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.
- Logistic regression ranking: SGD, Batch SGD
- MSE/RMSE loss
- LambdaMART: XGBoost, LightGBM and Catboost backends
- NDCG, MAP metrics
- Data formats: libSVM
- Native categorical features
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>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))
- 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
This project is released under the Apache 2.0 license, as specified in the LICENSE file.