Elasticsearch Learning to Rank 1.0 RC1 Released

We reached a major milestone over the weekend! We released the first release candidate for version 1.0 of Elasticsearch Learning to Rank. While our partners Wikimedia and Snagajob are currently using this to drive search in production, we still need community feedback to help make it bulletproof against the broadest number of use cases. So please try it out.

What is this plugin?

This plugin incorporates common machine learning methods for relevance ranking into Elasticsearch. This includes being able to incorporate ranking models from xgboost and ranklib, two common libraries used in ranking tasks. It also can use simple linear models to optimally weigh query boosts, used in simple linear models or SVM based models.

The functionality is inspired quite a bit by Bloomberg’s Solr learning to rank plugin. It works by

  • Letting you experiment with features (Elasticsearch queries) that correlate with relevance
  • Log feature values in production for live searches to use in training data
  • Store ranking models by name
  • Rank using the ranking model with a new sltr query

Can I still find the older version of Elasticsearch LTR?

Our original LTR plugin can still be used/found at this branch. Some of our old blog articles use this version of the plugin. This version of the plugin only stores and executes Ranklib models. It doesn’t contain all the bells & whistles of the 1.0 plugin.

What’s next?

If you have ideas for features beyond 1.1, please file an issue to start a discussion. And as always, feel free to get in touch with us at OpenSource Connections if we can be of help.