How Dialog Axiata used Amazon SageMaker to scale ML fashions in manufacturing with AI Manufacturing facility and diminished buyer churn inside 3 months

The telecommunications trade is extra aggressive than ever earlier than. With clients in a position to simply swap between suppliers, lowering buyer churn is a vital precedence for telecom firms who need to keep forward. To handle this problem, Dialog Axiata has pioneered a cutting-edge resolution known as the House Broadband (HBB) Churn Prediction Mannequin.

This publish explores the intricacies of Dialog Axiata’s strategy, from the meticulous creation of almost 100 options throughout ­10 distinct areas and the implementation of two important fashions utilizing Amazon SageMaker:

• A base mannequin powered by CatBoost, an open supply implementation of the Gradient Boosting Determination Tree (GBDT) algorithm
• An ensemble mannequin, making the most of the strengths of a number of machine studying (ML) fashions

About Dialog Axiata

Dialog Axiata PLC (a part of the Axiata Group Berhad) is considered one of Sri Lanka’s largest quad-play telecommunications service suppliers and the nation’s largest cellular community operator with 17.1 million subscribers, which quantities to 57% of the Sri Lankan cellular market. Dialog Axiata offers quite a lot of providers, akin to fixed-line, home broadband, cellular, tv, cost apps, and monetary providers in Sri Lanka.

In 2022, Dialog Axiata made important progress of their digital transformation efforts, with AWS taking part in a key position on this journey. They centered on enhancing customer support utilizing information with synthetic intelligence (AI) and ML and noticed optimistic outcomes, with their Group AI Maturity rising from 50% to 80%, in keeping with the TM Discussion board’s AI Maturity Index.

Dialog Axiata runs a few of their business-critical telecom workloads on AWS, together with Charging Gateway, Cost Gateway, Marketing campaign Administration System, SuperApp, and varied analytics duties. They use number of AWS providers, akin to Amazon Elastic Compute Cloud (Amazon EC2) and Amazon Elastic Kubernetes Service (Amazon EKS) for computing, Amazon Relational Database Service (Amazon RDS) for databases, Amazon Simple Storage Service (Amazon S3) for object storage, Amazon OpenSearch Service for search and analytics, SageMaker for ML, and AWS Glue for information integration. This strategic use of AWS providers delivers effectivity and scalability of their operations, in addition to the implementation of superior AI/ML purposes.

For extra about how Axiata makes use of AWS providers, see Axiata Selects AWS as its Primary Cloud Provider to Drive Innovation in the Telecom Industry

Challenges with understanding buyer churn

The Sri Lankan telecom market has excessive churn charges resulting from a number of elements. A number of cellular operators present comparable providers, making it straightforward for patrons to modify between suppliers. Pay as you go providers dominate the market, and multi-SIM utilization is widespread. These situations result in an absence of buyer loyalty and excessive churn charges.

Along with its core enterprise of cellular telephony, Dialog Axiata additionally provides plenty of providers, together with broadband connections and Dialog TV. Nevertheless, buyer churn is a typical situation within the telecom trade. Subsequently, Dialog Axiata wants to seek out methods to cut back their churn price and retain extra of their present house broadband clients. Potential options may contain enhancing buyer satisfaction, enhancing worth propositions, analyzing causes for churn, or implementing buyer retention initiatives. The secret is for Dialog Axiata to realize insights into why clients are leaving and take significant actions to extend buyer loyalty and satisfaction.

Answer overview

To cut back buyer churn, Dialog Axiata used SageMaker to construct a predictive mannequin that assigns every buyer a churn threat rating. The mannequin was educated on demographic, community utilization, and community outage information from throughout the group. By predicting churn 45 days prematurely, Dialog Axiata is ready to proactively retain clients and considerably scale back buyer churn.

Dialog Axiata’s churn prediction strategy is constructed on a strong structure involving two distinct pipelines: one devoted to coaching the fashions, and the opposite for inference or making predictions. The coaching pipeline is answerable for growing the bottom mannequin, which is a CatBoost mannequin educated on a complete set of options. To additional improve the predictive capabilities, an ensemble mannequin can also be educated to determine potential churn cases that will have been missed by the bottom mannequin. This ensemble mannequin is designed to seize extra insights and patterns that the bottom mannequin alone could not have successfully captured.

The combination of the ensemble mannequin alongside the bottom mannequin creates a synergistic impact, leading to a extra complete and correct inference course of. By combining the strengths of each fashions, Dialog Axiata’s churn prediction system beneficial properties an enhanced general predictive functionality, offering a extra sturdy and dependable identification of shoppers susceptible to churning.

Each the coaching and inference pipelines are run thrice per 30 days, aligning with Dialog Axiata’s billing cycle. This common schedule makes positive that the fashions are educated and up to date with the newest buyer information, enabling well timed and correct churn predictions.

Within the coaching course of, options are sourced from Amazon SageMaker Feature Store, which homes almost 100 rigorously curated options. As a result of real-time inference just isn’t a requirement for this particular use case, an offline function retailer is used to retailer and retrieve the mandatory options effectively. This strategy permits for batch inference, considerably lowering each day bills to underneath $0.50 whereas processing batch sizes averaging round 100,000 clients inside an inexpensive runtime of roughly 50 minutes.

