Using a taylor made version of the Recommendation Service to Improve Retention Campaigns and Customer Satisfaction

Challenge

The company needed to achieve a higher customer acquisition rate for those with a higher likelihood of no longer using the prepaid mobile service. Previously, specific offers were made to them based on their churn propensity level (High, Medium, and Low).

In the presented solution, there was no way to determine if that offer was the most suitable for each customer. In other words, customers with a High churn propensity were offered the offer designated for that group – is that the best offer for the group? There was also no way to know if these offers were indeed the most appropriate for that specific customer, meaning, does the customer within the group actually want that offer or a different one? Lastly, what would be the offer that the customer desires and that would be profitable for the company?

The main obstacle in doing this was how to create a model that could effectively predict what the customer might choose based on past decisions and within a specific context (such as available balance, time of offer, behavior, etc.).

Another challenge was how to increase customer monetization, meaning that while the first personalized offer had the highest likelihood of being selected, it might not be the most profitable one. Perhaps the second personalized offer, with a higher likelihood, could be more profitable.

Why AWS

For several years, CWP has used AWS services to store its information in data lakes, as well as for conducting analytics, predictive models, and visualizations. The client’s and Strata’s experiences and usage of AWS services built confidence in their use, deployment speed, and convenience. Implementing SRE on this platform was ideal because it did not require migrating to any services outside of AWS, as the entire SRE-MLOps service environment and development run on AWS platform services.

Why the Customer Chose the Partner

Strata has been working with CWP on various projects, including the implementation of different predictive models for both fixed and mobile segments, as well as various analytics to gain insights that add value to the business. Additionally, various dashboards were developed using QuickSight to visualize the effectiveness of campaigns, among other solutions. In other words, Strata has extensive knowledge of the potential and development within the AWS environment, and the client also trusts Strata and has data availability within the AWS environment.

Partner Solution 

The solution implemented by Strata was based on the creation of a framework that gathers data from three sources: Interactions, User Data, and Offer Data. Using this information, a model was trained using AWS Personalize, followed by a weighted ranking to prioritize monetization. Based on these results, personalized campaigns were created to visualize conversion outcomes. This entire process is automatically generated twice a week, using the most up-to-date information available to provide near-real-time data for training.

HL Architecture

The provided architecture consists of a first part with data from various sources, which are categorized into three domains (interactions, user information, and information about offered products). This information is then pre processed and used in Sagemaker, also leveraging AWS Personalize service, and the inferences are made available in an S3 bucket. These inferences are then utilized with Lambdas to trigger campaigns delivered to users. Logging services are also used to monitor the various steps in the process.

Results and Benefits 

The implemented solution is entirely automatic and does not require human intervention. It is also easily replicable for other markets or campaigns with a very short development time. Being a parameterized solution, if there is a need to change various parameters such as the number of offers to be made available or the time window for model training, it can be easily adjusted. The runtime of the entire process is approximately 2 hours.

Next Steps

The next steps involve continuously improving the model as more cases with additional interactions provided by SRE are trained, as it is a model that learns from its past recommendations. There is also progress being made in implementing this solution in other customer campaigns.


Posted

in

by

Tags: