Ranklib for .NET is a hardened port to .NET of RankLib, a popular open source learning to rank library written in Java. It maintains compatibility with input and output files of RankLib, allowing it to be used to integrate with systems that use RankLib, such as the Elasticsearch Learning to Rank and OpenSearch Learning to Rank plugins.
RankLib is available as both a command line tool for training and evaluating rankers, as well as a library for incorporating into solutions.
To add as a library to an existing project
dotnet add package SearchPioneer.RankLib --prerelease
To add as a global .NET command line tool
dotnet tool install -g SearchPioneer.RankLib.Cli --prerelease
To see all the commands supported by the command line tool
dotnet-ranklib --help
Learning to Rank (LTR) is a technique in machine learning that trains models to optimize the ranking order of items in a list based on relevance to a specific query or user intent. The goal is to improve the quality of search results, recommendations, and other ranked lists by understanding and modeling what users find most relevant or useful. LTR is widely used in search engines, recommendation systems, and information retrieval to enhance user satisfaction and engagement.
The following ranking algorithms are implemented
- MART (Multiple Additive Regression Trees, a.k.a. Gradient Boosted Decision Trees (GBDT))
- RankNet
- LambdaRank
- RankBoost
- AdaRank
- Coordinate Ascent
- LambdaMART
- ListNet
- Random Forests
The following metrics are implemented to measure the effectiveness of ranking
- Average Precision (
MAP
) - Best at K (
Best@K
) - Discounted Cumulative Gain (
DCG@K
) - Expected Reciprocal Rank (
ERR@K
) - Normalized Discounted Cumulative Gain (
NDCG@K
) - Precision at K (
P@K
) - Reciprocal Rank (
RR
)