Enable javascript in your browser for better experience. Need to know to enable it? Go here.
Marimekko

Decision factory creates tailored experiences for Marimekko's customers

Marimekko is a Finnish lifestyle design company, renowned for its original prints and colors. The company’s portfolio includes high-quality clothing, bags and accessories as well as home décor items ranging from textiles to tableware.

 

Marimekko Decision Factory

 

The Marimekko Decision Factory is a reinforcement learning -based personalization engine, running a factory abstraction for scalability. The factory recommends, personalizes, and optimizes virtually anything on Marimekko’s website to match users’ interest. While doing so, it learns from user behavior in real time and does not require past data assets nor expert data science skills to be scaled across the company's digital platforms. Most importantly, it generates new insights about what customers actually want - without data-induced bias. 

 

Just hours after its launch, the conversion rates on the site improved significantly and a double-digit increase in homepage-to-funnel conversion was achieved with a model that had only three hours of data in its hands.

Marimekko Decision Factory

Know thy customer - in real-time

 

Marimekko gave Thoughtworks the challenge to design and build a scalable personalization engine that strengthens Marimekko’s profitable growth through an increasingly broad customer base. The engine should enable compelling digital experiences - while increasing customer understanding.

 

Oftentimes, digital platforms are seen merely as another sales channel while leaving out possibilities to actually learn and support the business beyond incremental improvement and conversion optimization. This is where we wanted to go on a different path - learning as much as possible, in as many ways as possible, while improving the customer experience.

 

Furthermore, we wanted to create something that limits the need for endless AI point-solutions and reduces the total cost of ownership for AI technology. To achieve all of this, Marimekko and Thoughtworks created a cross-functional team to design and build the end-to-end decision factory within two months.

Creating truly personalized experiences

 

Generally speaking, personalization can be divided into two categories, illustrated in the image below. 

 

1.  Recommending from a finite action set like frontpage artwork, discount percentages or landing page assets

 

2.  Recommending from a potentially very large action set, such as thousands of products (SKUs)

Two categories of personalization Two categories of personalization
 

To provide a truly tailored customer experience, an organization should build both. Traditional product recommendations cover only a fraction of what can be considered a personalized experience and depending on the industry, it is either very hard or impossible to deliver true customer intimacy at scale using product recommenders only. Even more importantly, personalization systems are often built to leverage past data assets for delivering future experiences. In order for us to learn something new while delivering optimal experiences, we have to rethink the logic of personalization systems.

 

As a concept, the decision factory generalizes choosing the best action for any possible business decision. To learn what works best, it balances exploration and exploitation of different actions with real-time incremental learning. Therefore it can be used to personalize and optimize various items, such as:

 

  • Homepage assets and their order

  • Discount prices

  • Artwork assets

  •  Individual content items

  • Product listing pages

  • Product assets such as images

 

Importantly, it can be used in circumstances where the possible actions at any given time have a short lifespan - which is often the case in high-velocity and high-variety contexts such as in homepage assets. Humans create the assets, and the decision factory chooses which ones get on display, at a given time, to a given user.

Under the hood 

 

Technically speaking, a set of possible actions is abstracted into decision containers. This allows for a generalization pattern - any decisioning problem can be treated as a problem of picking the best action from a set of actions, given data about the customer and the context (if available), so that it optimizes against a well-defined business outcome. After being built, Marimekko doesn't need data scientists to scale it for additional use cases.

 

In order for a decision factory to learn, it ingests a real-time event stream from the digital platform. It triggers a learning process every 10 minutes to train all possible models inside all available decision containers that have new data points from which to learn. This learning happens against a context - behavioral and contextual segment-level understanding about customers.

 

Decision factory runs on Microsoft Azure services. The team built a serverless architecture with Azure Functions to achieve a highly scalable API backend for serving the personalizations. Micro batch training of the prediction models uses the same Azure Functions backend to keep the architecture lightweight and concise. These architectural choices allowed us to reduce service management overhead while keeping the costs on a reasonable level.

 

The decision factory does not only optimize and personalize, it creates new data products for strategically relevant use cases. This is the critical differentiation between traditional recommendation systems and next generation decision factories.

Past data alone doesn’t reveal what customers want next 

 

The decision factory can teach us a lot about customer behaviour. What content and products are interesting to new customers? What actions drive conversions for different segments? What segments seem to be most attracted to new designs and products? Do our beliefs about customer behavior actually hold true?

 

After being live for just a few hours, a whopping double-digit increase in homepage-to-funnel conversions had been achieved. This high performance is due to the fact that the model can accurately optimize what to show, and to which customers. For example, it may seem obvious to promote discounts first, but in some scenarios, it is more effective to show new collection items first. 

 

Whatever happens in Marimekko's business context, the real-time exploration mechanisms ensure the system adapts. Most importantly, the adaptation happens without the need to reconfigure it, retrain it, or worry that it will break down, even as customer behavior, competitor behavior, or operating environments inevitably change. 

Marimekko has always strived to offer best-in-class experiences in all of our channels. Adopting new technologies - like this decision factory - enables Marimekko to offer unique experiences in delightful ways and to do that at scale.
Kari Härkönen
Chief Digital Officer, Marimekko

About Marimekko

Marimekko is a Finnish lifestyle design company renowned for its original prints and colors. The company’s product portfolio includes high-quality clothing, bags and accessories as well as home décor items ranging from textiles to tableware. When Marimekko was founded in 1951, its unparalleled printed fabrics gave it a strong and unique identity. Roughly 150 Marimekko stores serve customers around the globe. The key markets are Northern Europe, the Asia-Pacific region and North America. The company’s share is quoted on Nasdaq Helsinki Ltd.

Let's talk about your next project