Feature selection for continuous attributes in physical-activity clasification tasks

Abstract

Mobile devices include many sensors capable of sending data used in decision making activities;
an example is the classification of physical activity based on the use of accelerometers and
gyroscopes. Sensor’s signals could be processed previously applying different techniques that
extract many attributes that would be used in the development of classification tasks. The
optimization of the classification system required the reduction of the number characteristics in
order to synthesize the set dimension and the learning time. This article uses metrics of
information gain based on continuous attributes, these metrics reduced the uncertainty and
extracted only the most significant across the processed information. The analysis of the results
obtained in the classification of physical activity using neural networks showed not only a
decrement on the number of characteristics, but also, an error less than 5 % and reducing
processing time in approximately 55%.

Full text

Revista Iberoamericana de las Ciencias Computacionales e Informática ISSN: 2007-9915
Vol. 5, Núm. 10 Julio - Diciembre 2016 RECI
Selección de características para atributos continuos en tareas
de clasificación de actividad física
Feature selection for continuous attributes in physical activity classification tasks
Seleção de recursos para atributos contínuos em tarefas de classificação de
atividade física
Enrique V. Carrera
Universidad de las Fuerzas Armadas ESPE
evcarrera@espe.edu.ec
Jefferson Stalin Rodríguez Páramo
Universidad de las Fuerzas Armadas ESPE
jsrodriguez2@espe.edu.ec
Resumen
Los dispositivos móviles contienen diversos sensores con capacidad para enviar datos que se
utilizan en la toma de decisiones, un ejemplo es la clasificación de actividad física basada en el
uso de acelerómetros y giroscopios. Las señales de los sensores se procesaron previamente
aplicando diferentes técnicas que extrajeron un sinnúmero de atributos, los cuales sirvieron para
el desarrollo de tareas de clasificación. La optimización de sistemas de clasificación requirió la
disminución del número de características de entrada con la finalidad de sintetizar la dimensión
de su conjunto y tiempo de aprendizaje. Este artículo empleó métricas de ganancia de
información para atributos continuos, que redujeron la incertidumbre y extrajeron únicamente
aquellas características más significativas a través de los datos procesados. El análisis de los
resultados que se obtuvieron en la clasificación de actividad física usando redes neuronales,
mostraron no solamente la disminución de características, sino también un error por debajo del 5
% y la reducción del tiempo de procesamiento en aproximadamente 55 %.

Revista Iberoamericana de las Ciencias Computacionales e Informática ISSN: 2007-9915
Vol. 5, Núm. 10 Julio - Diciembre 2016 RECI
Palabras clave: aprendizaje de máquina, actividad física, selección de características, atributos
continuos, ganancia de información.
Abstract
Mobile devices include many sensors capable of sending data used in decision making activities;
an example is the classification of physical activity based on the use of accelerometers and
gyroscopes. Sensor’s signals could be processed previously applying different techniques that
extract many attributes that would be used in the development of classification tasks. The
optimization of the classification system required the reduction of the number characteristics in
order to synthesize the set dimension and the learning time. This article uses metrics of
information gain based on continuous attributes, these metrics reduced the uncertainty and
extracted only the most significant across the processed information. The analysis of the results
obtained in the classification of physical activity using neural networks showed not only a
decrement on the number of characteristics, but also, an error less than 5 % and reducing
processing time in approximately 55 %.
Key words: machine learning, physical activity, feature selection, continuous attributes,
information gain.
Resumo
Dispositivos móveis contêm vários sensores capazes de enviar os dados utilizados na tomada de
decisões, um exemplo é a classificação de actividade física baseada na utilização de
acelerómetros e giroscópios. Os sinais dos sensores são processados através da aplicação de
diferentes técnicas extraídos inúmeros atributos, que serviram para o desenvolvimento de tarefas
de classificação. A optimização do sistema de classificação necessária a redução do número de
características de entrada, a fim de sintetizar a dimensão de tempo em conjunto e aprendizagem.
Este artigo usou métricas de ganho informações para atributos contínuos, o que reduziu a
incerteza e extraídas apenas as características mais significativas através dos dados processados.
A análise dos resultados obtidos na classificação de actividade física utilizando redes neurais, não