Dialog Axiata has meticulously chosen occasion varieties to strike a stability between optimum useful resource utilization and cost-effectiveness. Nevertheless, ought to the necessity come up for quicker pipeline runtime, bigger occasion varieties may be really useful. This flexibility permits Dialog Axiata to regulate the pipeline’s efficiency based mostly on particular necessities, whereas contemplating the trade-off between pace and value concerns.

After the predictions are generated individually utilizing each the bottom mannequin and the ensemble mannequin, Dialog Axiata takes motion to retain the purchasers recognized as potential churn dangers. The shoppers predicted to churn by the bottom mannequin, together with these completely recognized by the ensemble mannequin, are focused with customized retention campaigns. By excluding any overlapping clients between the 2 fashions, Dialog Axiata ensures a centered and environment friendly outreach technique.

The next determine illustrates the output predictions and churn chances generated by the bottom mannequin and the ensemble mannequin.

The primary desk is the output from the bottom mannequin, which offers priceless insights into every buyer’s churn threat. The columns on this desk embody a buyer identifier (Cx), a Churn Purpose column that highlights potential causes for churn, akin to Every day Utilization or ARPU Drop (Common Income Per Person), and a Churn Likelihood column that quantifies the chance of every buyer churning.

The second desk presents the output from the ensemble mannequin, a complementary strategy designed to seize extra churn dangers that will have been missed by the bottom mannequin. This desk has two columns: the shopper identifier (Cx) and a binary Churn column that signifies whether or not the shopper is predicted to churn (1) or not (0).

The arrows connecting the 2 tables visually characterize the method Dialog Axiata employs to comprehensively determine clients susceptible to churning.

The next determine showcases the great output of this evaluation, the place clients are meticulously segmented, scored, and categorized in keeping with their propensity to churn or discontinue their providers. The evaluation delves into varied elements, akin to buyer profiles, utilization patterns, and behavioral information, to precisely determine these at a better threat of churning. With this predictive mannequin, Dialog Axiata can pinpoint particular buyer segments that require instant consideration and tailor-made retention efforts.

With this highly effective info, Dialog Axiata develops focused retention methods and campaigns particularly designed for high-risk buyer teams. These campaigns could embody customized provides, as proven within the following determine, incentives, or personalized communication geared toward addressing the distinctive wants and considerations of at-risk clients.

These customized campaigns, tailor-made to every buyer’s wants and preferences, purpose to proactively tackle their considerations and supply compelling causes for them to proceed their relationship with Dialog Axiata.

Methodologies

This resolution makes use of the next methodologies:

• Complete evaluation of buyer information – The muse of the answer’s success lies within the complete evaluation of greater than 100 options spanning demographic, utilization, cost, community, bundle, geographic (location), quad-play, buyer expertise (CX) standing, grievance, and different associated information. This meticulous strategy permits Dialog Axiata to realize priceless insights into buyer conduct, enabling them to foretell potential churn occasions with outstanding accuracy.
• Twin-model technique (base and ensemble fashions) – What units Dialog Axiata’s strategy aside is the usage of two important fashions. The bottom mannequin, powered by CatBoost, offers a stable basis for churn prediction. The brink likelihood to outline churn is calculated by contemplating ROC optimization and enterprise necessities. Concurrently, the ensemble mannequin strategically combines the strengths of assorted algorithms. This mix enhances the robustness and accuracy of the predictions. The fashions are developed contemplating precision because the analysis parameter.
• Actionable insights shared with enterprise models – The insights derived from the fashions will not be confined to the technical realm. Dialog Axiata ensures that these insights are successfully communicated and put into motion by sharing the fashions individually with the enterprise models. This collaborative strategy implies that the group is best outfitted to proactively tackle buyer churn.
• Proactive measures with two motion varieties – Geared up with insights from the fashions, Dialog Axiata has carried out two important motion varieties: community issue-based and non-network issue-based. Through the inference section, the churn standing and churn purpose are predicted. The highest 5 options which have a excessive likelihood for the churn purpose are chosen utilizing SHAP (SHapley Additive exPlanations). Then, the chosen options related to the churn purpose are additional categorized into two classes: community issue-based and non-network issue-based. If there are options associated to community points, these customers are categorized as community issue-based customers. The resultant categorization, together with the anticipated churn standing for every consumer, is then transmitted for marketing campaign functions. This info is effective in scheduling focused campaigns based mostly on the recognized churn causes, enhancing the precision and effectiveness of the general marketing campaign technique.

Dialog Axiata’s AI Manufacturing facility

Dialog Axiata constructed the AI Manufacturing facility to facilitate working all AI/ML workloads on a single platform with a number of capabilities throughout varied constructing blocks. To deal with technical elements and challenges associated to steady integration and steady supply (CI/CD) and cost-efficiency, Dialog Axiata turned to the AI Manufacturing facility framework. Utilizing the ability of SageMaker because the platform, they carried out separate SageMaker pipelines for mannequin coaching and inference, as proven within the following diagram.

