Order from us for quality, customized work in due time of your choice.
Introduction
Today telecom industry is having ferocious competition where telecom service providers are trying to satisfy the customers. Despite various methods to attract and retain customers, customer churn remains a major challenge. In a more simplistic term, churn rate is the percentage of total customers who no longer use the telecommunication service of a particular operator after sometime as they turn towards rivals services. Thus, having a high churn rate can adversely affect the business.
Usually the customers churn due to several reasons; firstly, customers sign up for a product which dont fully meet their requirements and/or they themselves dont know what they want, thus they land up finishing with competitors. These customers dont usually form a fidelity with the company. At the slightest issue, they tend to react negatively. On the other hand, if customers are explained and made conversant with requirements, they tend to stay happy with the products and services at their disposal.
Secondly, buying a product/service implies that some customers are looking for specific outcome. In case the key features dont seem to work right, clients tend to get frustrated. So, the best way to counter this is to make sure that the company helps the clients in achieving their goals. Thirdly, frustrated clients dont want to wait or chat with bots; neither do they want to wait for long before they can talk to someone who can help to solve their problem. Putting customer first is key to differentiate a brand and retain customers. A good customer service will enable to help those clients.
Whenever, product/services are not unique, leaving for competitors product becomes so easy that customers tend to believe that the other product(s) meet(s) the needs much better. To counter this, unique products or services tend to attract clients to that particular service. Besides, product reliability is of primordial importance. In case of slightest doubt, the clients tend to lose confidence in the product/service. To avoid this, we need to make sure that the service/product is bug free with minimal downtimes. The trust can be reinforced by clear communication towards the clients regarding any issue and also clarify their doubts if any.
Another churning reason is that clients tend to be fed up with a particular service/product. Customers do feel the need to innovate and cater for their future requirements. This can be done by tracking customer satisfaction over time. Building such rapport will have a much closer understanding and capture room for improvements. Likewise, pricing of any service/product needs to be balanced, where customers need to see value in what they are being proposed. Proper balancing in the willingness to pay and the value of the service is important.
Digging the churn predictors, we can try actions to retain those potential users. One way to do this is to recommend personalized product/services to clients using an AI based recommendation system. It can allow proposing customized solutions to customers. Furthermore, the AI system can be enhanced by a 360 feedback to track the performance and efficacy of the recommendations to learn how to make better and more precious prediction analysis.
Predicting churn with an AI recommender system will undoubtedly provide personalized user product so that retention probability is considerably increased. The recommender system examines the client-product relationship along with the conditions of the client. The primary goal of building such a system is to simplify the clients search of products or content. It restricts the variety of choices, so people can focus only on those products which they are really interested in.
Problem Statement /Research Problem
In telecom industry, to acquire new customers it comes with high cost. In order to decrease this cost, working towards retention of existing customers is important. Customers leave the network due to several reasons and they give an indication beforehand if they want to voluntarily or involuntary leave. Analyzing the existing data will help in predicting potential churners and actions can be taken accordingly in an attempt to retain these potential clients.
Whilst it is good to have prediction, the loop remains open as there should be actionable items so that the potential churners are retained in the telecom network. So following the churn prediction, the actions will be derived from an AI recommender system which will work in a hybrid model i.e using both collaborative and content-based filtering recommendation. This allows for predicting a better recommendation which can help in winning back the customers.
Today, even if the company is trying to manually do some kind of prediction over churn, they dont have any digital platform which can regroup and provide recommendation of the actionable items. Everything today is being done with large percentage of error rate. Even defining some kind of business rules, churn prediction didnt function as expected due to large data volume processing.
Voluntary churners are those who consciously stop using a particular service or product. Involuntary churners are also known as passive churners who leave the service or product due to some unavoidable reason such as insufficient funds, server errors etc.
So, the problem statement can be summarized as to reduction of cost for the company to acquire new clients, and provide recommendation for retention of the potential churners. This in turn will definitely reduce the OPEX of the company in the long run and increase profits.