Revista Iberoamericana de las Ciencias Computacionales e Informática ISSN: 2007-9915
Vol. 5, Núm. 10 Julio - Diciembre 2016 RECI
só mostraram diminuição características, mas também um erro inferior a 5% e o tempo de
processamento reduzido em cerca de 55%.
Palavras-chave: aprendizagem de máquina, atividade física, seleção de características,
atributos contínuos, ganho de informação.
Fecha recepción: Marzo 2016 Fecha aceptación: Junio 2016
Introduction
Mobile devices contain various sensors that are currently used in various fields and countless
applications around the world (Das, Green, Perez, and Murphy, 2010). The "smart" mobile
devices due to their small size, ability to send and receive data and computing power, can store
information that can be manipulated (Kwapisz, Weiss, and Moore, 2011).
All information collected by these electronic devices provides a significant contribution to the
development and monitoring, aspects related to health care, rehabilitation, disease diagnosis,
safety of people, among others (Mitchell, Monaghan, and O'Connor, 2013 ).
The signals that are emitted by the sensors can not be classified with standard algorithms, so in
the first instance must be transformed state data pure information, the processing is simpler
function of time or frequency (Weiss and Hirsh, 1998). Thus, processing is achieved and
extracted a number of features based on different metrics.
The large amount of data from existing input causes the processing time increases (Han, Kamber,
and Pei, 2011), which causes the classification system optimization reducing these demands. This
requires using an algorithm that allows the selection of features, so that the dimension of time
together and learning (Yang and Wang, 2011) is synthesized.

Revista Iberoamericana de las Ciencias Computacionales e Informática ISSN: 2007-9915
Vol. 5, Núm. 10 Julio - Diciembre 2016 RECI
The classification or automatic selection of features is one of the most common tasks where
artificial neural networks have proven effective, as they perform an automatic data processing
and are based on the biological nervous system (Isasi and Galvan, 2004).
It is noteworthy that the Artificial Neural Networks, from its appearance and its rapid
development have had a significant use as a technology for data mining, it because the
technology has attributes for effective and efficient modeling of complex problems (Lu, Setiono,
and Liu, 1996).
This research is based on information processing continuous type using metrics to gain
information, which by means of an algorithm will be quantized. Accordingly, the feature
selection process for continuous attributes in classification tasks physical activity will be
possible; that is, the most important characteristics for the classification process will be
identified, by reducing uncertainty and get only those most significant.
Selected characteristics must specify the physical activity of a person (walking up or down stairs,
sitting, standing, lying down).
It is worth mentioning that the criterion of maximizing the information gain produces a bias
towards the attributes that have lots of different values, which solves this problem by using the
ratio of profit as separation criteria (Hong, 1997). This measure takes into account both the
information gain as the probabilities of the different values of the attributes; in turn, these
probabilities are collected by the separation information, which is not more than the entropy of
the data set from the values of the attributes.
The results of the classification of physical activity using neural networks, as described later, it
shows that by using: information gain, breakpoints for five groups of selection intervals and error
rate in each, he was achieved decrease the set of features (561) in 86% (78), so optimization is
perceived in the time data processing.

