Skip to content

Latest commit

 

History

History
63 lines (49 loc) · 2.14 KB

README.md

File metadata and controls

63 lines (49 loc) · 2.14 KB
Raw

libLTR: a Learn-to-Rank algorithm library

CI Status License: Apache 2 Last release

A Java/Scala library to wrap and implement basic learn-to-rank ML algorithms under the same human-friendly API. Currently, is under an active development.

Supported features

  • Logistic regression ranking: SGD, Batch SGD
  • MSE/RMSE loss
  • LambdaMART: XGBoost, LightGBM and Catboost backends
  • NDCG, MAP metrics
  • Data formats: libSVM
  • Native categorical features

Installation

libLTR is published to maven-central for scala 3.x, 2.12 and 2.13, so for SBT, add this snippet to build.sbt:

libraryDependencies += "io.github.metarank" %% "ltrlib" % "0.2.2"

For maven:

<dependency>
  <groupId>io.github.metarank</groupId>
  <artifactId>ltrlib_2.13</artifactId>
  <version>0.2.2</version>
</dependency>

Usage

Get LETOR dataset from https://github.com/dmlc/xgboost/tree/master/demo/rank, then:

val loader  = LibsvmInputFormat(new GZIPInputStream(new FileInputStream("<path_to_file.gz>")))
// a dataset descriptor, if you want to have access to feature metadata like names
val spec    = DatasetDescriptor((1 to 46).map(i => SingularFeature(s"f$i")).toList)
// the dataset itself
val dataset = Dataset(spec, loader.load(spec))

// configured booster 
val lm      = LambdaMART(dataset, LightGBMBooster)
// trained model
val model   = lm.fit(LightGBMOptions())
// NDCG error with cutoff on 10th position
val error   = lm.eval(model, dataset, NDCG(10))

Roadmap

  • wrap RankLib family of algorithms
  • Plackett-Luce weighting for cascade model
  • Reinforcement learning: LinUCB, hLinUCB, CPR
  • dataset support: webscope R6B, C14
  • unbiased algorithm evaluation: Li, IPW

License

This project is released under the Apache 2.0 license, as specified in the LICENSE file.