Rationale of the Study
Predictive analysis of potential churners will help the company to decrease cost and increase customer satisfaction. It can vary from analyzing the number of calls to call centers, call drops, cell site overloads, customer trends in using the network etc. If we can predict what the customers are dissatisfied about, appropriate actions can be taken by specific departments in the company.
There are several tools already available in the industry for this. But due to data sensitivity, cost etc., the company doesnt want to use any third party tool. Thus, my company is trying to build its own predictive AI model using its data lake available. Moreover, there are different algorithms to create churn. This thesis proposes at least 5 mechanisms which can be used for prediction in order to bring the most precise version of churners.
The prediction output will be at least two output columns; the first will be a binary value which will highlight whether the customer is a churner or not. The second column will highlight the probability which will determine the possibility of the customer to be a churner.
The study will also have an AI recommender system which will generate recommended actions for each potential churner. This will be based on previous peer-to-peer data and the users previous usage history. In this way, the recommendation will target the return on the network for the mobile user. Prior to providing recommendation, the system can create clusters of the potential churners which will regroup them based on certain criteria.
For the purpose of this thesis, the scope will be limited to the prepaid customers only.
Research Questions
For the study of this proposal, there are different research questions which will be addressed.
RQ1: What are the different AI techniques which can be used to predict prepaid telecommunication churners using deep learning?
Different techniques will be explored to identify which combination(s) can yield better results in this prediction. The implementation and comparison of different algorithm (at least five in my case) will enable the comparison so that we can know which one better analyses the data for churn prediction. For example (Alamsyah and Salma, 2018) predicted that employee churn varies based on the different evaluation model. The researchers found that the three popular algorithm Naïve Bayes, decision tree and random forest have yielded different accuracy level. However, it can be that deep learning will be more of use in order to unearth unseen tendencies and trend.
RQ2: Can different AI recommender system techniques be used to retain, potentially churning, telecom customers by providing solution(s)?
Based on the first request question, the output will be used in the second part of the thesis. So, once the potential churners have been identified and to produce actionable items, an AI Based recommendation system can be used. First and foremost, the potential churners can be clustered based on different like network-related churners, service-related churners, social-impact churners etc. Once this is done, the AI recommender system can make use of any kind of technique in the recommender system (collaborative, content-based, hybrid) to recommend the retention of the customers. For example, with peer-to-peer data and the users previous usage trend analysis, an AI recommender system can be optimal in providing a solution to win back the mobile user. He/she will only feel valued by the organization to be provided with such kind of approach. Today customers are at the center of each business and solution(s) is the key for their satisfaction.
The different recommendation techniques are:
-
Collaborative filtering: recommendations based on similar preferences defined by different users in the past.
-
Content-based filtering: recommendations based on previously bought items/services by the user. It can be based on name, location, preference etc. of the user previously.
-
Hybrid filtering: a combination of both types of filtering defined previously.
Aims
How can prediction of potential churners help in retention of customers in telecom industry? The primary aim of the project is to have a cost reduction in the OPEX of the company while targeting an improved customer experience throughout the lifecycle of the customer. In this world of fierce battle of mobile operators, it is vital to not only meet the expectation but also exceed their needs. This, in term, only increases the company profitability metrics. While attempting to have a predictive churn model, it is important to define and understand what is churn for the company. Churners can be voluntary or involuntary. However, a clear distinct is important.
Scope
To begin with, the scope of the project will be based on prepaid deliberate churners. This subscriber base experiences the highest churn rate. In this category as per (Santharam and Krishnan, 2018), there can be incidental and deliberate churn. Deliberate churn occurs due to number of reasons which can be related to price, competitors offerings, quality of service, number of incidents. Incidental churners are those who leave due to some location changes or financial position.
Model Aim
The aims will be to identify the churners who will potentially stop using the service/product after some time. The time definition is important here so that the track records (CDRs) and other information can be used. In the current context, the churn prediction will be for the forthcoming two months. During period, the company has ample time to work on the retention period and try to reassess the prediction model.
