JarCastingfeatran
Featran, also known as Featran77 or F77 (get it?), is a Scala library for feature transformation. It aims to simplify the time consuming task of feature engineering in data science and machine learning processes. It supports various collection types for feature extraction and output formats for feature representation.
Introduction
Most feature transformation logic requires two steps, one global aggregation to summarize data followed by one element-wise mapping to transform them. For example:
- Min-Max Scaler
- Aggregation: global min & max
- Mapping: scale each value to
[min, max]
- One-Hot Encoder
- Aggregation: distinct labels
- Mapping: convert each label to a binary vector
We can implement this in a naive way using reduce and map.
case class Point(score: Double, label: String)
val data = Seq(Point(1.0, "a"), Point(2.0, "b"), Point(3.0, "c"))
val a = data
.map(p => (p.score, p.score, Set(p.label)))
.reduce((x, y) => (math.min(x._1, y._1), math.max(x._2, y._2), x._3 ++ y._3))
val features = data.map { p =>
(p.score - a._1) / (a._2 - a._1) :: a._3.toList.sorted.map(s => if (s == p.label) 1.0 else 0.0)
}
But this is unmanageable for complex feature sets. The above logic can be easily expressed in Featran.
import com.spotify.featran._
import com.spotify.featran.transformers._
val fs = FeatureSpec.of[Point]
.required(_.score)(MinMaxScaler("min-max"))
.required(_.label)(OneHotEncoder("one-hot"))
val fe = fs.extract(data)
val names = fe.featureNames
val features = fe.featureValues[Seq[Double]]
Featran also supports these additional features.
- Extract from Scala collections, Flink
DataSets, ScaldingTypedPipes, ScioSCollections and SparkRDDs - Output as Scala collections, Breeze dense and sparse vectors, TensorFlow
ExampleProtobuf, XGBoostLabeledPointand NumPy.npyfile - Import aggregation from a previous extraction for training, validation and test sets
- Compose feature specifications and separate outputs
See Examples (source) for detailed examples. See transformers package for a complete list of available feature transformers.
See ScalaDocs for current API documentation.
Presentations
- Featran - Type safe and generic feature transformation in Scala - NABD Conf Palo Alto 2017 talk
Artifacts
Feature includes the following artifacts:
featran-core- core library, support for extraction from Scala collections and output as Scala collections, Breeze dense and sparse vectorsfeatran-java- Java interface, see JavaExample.javafeatran-flink- support for extraction from FlinkDataSetfeatran-scalding- support for extraction from ScaldingTypedPipefeatran-scio- support for extraction from ScioSCollectionfeatran-spark- support for extraction from SparkRDDfeatran-tensorflow- support for output as TensorFlowExampleProtobuffeatran-xgboost- support for output as XGBoostLabeledPointfeatran-numpy- support for output as NumPy.npyfile
License
Copyright 2016-2017 Spotify AB.
Licensed under the Apache License, Version 2.0: http://www.apache.org/licenses/LICENSE-2.0
