Showroomprivé: Putting ML-Powered Targeting in the Hands of Marketers

In the retail sector, data (particularly machine learning and, increasingly, deep learning) is poised in the coming years to open up huge opportunities in the way stores — both physical and online — fundamentally operate and serve customers. The global health crisis of 2020 has only strengthened the need for retailers to adopt AI systems that make them more agile and able to adapt quickly to changing market needs.

Showroomprivé has been at the cutting-edge of this wave, working from the ground up to develop its capacity to use data for improvements both in the product and in customer service as well as on the operational and business side. This story will go in-depth on just one of Showroomprivé’s many advanced use cases: leveraging machine learning-based targeting for marketing campaigns that are 2.5x more effective. Ultimately, with this use case, Showroomprivé was able to:

• Empower people, providing a webapp (known as Targetor, which is powered by Dataiku and used 2-3 times per day by the business) that allows marketers to generate their own machine learning-powered targeting recommendations. 
• Hone data processes, going through three iterations of Targetor to perfect the system and add additional features over the course of four years.
• Harness technology to make it all happen, using Dataiku from testing to model development to delivery and among both data scientists and marketers alike for true vertical and horizontal collaboration around the initiative.

Company Fast Facts

• ~950 employees
• Present in 7 countries across Europe
• 16 data team members across 4 data teams
• 15.7 million orders (& counting)

Tools & Tech Stack

• Dataiku (design + automation nodes)
• Amazon Web Services (AWS)

Background & Challenges

Showroomprivé is an e-commerce retailer specialized in flash sales. Naturally, in order to generate awareness, the marketing team sends emails to their customer base about both ongoing and upcoming sales. Until 2016, the team selected the target audience for these marketing emails more or less manually based on what they know about the brand (e.g., men vs. women, age range of the brand’s appeal, etc.), asking IT for an extract of user data that corresponds to their ideas.

However, this approach presented several challenges:

• Brands often have overlapping or broad audiences — for example, lots of sales applied to 20-30-year-old females — which meant touching some prospective buyers multiple times, while others not at all.
• This also meant casting out a wide net, potentially sending emails to people who were not interested in that particular brand (i.e., people who fit the targeting were not necessarily the people who were most likely to buy).

The ultimate goals of the project was for the marketing team to be completely autonomous (e.g., no help needed from data science, IT, etc.) in targeting and sending these emails — not a small feat given the complexity of the system that they were looking for.

First Iterations of Targetor

At first, Showroomprivé turned to an external predictive analytics tool to create the groups of users for each brand. However, the team realized that they had all the data and the skills to create Targetor themselves, so for several months, Showroomprivé tested versions of Targetor built in the out-of-the-box predictive solution, built with Dataiku, and built by hand (i.e., building data pipelines from scratch without a data science and machine learning platform — this was quickly scrapped as the results were not on par with the other two solutions).

In the end, though the initial results of the project using the out-of-the-box predictive solution and using Dataiku were similar, they ultimately understood that pivoting to Dataiku — one of the world’s leading AI and machine learning platforms — would allow for more granular control in the long run and could provide them with a data product that was a real differentiator for the business and in the market.

Following these initial tests, marketing started to work completely with Targetor and with Dataiku right away, thanks to the platform’s usability both on the data science side (with complete flexibility for coders) and on the business side (with webapps, a point-and-click interface, and other features for non-data practitioners). The marketing and data teams were even able to make improvements for a second iteration of Targetor that provided a ranking of users instead of just a homogenous group, allowing marketers to better prioritize their campaigns.

The Evolution of Targetor

After about a year and a half of use by the marketing team, the data team at Showroomprivé was adding so much functionality and so many features that the model was becoming complex. They decided to undergo a project that would streamline the model, allowing them to continue to scale and add functionality for many years to come.

Showroomprivé released a revamped version of Targetor for the marketing team in early 2020 that:

• Takes a more modular approach, cleaning up some complexity that evolved over time.
• Adds better monitoring features that allow Showroomprivé’s data team to A/B test new model versions more easily.

• Expands the scope, incorporating ideas for improvement from the data science team not specifically requested by marketing (e.g., ensuring that newer users without historical data don’t continually go undiscovered or untouched by flash sale marketing emails).
• Still leverages Dataiku, which has grown with the team, the project, and the company to fit ever-expanding needs.
• Delivers on their goal of an autonomous, empowered marketing team.

As a next step, the team is working on adding Kubernetes for scalability in the execution and scoring of the model. With an uptick in usage to two or three jobs per day, they needed a more scalable solution for processing that would allow for multiple jobs to be run in parallel; Kubernetes provides this elasticity and can be easily leveraged with Dataiku.

To run data science pipelines in a scalable way and to handle complex computation steps across a large amount of data (without breaking the bank), Kubernetes is becoming an increasingly popular choice.