Neural Network Model for Detecting Rare Patterns in Data: a Telecommunication Example


Since the liberation of telecommunications industry in the last few years, the industry has changed dramatically. Competition among long distance service providers has become intense. Better services and promotions are offered to customers. Because of the increase in customer choice, there has been an increase in churn rates. Churn, or customer abandonment happens when a customer switches providers for any reason, such as lower rates or better customer service. Churn in telecommunications is an expensive problem. According to [1], the average of the annual rate of churn is approximately 30% of total telecommunication customers in the U.S. the estimated cost of finding a new customer is approximately $400 per customer in wireless communications. It is obvious that spending money to maintain current customers is more efficient than acquiring new customers. to reduce customer churn, telecommunications companies need to be able to identify the customers who are likely to switch before the customers take any action. in this paper, the artificial data set based on claims similar to the real world data set of telecommunication customer churn information was used to train the probabilistic neural network to be able to classify the customers into two groups: churn and non-churn. the churn groups consists of 14.14% of the total in which the behaviors of data in this group are categorized as rare patterns. the data set contains 20 fields, and one class attribute. Some of the fields are significant and insignificant to the output of the network. Thus, two approaches, sensitivity analysis and weight analysis, were applied to identify and get rid of those insignificant fields. the results reveal interesting information that the customers in the churn group are those with “high international calls, high international plan, and high customer services calls,” and customers in the non-churn group are those with “low international calls and low customer service calls.”

Meeting Name

1999 Artificial Neural Networks in Engineering Conference, ANNIE'99


Engineering Management and Systems Engineering

Keywords and Phrases

Data Mining; Probabilistic Neural Networks; Sensitivity Analysis; Weight Analysis

Document Type

Article - Conference proceedings

Document Version


File Type





© 1999 American Society of Mechanical Engineers (ASME), All rights reserved.

This document is currently not available here.