Telecom Churn Prediction from Customer Usage Data (Igor Tymchuk)
Download as PPTX, PDF0 likes740 views
This document discusses building machine learning models to predict customer churn for a telecommunications company using customer usage data. It compares two models: a two-class boosted decision tree model and a support vector machine (SVM) model. The goal is to predict which customers who have been active for over a month will suddenly stop using the company's services. Both models are able to predict churn with high accuracy, with the boosted decision tree performing slightly better. The document also outlines the data used, challenges faced, and potential next steps to improve the model.
Telecom Churn Prediction from Customer Usage Data (Igor Tymchuk)
- 1. Telecom Churn Prediction from Customer Usage Data Two-Class Boosted Decision Tree vs SVM models. Igor Tymchuk, tymch777@gmail.com Goal To predict Customer Churns of the customers who were active for more than 1 month in a row (day by day) but then suddenly left the company. Prediction should be based on activity data (customer usage and top-ups only) Approach To Use Azure ML is the main idea for this project to show customers that it is very easy to start with Data Science solution and implement it within existed BI solution. Challenges • Defining sensors and observed raw data • Absence of memory of the previous states Idea: To create easy applicable solution to predict telecom (mobile) churns: Sophisticated handling of churn is a sign of telecommunications industry where subscribers are known to frequently switch from one provider to another. This voluntary churn is of prime concern. Moreover, providers have accumulated significant knowledge about churn drivers, which are the factors that drive customers to switch. Specifics: Mobile Virtual Operator customer Churns occur very often b/c people usually use it as 2nd sim only for International calls. Results The system is able to predict customers churns within low Threshold (< 0.1) with very high accuracy (every second record was real churn in nearest week) Remaining Work - Apply current created webservice into existed solution. - Get more data for Model raining - Investigate other data which can help us to predict Churns. (for example: CRM data) Input: 484 Hidden: 50 Visualized features learned “walls” part of weights “food” part of weights “predator” part of weights Comparing Models for Churn prediction “sand” part of weights Practical Value Above mentioned approach could be useful for mobile operators to predict churns. Usually we need to predict churns to apply for users special treatment to keep those customers within the company. The most important metric in churn is the misclassification rate: that is, of the top N churners as predicted by the classifier, which of them actually did not churn, and yet received special treatment. Raw data is: usage traffic data - Call type , - Roaming/non roaming calls ration - Actual call duration - Billed duration - Average cost - Amount of calls for last month Problem: raw data is transaction and is not prepared for Machine Learning Solution: To agregate data and to learn features by unsupervised feature learning techniques Two-Class Support Vector Machine module to create a model that is based on the support vector machine algorithm. The classifier that this module initializes is useful for predicting between two possible outcomes that depend on continuous or categorical predictor variables. Two-Class Boosted Decision Tree module to create a machine learning model that is based on the boosted decision trees algorithm. A boosted decision tree is an ensemble learning method in which the second tree corrects for the errors of the first tree, the third tree corrects for the errors of the first and second trees, and so forth. Predictions are based on the entire ensemble of trees together that makes the prediction Conclusion: This project describes a sensible approach to tackle with the common problem of customer churn by using a generic framework. I considered a prototype for scoring models and implemented it by using Azure Machine Learning. Compared results of two models Two-Class Boosted Decision Tree and SVM.