Solution:
Data Analytics & Machine Learning
Industry:
Retail
Personalized Recommendations on a Global Scale
The Challenge
Nutridome is a marketplace with a selected and comprehensive range of make-up and care cosmetics, with a wide selection of accessories and supplements. The company has its own brand, which it is effectively developing. A mobile app is also available for customers. The online drugstore operates in 15 markets, including Poland, the Czech Republic, Austria, Germany, Ireland, the United States, France, Italy and Spain.
Nutridome needed accurate product recommendations on their website and app. While the store's customers had already seen the recommendations while shopping, they were static and the same regardless of context or location. Nutridome turned to Univio with this need.
- The client operates on several foreign markets, so the recommendation system needed to work equally effectively in each of these regions. Customer behavior and product requirements in each market are different.
- It was essential that the recommendations were aligned with the local availability of the assortment, which would prevent customers from becoming frustrated by recommending products that were out of stock.
- The recommendation system had to gain access to the current inventory of products on an ongoing basis. Sold out products are removed from the generated recommendation sets.
The Solution
- In response to Nutridome's needs, we implemented a recommendation engine using models in Amazon Personalize. AWS provides models prepared specifically for e-commerce challenges.
- The first step in the project was to analyze transaction data and actions performed by customers in the store, which Nutridome collects in its database. Next, we prepared the right dataset to power and train the Amazon Personalize models.
- With the help of the AWS Lambda service, we implemented a mechanism that reacts to changes in the Nutridome database and forwards them to the recommendation system. Thanks to constant access to new data, the preferences of drugstore customers and changes in the assortment are taken into account in recommendations on an ongoing basis.
- To ensure the relevance of Nutridome recommendations, we used three types of models to place them in the appropriate places on the page. Each of them has a different function: "Best Sellers", "Frequently Bought Products Together" and "Selected for You".
- To achieve our business goals for promoting specific product groups, we tested two optimized versions of recommenders. We conducted A/B testing to determine which version was more popular with customers. We used Evidently from AWS CloudWatch to support the experiment.
- After the implementation of the recommendation system, we conducted training for the client, enabling independent monitoring and management of the system.
The Result
What We Learned
Servicing many foreign markets required close cooperation between our team and the client. The client regularly updated the product availability database, and it was our job to use this data to keep Amazon Personalize up to date. We implemented a solution that applies a filter to the generated recommendations, eliminating products that are not available in a given region.