Revista Iberoamericana de las Ciencias Computacionales e Informática ISSN: 2007-9915
Vol. 5, Núm. 10 Julio - Diciembre 2016 RECI
The structure of this article is as follows: in Section 1 the description of materials and methods
adopted is shown, which expose experimental development for capturing data, the mathematical
description of the proposed algorithm and the process of training network. In section 2 the
experimental results and their analysis are displayed; and finally in section 3 the conclusions.
1. MATERIALS AND METHODS
Dataset and sensors
In the market there is a wide variety of mobile devices which have been developed different
operating systems such as Apple's iOS and Google's Android. database "Human Activity
Recognition Using Smartphones Data Set" (UCI HAR Dataset) Repository of Machine Learning
at the University of California, same as working with a Smartphone (Samsung Galaxy S II) was
used in this paper placed at the waist. Through its accelerometer and gyroscope embedded, linear
acceleration and angular velocity in three XYZ axes it is obtained. The experiments were
videotaped to label the data manually. The dataset is divided randomly into two groups. There,
70% (21 people) of volunteers was selected to generate the training data and 30% (9 people)
provided test data.
Selected for this database characteristics derived from the raw signals of the three axes
accelerometer and gyroscope, which in the time domain were captured at a constant speed of 50
Hertz (Hz) and were sampled with sliding windows wide fixed 2.56 seconds (s) and 50% overlap
(128 readings / window). The acceleration signal has two components: gravitational and body
movement; which they are separated and refined in acceleration and gravity of the body, using a
bandpass filter and a third-order Butterworth lowpass, both with a cutoff frequency of 20 Hz to
eliminate noise. Gravity has only low frequency components, therefore a lowpass Butterworth
cutoff frequency of 0.3 Hz filter was used. From each window feature vector is obtained by
calculating the variables of time and frequency domain.
Body acceleration and angular velocity derived function of time, for the jerk signals, and the
magnitude of these signals was calculated dimensional handling the Euclidean norm (distance
from the origin).

Revista Iberoamericana de las Ciencias Computacionales e Informática ISSN: 2007-9915
Vol. 5, Núm. 10 Julio - Diciembre 2016 RECI
a Fast Fourier Transform (FFT) it was used in some of the signs that were used to estimate
variables vector characteristics, which provided a data matrix of 10,299 samples and 561 features
in time domain and frequency (Linchman, 2013 ).
Metrics
The Database Machine Learning Repository at the University of California has 33 variables
obtained from the signals in the three-axis accelerometer and gyroscope, which were processed
with 17 metric. This gives a total of 561 features, resulting from the multiplication between
variables and metrics. Then metrics and variables corresponding tables are observed.
Table 1 contains some of the same but in different axes variables, why the variable is counted
three times for the three axes (X, Y, Z).
Table 1. Set of Variables
#
Descripción
1,2,3
Aceleración del cuerpo en los tres ejes (XYZ), en función del tiempo.
4,5,6
Aceleración de la gravedad en los tres ejes (XYZ), en función del tiempo.
7,8,9
Derivada de la aceleración del cuerpo en los tres ejes (XYZ), en función del tiempo.
10,11,12
Velocidad angular del cuerpo en los tres ejes (XYZ), en función del tiempo.
13,14,15
Derivada de la velocidad angular del cuerpo en los tres ejes (XYZ), en función del
tiempo.
16
Magnitud de la aceleración del cuerpo, en función del tiempo.
17
Magnitud de la aceleración de la gravedad, en función del tiempo.
18
Magnitud de la derivada de la aceleración del cuerpo, en función del tiempo.
19
Magnitud de la velocidad angular del cuerpo, en función del tiempo.
20
Magnitud de la derivada de la velocidad angular del cuerpo, en función del tiempo.
21,22,23
Aceleración del cuerpo en los tres ejes (XYZ), en dominio de la frecuencia.
24,25,26
Derivada de la aceleración del cuerpo en los tres ejes (XYZ), en dominio de la
frecuencia.
27,28,29
Velocidad angular del cuerpo en los tres ejes (XYZ), en dominio de la frecuencia.
30
Magnitud de la aceleración del cuerpo, en dominio de la frecuencia.
31
Magnitud de la derivada de la aceleración del cuerpo, en dominio del tiempo.
32
Magnitud de la velocidad angular del cuerpo, en dominio de la frecuencia.
33
Magnitud de la derivada de la velocidad angular del cuerpo, en dominio de la
frecuencia.