A main benefit lies in value discount by way of the implementation of CI/CD pipelines. By conducting experiments inside these automated pipelines, important value financial savings might be achieved. It additionally helps keep an experiment model monitoring system. Moreover, the mixing of AI Manufacturing facility elements contributes to a discount in time to manufacturing and general workload by lowering repetitive duties by way of the usage of reusable artifacts. The incorporation of an experiment monitoring system facilitates the monitoring of efficiency metrics, enabling a data-driven strategy to decision-making.

Moreover, the deployment of alerting methods enhances the proactive identification of failures, permitting for instant actions to resolve points. Information drift and mannequin drift are additionally monitored. This streamlined course of makes positive that any points are addressed promptly, minimizing downtime and optimizing system reliability. By growing this undertaking underneath the AI Manufacturing facility framework, Dialog Axiata may overcome the aforementioned challenges.

Moreover, the AI Manufacturing facility framework offers a strong safety framework to control confidential consumer information and entry permissions. It provides options to optimize AWS prices, together with lifecycle configurations, alerting methods, and monitoring dashboards. These measures contribute to enhanced information safety and cost-effectiveness, aligning with Dialog Axiata’s aims and ensuing within the environment friendly operation of AI initiatives.

Dialog Axiata’s MLOps course of

The next diagram illustrates Dialog Axiata’s MLOps course of.

The next key elements are used within the course of:

• SageMaker because the ML Platform – Dialog Axiata makes use of SageMaker as their core ML platform to carry out function engineering, and prepare and deploy fashions in manufacturing.
• SageMaker Function Retailer – By utilizing a centralized repository for ML options, SageMaker Function Retailer enhances information consumption and facilitates experimentation with validation information. As an alternative of immediately ingesting information from the information warehouse, the required options for coaching and inference steps are taken from the function retailer. With SageMaker Function Retailer, Dialog Axiata may scale back the time for function creation as a result of they may reuse the identical options.
• Amazon SageMaker Pipelines – Amazon SageMaker Pipelines is a CI/CD service for ML. These workflow automation elements helped the Dialog Axiata group effortlessly scale their means to construct, prepare, check, and deploy a number of fashions in manufacturing; iterate quicker; scale back errors resulting from handbook orchestration; and construct repeatable mechanisms.
• Reusable elements – Using containerized environments, akin to Docker pictures, and customized modules promoted the convey your individual code strategy inside Dialog Axiata’s ML pipelines.
• Monitoring and alerting – Monitoring instruments and alert methods offered ongoing success by retaining observe of the mannequin and pipeline standing.

Enterprise outcomes

The churn prediction resolution carried out by Dialog Axiata has yielded outstanding enterprise outcomes, exemplifying the ability of data-driven decision-making and strategic deployment of AI/ML applied sciences. Inside a comparatively quick span of 5 months, the corporate witnessed a considerable discount in month-over-month gross churn charges, a testomony to the effectiveness of the predictive mannequin and the actionable insights it offers.

This excellent achievement not solely underscores the robustness of the answer, it additionally highlights its pivotal position in fortifying Dialog Axiata’s place as a number one participant in Sri Lanka’s extremely aggressive telecommunications panorama. By proactively figuring out and addressing potential buyer churn dangers, the corporate has bolstered its dedication to delivering distinctive service and fostering long-lasting buyer relationships.

Conclusion

Dialog Axiata’s journey in overcoming telecom churn challenges showcases the ability of modern options and the seamless integration of AI applied sciences. By utilizing the AI Manufacturing facility framework and SageMaker, Dialog Axiata not solely addressed complicated technical challenges, but in addition achieved tangible enterprise advantages. This success story emphasizes the essential position of predictive analytics in staying forward within the aggressive telecom trade, demonstrating the transformative influence of superior AI fashions.

We admire you for studying this publish, and hope you discovered one thing new and helpful. Please don’t hesitate to go away your suggestions within the feedback part.

Thanks Nilanka S. Weeraman, Sajani Jayathilaka, and Devinda Liyanage on your priceless contributions to this weblog publish.

In regards to the Authors

Senthilvel (Vel) Palraj is a Senior Options Architect at AWS with over 15 years of IT expertise. On this position, he helps clients within the telco, and media and leisure industries throughout India and SAARC nations transition to the cloud. Earlier than becoming a member of AWS India, Vel labored as a Senior DevOps Architect with AWS ProServe North America, supporting main Fortune 500 firms in america. He’s captivated with GenAI & AIML and leverages his deep data to supply strategic steering to firms trying to undertake and optimize AWS providers. Outdoors of labor, Vel enjoys spending time together with his household and mountain biking on tough terrains.

Chamika Ramanayake is the Head of AI Platforms at Dialog Axiata PLC, Sri Lanka’s main telecommunications firm. He leverages his 7 years of expertise within the telecommunication trade when main his group to design and set the muse to operationalize the end-to-end AI/ML system life cycle within the AWS cloud atmosphere. He holds an MBA from PIM, College of Sri Jayawardenepura, and a B.Sc. Eng (Hons) in Electronics and Telecommunication Engineering from the College of Moratuwa.