Activate 2018

Learning-to-rank (LTR) is a powerful technique which utilizes supervised machine learning to address the problem of search relevancy. While recent versions of Solr include an LTR component, there are still significant practical barriers to using LTR. This talk will demonstrate both the engineering and the data science necessary to build a production-grade, end-to-end LTR system on a real world dataset.

The talk is divided into three parts: First, I will show how to set up, configure, and train a simple LTR model using both Fusion and Solr. Secondly, I will demonstrate how to include more complex features and show improvement in model accuracy, in an iterative workflow that is typical in data science. Particular emphasis will be given to best practices around utilizing time-sensitive user-generated signals. Lastly, I will explore some of the tradeoffs between engineering and data science, as well as Solr querying/indexing strategies (sidecar indexes, payloads) to effectively deploy a model that is both production-grade and accurate.