Revista Iberoamericana de las Ciencias Computacionales e Informática ISSN: 2007-9915
Vol. 5, Núm. 10 Julio - Diciembre 2016 RECI
Table 2. Set of Metrics
#
Métricas
1
Media
2
Desviación Estándar
3
Desviación Media Absoluta
4
Valor Máximo
5
Valor Mínimo
6
SMA
7
Energía
8
IQR
9
Entropía
10
Auto regresión
11
Correlación
12
Máximo Índice
13
Frecuencia Media
14
Skewness
15
Kurtosis
16
Energía de un intervalo de frecuencia
17
Ángulo entre vectores
Information Gain
He mentioned that the criterion of maximizing the information gain is based on the entropy of
information theory, ie it is a measure of uncertainty of a random variable (Roobaert, Karakoulas,
and Chawla, 2006).
To determine the information gain of this study attributes are discretized, from which the gain for
the 5 groups ranges as calculated; thereby an ordered list was generated and removed those
attributes with lower results.
The 5 groups were used: 4, 6, 8, 10 and 12 intervals; the reason for using only these groups is
because intervals show a general tendency, that is, if a greater or lesser number of intervals (less
than 4 or more than 12), the average remains the same is used, which does not change the data.
In conclusion, the expected reduction in the entropy of the data to know the value of continuous
attributes in classification tasks of physical activity was performed.

Revista Iberoamericana de las Ciencias Computacionales e Informática ISSN: 2007-9915
Vol. 5, Núm. 10 Julio - Diciembre 2016 RECI
Neural networks
Its main function is learning scheme and determines the type of problems that will be able to
solve (Isasi and Galvan, 2004). On the other hand, researchers from artificial intelligence and
statistics have been interested in the most abstract properties of neural, such as its ability to
develop distributed computing and tolerate the noise at the input of the (Cazorla, Colomina Pardo
network networks, and Old Hernando, 2011). Currently it is understood that other types of
systems (including Bayesian networks) have these properties, however, neural networks are
worthy of study because they remain as one of the most popular and effective when building
learning systems (Russell and Norving forms, 2004).
In the present work it is to know what type of neural network offers greater efficiency when meet
the proposed requirements. It was decided networks feedforward (feedforward), which contain a
series of layers: one with input connection to the network, a backing layer that has a connection
to the previous layer and a layer that produces the output of the network . These networks with
enough neurons in the hidden layer can be adapted to any problem of input-output mapping
finite. Two types of feedforward networks known and used in the mathematical tool MATLAB®
are FITNET and Patternnet.
FITNET is a feedforward network with two layers sigmoid activation function: a layer of hidden
neurons and other neuron linear output, which engage multidimensional problems with consistent
data allocation. The network will be trained with backpropagation algorithm "Levenberg-
Marquardt", and if the memory is not sufficient propagation algorithm is used back scaled
conjugate gradient (Monar, 2014).
Networks pattern recognition (Patternnet) is a feedforward network of two layers: one with a
hidden and output neurons Softmax (net pattern) with transfer function sigmoid type. This
network is trained with the algorithm backpropagation scaled conjugate gradient for sorting the
outputs according to the inputs, and the target data should consist of vectors of all zero values
except for a 1 in the (i) which it is the class to represent (Monar, 2014).

