Pomelo Fashion, a global fashion e-commerce service based in Southeast Asia, had been displaying items on its website in much the same way since it was founded in 2013. The setup had grown stale, not to mention that the algorithm for displaying the items relied on old data streams with limited inputs and spotty accuracy. So as a fast-growing, innovative startup, Pomelo Fashion set out to create personalized customer experiences that would improve the discoverability of new items and increase revenue—and it needed a solution that would do so at scale.
Pomelo Fashion turned to Amazon Web Services (AWS) and used Amazon Personalize, which enables developers to build applications with the same machine learning (ML) technology used by Amazon.com for real-time personalized recommendations. By using Amazon Personalize—and the services of AWS Advanced Technology Partners Segment and Braze—to build fresh sorting and categorizing features, Pomelo Fashion created a unique, personalized shopping experience that boosts customer engagement and more efficiently converts it into sales.
Updating a Years-Old Algorithm Using Amazon Personalize
Pomelo Fashion sells apparel online and in 18 retail locations throughout Southeast Asia. Shipping to nearly two million customers in more than 50 countries, the company currently employs 500 staff members across its corporate offices, retail stores, and warehouses. Its gross revenue tripled from 2017 to 2018, doubled from 2018 to 2019, and is on track to double in 2020 despite the overall global economy being down—in July 2020 alone, the company reported $7.5 million in revenue. For years, Pomelo Fashion relied on an algorithm that ranked products on category pages—such as "Dresses," "Blouses," and "Pants & Bottoms"—based on page views and sales, blending the trends of the past 30 days with lifetime behaviors, product price, and newest releases. The rank was calculated daily and stored in a database, providing an identical experience for every user by country.
But as Pomelo Fashion grew, it recognized that enhancing the algorithm with ML would improve the quality of recommendations on category pages for customers, leading to higher digital user engagement and conversion. Category pages generate the largest portion of sales for Pomelo Fashion: 38% of purchased products are discovered by customers on category pages. Increasing the relevance of the products shown on these pages had huge potential to uplift revenue. Even if customers wouldn't purchase the recommended product, they would enter a funnel to see other products on pages like "Color Swatch," "Shop the Look," and "Just for You," which generate 30% of Pomelo Fashion's revenue.
That's when the company, which had always been an AWS customer, first heard about Amazon Personalize at an AWS-hosted workshop. "When you think of e-commerce, you think of AWS," says Shane Leese, business intelligence director at Pomelo Fashion. "New services are always coming out on AWS, and support is very good." Using AWS would also provide regional availability and help Pomelo Fashion set up the new logic to personalize its categories and sorting to each shopper.
Customizing the User Experience and Boosting Sales
Pomelo Fashion was already working with Segment—a customer data system that collects, schematizes, and loads sales data from Pomelo Fashion's mobile app, website, and kiosk services on AWS to enable a 360-degree view of customers and real-time personalization, all without complicated setup or maintenance—when the Amazon Personalize private beta was released in June 2019. So, because Pomelo Fashion didn't have the infrastructure to create personalized experiences at scale to help with product discoverability, it decided to integrate Segment and Amazon Personalize. "Without Segment, we would not have gotten this off the ground," says Leese. "We were trying to build in-house event tracking but were looking at a pretty messy set of event data. Our AWS solutions architect could see this would be a long road, so he suggested getting Segment on board to save more developer time than it would cost. With the data flowing from Segment, we didn't have to build a lot of infrastructure to make this happen."
The new logic sorts products on category pages based on individual shoppers' preferences. Customers' product interactions—their clicks, add-to-cart selections, wish lists, purchases, and more—are used to predict which products they are most likely to find interesting. Product details, such as price, color, and category, are correlated with customer details, such as their location, so that the ML model can better find similar products and customers. The more product and customer data the ML model processes, the more accurately it makes recommendations. New Pomelo Fashion shoppers are first shown a popular sorting of items, and in as little as a few minutes, the ML model personalizes the sorting based on their predicted preferences.
Using Amazon Personalize to optimize recommendations, Pomelo Fashion significantly boosted sales. "After a beta implementation proved stable, we began to realize the full potential of the service and made it a central part of our personalization roadmap," says Leese. "Within a month, our return on investment increased by 400% for our 'Just for You' recommendations carousel by means of hyperparameter optimization and additional metadata. After that, we began to apply other 'recipes' or models to other parts of our site." As a starting point, Pomelo Fashion trained and applied a personalized-ranking recipe to its dresses category, leading to a 10% increase in click-through rates from a category page to an individual product page and an 18.3% increase in revenue. After fine-tuning the solution based on the data from the dresses category, Pomelo Fashion expanded it to other categories.
Pomelo Fashion is currently using its personalized ranking algorithm on all its categories, except for new arrivals and select collections. As of November 2020, 60% of product views come from Amazon Personalize–fueled recommendations. Pomelo Fashion has increased gross revenue from category pages by up to 15%, click-through rates from category to product pages by up to 18%, and add-to-cart clicks from the category page by up to 16%. This expansion enabled the company to unlock an 8% gain in incremental gross revenue.
Pomelo Fashion also enlisted Braze, a leading customer engagement service that delivers messaging experiences at scale. Braze's Connected Content feature uses recommendations from Amazon Personalize to customize Pomelo Fashion’s cross-channel campaigns—those sent through email, in-app, and more. Connected Content saves Pomelo Fashion's staff time by pulling content directly from Amazon Personalize to populate messages to users in real time, up to the minute. When Pomelo Fashion sends emails to its customers, for example, they receive recommendations based on their browsing history and behavior. Emails with Braze Connected Content showed click-through-rate increases of up to 50% in some segments and an average increase of around 20%.
Further Personalizing the Shopping Experience on AWS
Pomelo Fashion plans to continue working with Segment to customize the shopper experience. Its first major initiative is to improve the relevance of its category pages by taking customer size preferences into account—currently, many products are not available in the most common sizes, which results in a high number of clicks without any conversions. Using its existing personalization structure, Pomelo Fashion plans to add tracking for size selections on its product detail page, ask basic sizing information at key points of the customer journey, and iterate a series of filters to remove less relevant products from the category pages based on a customer’s purchasing history.
The company also wants to improve discoverability and make sure customers don't repeatedly see the same products. It expects to use Amazon SageMaker to build additional ML models for forecasting and is also considering using AWS Lambda, a service that lets companies run code without provisioning or managing servers, to create a more scalable infrastructure.
By using Amazon Personalize and AWS Partners Segment and Braze, Pomelo Fashion is able to provide a dynamic and always-improving customer experience that also significantly increases revenue.
Amazon Personalize custom recommendation and ranking inference runs on Amazon EC2 C5 instances featuring the latest Intel® Xeon® Scalable processors and AVX 512. Amazon EC2 C5 instances deliver cost-effective high performance at a low price per compute ratio for running advanced compute-intensive workloads like machine/deep learning inference.