Results in Control and Optimization 14 ( 124) 100242 Contents lists available at ScienceDiree P pa ore Results in Control and Optimization fae ° ae ELSEVIER journal homepage: vw.slsevier.convioeate Customer churn prediction in telecom sector using machine col learning techniques Sharmila K. Wagh®”, Aishwarya A. Andhale”, Kishor S. Wagh‘, Jayshree R. Pansare’, Sarita P. Ambadekar , S.H. Gawande “prot of Comper Bagmang MAS. Calo xgnerng SP Pane Urey, Pins Moar 4100), nda aaron of norman Tec, MRSS + Cris Calg of Engen fr Wren, ne 4105, Inka “beprtent of Compce’bgeeog AISMS ns of frat nl Pane Ue, Pe, Mahara 1100, na "ogame of Camper gern Sra ae of Teco Sian Murda a “inartarig atraeyernet f Monil rce MS Clef gman SP ae nent, Pn Mohr ARTICLE INFO ABSTRACT Keyword: Inthe telecom industry, large-scale of data is generated on daly basis by an enormous amount of hae customer base Her, geting & new stoner bas is eos thas holding the cue etstomers ‘rtmer chum rdion where chur the pres of cstomers switching (om ae im to another na given silted Vosenpling time. Telecom management and analyts ace finding the explanations behind customers leaving subscriptions and behavior activites of the holding churn customers’ data This system uses ‘lasification techniques to find out the leave subserptions and collet the reasons behind the leave subscription of exstomers inthe telecom industry. The major goal of this system is t0 analyze the diversified machine learning algorithms which are required to develop customer chur prediction models and identify churn reasons in order to give them with retention stra tegies and plans. In ths system, leave subseriptons collects customers’ data by applying classi Seation algorithms such as Random Forest (RF, machine letring techniques sich as KNN and decision tree Clasifer. It offers an efficient busines model that analyzes customer churn data, land gives accurate predictions of chur customers so that business management may take action ‘within the churn period to stop chur as well as loss in profit. System achieves an accuracy of 99, '% using the random forest clasifier for churn predicts, the classifier matrix has achieved 3 precision of 95 with a recall factor of 9 9 alongwith received overall securacy of 98.09 % Likewise, our research work improves churn prediction, scope athe Busines elds, and provide prediction models to hold thir existing customers customer service and avoid chur effectively. 1. Introduction Role of predictive mode! is to bring the churned customers to light. The proposed model's purpose is to ring churned customers to light. Ina targeted approach industry try to identify which customers are likely to churn. The industry then targets those customers or clients and provides them with special incentives, offerings, and plans except normal customers. This approach can bring a huge oss to industry, ifchumed measures are inaccurate because the industries are wasting alot of money tothe customers who would have stayed * Conesponding author Bail adress: sag 5.apail.com (SK. Wagh). Received 30 March 2023; Received in revised form 24 October 2023; Accepted 11 November 2023 ‘Avallable online 13 November 2023, 2666-7207/6 2028 The Author(s). Published by Elsevier BV. This is an open access article under the CC BY-NE-ND license (utp /ereativecommansong/licenses/by-ne-nd/4.0/SK. Wah a Ress Gono and Opinion 14 (2024) 100842 anyways, ieespective of short or long distance. Communication has become an important par of today’s life It's being used in every field [1-3], The telecom industry needs to build the best predictive model for churning customers. ‘Churn Customer refers tothe number of existing customers who may leave the service provider over a given period. These cus- ‘tomers can be called as churners, The main aim of chuen is to predict the churnable customers atthe earliest, to identity the reason for ‘churning. The primary goal of churn analysis sto identify ané anticipate churnable consumers as soon as possible. This will help us to rectify the issues of the customer, This wil be helpful to satisfy the customer needs and will continc to use that service. This will help tomeet the needs ofthe customers, and they will continu to utilize the service. There are promotional costs known as acquisition costs and retention costs in a telco company. The acquisition costs the price a company pays to gain new consumers. Retention costs, on the other hand, are the costs of keeping existing clients. I is very difficult to predict which customers would churn and which customers will be maintained due to human limitations. Asa result, the allocation of money may be incorrect, resulting ina higher amourt of cash being issued, Furthermore, according to some reports, the acquisition cost is S times that of the retention cost. If i is incorrect in projecting a client who will churn, but it turns out that we are correct in anticipating a customer who will be kept, we will have to spend more than. we should. This paper attempts to develop a Machine Learning model that can predict customer churn prediction and retention. Retaining cost of the company for the existing customers is far economical as compared to addition of new customers within the network. The customer churn is the direct loss in terms of revenue tothe company. Ifthe information ofthe development of churning of ‘customers s known wel in advance, then appropriate steps can be taken, and a better service ean be provided to such customers. It is ‘observed that the long-term customers add more revenue to the company as they are not much responsive to slight changes. Today's ‘most challenging and ertieal problems faced by the telecommunication industry is the management of churning customers. Recent studies show that the main objective isto identify the valuable ehurn customers using a huge amount of data received from the telecommunication industry. Practically there are so many restrictions in using current models, which faces lots of difficulties & hurdles towards the problem of churning in today’s present environment. While modeling development, a lot of information-rieh features are neglected in In Feature selection Process. Mosiy statistical methods are being used in a diverse domain, which tends to give undesirable results of the present predictive models, The feature selection is