Revista Iberoamericana de las Ciencias Computacionales e Informática ISSN: 2007-9915
Vol. 5, Núm. 10 Julio - Diciembre 2016 RECI
RESULTS AND ANALYSIS
In order to select the neural network that offers improved efficiency relative to processing time
and error rate, three training trials were performed with 1, 70 and 561 features respectively, as
shown in Table 3.
Table 3. Training neural networks.
Red
Neuronal
Número de
Características
Número de
Muestras
Tiempo de
Procesamiento (s)
Porcentaje de
Error Train (%)
Porcentaje de
Error Test (%)
Fitnet
1
7 352
6.296
15.821
15.845
70
7 352
97.340
1.352
4.515
561
7 352
325.761
0.4528
1.8709
Patternnet
1
7 352
9.647
29.491
29.583
70
7 352
44.226
1.990
5.417
561
7 352
112.262
0.701
2.322
Referring to the results of the training of each network, it is observed that the Patternnet network
has less time for data processing and a percentage of considerable error compared to the FITNET
network, allowing the status of physical activity a or more people simultaneously and in real
time; those reasons support the use of this network.
Subsequently gain characteristic information that each dividing the set of training data into 5
groups containing different numbers of intervals is calculated. These were structured as follows:
the first group divided into 4 intervals, the second in 6 intervals, the third in 8 intervals, the fourth
in 10 intervals and the fifth in 12 intervals. This was done because the data are continuous
attributes (Cao, Ma, Liu and Guo, 2012).
Features accelerometer and gyroscope
The features were ordered from one that provides as much information up to the least contributes,
as shown in Figure 1.

Revista Iberoamericana de las Ciencias Computacionales e Informática ISSN: 2007-9915
Vol. 5, Núm. 10 Julio - Diciembre 2016 RECI
Figure 1.- Gain information for 5 groups of intervals (561 features).
When calculating the derivative gain information you can appreciate the break points in each
group of intervals, for example, about 200 features is clearly the fall of the curve from this the
mean (average) is obtained, which later serve to set the number of features and the respective
error mark both training (train) and testing (test), which can be seen in table 4.
Table 4. Selecting break points.
INTERVALOS
INTERVALOS
4
6
8
10
12
Promedio
4
6
8
10
12
Promedio
NÚMERO DE CARACTERÍSTICAS
280
276
278
280
278
278
133
131
139
129
133
133
267
266
263
263
254
263
127
123
125
125
126
125
261
259
259
254
253
257
103
102
104
101
109
104
250
248
252
247
248
249
85
94
93
86
89
89
234
234
232
231
229
232
85
83
85
81
78
82
230
229
224
227
229
228
72
71
69
72
70
71
211
214
213
215
213
213
61
63
60
64
61
62
206
205
208
204
203
205
51
52
47
50
50
50
198
201
195
192
191
195
34
43
40
38
39
39
193
188
184
185
181
186
24
25
30
25
22
25
178
179
170
179
171
175
16
14
13
16
14
15
171
169
161
164
168
167
7
7
7
7
7
7
157
150
150
152
151
152
5
4
4
4
3
4
137
142
142
140
141
140
1
1
1
1
1
1

Revista Iberoamericana de las Ciencias Computacionales e Informática ISSN: 2007-9915
Vol. 5, Núm. 10 Julio - Diciembre 2016 RECI
Through the application all the attributes identical characteristics were determined in the 5 groups
of intervals, identifying that which way are contributing more information relevant to the process
of classification of physical activity for the specific case.
Clearly the error rate for both the training set and for the test, is gradually increased while the
characteristics are reduced. Thus the number of minimum data capable of generating an error less
than 5% for the neural network classifier can learn and detect patterns that determine physical
activity being performed is chosen.
Table 5. Reduction accelerometer and gyroscope features with identical characteristics.
Número
Características
Número
Características
Idénticas
Train
(%)
Test
(%)
Número
Características
Número
Características
Idénticas
Train
(%)
Test
(%)
561
561
0.7743
2,9446
133
133
1.6391
3.6263
278
235
1.4187
3,1113
125
125
1.7344
3.6958
263
224
1.587
3,2117
104
104
1.6841
3.8639
257
221
1.4117
2,9988
89
89
6.6016
7.0473
249
218
1.5557
3,1409
82
82
6.9032
7.3374
232
213
1.4177
3,1252
71
71
7.0418
7.6495
228
211
1.5069
3,0083
62
62
7.204
7.6347
213
205
1.738
3,3246
50
50
7.3649
8.0591
205
201
1.547
3,0004
39
39
8.8933
8.7641
195
194
1.5908
3,2094
25
25
10.197
10.2485
186
185
1.3913
3,1513
15
15
12.2387
13.3877
175
174
1.4069
3,1168
7
7
13.216
14.8047
167
167
1.4222
3,0931
4
4
13.2696
14.865
152
152
1.5395
3,3428
1
1
15.2551
15.9221
140
140
1.6762
3.5242
As evidenced in Table 5, with decreasing the number of features tends to be compared with the
number of identical characteristics. Thus, if observed for the number of features 167, it has the
same number of number of identical characteristics, and these characteristics diminish as the
error tends to increase. Therefore, by making the combined calculation for the accelerometer to

