Indonesian Interconnect Telco’s Traffic Forecasting for Anomaly Detection Using Hybrid (EEMD and BPNN) Method

Abstract

Interconnection is a connectivity between telecommunications operators that enables
customers of one operator communicate with customers of another operator. This gives a good
business opportunity to all the operators engaged in interconnection. For the Telco Provider,
Outgoing calls are the “expense” and Incoming calls are the “revenue”. Periodically, operators
make interconnection settlement to calculate the amount to be paid and received. There are two
issues related to settlement process. First, the total of transaction value is huge, approximately
IDR 5 trillion/month. Second, each operator should complete their obligations to other
operators at the agreed time; subsequent payment is not allowed. Based on them, validity of data
is an important aspect that must be ensured. Validation has purpose to detect anomaly. Anomaly
should be correctly identified because it has the potential to cause loss.
A conventional method to find outliers is by comparing time-series traffic forecasting result
against actual traffic. This study proposed a “Hybrid” model that combines data decomposition
using EEMD followed by time-series traffic forecasting using BPNN. Data decomposition is
needed to be done because of traffic data characteristics that are non linear and non stationary.
After being decomposed into simpler components, traffic forecasting process is expected to be
more efficient.
This research simulates data from an operator in Indonesia that acts as terminating side. Dataset
used in this experiment was a 7-months dataset from billing process, that consists of 6 months
data for training and 1 month data for testing.
The experiment result shows that Hybrid model (EEMD and BPNN) is more efficient because it
has smaller iteration number compared to BPNN model only. Moreover it also finds that BPNN
model gives a slightly more accurate result compared to Hybrid model. Unfortunately BPNN
model needs bigger effort that is represented by bigger iteration number. Both Hybrid and
BPNN models have agreement and same abilities in detecting anomaly that is defined by daily
deviation more than 2%. The presence of anomaly has business impacts for terminating
operators. As illustration, this research finds two anomalies in a month that are equivalent to
company’s loss as much as IDR 7.9-11.7 millions.