another huge problem with the existing models Every customer may be an individual or a group and have different churning reasons. Classification of a churn customer can be as a cchurmer, irespective observing reasons and factors for chuen. There are different patcerns of behavior during the churn process and should not treat all of them inthe same manner. Some customers may not churn easily than others. Today's need is fr a more realistic prediction mode! which can predict churn customers in advance, This will be of great help to provide strong retention strategies for the different groups of churmers may it be different promotions depending upon the chur factors for the different group of churners. Encouraged by the above-mentioned facts & observations, in this study, a model for prediction of churners with the help of different ‘machine learning algorithms is proposed. Ullah, et al. [1] provided a customer attrition model for data analytics that is validated using common assessment measures. The results demonstrated that utilizing machine learning techniques improved the performance of their proposed churn model, The F-measure result from Random Forest and J48 was 88 % better. The authors used the dataset to identify the primary churn variables and did cluster profiling based on their churn risk. Finally, the authors presented recommendations for telecom decision-makers on client retention. Ammar, et al. [2] presented a hybridized algorithm technique to predict churners that was quite efficient. When compared between the proposed algorithm and normal firefly algorithm, the accuracy obtained by both is found to be similar. Yet the hybrid firefly algorithm performed better than the normal firefly algorithm in terms of time latency which is very low. The study was carried out ofthese algorithms with respect to F- Measure, Accuracy, Time, PR and ROC. The prediction model inthe telecommu- nication companies should be as efficient as accurate [3]."The efficiency of the model can be gained by making the bes training model with reduced dimension and size. The extraction and feature selection techniques used, helps to get the efficient features along with, predicting accurately. Thus, this paper shows that being smaller in size .e., 232 KB this prediction model performs the similar tasks ‘even with more accurate results e., 92 % than the initial model of 303 KB size. In many researches, was represented in different ways and data mining is most usable in churn prediction in telecommuniation industry (4). The concept of multilayer perceptron neural network was used to proposed churn classification model and showed that the logistic boost and logical regression both are useful to build a chur predietion model. Black box models are complicated, but actual work tell only logistic regression and using logit boost training example for weak prediction that play an important role in chum prediction, Effective classifiers like KNN, Decision tree classifier, random forest are studied (5). Classiflers consist of four stages: data access, data wrangling, training, and acquiring insights that all aspects play an important role to analyze churn prediction data. Here, data acces is gves input tothe proposed model. Data ‘wrangling isthe process of collecting distributed data [6]. The training process is constructed by model and test the data set for processing and acquiring insights i useful to predict final output and analyze the final result of churn prediction, ‘Machine learning techniques used [7,8] in developing this model for this work. The telecom company can predict churn customers with the help of electronic learning technology with developed model. Industries can provide best services so as to reduce the churn level. Such types of models help telecom services for making them proftable. Random Forest and Decision Tree are used for this model. In (9) the PMM (Predictive Mean Matching) algorithm helps to manage lost values, rather than feature removal or recognition of ‘missing data. The combination of two Ensemble classifiers is embedded within the customer speculation model to manage large databases, time-dependent feature labels, and the distribution of inequitable data in the Telecommunication industry. To evaluate the effectiveness of customer churn uplift models, the author presented a novel profit-driven assessment method dubbed the maximum, profit uplift measure in| 10). The proposed MPU measure extends the maximum profit measure for customer churn prediction models, allowing for evaluation ofa customer churn uplift model's performance in terms of profit per customer in the customer base earned ‘wher a retention campaign targets the optimal proportion of customers withthe highest uplift scores. Maximizing the profit generatedSK. Wah a Ress Gono and Opinion 14 (2024) 100842 by the retention campaign determines the best proportion of customers to target, which is proven inthis article to be strongly tied to ‘the uplift model's capacity co identity the so-called persuadable i., Customers who are about to chun but wll be kept ifthe campaign is targeted. Saran Kumar et al. (11) gives a thorough examination of the strategies used to anticipate client chum. Fach of their proposed churn prediction models has a low forecast accuracy. To avoid the problem of client turnover, a solid prediction model is essential. Kraljevié and Gotovac (12) given a logical foundation for developing data mining applications by defining an enhanced technique for designing applications based on Data Mining technologies, The methodology proposed has been deployed and proven in the telecom industry using data mining applications to predict prepaid subscriber attrition. Real-life examples of customers are used who decided to go and learn the qualities and behaviors that precede customer profits (13). There are results that show a clear height for advanced versions of models compared to plain (non-expanded) versions. The best separator was SVM-POLY using AdaBoost with approximately 97 % accuracy and F-rating over 84 %. Alzubaii, etal. 1), used CDR attributes to determine the strength of social tes between users for each identified community, and then used a model to propagate ‘churn influences on the call graph to determine the net aggregated influences from the chumer node. This effect was employed in the