Department of Computer ScienceFederal Urdu University of Arts, Science & Technology (FUUAST) Islamabad, Pakistan
Department of Computer ScienceFederal Urdu University of Arts, Science & Technology (FUUAST) Islamabad, Pakistan
TOC o “1-3” h z u INTRODUCTION PAGEREF _Toc4011424 h 1Data Mining PAGEREF _Toc4011425 h 1MOTIVATION PAGEREF _Toc4011426 h 2LITERATURE REVIEW PAGEREF _Toc4011427 h 3PROBLEM STATEMENT PAGEREF _Toc4011428 h 5METHODOLOGY PAGEREF _Toc4011429 h 6REFERENCE PAGEREF _Toc4011430 h 7
INTRODUCTIONChurn in the telecom industry is the number of customers that left all the previous subscribed firms and converted to all other organization (Bi, 2016).
The entire client is conscious that churn prediction is widely used to retrain its precious clients in the telecommunications industry as it usually costs quite less than creating new clients (Xiao, 2016). All the desire to exploit greater facilities that will lead to a fresh client churn and if we tend to lower the client churn level to ten percent development level of a business that will rise to thirty percent and it will be eight percent (Kumar, 2016).
Basically, the churning of all the customers and retention of the customer is normally the part of the customer management system. Here we have different researchers, and they worked for the implementation of customer churn management in the service industry by the help of using a different method which includes the data mining as well as machine learning algorithm and there will be any model which is based on statistical models (Jiang, 2016).
Data MiningIt is a process of extraction for all the previously unknown data which is having some useful information that may also be useful to take some action based on the information which is provided by the data mining. In other word data mining is a process of extracting useful data from huge dataset, obtained from unknown dataset after applying relevant patterns and methods, ma be helpful in designing process of various profitable businesses. All the association rules as well as classification, clustering, decision trees and artificial neural network, known as methods, are used for the mining. There are researchers, and they have all worked as well as proposed many models of churns. Machine language is a branch of computer techniques which is used for the construction of the whole algorithm which is developed from the research of pattern recognition and computational learning in artificial intelligence. All the machine learning algorithm have already been used many times by researchers to develop the prediction system (Tsai and Lu, 2010).
All these algorithms are divided into two main parts namely famous with the name of a supervised learning algorithm and unsupervised learning algorithm. Machine learning algorithms are divided into two main categories namely supervised learning algorithm and unsupervised learning algorithm. Algorithms like regression, classification are mostly supervised learning algorithm. Classification models get information from past customers then train classifiers and used it to classify unseen patterns. Clustering algorithms emphasis on similar features to assemble the data referred to as cluster so move the new data-set into anyone of the connected group/clusters. In any classification algorithm to build a customer churn classification model. We will train a classifier on the training data-set, and after training the chosen classifier, we will test the accuracy of the classifier to classify the customer churn rate (Coussement and Lessmann 2017).
From some more studies we may get to know about the different factors like as churning rate of customer, existing system prediction performance & retention capability which are the main drivers we can focus on predicting a customer behavior whether it would be going to churn or not, many researchers have projected various solutions which can work with efficiently for holding customers by smearing different techniques (Zhang, 2015).
Many data level and algorithmic approaches are working and making its best result in predicting churned customers. Hybrid oversampling and undersampling, Inductive algorithms, SMOTE oversampling technique, clustering techniques, (ANN) artificial neural networks, echo state network (ESN) with SVM training algorithm, random forest algorithms, PSO, decision trees, MRMR, GA and ensemble classifier are used for obtaining very best performance in churn prediction [(Zhang, 2015), (Jiang, 2016)]. Early prediction of customer churn will increases the efforts of the corporate to retain the customer. Thus this analysis work predicts the customer churn rate on time before a customer leaves an organization so the corporate might take prognostic measures to retain its loyal customers.