Once the predicted data is available, the potential churners will be clustered based on different parameters. Out of these parameters some of them are: number of call drops, number of wrongly billed clients, number of complaints, number of rejected SMS, number of failed payments, competitors price etc.
Over this clustered information, the AI recommender system will work towards providing recommendations for retention of the customers. For example, what products/services can be optimized to solve the issues encountered by the client. Besides, there can be priorities set in this recommendation system based on the number of clients in each of the cluster identified or based on revenue generation.
As a final step, the aim is to improve the predictive model by using the retention rate obtained as defined below:
Contribution towards sustainability
In Mauritius, most of the sim cards are plastic/electronic base. The project also aims to reduce the emission of these small electronic and plastic items by allowing clients to make use of it for a maximum duration which in term will reduce the plastic emission in our environment.
Literature Review
Several researchers have tried to solve the churn problem and use multiple techniques for recommender systems (Table 1). According to Weber and Schutte (2019), the use of marching learning is both practical and highly developed. According to Ammari (2022, p.317) and Sharma et al. (2022), artificial intelligence has become an integral part of marketing and social customer relationship management in this era of Web 4.0. For instance, Carrefour and Sirgul recently launched a small retail shop which relies on artificial intelligence to personalize the customer experience, reposition products, reduce food spoilage and provide real-time alerts (Zabala, 2018).
A decision tree model makes use of R programming to build a churn prediction model. (Bhadoria and Mathur (2018)) showed that with a decision tree model, churn prediction is better than the logistic regression model. The accuracy of the confusion matrix shows that the Decision Tree model outperformed the Logistic regression. Even (Dahiya and Bhatia, 2015) confirm that the decision tree performs better than the logistic regression technique. However, (Alamsyah and Salma (2018)) added another algorithm of Random Forest for the predictive analysis of the previous two algorithms. Based on the experiments, they concluded that random forest has a better accuracy of 97.5% and is more reliable for this kind of prediction. (Ahmad et al. (2019)) used four different algorithms, out of which the best accuracy results were from XGBOOST algorithms (93.3%). With 10-folds cross-validation, the experiment was performed. Though there were issues with the dataset, like an unbalanced dataset, they were tackled accordingly by methods like under-sampling. (Idris and Khan (2012)) proposed a genetic programming model with an AdaBoost-based methodology, which was advantageous for the forecast.
Churn Prediction
Some efforts by companies have proved retrogressive in maintaining their users. For example, Ascarza et al. (2016) carried out an experiment which revealed that recommending prices to customers can increase the number of potential churners from a system. While having customer churn prediction is good, it is even better to have a recommendation AI-based system that can help retain the potential churners. In this regard, Sree Buddha College of Engineering and Renjith (2015) proposed a model which is related to E-Commerce customer churn. The model firstly builds a forecast of churners; then, it uses clustering algorithms (K-Means and hierarchical clustering) to form groups based on their profile and behaviour. Lastly, the AI-based recommender system uses a collaborative filtering mechanism to recommend actions for the potential churners in the Business Customer (B2C) environment.
There are researchers who have used hybrid models to predict customer churn likelihood. Renjith (2017) proposed a framework founded on a support vector machine to help predict E-commerce customer churn. The author also proposed a hybrid recommendation strategy to suggest personalized retention actions. Dingli et al. (2017) demonstrated how a business could utilize its transactional data features to predict churn in the retail industry. The researchers demonstrated how extracting and analyzing data available within the Point of Sales systems can help in predicting customers buying patterns. The researchers obtained their data from a local supermarket, which they used to identify the churners using Convolution Neural Networks and Restricted Boltzman Machine learning techniques. The Restricted Boltzmann Machine achieved 83% accuracy in predicting customer churn. In another study, Ullah et al. (2019) proposed the use of the Random Forest (RF) algorithm, a classification and clustering technique to identify customers that churn, to predict the factors leading to their churning using information correlation and gain attribute ranking filler. The RF correctly classified the churning instances in 88.63% of the cases, proving slightly better than the Restricted Boltzmann Machine Learning Techniques at predicting customer churn.
