding different elements are still distinguishable. In addition, it is
also possible to define the corresponding equivalence relation, which divides the set into classes
of equivalences, with cardinality of each class as before coding.
Coding does not lose any information about the attribute. The coded data receives addi-
tional properties that enrich them. Before coding, the cardinality of the given value was as the
external feature. Now, through the module, the cardinality is an inherent property of the coded
value of the attribute. The module presents information about the statistical strength of a given
subset of elements. The phase contains the information about the number of equinumerous
classes. Additionally, coding with the use of complex numbers brings other properties previ-
ously unknown. Above all, on complex numbers all arithmetic operations can be performed.
Objects in data space can be viewed as vectors in a complex space. In this space, a scalar prod-
uct, norm, as well as metric can be defined [4]. Scalar product of two complex vectors x and y
is defined as follows:
Xn
(x, y) = xi y i (3)
i=1

This way the norm, which is generated by the above scalar product, can also be defined:
p
||x|| = (x, x) (4)

By using this norm, metric also can be defined:

ρ(x, y) = ||y − x|| (5)

Proposed methodology of coding can be employed for analysis of nominal data. In par-
ticular, in a natural way it may be used for clustering and classification, because of the metric
defined above.

5 Possible application of proposed approach – an example

The proposed approach can be used in k-means method, since the distance between objects
in data space can be calculated by taking into account complex ranks of data. Data from the
company that sells cars will be shown as an example [5]. The data set consists of ten objects
(Table 5). Each object is described by means of the seven attributes. The first two attributes
(number of doors, engine power) are numbers. The other five attributes (color, fuel, interior,
wheels and brand) take the nominal values. The first four of them were coded using the proposed
approach (Table 6).
The set of attributes will be divided into two subsets. The first six attributes are conditional
attributes. The last one is an attribute of decision-making. Based on conditional attributes the
set of objects will be clustered into three subsets. Afterwards, it is needed to check to what
extent these subsets are consistent with the decision attribute. In other words, it is necessary to
check whether the cars brand can be recognized from the car's description. In order to verify the

57
 Zenon Gniazdowski, Michał Grabowski

Table 5: Description of cars

No. Door Power Color Fuel Interior Wheel Brand
1 2 60 Blue Petrol Fabric Steel Opel
2 2 100 Black Diesel Fabric Steel Nissan
3 2 200 Black Petrol Leather Alloy Ferrari
4 2 200 Red Petrol Leather Alloy Ferrari
5 2 200 Red Petrol Fabric Steel Opel
6 3 100 Red Diesel Leather Steel Opel
7 3 100 Red LPG Fabric Steel Opel
8 3 200 Black Petrol Leather Alloy Ferrari
9 4 100 Blue LPG Fabric Steel Nissan
10 4 100 Blue Diesel Fabric Alloy Nissan

Table 6: Coded description of cars

No. Door Power Color Fuel Interior Wheel Brand
1 2 60 2 3 3.5 3.5 Opel
2 2 100 -2 2 3.5 3.5 Nissan
3 2 200 -2 3 2.5 2.5 Ferrari
4 2 200 2.5 3 2.5 2.5 Ferrari
5 2 200 2.5 3 3.5 3.5 Opel
6 3 100 2.5 2 2.5 3.5 Opel
7 3 100 2.5 1.5 3.5 3.5 Opel
8 3 200 -2 3 2.5 2.5 Ferrari
9 4 100 2 1.5 3.5 3.5 Nissan
10 4 100 2 2 3.5 2.5 Nissan

58
 Numerical Coding of Nominal Data

Table 7: The results of classification

Accuracy of classification
Considered Data Type 90% 80% 70% 60% 50%
Ad hoc – – – – 20
Only Numerical Data – – 12 1 7
Only Coded Symbolic Data 1 8 3 8 –
Numerical and Coded Symbolic Data 3 4 10 2 1

usefulness of the proposed complex coding, four classifications were made based on different
conditional attributes:

• Numbers and ad-hoc coded nominal data,

• Only numbers (number of doors and engine power),
• Only nominal data (color, fuel, interior, wheel) coded by the use of (2),
• Numbers as well as nominal data coded by the use of formula (2).