This research work will focus on introducing a new data mining technique to manage and predict the Customers churn on various churn rates on the subscribed data. The proposed prediction system will be helpful for the organization to take some prior steps in managing customer churn. However, by all the previous studies, we will consider the grey area where we can work and remove the shortcomings of the traditional classification methods. In this research work, we will introduce a certain deep learning algorithmic technique.
MOTIVATIONThe motivation of this research work is the prediction of churn customer before he/she left any organization. Churn prediction will be helpful for the organization to take the early steps to retain the customer with them. For this purpose, we shall use a dataset which has a high volume that will provide a significant prediction. We will use machine learning algorithms like a neural network, deep learning, auto MLP, etc that will be helpful to predict/classify churn customer from available data. Deep learning algorithms have shown excellent performance for huge datasets, and these algorithms contain auto feature learning capabilities. We will apply augmentation techniques like Ada-boosting/pooling to our dataset that will make it in the more standardized form.
LITERATURE REVIEWChurn prediction is one of the challenging problems and may cause serious loss/profit for any organization. The accurate churn prediction is significantly helped Customer Relationship Management (CRM) to minimize their loss and maximize their profit by providing better services to users. Echo state network training and support vector machine (SVM) algorithms are employed in predicting the customer churn by victimizing the info of telecom firms. Normally better quality of a corporation motivates the customer to switch the existing company accuracy to resist overfitting problem retaining existing customers is one of the challenges for any organization, lot of advertising and selling expense are needed to draw new customers (Hudaib, 2015).
To upsurge client retention and gym utilization, number of experiments are presented, aims for customer churn prediction. Azure ML & Big ML platforms are used for churn prediction by concentrating on user services supported demographic and behavioral data. Customer-centric platform Provides better results than random or means Monthly renewal of gym membership Limited availability of users demographic data. Additional engineered features to maintain consistency of predictor (Jas Semrl, Alexandru Matei, 2017).
A comparative study made on ten different techniques is worth important in analyzing the churn prediction models which work best and outperforms its results amongst each model. During the analysis studies show that ensemble-based techniques (ADA boost and Random Forest) outperform its result with an accuracy of ninety six on the data-set containing 3333 records and at the moment SVM and Neural Networks (Multi-layer perceptron (MLP)) shows the most effetive performance on the given data set. This comparative study provides us a strong and solid comparison of different churn prediction models which is a base knowledge for the future work to adopt the best performing model and enhance its accuracy. Small data sets used for the higher result and improved accuracy can be obtained if we use the hybrid or deep learning models for the churn prediction of customers in the telecom industry (Sahar F. Sabbeh 2018).
According to Verbeke and Martens (2011), profit and retention have an uneventful relationship for prediction. The higher profit is often achieved by enhancing the retention capability. The prediction and retention algorithms are helpful if some proper actions are taken, boosting is only used to improve accuracy. The sampling method used for all customers is based on Random sampling Profit maximization is affected by customer churn. The aptitude of maintaining constant retention is not practical.
According to Verbeke (2011) there are various causes due to which client leaves a corporation and try to predict churn customers by identifying client behavioral characteristics and by using classification of data mining methods such as pattern recognition, try to predict churn customers by defining behavioral attributes of customer and by using the data mining techniques classification such as machine learning, pattern recognition, statistics, support vector machine as well as applying to clusters such as DB Scan and K-Mean algorithm realize the correct behavior of client. Apply the prediction method on the bank customer regarding future churn. The results of this research work will apply to bank customers; only some of the algorithms provide optimal results while some algorithm does not provide optimal results (more use of memory for a tiny quantity of information).
A six-step data mining model selects the useful data and investigates it after that classification and clustering will be done, in the end, this model will give us useful knowledge. This study also refers us with a comparison of some traditional methods such as Regression Analysis (RA), Decision Trees (DT) and some of the soft techniques like Fuzzy Logic (FL), Neural networks (NN).In these Model SVM methods, the data set outperforms to predict the churners and non-churners that assist customer churn management (CRM) retaining their customer and by studying their mental operation. A deeper Analysis using deep learning algorithm can be used in future studies to obtain a better result (Karyakina and Melniko, 2017).
