Page 99 - ITU Journal - ICT Discoveries - Volume 1, No. 2, December 2018 - Second special issue on Data for Good

ITU JOURNAL: ICT Discoveries, Vol. 1(2), December 2018

The percent error (%Error) is computed as: modeling are illustrated in Table 2. Each sample is
divided into three parts of 70% training, 15% for
௉ିଵ
ͳͲͲ ேିଵ ห݀ݕ െ݀݀ ห validation and 15% for testing. The data set analysis
Ψܧݎݎ݋ݎ ൌ ෍෍ ௜௝ ௜௝ ሺͳͻሻ was carried out by NeuroSolutions software
ܰܲ ௜ୀ଴ ݀݀ ௜௝
௝ୀ଴ version 7.1 and Eviews software.
where, P is the number of output processing Table 4 shows the descriptive statistics of 3G
elements, N is the number of exemplars in the data downlink traffic under five classifications of hourly,
set, ݕ is the network output for exemplar i at daily, weekly, monthly and quarterly. The weekly
௜௝
processing element j, ݀ is the desired output for and quarterly data had negative skewness while
௜௝
exemplar i at processing element j, ݀ݕ is the hourly, daily and monthly data produced positive
௜௝
denormalised network output for exemplar i at skewness.
processing element j, ݀݀ ௜௝ is the denormalised
desired output for exemplar i at processing element. 3.2 CANFIS network training

3. ANALYSIS AND RESULTS Table 4 illustrates the training results of the CANFIS
network using the Momentum learning algorithm
3.1 Data collection and specification and an Axon transfer function. 7KH GDWD ZLWK
DUFKLWHFWXUH ZDV VHOHFWHG DV WKH
The data used for this study is obtained from a DOORFDWLRQ ZLWK PLQLPXP SHUFHQW HUURU RI DQG
mobile network operator in Ghana with a 06( RI
nationwide coverage of three hundred and eighty
(380) active NodeBs connected to three radio In Table 5, five different architectures of CANFIS
network controllers (RNCs). The network operator were analysed with varying learning rule and
recorded an estimated figure of 3,889,821 data transfer functions.
subscribers according to 2017 statistics [24].
The best network with minimum values of MSE,
The experimental 3G downlink traffic data was NRMSE and percent error of 0.000406, 0.0112 and
collected from 2015 to 2017. It consists of 957 12.33% respectively using Table 5 is the CANFIS
samples of hourly traffic, 707 samples of daily traffic, (5-2-5) model with an Axon transfer function and
178 samples of weekly traffic, 72 samples of Momentum learning rule.
monthly traffic and 15 samples of quarterly traffic.
The different samples of the 3G downlink traffic for

Table 2 – Training, testing and validation sample of 3G downlink traffic

Data Hourly Daily Weekly Monthly Quarterly
(957 samples) (707 samples) (178 samples) (72 samples) (15 samples)

Training 671 495 126 50 9
Validation 143 106 26 11 3
Testing 143 106 26 11 3

Table 3 – Descriptive statistics of 3G downlink data traffic
Statistic 957 samples 707 samples 178 samples 72 samples 15 samples
(Hourly) (Daily) (Weekly) (Monthly) (Quarterly)
Mean 413531.6 553128.8 1362434.0 3114591.0 13919290
Median 409333.4 431660.2 1619248.0 2383673.0 14087025
Maximum 1041087.0 3391268.0 2709330.0 10691053 20920417
Minimum 78563.67 205131.2 437722.6 270005.7 6388322.0
Std. Deviation 186704.9 329937.4 553089.7 2839062 6582633.0
Skewness 0.466702 1.819760 -0.105359 0.794602 -0.019465
Kurtosis 3.018888 10.54329 2.168778 2.295719 1.100051
Jarque-Bera 34.75510 2066.423 5.453709 9.064749 2.257075

© International Telecommunication Union, 2018 77