Revista Iberoamericana de las Ciencias Computacionales e Informática ISSN: 2007-9915
Vol. 5, Núm. 10 Julio - Diciembre 2016 RECI
the gyroscope a joint decreased to 104 characteristics is obtained, with an error of test assembly
3.8639%, it becomes the limit of decreasing characteristics before passing 5% error.
Figure 2.- Processing time (all features).
By decreasing the number of features as shown in Figure 2, shows an efficient reduction of
processing time reaching optimal results of 58.8267 (s) 104 for characteristics.
Features Accelerometer
The database used provides features 561, 345 of which are accelerometer and gyroscope 216.
Once this data analysis classified the same selection and attribute reduction is conducted, taking
the characteristics of the accelerometer and gyroscope separately as shown in the following tables
and figures.

Revista Iberoamericana de las Ciencias Computacionales e Informática ISSN: 2007-9915
Vol. 5, Núm. 10 Julio - Diciembre 2016 RECI
Figure 3.- Gain information for 5 groups of intervals (345 features accelerometer).
Through the derivative gain information you can appreciate the break points in each group
interval. In the case of accelerometer specifically a drop characteristics curve at approximately
120, as shown in Figure 3. Table 6 is set to the average number of features set with which the
respective error calculation marking will be made is denoted both the training set and the test.

Revista Iberoamericana de las Ciencias Computacionales e Informática ISSN: 2007-9915
Vol. 5, Núm. 10 Julio - Diciembre 2016 RECI
Table 6. Selecting break points (accelerometer).
INTERVALOS
INTERVALOS
4
6
8
10
12
Promedio
4
6
8
10
12
Promedio
NÚMERO DE CARACTERÍSTICAS
195
194
200
200
206
199
88
92
86
84
85
87
187
181
182
178
184
182
78
82
80
78
74
78
166
168
168
167
169
168
73
70
70
67
67
69
152
161
159
159
163
159
64
61
61
62
64
62
146
147
149
147
147
147
49
50
47
43
46
47
144
141
144
143
144
143
38
38
40
33
34
37
125
129
132
132
128
129
22
25
30
25
24
25
121
122
121
124
122
122
14
16
13
14
16
15
103
100
99
100
98
100
7
7
7
7
7
7
91
94
95
90
96
93
1
1
1
1
1
1
From points calculated break, shown in Table 7 that adequate decrease before exceeding 5%
error, corresponds to 78 characteristics, demonstrating that independently a greater decrease in
characteristics is obtained in the case of the accelerometer . It should be noted that applies the
same calculation and analysis for the case of the gyroscope.
Table 7. Accelerometer reduction characteristics with identical characteristics.
Número
Características
Número
Características
Idénticas
Train
(%)
Test
(%)
Número
Características
Número
Características
Idénticas
Train
(%)
Test
(%)
345
345
0.7721
3.0793
87
87
1.3014
4.1031
199
164
1.0944
3.3315
78
78
2.9037
4.1573
182
155
1.1282
3.244
69
69
6.8814
8.3506
168
148
1.1226
3.5088
62
62
7.3665
8.47
159
145
1.1337
3.4383
47
47
8.1803
9.1972
147
140
1.279
3.3502
37
37
7.5788
9.337
143
138
1.1135
3.2756
25
25
9.6996
11.576
129
129
1.3649
3.6542
15
15
11.1481
12.1537
122
122
1.0565
3.8856
7
7
12.7406
13.3385
100
100
1.4052
4.0535
1
1
15.228
15.8485
93
93
1.1843
3.9107

