Learning to Rank: What it is, how it's done, and what it can do for you.
At the intersection of information retrieval and machine learning, Learning to Rank (LTR) techniques have finally moved from academic curiosity to viable commercial offering. They offer a way to enhance the search experience by learning complex ranking models that can incorporate implicit feedback from users, explicit feedback from domain experts, features from knowledge bases, non-text signals from accompanying data assets or the environment, as well as signals from the underlying search engine itself. LTR will be part of every search/relevance engineers' toolbox in the near future. And with LTR plugins being offered for popular search engines like Solr and Elasticsearch, you don't have to be a machine learning genius to use them. This presentation will cover the intuition behind the different classes of LTR algorithms, the LTR software ecosystem, how to train and deploy your own LTR model, and discuss use cases in and around search where LTR can be useful.
This project contains some supporting code for a presentation on Learning to Rank (LTR) that I gave at the Search Summit 2018 organized by the Elsevier Search Guild. I demonstrate an implementation of LTR using The Movie Dataset (TMDB) using Solr and Elasticsearch LTR plugins. Both plugins use the RankLib library (containing a suite of LTR algorithms) from the Lemur project.
Look inside data/README.md and scripts/README.md for the download links for the TMDB dataset and the JAR files.
In addition, we have used Solr 7.4.0 and Elasticsearch 6.3.1 for the respective index backends.