Restricted Boltzmann Machines for Collaborative Filtering was one of the first related approaches in the neural network literature. Lin and Gao (2021) proposed a two-headed transformer-based network to predict unlocked sessions and to predict user feedback through multitasking with click behaviour prediction, session-aware re-weighted loss, and randomness-in-session augmentation. The method attained a categorization accuracy of 0.39224 on Kaggle. Hidasi et al. (2015) proposed an RNN-based technique for session-based recommendations that is able to model an entire session instead of basing recommendations on short session data. The model also introduced modifications to classic RNNs by introducing aspects like ranking loss function. Tan et al. (2016) proposed the application of two techniques for the improvement of RNN-based models for session-based recommender systems: data augmentation and a method for accounting for shifts. Sheikhoushaghi et al. (2022) proposed an RNN for oil forecasting. In a more recent paper, Wang et al. (2015) offer a more general technique in which a deep network is used to extract generic content features from all sorts of objects. Then these features are put in a typical collaborative filtering model to improve the recommendation performance. Situations with insufficient data on how users interact with products seem to benefit most from this strategy. Saias et al. (2022) proposed a churn risk prediction system that can help cloud server providers recommend adjustments at the service subscription level to avoid CSP customer loss and promote rational resource consumption. The researchers built a training data set from customers data, the service they subscribed to, and their usage history to predict churning trends using a machine-learning approach. The researchers built and evaluated classification models based on AdaBoost, multilayer neural networks and random forest algorithms. The forest-based model produced the churn prediction results, having 0.997 AUC value and a 0.988 accuracy, with 64 estimators.
Recommendation Systems
Recommendation systems are very important and are applicable in various online supported industries. Lü et al. (2012) highlighted the importance of the recommendation systems, noting that they have a great scientific depth emanating from different research fields. According to Singh et al. (2021), the online marketing recommendation tools help ease a users work by presenting them with the exact product they are searching for or products they can consider as alternatives. Some of the most popular techniques identified by the researchers include collaborative filtering, data mining algorithms like WebMining, and the use of the if-then statement, also known as the association rule. Another important technique for recommender systems is graph neural networks which entail high order connectivity, enhanced supervision signal, and structural property data (Gao et al., 2022).
One of the most vital fields of study in AI is text mining. Masood and Raha (2021) explained text mining, also known as text analytics, as a subfield of AI that makes use of natural language processing (NLP) to organize and normalize the free (unstructured) text included in documents and databases so that it may be analysed or used on machine learning (ML) algorithms. According to Hassani et al. (2020), text mining may be used to find insights, information, relationships, and claims that would otherwise be buried in big textual data. After the data has been sorted, it is structured for graphical representation in the form of clustered HTML tables, mind maps, and charts for further analysis and presentation. Text mining relies heavily on natural language processing (NLP) tools to process incoming text (NLP). Databases, data warehouses, and BI dashboards may all benefit from the addition of structured data obtained by text mining for descriptive, prescriptive, and predictive analytics. Emulating the human ability to understand a natural language like English, Spanish, or Chinese, Natural Language Understanding may help machines read text (or another input like speech). NLP and NLG are subsets of Natural Language Processing, which aim to replicate the human ability to understand and generate written text in natural language for purposes such as summarizing data or carrying on a discussion.
Natural language processing (NLP) has developed over the last decade to become a mainstream technology utilized by widely adopted applications like Siri, Alexa, and Google Voice Search to understand and respond to user requests. Modern text mining techniques have aided researchers in many fields, including healthcare, business (risk management), customer service (fraud detection), and contextual advertising. According to Mach-Król, et al. (2021), it is important to highlight the idea of customer insight regarding the relationship with these notations since these ideas have been structured and prioritized many times; scientific works show this from the previous century in the area of information understanding. Knowledge about consumers that is valuable, unique, difficult to mimic, and which the company is aligned to utilize gives it a competitive advantage.