Revista Iberoamericana de las Ciencias Computacionales e Informática ISSN: 2007-9915
Vol. 5, Núm. 10 Julio - Diciembre 2016 RECI
Similarly it can be seen in Figure 4 that the 78 identical characteristics are processed in 55.9645
(s). This generates a significant time reduction with respect to the processing of data from both
sensors.
Figure 4.- Processing time (accelerometer features).
Features Gyroscope
Figure 5 shows a gain of different information for the 5 groups of intervals, so that in calculating
its derivative the break points will be more apparent. In this case a drop of the curve around 80
characteristics is observed. Consecutively selection and analysis only attribute reduction
characteristics gyroscope is made.

Revista Iberoamericana de las Ciencias Computacionales e Informática ISSN: 2007-9915
Vol. 5, Núm. 10 Julio - Diciembre 2016 RECI
Figure 5.- Gain information for 5 groups of intervals (216 features gyroscope).
Table 8 shows an obvious decrease in the number of break points compared with previous results,
just as the mean number of characteristics is determined and then the error rate for the training set
and test is calculated.
Table 8. Selecting break points (gyroscope).
INTERVALOS
4
6
8
10
12
Promedio
NÚMERO DE CARACTERÍSTICAS
133
138
147
136
140
139
122
124
128
122
128
125
113
116
118
119
118
117
103
102
101
102
102
102
96
92
94
96
94
94
88
84
85
88
86
86
78
75
79
80
79
78
60
58
61
53
55
57
47
51
48
47
46
48
34
32
39
34
35
35
19
19
18
21
18
19
13
16
12
14
14
14
1
3
1
4
3
2

Revista Iberoamericana de las Ciencias Computacionales e Informática ISSN: 2007-9915
Vol. 5, Núm. 10 Julio - Diciembre 2016 RECI
The fewer points calculated break, as judged in Table 9, it shows that proper decrease before
exceeding 5% error, corresponds to 78 characteristics, demonstrating that independently
characteristics gyroscope can also be used for the classifier.
Table 9. Reduction gyroscope features with identical characteristics.
Número
Características
Número Características
Idénticas
Train (%)
Test (%)
216
216
0.758
3.1706
139
116
1.257
4.1913
125
110
1.2988
4.2884
117
109
1.299
4.2975
102
102
1.4572
4.2413
94
94
2.025
3.7185
86
86
1.9113
4.0126
78
78
2.0773
3.7174
57
57
5.8204
6.3557
48
48
6.3161
6.6827
35
35
7.4344
7.5665
19
19
9.6658
11.5729
14
14
10.1099
12.0147
2
2
16.0832
17.1085
Similarly, the graphics processing time for the characteristics expressed gyroscope efficient
reduction equal to the accelerometer values, as shown in Figure 6.

Revista Iberoamericana de las Ciencias Computacionales e Informática ISSN: 2007-9915
Vol. 5, Núm. 10 Julio - Diciembre 2016 RECI
Figure 6.- Processing Time (gyroscope features).
In the case of the accelerometer, in about 78 identical characteristics permissible error range
(3.7185%) is obtained before exceeding an error of 5% and time efficient processing of 42. 0811
seconds.
The analysis of the characteristics for each sensor checks that the selection process is more
efficient if working independently. While the mobile device has two sensors (accelerometer and
gyroscope), it is preferable that the classification process characteristics is made with one of the
two, so be achieved optimize the processing time. This result is not achieved efficiently if both
sensors work in tandem.