K–means algorithm was used for classification. The data were standardized, for this pur-
pose. Euclidean norm was used to measure distances. For this purpose, the adequate number of
starting points for the k-means algorithm was chosen randomly. All these tests were repeated
twenty times. In none of these twenty cases, the sequence of randomly selected points was not
repeated.
After completion of the experiments, the results obtained for different conditional attributes
were compared. Table 7 shows the comparison of classification results for different types of
used data. It can be seen that the classification using coded nominal data and both numerical as
well as coded nominal data is more effective than the classification, which uses ad hoc coding
or only numerical data. Based on obtained results it must be concluded that the information that
is contained in the coded nominal data is important for classification.

6 Conclusions
In this paper, a novel approach for coding nominal data was proposed. For the given nominal
data, it can be assigned rank in a form of complex number. The module of this rank presents
information about the statistical strength of a given subset of elements. The phase contains the
information about the number of equinumerous values of attribute.
Proposed methodology is unambiguous. After coding, different values of attribute are still
distinguishable. The method does not lose any information about the attribute. Additionally,
coded data receives properties previously unknown that enrich them. Above all, on complex
numbers all arithmetic operations can be performed. In complex space, a scalar product, norm,
as well as metric can be defined. It means that coded data may be used for clustering and
classification.

59
 Zenon Gniazdowski, Michał Grabowski

The well-known, folklore-type idea represents m–valued nominal domain as equidistant

points in Rm Euclidian space. Thus the classical Euclidean approach represents vectors of
nominal values as points in Rq space, where q is equal to the sum of cardinalities of nominal
domains under consideration. The proposed coding represents vectors of nominal values as
points in C s space, where s is equal to the number of nominal domains under consideration. We
think that our coding can be considered as one of alternatives when the analyzed continuous–
nominal data are sensitive to course of dimensionality (see [6], [7]), due to low (compared to
Euclidean coding) dimension of final space of codes.
It is plain enough that the proposed coding injects additional information. For instance, the
symmetries of codes of nominal data with equal frequencies are specific for the proposed coding
schema and it may happen that the symmetries of the original data differ from the symmetries
of their codes. Nevertheless, the other coding schemas involving nominal data frequencies
information, for instance Bayesian coding of nominal values by scoring (see [8]), are subjected
to this weakness as well. Moreover, Bayesian coding by scoring is applicable only to training
data sets with two decision categories. Our proposal is fully general and can be applied to data
sets with no decision categories information at all.
The analyzed in this article data set shows, that classification with the use of coded nominal
data or both numerical as well as coded nominal data is more effective than classification, which
uses ad-hoc coding or only numerical data. From here, it must be concluded that the information
that is contained in the coded nominal data is important for classification.
The presented proposal is preliminary proposal. Although the method looks interesting,
further investigations of this approach are necessary, for instance a serious experimental study.
These investigations could confirm the usefulness of the method. They could also show other
possible applications of the proposed method.

References
[1] M. Grabowski and M. Korpusik. Metrics and similarities in modeling dependencies be-
tween continuous and nominal data. Zeszyty Naukowe WWSI, 7(10):25–37, 2013.
[2] Z. Gniazdowski. Numerical coding of nominal data. Seminar, Warsaw School of Computer
Science, May 15, 2014.
[3] F. Wilcoxon. Individual comparisons by ranking methods. Biometrics Bulletin, 1(6):80–83,
1945.
[4] S. G. Krejn. Analiza funkcjonalna. PWN, Warszawa, 1967.
[5] L. Rutkowski. Metody i techniki sztucznej inteligencji. Wydawnictwo Naukowe PWN,
Warszawa, 2012.
[6] T. Hastie, R. Tibshirani, and J. Friedman. The Elements of Statistical Learning. Springer,
New York, 2001.
[7] R. Bellman. Adaptive Control Processes. A Guided Tour. Princeton University Press,
Princeton, 1961.

60
 Numerical Coding of Nominal Data

[8] J. Koronacki and J. Ćwik. Statystyczne systemy ucza̧ce siȩ. Akademicka Oficyna
Wydawnicza EXIT, 2008.

61

View publication stats