logistic regression method to calculate the churn tendeney for individual users. When building the Telecom social network, examine the relationship between users to improve the performance of the chura prediction model. CDR attributes were employed to char- acterize the relationship strength of social communication between edges in the graph. CDR attributes contain information about phone calls and SMS messages. In this work the quality of prediction measurements for various prediction models is compared. Khalid, etal. [15] looked at different prediction models and compared the quality of prediction models such classification algorithms and decision trees, They discovered that the decision tee’s accuracy is higher than the other techniques (3 6 higher than the second result and 6 % higher than the lowest-achieving algorithm), showing thatthe decision tree is an excellent churn prediction technique. The primary goal of study presented in [16] is to develop a method to predict high-value customer turnover based on existing research and customer attribute characteristics in the telecom industry. This study accomplishes customer churn prediction based on the telecom business based on the analysis of big data inthe telecom Industry and historial information estimation of customers, combined with logistic regression method. Ic identifies possible churned customers inthe customer library by studying the features of customer churn behavior in the telecom industry, and it assists organizations in taking targeted win back efforts based on the char- acteristics of the potential churned A review of customer churn prediction inthe elecommunications industry Is presented in [17]. The study demonstrates a huge number of atributes that are put into practice by a big number of paper reviewers to construct a customer churn prediction model. Lalwani etal, [8] presented a comparative study of customer churn prediction in the telecommunication industry using well-known machine learning techniques like Logistic Regression, Nave Bayes, Support Vector Machines, Decision ‘Trees, Random Forest, XGBoost Classifier, GatBoost Classifier, AdaBoost Classifier, and Extra Tree Classifier inthis research paper. The experimental results demonstrates that two ensemble learning algorithms, Adaboost classifier and XGBoost clasifer, have the highest accuracy when compared to other models, with an AUC scare of 84 percent for the churn prediction problem. They outscored other algorithms across the board in terms of accuracy, precision, F measure recall, and AUC score, In (19) an improved churn prediction for credit card system using supervised learning and rough clustering technique is presented, Whereas, Amin etal, (20) implemented ‘chum prediction using Naive Bayes for customers of telecommunication industry with focus onthe challenges to get new customers in telecommunication industry instead of handling old customers, Model to analyze customer behavior was developed along with {improvement in accuracy of prediction. In line with this work again Amin etal. (21), highlighted telecommunication sector's ehura prediction using Justin-Time approach. It handled various problems of customers intention that may switch from one service provider to another. In this context, different churn prediction methods are applied by practitioners to resolve issues related to preserving In other work, Amin et al. (22) presented churn prediction for customers of eross-company by using transformation of data techniques. It emphasizes target company with lacking data which may be used to source company to predict customer churn elfectively. ven, itis also including impact ofthese techniques on performance using various clasifiers needed in telecommunication sector. Amin etal [23] used rough set method to predict customer churn of telecommunication sector. Customer behavior was used to differentiate the chum fom non-churn eustomers. The major objective was to desperate the necessity of businesses in retention of existing customers, Moreaver, this work uses intelligent rule-based techniques for decision making along with rough set to detect. feustomer chura from non-chuen customers, Aled et al. [25] focused on multi dataset approach for churn predietion which is applicable fr telecommunication industry. In this work, major conteibution related to feature engineering phase for creation of custom. features with the help of machine learning is presented. Moreover, the big data approach along with comparative results of multiple algorithms of machine learning has been used. In the telecom sector, a number of customer retention tactics are being used to lower chun and boost loyalty. Customer loyalty In the telecom industry is fa fom guaranteed, Many companies combined its cellular and wite line customer loyalty programmers as part ‘of reorganization that aimed to give customers more rewards options. Customers can now take advantage of a variety of benefits for their loyalty, including: hardware enhancements include fiber optic service and internet content. To lower their churn rates, telecoms are investing more and more in cutting-edge technologies. Tools for customer analytics and insight can be used to forecast customer behavior, allowing service providers to decide which strategies will Increase retention rates the most effectively. Additionally cuting- edge technology uses artifical intelligence and machine learning techniques to pinpoint the clients most likely to churn. From this discussion itis observed telecom industry playing a very major role in our daily life and generating an enormous amount of customer bbase. New customer base generation is costlier than retaining the existing customers. Churning is the process of switching of customers in a given time, Telecom management and analysts find the explanations behind customers leaving subscriptions and behavior ac- tivities of the holding churn customers’ data. The model proposed uses classification techniques to find the leave subscriptions and collects the reasons behind it in the telecom industry. The aim of this model is to analyze the various machine learning algorithmsSK. Wah a Ress Gono and Opinion 14 (2024) 100842 required to develop customer chur predfetion models and identify churn reasons in order to give them with retention strategies and plans. The main motivation behind churn prediction in the telecom sector is to reduce chums and retain the existing customers Section 2 emphasizes system architecture of proposed system and proposed system model is presented in detail, Sequentially, Section 3 focuses on detailed dataset description based on single dataset including data pre-processing, and feature selection and also high- lighted multi dataset approach concisely. In Section 4, experimental analysis using various ways such as decision tree classifier, Random Forest algorithm, survival analysis, Cox proportional hazard model, and retention strategy is presented, Finally, concluding remarks are summarized in Section 5. 