Creating or extracting value from customer insights was a challenge in the 21st century for the enterprise. Mach-Król et al. (2021) proposed a framework with 12 successive stages to correctly extract value from customer insights. One of its most important elements is to listen to consumers and properly develop the proposed changes. A company should avoid deciding what change is most needed for its proper development based on its preferences. He also demonstrated that listening to consumers may disclose important information by gathering their insights and evaluating them. Song et al. (2016) proposed a deep neural network-based architecture that can model short-term temporal and long-term user preferences to improve the performance of recommendation systems. The researchers proposed a novel pre-train method to train the model efficiently for large applications by significantly reducing the number of free parameters. The resulting model was applied to a commercial News recommendation systems real-world data set. The model significantly outperformed the state-of-the-art when compared with established baselines in experiments.
There are many machine learning techniques that have been used in the telecom industry to help with churn prediction. Liyanage et al. (2022) used machine learning techniques like artificial neural networks and k-nearest neighbors and deep neural networks (DNN) to analyze 21 attributes of 7000 post-paid subscribers of a telecom company. The researchers found that the long short-term memory networks (LSTM), a DNN, produced better results than the machine learning algorithms with an accuracy rate of 82.46%, making it the method of choice for creating the companys final churn prediction model. Gharaei et al. (2021) proposed a content-based clothing recommender system using a deep neural network which eliminates the need to manually extract product features required to predict unobserved item ratings. It is a unique system that incorporates gender specifications as a feature in suggestion-making important demographic information. From the experimental results, the loss of the proposed system is lower than other baseline systems, helping in resolving the cold start challenge faced by new items. It also recommends new, relevant, and unexpected items.
It is possible to model the behavior of online users with respect to the time they spend on a platform. Wu and Yan (2017) proposed a list-wise deep neural network-based architecture for modeling online user behaviors within each session. The model was first trained using an embedding method that pre-trains a session representation by incorporating user behaviors like views and clicks. The researchers then proposed that the learnt lesson be used to create a list-wise ranking model, generating a recommendation result for each session of an anonymous user. They used quantitative experiments on an e-Commerce company to validate the results, revealing that the list-wise deep neural network performed better than state-of-the-art.
One benefit of deep learning is that it can automatically extract characteristics, which helps to both eliminate the need for clumsy artificial screening features and boost the quality of prediction models. A_iro lu et al. (2019) developed a cloth recommendation system that uses a single photo of a user to recommend to them clothe options without having to trace their previous shopping activities or digital footprints using CNN. The system had the following accuracies: color prediction was 98%, gender accuracy was 86% and cloth pattern prediction was 75%.
For the time being, deep learning has proved effective in areas such as image classification, voice recognition, network situational awareness, and high-dimensional time series modelling. According to Geetha and Renuka (2019) the tailored recommendation is a popular area of study and application for deep learning. An AI-powered collaborative filtering method was suggested in the literature. Information from the project side is encoded using the noise side automated encoder; the Person similarity score is then computed; the time SVD++ score is then added, and the final score value is the sum of the two.
Fake users can be a head-ache to online platforms, causing system poisoning. Wu et al. (2021b) proposed the adversarial poisoning training (APT) that simulates the poisoning process, injecting ERM (fake users), dedicated to minimizing empirical risk for the sake of building a robust recommender system. The APT also estimates the influence of each fake user on the empirical risk. APT outperformed baseline models in real-world dataset poisoning attacks, showing its robustness. It also improved the model generalization in most of the cases during the experiments. Lin et al. (2020) studied the shilling attack on recommender systems. They proposed the Augmented Shilling Attack framework (AUSH) with the ability to tailor attacks against RS according to complex and budget attack goals that target a specific user group. The researchers demonstrated experimentally that the frameworks attack was noticeable in diverse RS, based on either modern or deep learning RS. In contrast, the state-of-the-art attack detention model could not detect it.
