Introduction
One of the leading power distribution companies in the Philippines, Meralco, is undergoing a digital transformation, which includes implementing various Generative AI and Machine Learning initiatives. A key focus of this transformation is enhancing customer engagement, particularly through digital channels. To support this goal, Meralco aimed to encourage more customers to adopt digital payment methods, as a significant portion of their customer base continued to use offline channels for monthly utility bill payments. This reliance on offline transactions required Meralco to maintain physical stores, driving up operational costs. Additionally, customers faced the inconvenience of traveling to make payments, often resulting in delays. Despite having a website and a mobile app, a substantial number of customers were not utilizing these digital platforms.
Solution
To encourage digital payment adoption among their customers, we implemented targeted promotional campaigns with personalized messaging. Using the AWS technology stack, we developed several key subcomponents for development, prototyping, and deployment.
The solution included the following subcomponents:
- Data & Feature Engineering Pipelines : We built pipelines to combine data from multiple sources and prepare features necessary for training and scoring models
- Propensity Model: To maximize the effectiveness of marketing campaigns, we developed propensity models using machine learning algorithms. These models helped identify customers most likely to switch to digital payments when offered an incentive.
- Customer Segmentation: We created customer segmentation to define key personas. This segmentation was layered on top of the propensity models, allowing us to tailor communication and offers for each key customer group.
- Customer Scoring: Each customer is scored based on these models on a daily basis to decide how best to engage with them. The scoring helps us identify which customers would be sent which offers (if any), and this is communicated to the marketing automation platform.
- A/B Testing Framework: We evaluated campaign performance through an A/B testing framework, where each campaign was launched with a control group to measure effectiveness.
Implementation
The different steps in the end to implementation were as follows:
Data Pipeline Construction: Given that data was sourced from multiple databases and systems, our initial task was to consolidate the data, apply necessary aggregations, and conduct data quality checks. We utilized PySpark running on AWS Glue notebooks to process the data, ensuring it was ready for modeling.
Propensity Model Development Using AWS SageMaker: Once the input data was prepared, we conducted thorough Exploratory Data Analysis (EDA), addressed outliers, applied sampling techniques, and performed feature engineering using AWS SageMaker notebooks. We experimented with several tree-based machine learning algorithms and fine-tuned various hyperparameters to develop a scalable propensity model. To ensure model transparency, we used SHAP to create feature importance plots and back-tested the models across different validation windows. The final model selected was a fine-tuned LightGBM model, achieving over 70% precision and recall at the designated probability threshold for the problem.
The probability threshold determines whether a customer is predicted to adopt digital payments, with different business implications:
- Lower Threshold: Predicts more customers as likely adopters (higher recall, lower precision), prioritizing reaching a larger pool of potential digital payment users, even if some do not convert.
- Higher Threshold: Predicts fewer customers as likely adopters (higher precision, lower recall), focusing on a smaller group with greater confidence in their conversion to digital payments.
Customer Segmentation: We identified key customer personas through customer segmentation to craft tailored offers and customized communications for the marketing campaign, maximizing relevance and engagement. To improve the effectiveness of the digital payment campaign, we applied K-Means Clustering to segment customers into distinct groups based on their behavior and preferences.
Examples of segments we uncovered include:
- Tech-Savvy Users: Customers who are familiar with digital platforms but only use them occasionally.
- Offer: Cashback incentives for making payments digitally.
- Traditional Offline Payers: Customers who prefer physical payment methods due to a lack of trust or familiarity with online systems.
- Offer: Exclusive discounts and personalized support to encourage their first digital payment.
- Occasional Digital Users: Customers who alternate between offline and digital payments, needing encouragement to fully adopt digital methods.
- Offer: Loyalty points for consistent use of digital payments.
Campaign Launch: We collaborated closely with our client's marketing team to roll out both digital and offline campaigns. The first step was scoring customers using the propensity model. Based on these scores, we identified the top deciles of customers with the highest likelihood of switching to digital payments for their utility bills.