2. System architecture In this section, system architecture and proposed system model are discussed subsequently. 2.1, System architecture of the proposed system ‘The implementation for churn will require the latest version of Anaconda with builtin features that consist of Jupiter notebook for training and testing data, The latest version of Anaconda with builtin functionality, like Jupiter notebook for training and testing date, will be required forthe churn implementation, Chum predictions forthe telecom industry have been carried out using literature with various methods that includes machine learning algorithms, data mining techniques and retention strategies. These techniques effectively support many companies for predicting, identifying, and retaining churners which help in CRM (Customer relationship management) and decision making. CRM deals with the data to identify 2 loyal customer for industry. High revenue generating ceustomers (loyal customers) for a company have no impact on the competitor companies, Such loyal eustomers help to grow profit ability ofa company by referring to the other people such as their family members, colleagues, and friends. Hence, the role played by (CRM is very important in churn prediction and it also helps to retain the churning customers. Pig. 1 depicts a cycle through which ‘chur prediction can be made. For prediction there are many algorithms such as Support vector machine, K-nearest neighbor, 148, naive Bayes, logistic regression, LWL, Random Forest, Decision tree classifier which are used to resolve classification problems. Random forest and Decision tree classifier are considered relevant with better accuracy and performance. 2.2, Proposed system model Let $ isthe proposed system. S=(, 0, DP, Fs, EF) Where 1) 1 Gnpub: Dataset 2) (Output): CM (Confusion matrix) —{R-> Actual False, Actual True} {C->Predicted False, Predicted true) 3) DP (Data Processing) 4) FS (Feature Selection) is measured by Pearson correlation formula (Ge) where set of numerical attributes is taken X' ranging from X1, Xo) Xow nny Kn considera set of attributes X, form a subset of two attributes each, X= {€X, Xi. Oy Xo Nay Nab eo OX os Kad Csi T Fig. 1. Proposed systemSK. Wah a Ress Gono and Opinion 14 (2024) 100842 If each subset = X then evaluates 2X), SX? & r= VE OK * XY = DX? EXP ‘Where Cr~ 1 (+ve comelation, Cr = 0 (no cortelation), Cr = ~1 (ve correlation) Conver string values into numerical (S-+N) ‘Then apply MIL algorithm on new Dataset (consider as C) 5) FF ficiency of Proposed Model “The proposed system for chur prediction is derived using accuracy, precision, recall, Feasure. Accuracy calculates the ac curacy metre ‘TP—True positive, TN—True negative, FP—False postive, FN~False negative values, AP—Actual postive, AN-Actual negative 2) Aceuracy~ (IP + TN) + (TP + TN + FP ~ FN) 'b) True postive and true negative are values ofthe confusion matrix after applying classification algorithms. True postive rate is the value that shows us which pat of data is classified as correct and false positive classifies incorrect values. ©) (IP) rate ="TN + TN +P &) (FP) rate= EP + FP=TN «) Precision = TP = (IP + FP) £) Recall —"1P + (IP + EN) ) FeMeasure ~ (2 x Precision » Res Kt XD I) / (Precision + Recall) Constraints: The time isthe key or predicted churners. IC (customer) churns after one month, but sometimes dataset shows that it ‘would churn after one week that will lead to incorrect churn. Henee, C depends upon T. Incorrect dataset (1D) with irelevant features, will be other constraint to predict churn model Pailue ofS: eustomer (C) gets churned out for one month time (1) but then customer rejoins again would lead to loss of company and system (S). 3 Dataset description Here, specifically focus ison Single Dataset presented in sub Section 3.1. Sequentally, Section 3.2 emphasizes multiple datasets to validate proposed system. 3.1, Single darasee ‘The dataset used for experiments inthis paper, contains results of Telco-Customer-Churn dataset obtained from Kaggle website (itis also known as IBM Watson dataset which was released in 2015).Fach row represents a customer, each column contains attribute described on the column Metadata. It consists of 7043 customer information, Every customer has 21 features and the “Churn” it contains 11 missing values in the Total Charges column. The last attribute contains labelled data with two classes where 26.53 % of total customers are labelled as "indicating true customers Lc. categorized as churning customers and the remaining 72.46 % customers are labelled as“ indicating false customers i. sorized as non-churning customers. The attribute selection depends fon the results of techniques of feature selection that find useful, the most similar and effective attributes to predict the churning customers, A total of 5174 are non-chumers and 1969 are churners, The dataset contains 16 categorical columns and 5 numeric columns. The dataset helps to figure out customer prophecy and build retention possibilities. Figs. 2.and 3 shows details of database. 3.1.1. Data preprocessing ‘This step is required to remove all the irrelevant and dirty data of real world, As the data is congregated from many resources its Important to overcome this issue. Without execution of this step decision makers cannot predict outcomes even ifthey did it won't be Fig. 2. Count of target Variable per category.SK. Wah a Ress Gono and Opinion 14 (2024) 100842 apes Fig. 8. Database attribute detail correct that leads to no quality data and no quality mining. To solve these issues following methods are considered for cleaning the data, Likewise, there are many such features taken as input to predict model. Classification of data and performing pre-processing cleans the data and makes it easy to use. Data classification and pre-processing clean the daca and make it easier to use. There are three steps in data preparation.SK. Wah a Ress Gono and Opinion 14 (2024) 100842 ‘+ A data comparison is carried out in the first stage to identify redundant data. The properties that are repeated are immediately removed ‘Fach instance searches for the missing value in step two. If missing values are detected, replacement procedures are used; otherwise, missing values are replaced with the most suited value. When there is no way to replace the instances, they are discarded. + The final stage entails defining the type ofeach valuc inorder to eliminate extraneous data Ifthe information js no longer useful it is discarded, In other words, noisy data is deleted from the data that has already been pre-processed. Check the data types ofeach column frst. ecause the total charges column’s datatype is object type, we convert it tonumeric and make a duplicate ofthe underlying data to ‘manipulate and process. This conversion is performed using panda's library imported inthe program and using to-numeric function of library. Next, check for duplicate values in dataset, but there are no duplicate values present in dataset. Further continue checking for ‘missing values. So according to general thumb rule the feature with less missing values there to fill means values or simply use regression to predict the missing values based on the particular feature. Similarly, in case of feature with high number of missing ed, Now by using is null and sum function there are 11 missing values present inthe Total Charges column so drop ths missing value. The tenure consists of maximum value of 72 and also dividing the tenure into 6 categories of 1-12, 13-24, 25-36, 37-48, 49-60, 61-72 make easy in visualization of column data of tenure period. Dropping the columns that are not required for processing. It was found the ‘customer ID isnot useful and contains unique value which won'taffect the prediction results, so drop customer ID and similarly drop the tenure column which is not necessary to evaluate the results values, it would better to drop those columns due to less analysis and insights. Also, the columns with more than 30-40 % are delet ‘Name: tenure.group, dtype: int64 Univariate analysis - In data exploration, the distribution of individual predictions by churn willbe determined, and the number of. ‘chumers representing non-seniors will be determined. Here, 1 represents senior citizens, and the plot demonstrates that ifa client isa senior citizen, they are more likely to churn. Similarly, when a candidate is single and does not have a partner, the partner churn ratio is high, People with phone service are more likely to chura, similarly payment method, if there is an electronic check appears higher than the case of credit card appears lowest churmer because of this they could be having auto debit features, which is one of the {important features for churn, similarly remaining people with phone service are more likely co churn, similarly remaining people with phone service are more likely to churn, similarly remaining people with phone service are more likely to churn, remaining figures also shows the churn count respectively above selected features. ‘Bivariate analysis - Its used to find a valve prediction fora single variable, Correlations berween variables are simple to find. A relationship between two variables is defined as bivariate. There are numerous features in our dataset, and we presented the results, ‘Two variables were examined, and two new data frames for churners and non-churners were generated. A function is created that maintains a data frame that is passed with column, title, and hue information for each feature, similar to how bar graphs for different features may be shown in diagrams. Gender characteristics considers 2500 female and male participants, with a churner/non-churner ratio of around 50 % foreach gender. According to gender feature analysis, females are more likely to chum ifthey have a relationship, bbut males are more likely o churn if they do not have a pariner. The groups are classified into churers and non-churers. 31.2, Feature selection This is an important step in achieving our model's goal. Unnecessary data is discovered in datasets while training the model, resulting in a reduetion in model accuracy. As a result, feature selection on a dataset is used to solve these issues. The following are the advantages of feature selection: + Reduced over iting means less chance of making conclusions based on noise «+ Accuracy is improved because there is fewer misieading data, ‘Training time is reduced providing lesser algorithm complexity with algorithms that tain faster «Applying Machine Learning Algorithms [1] Decision tee: Decision tree are used to solve both classification and regression problem in the form of trees that ean be incrementally updated by splitting the dataset inco smaller dataset, where the result are represented by the leaf node. Each branch represents the possible deision outcome or reaction. Iti like a lowehart diagram that shows the various outcomes from a series of decisions. I can be used as a decision-making tool, Decision tree has some series of same craft questions regarding attributes of test data record and its use to solve cassification-based problems, every ime it gets solution from itand follows until the final conclusion of clas label record, Then visit several decision trees for achieving target value. It can be either true or false, Naw pick # majority vote of tres or count the target values provided, then based upon decision tees predictSK. Wah a Ress Gono and Opinion 14 (2024) 100842 ‘feustomer churn i true or false for research analysis, or for planning strategy. A primary advantage for using a decision tree Is that itis easy o follow and understand. Decision tree classifiers simple and adaptive classification technique, ths is basically implying a straightforward process to analyze and solve the problem. [2] Random Forest tree: The random forest i a classification algorithm consisting of many decision trees. More numbers of tres in the forest led to more robustness in prediction with higher accuracy. I¢ uses bagging