AWS Machine Learning Blog
Increasing customer engagement and loyalty with personalized coupon recommendations using Amazon Personalize
This is a guest blog post by Sungoh Park, a big data analyst at Lotte Mart. In their own words, “Lotte Mart, a division of Lotte Co., Ltd., is a leading South Korean retailer that sells a variety of groceries, clothing, toys, electronics, and other goods.”
Consumers today have many options for purchasing daily necessities; they can shop across multiple channels such as hypermarkets, e-commerce, convenience stores, and supermarkets. Lotte Mart, a Korean hypermarket, uses Amazon Personalize to offer personalized recommendations to frequent customers to increase engagement, increase purchase rates of new products, and ultimately further build customer loyalty. This post shares the difficulties that Lotte Mart faced before using Amazon Personalize and how they improved their product recommendations and increased new product purchases.
Lotte Mart sells a variety of groceries, clothing, toys, electronics, and other goods to over 600,000 daily shoppers across 189 stores in Korea, Indonesia, and Vietnam, with $5.1 billion in revenue in 2019.
Lotte Mart uses M-coupon, a proprietary mobile coupon system, to encourage customers to shop by promoting in-store savings. Because hypermarket customers spend an average of $50 to $200 per visit, customer visit frequency directly impacts Lotte Mart’s business performance.
Traditionally, M-coupon made recommendations based on a customer’s purchase history; for example, recommending a particular brand of instant noodles if the customer had bought that product before. These rule-based recommendations that rely on purchase history are meaningful because customers repurchase items with timely issued coupons. This helped to drive repeat purchases and maintain customer loyalty, but didn’t drive demand for new products or create a personalized user experience that adapted to customers’ evolving needs. With new products added daily, hypermarkets must generate demand for those products quickly. However, to provide an optimal customer experience, hypermarkets can’t bombard customers with every new product. Such tactics can quickly overwhelm customers with irrelevant recommendations. Lotte Mart needed to figure out a long-term strategy to increase traffic in stores and influence purchase decisions of new products. Lotte Mart turned to Amazon Personalize as a solution to provide its M-coupon users highly curated and personalized product recommendations to increase customer loyalty and demand for new products.
Using a statistical approach to generate recommendations at scale
Traditionally, Lotte Mart used sales history and user preferences in a customer’s profile to execute targeted product recommendations via coupons. This process worked when target conditions like interval of repurchases and favorite brands were set correctly, but it wasn’t enough to promote individualized recommendations for each shopper, and only worked on previously purchased products. When recommending new products to shoppers, the coupon utilization rates, which provided an indicator into user interests of the offers, and actual purchase rates were extremely low.
Furthermore, building and maintaining a statistical target marketing engine for new offers was time-consuming and a drain on the highly valued big data engineering resources. The process required them to calculate the purchase cycle for each product manually and, to estimate the degree of impact by the coupons, perform association analysis with related products.
Despite all the effort and investment of time, performance didn’t meet expectations for new product adoption. For Lotte Mart, the metrics that mattered were the number of coupons viewed, usage of coupons, and buying ratio of repeat purchases and purchases using personalized coupons. The increased ratio is can indicate customers’ hidden needs. That is why Lotte Mart started to explore Amazon Personalize.
The following diagram illustrates Lotte Mart’s past architecture for statistical recommendations.
Improving customer experience with tailored recommendations
With Amazon Personalize, Lotte Mart could cost-effectively recommend new products that were difficult to promote and drive demand for through traditional methods. The coupon hit ratio, which is the response rate for the promoted coupons increased and started to significantly contribute to monthly sales.
There was an immediate value realization as well. Compared to the prior approach, Amazon Personalize eliminated the need for tedious and complex manual data analysis, and reduced development time by 50%. This saved time because Lotte Mart only had to provide predefined Interactions, Users, and Items datasets. The engineering team could generate test results in half the time compared to the prior approach.
However, saving time was only part of the success metrics needed to determine that Amazon Personalize was the right solution for them. They needed customers to engage and purchase new products at a greater frequency while still ensuring an optimal customer experience. Amazon Personalize made this easier by having the referential schemas of each type of dataset readily available. Schemas kept evolving along with the trial, such as adding or replacing features to improve the overall coupon response rate. After Lotte Mart prepared the data, they tested various algorithms.
Finally, Amazon Personalize enabled custom recommendations for each customer rather than the traditional rule-based recommendations for everyone. The whole process greatly improved productivity by reducing the initial development time and removing the need for managing and maintaining custom models.
Personalizing coupon recommendations
Lotte Mart’s goal was to increase the participation rate for customers who hadn’t used coupons before to drive the demand of new products. Customers’ interests and requirements are constantly changing, and the competitive landscape is continuously evolving and getting fiercer. Lotte Mart could increase customer retention and loyalty by actively discovering customers’ unknown needs and responding to changes in intent.