Revista Iberoamericana de las Ciencias Computacionales e Informática ISSN: 2007-9915
Vol. 5, Núm. 10 Julio - Diciembre 2016 RECI
Conclusion
14% of continuous attributes sampled is used, and it is determined that one can obtain a reduction
in processing time by about 55% and an error less than 5% in the selection process characteristics
without affecting the classification of activity physical.
In the course of the investigation it was established by reducing features for continuous attributes,
which can meet the physical activity and / or status of a person (walking, jumping, running, etc.)
so more efficiently by Gyroscope or accelerometer independently.
This article serves as a foundation for future work with the approach of other methods. One can
cite as examples the improvement information gain algorithm through the introduction of the
dependence of attributes. So, you get to collect data not only from one person, but several
simultaneously and processing information in real time.

Revista Iberoamericana de las Ciencias Computacionales e Informática ISSN: 2007-9915
Vol. 5, Núm. 10 Julio - Diciembre 2016 RECI
Bibliography
Cao, D., Ma, N., Liu, Y., & Guo, J. (2012). A Feature Selection Algorithm for Continuous
Attributes Based on the Information Entropy. Journal of Computational Information
Systems, 1467-1475.
Cazorla, M., Colomina Pardo, O., y Viejo Hernando, D. (19 de mayo de 2011). Presentaciones de
la asignatura Técnicas de Inteligencia Artificial (Curso 2010-2011). Obtenido de
http://hdl.handle.net/10045/17323
Das, S., Green, L., Perez, B., & Murphy, M. (30 de julio de 2010). Detecting User Activities
using the Accelerometer on Android Smartphones. TRUST REU The Team for Research
in Ubiquitous Secure Technology, 29.
Han, J., Kamber, M., & Pei, J. (2011). Data Mining: Concepts and Techniques, tercera edición,
USA: Elsevier.
Hong, S. J. (1997). Use of contextual information for feature ranking and discretization. IEEE
transactions on knowledge and data engineering, 9(5), 718-730.
Isasi, P., y Galván, I. (2004). Redes De Neuronas Artificiales. Un enfoque práctico, primera
edición, Madrid, España: Pearson.
Kwapisz, J., Weiss, G., & Moore, S. (2011). Activity recognition using cell phones
accelerometers. ACM SIGKDD Explorations Newsletter, segunda edición, vol. 12, New
York, USA.
Linchman, M. (04 de abril de 2013). UCI Machine Learning Repository. Recuperado el 10 de
noviembre de 2015, de http://archive.ics.uci.edu/ml
Lu, H., Setiono, R., & Liu, H. (1996). Effective data mining using neural networks. IEEE
transactions on knowledge and data engineering, 8(6), 957-961.
Mitchell, E., Monaghan, D., & O'Connor, N. (19 de abril de 2013). Classification of Sporting
Activities Using Smartphone Accelerometers. Sensors, 13, 16.
Monar, W. L. (octubre de 2014). Repositorio Digital - Escuela Pilitécnica Nacional. Recuperado
el 30 de julio de 2016, de http://bibdigital.epn.edu.ec/handle/15000/8711

Revista Iberoamericana de las Ciencias Computacionales e Informática ISSN: 2007-9915
Vol. 5, Núm. 10 Julio - Diciembre 2016 RECI
Roobaert, D., Karakoulas, G., & Chawla, N. (2006). Information gain, correlation and support
vector machines. Feature Extraction. Springer Berlin Heidelberg, 207, 463-470.
Russell, S., y Norving, P. (2004). Inteligencia Artificial un enfoque Moderno, segunda edición,
Madrid, España: Pearson.
Weiss, G., & Hirsh, H. (27 de agosto de 1998). Learning to Predict Rare Events in Event
Sequences. Knowledge Discovery and Data Mining.
Yang , B., & Wang, L. (2011). The Construction Method of Knowledge Discovery Theory System
Based on Cognitive. (C. a. Circuits, Ed.) Wuhan: IEEE.