Project Pearl – JYSK Programmatic Performance boosted by AI and Automation

Baggrund

As JYSK has grown from just 1 store in Aarhus to 2.877 stores across 52 markets, the ability to deliver “a great offer” to all potential customers is still the very core of the business strategy. But as the traditional retail power media as Door drops and TV have lost power in reaching the consumers effectively - this task is getting impossible and the need of cost efficient global scaling is crucial to be able to support all markets simultaneously.
But that is in its simplicity exactly what Project Pearl has made possible for JYSK – while delivering outstanding performance and business impact across markets!

Løsning

1. Learning from data. Counting all parameters available in Google’s Demand Side Platform (DV360), there are over 33 million combinations. It is impossible for a human to exhaustively learn from this large search space.
We built a machine learning algorithm to learn from past campaign log files and explore all possible combinations. Past campaign and sales data consist of 90 columns and millions of rows (impressions, clicks, and conversions). The algorithm uses this as training data to learn what drives the desired outcome and then generates a set of effective combinations.

2. Translate learnings into action. This is done using a decision tree that builds out the different pathways and combinations that the machine learning has identified in the data. A single “leaf” in the tree represents a suggestion for a digital targeting strategy, also called a “line item”. Some paths are much deeper than others and show how the machine can unpack the data into various dimensions and depths in order to find the most efficient combinations. With enough data trees can have thousands of leaves/line items. However, we have found a sweet spot for system and campaign performance around 200-500. To put that into perspective a human trader will normally work with 10-20 line items!

3. Activate and optimize. A trader optimizes a campaign by changing line item settings, such as budget, bid value, and audience segments, which is done on a weekly basis.
Project Pearl applies a reinforcement learning algorithm several times a day on the newest log file data to understand how to optimize the line items. Another algorithm (a multi-armed bandit) then estimates how much we can further lower down the bid value for each line item while still maintaining the consistent pacing.
At the beginning of the campaign, each line item is assigned equal weight and budget. Based on performance the algorithm will shift budget from low performing line items to high performing ones. Changes to the line items are automatically pushed to DV360 via API. Over time this typically leads to 10-20 strategies being the absolute best and delivering top performance.

4. Automate the workflow. All recurring parts of the workflow are automated, and the human trader only needs to brief the system on campaign goals, timeline and budget. From there on Project Pearl will do the rest with unparalleled scale, frequency and precision.