idealo Doubles Click-Through Rate through Personalized Recommendation Engine Developed Using Amazon SageMaker

With 2.5 million daily page views and over 76 million monthly visits, idealo offers an online portal for customers in six countries across Europe to compare prices for over 500 million products from about 50,000 vendors. User traffic drives revenue from advertisers who closely track certain key performance indicators (KPIs) in the highly competitive retail industry. These KPIs include click-through rates, a measure of how often a customer visits a website to make a purchase, and session rates, the amount of time a user spends on a website. “We wanted to improve what we were already doing as a company and explore other business opportunities,” says Luiz Davi, ML product manager at idealo. “The goal was to build a central offering for the whole company for product recommendations and user-based personalized recommendations.”

In 2021, idealo, a subsidiary of Axel Springer SE, decided to migrate all in to the cloud. It wanted to remove the operational risks of its aging on-premises data center, improve the scalability and reliability of the idealo solution for its customers, and boost KPIs. The MLE team was an early adopter of AWS services within idealo, identifying several use cases that it wanted to explore to enhance CRM. The team decided to build a small prototype in the cloud that could drive immediate value, and then iterate through A/B testing to evaluate the impact of the ML model and use the insights to steer business decision-making.

In early 2022, the team developed an ML model that provides complementary recommendations: items that correspond to a purchase, such as a case for a purchased mobile phone. In 3 months, the MLE team released the initial model into production. “That first model showed an impressive improvement from our past benchmark,” says Davi. “That opened multiple doors inside the company so that we could move forward and try more.” The team quickly built upon its success with another model that recommends similar products, which are items that are comparable to a purchased item. The team then created an even more sophisticated model, using data about complementary and similar purchases to deliver personalized recommendations to customers. idealo promotes items of interest to customers based on information collected automatically—with permission—about their shopping history.

The team uses solutions from AWS to alleviate much of the manual work involved with the orchestration of data so that it can experiment fast, iterate on models in development, and push useful ML models into production twice as fast as it previously could. It built a pipeline using Amazon SageMaker, which developers use to build, train, and deploy ML models for nearly any use case with fully managed infrastructure, tools, and workflows. “Using Amazon SageMaker really speeds up the whole iteration cycle,” says Arjun Roy, idealo ML engineer. “When I think of innovation, I think about playing around with the data and trying different models. And as an extension to that, the pipelines are very flexible.”
For example, ML engineers could run one of their models in one-sixteenth of the time by using a technique called parallelizing. The team spun up 16 compute instances to speed the process of running the model on AWS. “If we had to run the servers and host the applications ourselves, that would require much, much more time,” says Davi. “Now, we can be agile and try different approaches as we go.” Furthermore, idealo allocates costs granularly to certain workloads using the cost transparency of AWS services.

The team delivers additional functionality to the CRM team through the use of AWS Lambda, a serverless, event-driven compute service that lets organizations run code for virtually any type of application or backend service without provisioning or managing servers. Through AWS Lambda functions, customized bargains automatically generate as part of the CRM team’s monthly email campaign. “We have automated the process so that we don’t have to do manual work to keep it running,” says David Rosin, idealo ML engineer. “We set it up once, and ideally, it runs every month.” Customers who receive the emails see bargains that have been automatically selected specifically for them. “Using the MLE team implementation versus our old top-sellers’ logic, we achieved a conversion rate increase of 154 percent,” says Felix Gehlhaar, idealo’s CRM manager, who closely collaborated with the MLE team. “This is exciting for us.”

The automated user-recommendation engine has also generated success for the idealo website, which has seen a 111 percent rise in click-through rates and has increased session rates by 151 percent. “We’ve made a huge leap,” says Davi. “We can see the impact that it’s generating for our internal users. And then we see that impact on the website. People are deciding to buy specific items because they found what they wanted.”
 

As the entire company continues its migration to AWS, internal idealo teams share data and collaborate more effectively. “One of our CRM managers told us that the ability to share information makes his life much simpler,” says Davi.Throughout 2023, the MLE team plans to explore using near-real-time data to continue to improve KPIs by driving recommendations, a process that builds upon its strong ML pipeline. “There’s a lot to build,” says Davi. “We have never tried something like this before, but we see great potential as we advance this initiative. Using AWS, we create products that support us as a company moving forward.”