Adnan Anjum et al (2017) suggested a fresh model of E-Churn improve the customer churn forecast by increasing recall factor. This model using the ensemble technique for churn prediction. The various finding is achieved by combining various algorithms such as QUEST, C5, CHAID, logistics regression, and CRT, but the highest outcomes are achieved by combining C5 and QUEST, which was 93.4 percent. The E-Churn model surpasses its outcomes among many classifications and clustering algorithm. There was No clear data-set used for the experiments. If we have very clear churn information history, the precision of this algorithm can be improved very effectively (Adnan Anjum et al, 2017).
Adnan Idris et al suggested a Genetic Programming (GP) algorithm and AdaBoost churn prediction model. GP programming worked effectively to search information and AdaBoost was used to define various factors for client churn behavior in an iterative method. Particle Swarm Optimization (PSO) was used to balance the dataset using the sampling method. To solve many complicated issues, this model worked very effectively. Using machine learning or deep learning to improve efficiency, effective feature selection methods can be introduced (A. Idris, A. Iftikhar, Z. Rehman).
The current clustering algorithm described by Rathore and Bhardwaj (2017) research involves semantic driven subtractive clustering method (SDSCM) relying on subtractive clustering method (SCM) and axiomatic fuzzy sets (AFS). The fresh algorithm is useful in improving precision and reducing the amount of risk. Compared to the K median clustering matrix, the suggested SDSCM delivers effective outcomes. Attributes are selected in the proposed SDSCM, then the neighborhood radius is determined and the cluster number is calculated. The proposed SDSCM automatically determines neighbor radios and also set the termination state. SDSCM stronger strength than SCM and AFS Fast running speed Increase accuracy of K-Mean and AFS, SDSCM is based on two other techniques SCM and AFS, and it depends on the centroid of the cluster (Rathore and Bhardwaj, 2017).
Reference Publish Year Technique Dataset Accuracy
Sahar F. Sabbeh  IJASCA 2018 SVM
decision tree Public dataset telecom company 96%
Anjum, and Usman 
Jas Semrl, Alexandru Matei  IEEE 2017 Neural Network,
Decision Tree Azure ML, BIG ML DT=70.8%
Idris, A. and Iftikhar, A.
 Springer 2017 GP-Ada Boost
Ibrahim M.M.Mitkees, Assist. Prof. Sherif M Badr, Dr. Ahmed Ibrahim Bahgat ElSeddawy  IEEE 2017 Decision tree (DT)
Xiao, J.  IEEE 2016 DL,LR BUANBXI mobile 40mm ltd 76.8%
Zhaojing Zhang, Regina Wang, Weihong Zheng, Shizhan Lan, Dong Liang, Hao Jin  IEEE 2015 decision trees
Regression Guangxi Mobile Communication Company Limited real dataset 69%
Ali Rodan and Hossam Faris  IEEE 2015 Support Vector Machine (SVM)
E. Shaaban, Y. Helmy, A. Khedr, and M. Nasr  IJERA 2012 SVM
DT 23 att &5000instances 83.7%
Table SEQ Table * ARABIC 1: Stat of Art
Most of the existing machine learning techniques face the problem of local optima that reduces the accuracy and does not decreases error rate. Although, there is still space for improvement because some techniques work for small data sets while others do not achieve the acceptable levels of accuracy
PROBLEM STATEMENTPropose a novel approach to acquire a low-cost model for retaining the existing customer than making new customer, also improve the accuracy and reduce the error rate by using some particular algorithms for machine learning.
METHODOLOGYThe proposed methodology for the churn prediction is using feature selection with deep learning for the majority. The researcher will normally work for the evaluation of the ensemble methods in the proposed methodology. It will involve the number of all the steps for the formulation of all the preprocessed balance data.