and feature randomness, when building each individual tree to create an uncorrelated forest of tees whose prediction by committee is more accurate than that of any individual tee. In this algorithm, each tree will have its own output from a dataset provides, such that output generated willbe considered from the majority of tres. Decision trees made inthis algorithm are of numeric type in which the tree picks any random attributein the dataset. Advantages of Random Forest that ithelps to solve both regression and classification problems. (3) Important Features or factors responsible for churning according to Decision Tree model: Finding the important factors that are responsible for churning makes it possible to find the service that is required to customer to prevent from attrition, This can be done using feature importance. Here the feature is ranked according to their importance. The most important feature atthe top of lst, while leat important are at end of list In Fig. 4 it is seen that Contract-Month-To-month ranked frst in the list with its importance as 0.517, Total Charges with {importance as 0.104, No internet Service i 0,093, DSL Internet Service as 0.0795, Monthly Charges as 0,0517, Contract of Two years as 0.0464. Contract of one year as 0,0419, tenure group of 1-12 as 0.0397, No Streaming Movies as 000696, Fiber Optic Internet Service as0.00543, and last in list is with No Paperless Billing as 0.004727. (4) Important Features or factors responsible for churning according to Random Forest Tree modek In this model from ig. S the features Importance ranked from 1st co last as, ‘© Contract month to month = 0.1245 «Tenure group = 0.10518 «Internet Service Fiber Optic = 0.07693 ‘Total Charges = 0.07000 ‘= Contract Two year ~ 0.06123, ‘ Tenure group 61-72 ~ 0.04302 + Online Security yes = 0.03977 ‘+ No tech support = 0.03943 ‘Online Backup No Taternet Service ~ 0.0360 + Streaming Movies No Internet Service = 0.03479 + Internet Service DSI, = 0.0338, '» Monthly Charges = 0.0309 ‘+ Online Security No Internet Service + Contract One year = 0.02819 ‘+ No Online Security = 0.02620 ‘Tech Support Yes = 0.02491 ‘Tech Support No Internet Service ‘Partner Yes = 0.02422 ‘+ No Muliple Lines = 0.001289 024313 Feature importances ¥ z g i 3 i Fig. 4. Feature importance for decision teeSK. Wah a Ress Gono and Opinion 14 (2024) 100842 dV) ih i a Fig. 5. Feature imporance for Random Forest classifier. Gender Male = 0,0011822 ‘Multiple Lines No phone service = 0.0011689 No phone Service ~ 0.001107 Gender Female = 0.001101 3.2, Multiple dataset ‘There exist multiple types of data in SyriaTel [25] as classified below which may be applied to construct our churn prediction model used in telecom sector. I is observed to present dataset structure using spark engine, its required to include phase-based exploration ‘with suitable pre-preparation for algorithms based on clasifcation. 3.2.1. Towers and complaints database Inthis database, the detailed data of actual location is shown asin the form of digits. The serial digits are mapped with database of towers which offers the actual location of the transaction and providing stat, city, area, sub-area, latitude and longitude. Database including various complaints provides all submitted complaints along with inquires statistics based on coverage, any problem related to the telecom business and issues related to packages and offers, 3.2.2, Customer data Customer data comprise the data-based contact information and services of customers. Moreover, various packages, services and offers taken by customer. Also, it contains CRM system including information generated from all customer GSMs such as gender, birth date, the location and type of subscription ete 3.2.3, Network logs data Network logs data includes sessions related to calls, SMS and internet for each transaction used by telecom operator such as required time to Initialize a session for eal ending and to check internet status. Moreover, it represents whether session is expired or not due to error occurred in the incernal network, 3.2.4, Mobile IMEI information ‘Mobile IMEI information comprises the model, type, model of the mobile phone and whether mono or dual SIM device. Data may have large size which may require information in detailed. It requires aot of time for understanding. It also needs to know the original sources along with format for storage. Moreover, related to these records, data must be linked to each other logically using relational databases that actually represent customers detailed information, 3.2.5, Gall details records Call Details Records (CDRs) includes modifiable data related to MMS, calls, SMS ete. Also transaction made by customers using {internet which is ultimately generated in the form of text files. 4. Experimental analysis [After preprocessing on dataset following are the observations:SK. Wah a Ress Gono and Opinion 14 (2024) 100842 + Strong correlation exists between tenure and total charges, means as tenure increases so does total charges. + Strong correlation exists between monthly charges and total charges as well. + Tenure and Contract duration seems to be strong factors in éetermining churn. ‘+ Among service types, phone service seems to be most popula. + CSP should investigate if customers receiving digital invoice have any concern with understanding the bill details. + Also, they should encourage customers to move to automated payment modes to improve customer experience. + Gender does not play an important role. However, CSPs should take care of the experience of senior citizens. From Figs. 6 and 7 itis sean that churn is higher when monthly rates are high. Even with modest overall charges, there is large ‘churns, as shown inthe Fig. 7. When all three parameters (Total Charges, Monthly Cost, and Tenure) are combined, higher monthly charges at low tenure result in lower total charges, implying that ll ofthese characteristics are associated to higher churn Fg 8 shows used all features. 