Campaign Execution:
- Customer Selection: Using the propensity model, we scored all customers and selected those in the top deciles, indicating the highest probability of adopting digital payments.
- Offer Delivery: Offers were personalized based on customer segments and delivered via channels such as SMS and email for digital communications. For offline segments, traditional methods like flyers and in-person support were also used.
- Control Group: A control group was established to evaluate the effectiveness of both digital and offline campaigns, allowing for a direct comparison between those who received offers and those who did not.
This approach enabled us to target the most relevant customers while ensuring the campaign's success was measurable.
Architecture on AWS
Data Sources: The following are key data sources utilized for the project:
- Customer Demographics Data: Information such as age, gender, income, location, and occupation.
- Source: Internal CRM systems.
- Customer Transaction History: Records of payment activity (both offline and online), bill amounts, payment delays, and payment frequency.
- Source: Utility billing systems or payment gateways.
- Customer Interaction Data: Data on website visits, app usage, clicks, and engagement with marketing campaigns (emails, SMS, etc.).
- Source: Customer portals or CMS (Content Management System).
- Marketing Campaign Data: Data on historical campaigns, offer types, and their performance (response rates, conversions).
- Source: Marketing automation systems (e.g., Amazon Pinpoint, third-party platforms).
- Other Public Datasets: External data such as macroeconomic trends (inflation rates, employment rates), local business growth, and environmental factors.
- Source: Government data repositories and public databases.
All of the above data sources are imported into AWS Glue via an SFTP server.
Further Steps:
- Data Processing with AWS Glue: AWS Glue ETL jobs are used to clean, transform, and aggregate data from the various sources, making it ready for customer segmentation and propensity modeling. The processed data is stored in an S3 bucket for further use.
- Training Job Automation with AWS Lambda: Amazon EventBridge is utilized as a scheduler to trigger AWS Lambda functions. Jobs are triggered either on a scheduled basis (cron job) or by event notifications when new data is available in S3. This process initiates a SageMaker training job for both the propensity and customer segmentation models using AWS SDK (Boto3). The Lambda function is parameterized to manage different datasets and training configurations for each model.
- Inference Automation with AWS Lambda: Once the models are trained, batch inference jobs can be scheduled using EventBridge and AWS Lambda. EventBridge periodically triggers inference jobs on new customer data batches. Predictions are stored in an S3 bucket, which is crawled by a Glue Crawler into an Athena database. A push notification is sent to the campaign automation platform to use the latest predictions for new campaigns.
- Monitoring and Logging: AWS CloudWatch monitors job statuses, including errors, successes, and key metrics.
Results
Both the offline and online campaigns were deployed and they delivered promising outcomes. The initial online campaign resulted in a 10% increase in digital payments compared to the control group, while the offline campaign successfully encouraged 20% of store visitors to switch to online payments. Based on these positive results, the management team decided to continue running additional campaigns, with the goal of gradually transitioning offline customers to adopt online payment methods more consistently.
WE ARE ONEBYZERO:
Headquartered in Singapore with local presence in Asean nations, we are a modern data & AI consulting firm. We focus on transforming enterprises with cutting edge solutions to generate value from data. We specialise in serving the Telecommunications, Banking & Financial Services, Retail and Ecommerce industries.
We focus exclusively on AI/ML, Data & Martech
AI & ML, including Generative AI
We help organisations define their AI/ML strategy, develop and operationalize AI/ML & generative AI models, and implement MLOps to streamline operations. We have experience doing a variety of work in classical data science and cutting edge GenAI for Telcos
Modern Data Platforms
Our team of experts build robust data pipelines, design data warehouses and lakes with strong focus on data quality and lineage, to build C360, marts, support reporting, dashboards to help organisations uncover insights in their data using a modern data stack.
Martech & Personalization
We enable customers to modernise their digital platforms to deliver omni channel personalization use-cases. We have deep experience with marketing use-cases & personalised content generation, automated campaign grid optimization and next-best offers & actions.