The selection of all the data is very tricky and complex, Telecom data set is also not available freely. Here I found two data sets which are available freely for the experiments. Both data sets are obtained from the GitHub from the links. Link1:
Once we get the pre-possessing techniques of information pairs that are added to it, we get the greatest efficiency on it. In prepossessing methods that we handled in all the missing values for removing the duplicated and setting the roles for giving the attributes as well as for removing the outliers. We will use it for our experiments and then we will be able to convert it into the balance data set.
Once we will have applied the data set of 10 fold cross-validation technique on it, the cross-validation in our data set will be divided into training and testing data sets in the specific ratio that set we set like for the more than 60 percent and more than 30 percent. It means that the data will be trained and we will execute the test 10 times for the complete test operation. We have implemented three distinct classifiers in cross-validation, one by one, which is called as deep learning as well as neural network and AUTOMLP.Figure SEQ Figure * ARABIC 1: Proposed Model Diagram
In the proposed technique, the performance will be evaluated by using the following parameters:
Where true positives are correctly classified actual positives and false positives are actual negatives that are incorrectly labeled as positives.
Where false negatives are incorrectly labeled data as negatives that are actual positives.
The F-Measure is the weighted average of Recall and Precision.
Accuracy is the rate of correct predictions by a proposed model in comparison with actual measurements executed on the test data
REFERENCEBi, W. (2016). A Big Data Clustering Algorithm for Mitigating the Risk of Customer Churn: IEEE Transactions on Industrial Informatics, 12(3), pp.1270-1281.
Xiao, J. (2016). Churn Prediction in Customer Relationship Management via GMDH-Based Multiple Classifiers Ensemble: IEEE Intelligent Systems, 31(2), pp.37-44.
Tsai, C. and Lu, Y. (2010). Data Mining Techniques in Customer Churn Prediction: Recent Patents on Computer Science, 3(1), pp.28-32.
Coussement, K. and Lessmann, S. (2017). A comparative analysis of data preparation algorithms for customer churn prediction: A case study in the telecommunication industry. Decision Support Systems, 95, pp.27-36.
Zhang, Z. (2015). Profit Maximization Analysis Based on Data Mining and the Exponential Retention Model Assumption concerning Customer Churn Problems: IEEE International Conference on Data Mining Workshop (ICDMW).
Xiao, J., Jiang, X. and He, C. (2016). Churn Prediction in Customer Relationship Management via GMDH-Based Multiple Classifiers Ensemble: IEEE Intelligent Systems, 31(2), pp.37-44.
Karyakina and Melnikov (2017). Comparison of methods for predicting customer churn in Internet service provider companies: Machine Learning and Data Analysis, 3(4), pp.250-256.
Idris, A. and Iftikhar, A. (2017). Intelligent churn prediction for telecom using GP-AdaBoost learning and PSO undersampling: Cluster Computing.
Anjum, A. and Usman, S. (2017). Optimizing Coverage of Churn Prediction in Telecommunication Industry: International Journal of Advanced Computer Science and Applications, 8(5).
F., S. (2018). Machine-Learning Techniques for Customer Retention: A Comparative Study: International Journal of Advanced Computer Science and Applications, 9(2).
Verbeke, W. (2011). Building comprehensible customer churn prediction models with advanced rule induction techniques: Expert Systems with Applications, 38(3), pp.2354-2364.
Verbeke, W. and Martens, D. (2011). Building comprehensible customer churn prediction models with advanced rule induction techniques: Expert Systems with Applications, 38(3), pp.2354-2364.
Beyaz, Y. and Bak?rta?, P. (2017).The economic theory of internal conflicts on the basis of neural resources: Anadolu University Journal of Social Sciences, 17(2), pp.27-42.
Rathore, A. and Bhardwaj, A. (2017). Prediction model for telecom postpaid customer churn using Six-Sigma methodology: International Journal of Manufacturing Technology and Management, 31(5), p.387.
Mathworks.com. (2018). What Is Deep Learning? How It Works, Techniques & Applications: Retrieved from