4.1, Experiment analysis using decision tee classifer So, in our proposed model using a Decision Tree classifier, the obtained accuracy of the model is 78 %, which is very low, and. printing the classification repors led to the dataset unbalance, which results in less accuracy. Decision Tree clasifer took Number of Leaves - 331, Size of the tre - 552, Time taken to build model-1.06 s for execution. ‘As a result, the accuracy of the Decision tree classifier model before up-sampling & ENN should not be used as a meaningful ‘measure because it leads to unbalanced datasets. Asa result, when checking recall, precision, and Fl scores fr the minority class, t's clear that the precision, recall, and FI ratings for Class 1, ie. churned consumers, are far too low. For up-sampling training data into a decision tree classifier that differentiates into x train and y train and creates a prediction variable and calls the classification input to process input to produce output with accuracy, and using SMOTE (synthetic minority oversampling technique) by performing oversampling and cleaning using ENN (Edited Nearest Neighbors), the dataset is balanced with dataset values of 492, 40, and 599, 24, and it provides a solution. Tables | and 2 shows details results of Decision tree classifier ‘model before and after up-sampling & ENN. The precision of the clasifier matrix is 93 %, the recall factor i 98 %, and the F1 score is 93 % and provides 93.85 % accuracy. 4.2, Experimental analysis using random forest algorithms Missing values are handled with carefully, and accuracy is maintained. 1's even capable of hendling big, multi-dimensional datasets. So, the accuracy of our model using the Random Forest Tree classifier is 98.91 %, Table 3 shows results of Random Forest Classifier model before Up-sampling and ENN. There are several techniques used in random forests. Generally bagging technique is used known as ensemble classifier. or the minority class of churned consumers, employ a classification matrix to increase the model's performance. The imbalance database and its merely oversampled minority class can be addressed by using SMOTE (synthetic minority oversampling technique) and ENN (edited nearest neighbors). In the required minority class, it would include duplicate examples. ‘The results of the Random Forest classifier model after up-sampling and ENN are shown in Table 4. Asa result, the random forest classifier predicts churn with an overall accuracy of 99 %. The classifier matrix has a precision of 99 %, a recall factor of 99%, and an accuracy of 99.09025616471152 %. 43. Survival analysis ‘The survival analysis technique is a valuable statistical technique for predicting how long a client would keep a subscription when. Monthly charges by churn Monty hares Fig. 6. Monthly charges by ChurnRess Gono and Opinion 14 (2024) 100842 SS wown ‘wal charges Total charges by churn Fig. 7. Total charges by Churn 0004 SK. Wagh eo Suppor a6 Perarane 220077% Fig. 8. All earures in bar graph. Neo ances 150, reion| bet ‘one Gas stance Income lassie stances Gas 0) as 1088) Weighted ag Results of decision tee elstifer model before up-sampling & ENN, ‘TableSK. Wagh a Ress Gono and Opinion 14 (2024) 100842 ‘Table 2 Results of decision tee classifier model ater up- sampling & Gas 0 Wo) 095 092 oss Cas 1%) se ase aot Weighted ag 095 095 oss Coney Gasiied stances esis sah ‘Table 3 Results of Random Forest ce clasier model before Up Gass Wo) 099 0 89 Corey Gasitied stances on sas1% ‘Table 4 Results of Random Forest laser model after Up-sampling and ENN. Gao Wo) 098 om as 104) 00 095 099 they churn, "Time to event analysis" is another name for survival analysis. Customer retention is heavily influenced by survival analysis, To avoid churn, we concentrate on a large number of consumers witha short survival span, This analysis determines the value ‘of a customer's lifetime. ‘The event is defined as the precise time when a customer cancels or leaves a subscription, and the time is specified as the time when the consumer joins the service. ‘Survival function: = st Pr (tt) =1-FO) =dr Here T = event time, (= density funetion 44. Cox proportional hazard model Time-to-event’ data is analyzed using the Kaplan-Meier (KMD approach, Al-cause mortality sa common outcome in KM analyses. However other outcomes such a the occurrence of a cardiovascular event could also be included. The Cox Proportional Hazard model {s useful to predict beter survival probability of individuals. n this model, some characteristics inelude partner, monthly casts, phone service, gender, and remaining variables are covariates, which impute onthe survival probability taking into aecount each customer's tenure a the time they churned. All variables and survival funetions are likewise included in this model ‘The log-hazard of an individual is linear function of thelr variables and a population-level baseline hazard that changes with time, according to Cox's proportional hazard model. Mathematically: ‘There area few things to notice about this model: tne baseline hazad,b0, has the only temporal component (1). The partial hazard {sa time-invariant scalar element in the preceding equation that solely raises or éeereases the baseline hazard. As a result, changes in variables will only affect the baseline hazard, Fig. 9 show the coefficient in another way. For instance, the coefficient for PhoneService Yes (having a phone service) is around 0.69. In the Cox proportional hazard model, @ one-unit increase in PhoneService Yes increases the baseline hazard by a factor of exp (0.68) =2.00, or nearly 20. A greater hazard indicates that the event is more likely to occur. The hazard ratio is define as exp (0.69) divided by 1 ‘The key thing to notice here is that although though the (coef) values for covariates MonthlyCharges and gender Male are close toSK. Wagh et Ress Gono and Opinion 14 (2024) 100842 ener Yor reeng Nota nce +--+ 0s . Fig. 9. Significance of covariate in predicting churn risk, zero (~0.01), the former stil has a substantial impact in forecasting chura, whereas the later is inconsequential, The reason for this is because MonthlyCharges is set of continuous values that can change from one month to the next. "ig. 10 shows survival curve forthe selected customers. So, based on the Fig. 10, itis concluded that customer 2 has the highest chance of churning. Creating survival curves at the customer level lloves us to develop a proactive plan for high-value clients for various survival risk segments slong the timeline, 45, Retention stracegy Some high-effort interactions to support customer retention in the telecom industry include Customers Must Be Educated, attractive 10 09 08 o7 06 os os 03 02 é © ® 2 o o @ 7% Fig. 10. Survival curve forthe selected customers (Customer 1,2, 9 and 4).SK. Wah ra Ress Gono and Opinion 14 (2024) 100842 Offering Plans, repeat contacts, an emphasis on wasting customers’ time, shoddy self-service, avoiding unneeded robotie service and giving complicated instructions. 