The following diagram illustrates the new architecture for recommendations and personalized coupons with Amazon Personalize.
Lotte Mart targeted over 700,000 of their highest-spending frequent customers for the coupon recommendation service. This cohort of shoppers is also highly susceptible to discounts or promotions from other shopping channels. They also actively engaged this cohort in the initial statistical recommendation approach, which made them the ideal group to do A/B testing of the two recommendation engines.
For this test, Lotte Mart modeled on products such as processed food, bathroom items, detergent, and other daily household products with no seasonality, so they could mitigate the effects of events such as back-to-school or holidays. As a result, they only needed one month of sales history as input data. The dataset was comprised of tens of millions of transactions. Each distinct item purchased on a receipt became an interaction in the dataset.
For modeling in Amazon Personalize, Lotte Mart relied on three datasets: sales history (few months), product metadata, and user profile, which they extracted through the legacy statistical system. They uploaded the extracted data to Amazon S3 and performed preprocessing to remove irrelevant or noisy data, such as discontinued products, and anonymized user profiles. They imported the datasets as Interactions, Items, and Users, which contained the following information:
- Interactions – All sales history over a certain period
- Items – Product meta-information such as category and SKU
- Users – Anonymized user profile data with multiple categorical variables
Integrating product metadata led to significant improvement in the recommendations.
Before Lotte Mart could model in Amazon Personalize, they had to decide which product metadata to include. There is no shortcut for finding the best product metadata to model on, and it’s heavily dependent on domain expertise. The same applied in this scenario; metadata was constantly updated and improved based on domain experts’ experience and knowledge. One tip is that the categorical feature is best suited for low cardinality attributes. For example, SSN, email, or userid aren’t appropriate for the categorical type because these are unique and have high cardinality.
To evaluate the impact of features on Items or Users datasets, Lotte Mart explored HRNN-meta and HRNN recipes. HRNN is more straightforward than HRNN-meta in terms of data preparation because HRNN looks at the Interactions dataset. However, HRNN-meta uses more features of the Items metadata or Users dataset in addition to interactions. A solution version is a kind of trained machine learning model, trained with a selected recipe, dataset group, and other parameters. You can build multiple solution versions over the same dataset group, which allows you to evaluate or compare models trained with different recipes. This way, you can find the recipe that works best for the task.
Coupon recommendations are performed every other week so that Campaign, compute providing recommendation via API, doesn’t need to run all the time. Results are gathered by invoking
GetRecommendations API with
userid. Then responses that contain recommendations for users are written to a file. The on-premises coupon delivery system, M-coupon, downloads the result from Amazon S3, at which time post-processing is done, and finally sends personalized coupons to the customer. Lotte Mart can gather business metrics such as the revenue of coupon-influenced purchases and coupon consumption rate.
The following screenshots show the M-coupon mobile app and recommended coupons from Amazon Personalize.
Lotte Mart operates Amazon Personalize cost-effectively by deleting campaigns after retrieving all recommendations so they only incur costs for runtime and TPS of campaigns used. The TPS of a campaign is similar to a number of concurrent transactions. Therefore, a campaign that has TPS 1 can produce multiple recommendations in a second if the response time is less than 500ms. For example, 50 recommendations are retrieved in a second if a recommendation API call only takes 20ms. Hyperparameter optimization (HPO) helps to build optimized models by figuring out the best hyperparameters and applying them to the training to get the best possible model. Because there are interim trainings for optimization, it does come with cost implications when compared to single model training. However, after HPO is complete, you can reuse hyperparameters on future trainings by specifying
algorithmHyperParameters, which helps to effectively manage cost and performance.
Lotte Mart is constantly striving to increase its coupon hit ratio, and with Amazon Personalize, coupon usage has more than doubled since the introduction of personalized recommendations. They also saw a growth of 1.7 times in frequency of new product purchases—a drastic improvement compared to their earlier statistical approach. This increased ratio indicates that Lotte Mart is discovering its customers’ hidden buying needs successfully. Consequently, the improvement in KPIs with personalized coupons has significantly impacted Lotte Mart’s monthly sales. Learn how retailers like Lotte Mart leverage Smart Store Solutions on AWS to provide a fast, frictionless shopping experience that delights customers while driving operational efficiency and IT agility. For more information about how you can start using Amazon Personalize to help improve your product recommendations and increase customer engagement, visit the product webpage.
About the Authors
Sungoh Park is a Big Data Analyst at Lotte Mart who has worked on target marketing using a statistical approach for over 7 years.
Kyoungtae Hwang is a Solutions Architect with AWS. He works with enterprises to build workloads that realize customers’ business outcomes.