Guo et al. (2017) proposed an item-based top-N recommendations model that works by learning the item-item similarity matrix, a product of two low-dimensional latent factor matrices learned through the structural equation modelling approach to handling sparse datasets. Experimental results on datasets showed that the model outperformed baseline top-N recommendation models, with the relative performance gains increasing as data becomes sparse. These findings corroborate the experimental results of Kabbur et al. (2013). Another study by Munemasa et al. (2018) proposed deep reinforcement learning based recommender system that uses a multilayer neural network for updating a datas value function to handle sparse datasets.
According to Zhang et al. (2019), deep learning has become the choice for recommender systems due to its effectiveness in retrieving information in recommender systems research. Covington et al. (2016) reviewed YouTube, a large-scale and highly sophisticated industrial recommendation system and found how deep learning has brought dramatic performance improvements. The site uses a two-stage information retrieval dichotomy of a deep candidate generation model and a deep ranking model, underscoring the importance of deep learning in recommender systems. In their study, Rahmani et al. (2022a) found out that the definition of disadvantaged/advantaged user groups played a vital role in making the fairness algorithm and improved the performance of base ranking models in fairness-aware recommender systems.
There are models that have been created for the sake of collecting free-text data. (Tarnowska et al. (2020)) proposed a recommendation system to collect feedback from customers in free text format. Thus the authors base their research on working with unstructured data. In future work, sequential pattern mining has been proposed to be researched in the customer attrition problem. Naghiaei et al. (2022) proposed an optimization e-ranking approach that seamlessly integrates fairness constraints from the producer and consumer sides in a joint framework. The proposed approach was tested against eight large-scale data sets. It showed that it could improve producer and consumer fairness while maintaining the overall recommendation quality. This algorithm has an important role to play in minimizing data biases.
There are certain scholars who have preferred hybrid systems for their robustness in giving results. Saravanan and Sathya (2019) proposed a hybrid feature extraction method involving t-Distributed Stochastic Neighbor Embedding (t-SNE) and Principle Component Analysis (PCA) with Support Vector Machine (SVM) to isolate and list online products according to the products high positive reviews. The researchers predicted that the proposed method would have the complexity, recall, accuracy, and precision necessary for the systems entire accurateness. Abdulla et al. (2019) proposed a personalized size recommendation system to predict appropriate users size based on the product data and their history. The system uses a skip-gram-based Word2Vec model to embed users and products in size and fit space and employs the GBM classifier to predict fit likelihood. Nevertheless, this system is still inferior to other content-based recommendations that exploit recurrent neural networks. One of those systems is the one proposed by Suglia et al. (2017), which entails a deep architecture adopting Long Short Term Memory (LSTM) networks that represent the user preferences and the terms to be recommended.
The nonlinearity attentive similarity model has also been modified to enhance recommender systems. Shan et al. (2019) proposed a nonlinearity attentive similarity model (NASM) using locally attentive embedding for item-based collaborative filtering (CF), based on the neural attentive item similarity (NAIS) model. The researchers introduced novel non-linear attention and location to simultaneously capture global and local items information. Compared with the other state-of-the-art recommendation models, the NASM attained superior performance in normalized discounted cumulative gain (NDGC) and hit ratio (HR). Lamche et al. (2014) also proposed a shopping recommender system that combines critiquing and active learning for the exploratory mobile context. According to the results from the test, conversation Active Learning improves the users experiences. One of the weaknesses of a collaborative topic regression (CTR), which learns from users ratings for items, and item content information, is that the latent representation it learns may be ineffective when the auxiliary information is sparse. Wang et al. (2020) proposed a hierarchical Bayesian model called deep collaborative learning (CDL). The CDL can perform deep learning for the content information and the collaborative filtering feedback matrix. Experiments showed that CDL has a s
Order from us for quality, customized work in due time of your choice.