5. Conclusion ur system works effectively for achieving the major objective to analyze the various machine learning algorithms required to develop customer churn prediction models and in identification of churn reasons in order to give them with retention strategies and plans. Our system focuses on resolving an wnavoidable issue Customer churn arise in the telecommunications sector for a variety of reasons. The most perplexing aspect about client rumover is that churn is impossible to manage. Moreover, customer turnover has several causes, some of which are visible and others which are not. On the other hand, operators in the telecommunications industry should be aware that client loss will occur sooner or later and they must be prepared to respond. In this context, our system plays ‘important role on solving the problem of customer churn has become critical for telecom companies’ survival ‘A customer chur system is based on machine learning methods, a decision tree classifier, and a random forest algorithm. Here, prior to and after up sampling and FNN, both techniques are applied tothe system. Decision tree classifier models initially produced poor results on an unbalanced dataset that did not take ultimate accuracy into account when matrices were used to evaluate the model. In comparison to a decision tree classifier, a random forest classifier produces better results. With an overall accuracy of 99 %, the random forest classifier predicts churn. The classifier matrix has a precision of 99 %, a recall factor of 99%, and an accuracy of 99.09, "6, System comprises churn prevention that is based on survival analysis using Cox Proportional Hazard model and retention plans ‘Survival curve is used forthe selected customers plays an important role in customer churn prediction. Inthe telecom sector, it appears ‘obvious that lowering customer effort is a method to boost customer retention. This churn prevention system may be made more complex and sophisticated in the future to provide better and more precise recommendations. More advanced algorithms such as deep learning recurrent neural networks that help to identify nonlinear complex relationships between data variables which may be applicable in future study to estimate survival likelihood. Moreover, eustomer chur prediction in the telecommunication sector is possible a rough set approach and data certainty (24) in future. Financial and ethical disclosures ‘This work is not supported fully or partially by any funding organization or agency. Declaration of Competing Interest ‘The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Data availability ‘No data was used for the research described in the article. References [1] Uh, Raza B Malik AK, van Ms SU, Kim SW, A chur prediction md! wing random forest analysis of machine laring ecniqus fr churn (2) Almed AAO, Mabesvad D. Chum prediction huge tecom dra sig hyd rely basd casseaon. Eyypafon 32017183): 215-20, s/o (a1 EV, Ravamar 5G HSK As fetter for ature selon Lo edit csimer hur nace indus: Proceedings ofthe It rural (Gonfereace on Advancs in nfaton Techology (GAIT 2019. p. 174-9 ps /do or 101109 eaTT47O88 2019 898797 (4) sin, Khana A, Sra Chur pedi in feleommuncaton sang logic regres a logt boos. Procodia Carp St 202167101-12, {5} Pamina J, Boshi R,SathyeBama , Soundarya 8, Srathi MS Kiathika S,Aiswaryadci V3, Priyanka 6. An efectve classifier for proietng chur in teecommuneson J Ade Res Drm Control Syst 201%11(1-Speil Inte) ip arcu mote 99917 (et Berger, Kampen. User modeling fr chur pedicton in ecommerce, EEE Intl Spt 2018,4(2}-44-52 ps//de.org0.1109-MIS. 20192855788 Naeh-api (71 Gaur A, Dubey R reiting cuter chur peicion in elec actor ning wir machine learing ecigues I: Proceeding ofthe interasonal fenference on edvanced computation and tlaomamuniation UCACAT); 2018p nip ofp 10 1108/UACAT 2018895978 {a Bora Reearing Ale Ried G, Dad MM The srbitecrre of charm pedi sat aed on srs mining Pron Ari itl App 2012:157-6 (9) ‘dem, var S. Customer chur pelo in eeeemmuneation. fo: Procedings ofthe 25 sina roesing and communicsionseplcans conference, SIU 201% proceeing, 2015. p 23605. ip do np 10.1109/SI 20157 25808, 10) Devsendt Berevodts J, Veotke W. Why You sould kp reiting customer hart ad start sing wplt mel Se 2021:548:497-515, boson tovtove/ ns 201912073 09) (U1) San Kamar A, Chandra D- sarey on customer chur redition using machine learning techniques. nt. Compt Ap 20165154136. ips/eons/ 112) Kralove 6, Gotvee 8. Moding data mining applications for pedcion of prepa chur in tlecommunication serves. Automatica 201, S191275-88, 11) Hg eB, Wang HY. Aig dt mining eco chum mangement Ep SAP AMEN) S124, hpSK. Wah a Ress Gono and Opinion 14 (2024) 100842 (14 Ani 80, ALS 8 Preditng stone cnn ina tre on ion eh nme hn rig tg Th doiapssrontnon raoqyrosattojerne gar (15) Kall on Annes Zee Da ed FZ DA Catone har edn nelson ay Do at nn Frnt he aS pens tnd! cnn & opis (SHAY BY poe’ /orne O09 Go SN tas) tant eg chang oa 31 Apo moe of aoa cco cstv an mpi ea of lcm indy na ice yea oe BLA he anne 1D Ses n: Aee 160. (17) Nar AN Ur stb NSA omer hora rion ang ong cosa 86 Teil 20181121. (1) ln a Chad Coercion tes mahine eng proc, Gpting 202042714 p/or/0100/ U9) Rajamobamed , Mapokaran J mprovd rit card chur precton bas on ough clustering and supers learning techniques Caster Comput 2018;21 Gnas maps tan ive roasts (a0) Ane Adan Ahaha dive ang spac fr custome hur radon nthe elconmunitn Indy wg ear compton {nina faye Alsen Conparanag ay. tse) cong oon me asso tan hain ALobiin 7, Sah hie Ks Us Ras Daa Aware ner hun edison i te lemon me. 1 Stpesmper mosreapneas. on hkoy/toaoay a os (2) hate a shen Kotak, Lipson A, Raha, Arwor 8 Crome coe chm icon nema: compute of A ndorton nda tn Manag tioeSOe. sia oe jeanne (2 Nin do Naa, Alt kn ng K Comer carn peo he common ec sng 04h 8 pei SSincoaputse NPIL Nine oa ele eter te 2a (24) Rado Shah Rati aps Marat, A GR, na Cater hm rin coin inks ng ate Sun. mr seeabi9aa-901 re tales To note fre as) Sn fr Asan Coren elon lectin ing aching ni at ator. Jig Da 2016-24 s/