From Nothing to Something: Klarna’s Journey With Recommendation Systems

About Klarna's journey from zero recommendation models to our current state (five use cases) in one year. Challenges, wins, and learnings. Also a sneak peek at where we are headed.

Anil Sharma, data science manager at Klarna, told how a newly merged team went from no recommendation models in October 2020 to personalisation across the Klarna app in about a year, by focusing on customer value, speed of iteration, and testing hypotheses rather than on new algorithms.

Before any model could be served, the data scientists shipped lookup tables as CSV files to the existing rules-based service, which let the team run seven A/B tests in three months and raise click-through in the shopping app by 10 per cent. Automated retraining and deployment followed: with fresher data each day, the same collaborative-filtering model gained 5 per cent in clicks, and Thompson sampling, used to rank the curated shopping lists, raised click-through by 25 per cent and replaced 140 hours a year of manual work. The same simple models were then reused for product recommendations in transaction emails, on the home page, in marketing campaigns, and on store pages, and a model for deals reached production in under a week.

Sharma closed with the next step, a two-stage system for real-time recommendations, and in the questions explained that each market has the same model trained on its own data.