Learning Learning To Rank

Torsten Köster & Fabian Klenk & René Kriegler • Back to Haystack Europe 2018

At Shopping24, we have recently started to apply machine learning to the search result ranking on our Solr-based product search platform. We could easily train a ranking model using open source software and deploy it to Solr. However, we soon realised that this was only the easier part of it and that we had to put our efforts into the tasks and processes that empower us to train a successful model, such as: gathering valid training data, preparing judgement lists, feature engineering, expectation management, computing offline search quality metrics and- connecting offline and online metrics through A/B testing. We finally rolled out a search result ranking based on learning-to-rank. During the eventually successful journey, we shed a lot of tears. In this talk we are going through the project’s pitfalls, key moments, lessons learned from three different perspectives: product management, engineering and information retrieval.

Download the Slides
Torsten Köster

Shopping 24

Torsten has been working with Lucene/Solr for more than ten years, the last eight years focused on e-commerce searches. He co-organizes the Search Technology Meetup Hamburg and maintains several open source Solr extensions. Currently he is the CTO of the shopping24 internet group working on their e-commerce search engine. As a consultant within the Otto Group he enables other companies to focus on e-commerce search as a first class citizen and use Solr as a search engine.

Fabian Klenk

Shopping 24

Fabian has been working as product manager in e-commerce for about ten years. Currently he is working at shopping24 internet group as a product owner for search & tracking.He is interested in search, analytics, tableau and data driven structures.

René Kriegler


René has been working as a freelance search consultant for clients in Germany and abroad for more than ten years, often contracting for OpenSource Connections. Although he is interested in all aspects of search and NLP, key areas include search relevance consulting and e-commerce search. His technological focus is on Solr/Lucene. René co-organises MICES (Mix-Camp E-Commerce Search, Berlin). He maintains Querqy - an open source library for query pre-processing.