Merge pull request #127 from metarank/fix/libsvm-label-format

handle doubles as label ids in libsvm format

Author: shuttie

build.sbt

@@ -2,7 +2,7 @@ import Deps._
 
 name := "ltrlib"
 
-version := "0.1.23-M5"
+version := "0.1.23-M7"
 
 scalaVersion := "2.13.10"
 

src/main/scala/io/github/metarank/ltrlib/booster/CatboostOptions.scala

@@ -12,5 +12,6 @@ case class CatboostOptions(
     randomSeed: Int = math.abs(Random.nextInt()),
     objective: String = "QueryRMSE",
     loggingLevel: String = "Verbose",
-    earlyStopping: Option[Int] = None
+    earlyStopping: Option[Int] = None,
+    boostingType: String = "Plain"
 ) extends BoosterOptions

src/main/scala/io/github/metarank/ltrlib/input/LibsvmInputFormat.scala

@@ -62,7 +62,7 @@ object LibsvmInputFormat extends InputFormat {
       throw new IllegalArgumentException(
         s"LibSVM format requires at least two columns: label and qid, but got ${tokens.length} on row $index"
       )
-    val label = tokens(0).toInt
+    val label = tokens(0).toDouble
     val qid = tokens(1) match {
       case queryPattern(_, id) => id.toInt
       case _ => throw new IllegalArgumentException(s"qid format for item '${tokens(1)}' is not supported on row $index")
@@ -88,6 +88,6 @@ object LibsvmInputFormat extends InputFormat {
       values(featureIndex - 1) = value // libsvm indexing starts from 1
       i += 1
     }
-    new LabeledItem(label, qid, values)
+    LabeledItem(label, qid, values)
   }
 }

src/main/scala/io/github/metarank/ltrlib/output/LibSVMOutputFormat.scala

@@ -10,8 +10,9 @@ object LibSVMOutputFormat extends OutputFormat {
       query <- ds.groups
       rowid <- 0 until query.rows
     } {
-      val row  = query.getRow(rowid).zipWithIndex.filter(_._1 != 0).map(x => s"${x._2 + offset}:${x._1}")
-      val line = s"${query.labels(rowid)} qid:${query.group} ${row.mkString(" ")}\n"
+      val features = query.getRow(rowid).zipWithIndex.filter(_._1 != 0).map(x => s"${x._2 + offset}:${x._1}")
+      val label    = math.round(query.labels(rowid)).toString
+      val line     = (List(label, s"qid:${query.group}") ++ features).mkString("", " ", "\n")
       data.write(line.getBytes())
     }
   }
@@ -24,8 +25,9 @@ object LibSVMOutputFormat extends OutputFormat {
       for {
         rowid <- 0 until query.rows
       } {
-        val row  = query.getRow(rowid).zipWithIndex.filter(_._1 != 0).map(x => s"${x._2}:${x._1}")
-        val line = s"${query.labels(rowid)} ${row.mkString(" ")}\n"
+        val features = query.getRow(rowid).zipWithIndex.filter(_._1 != 0).map(x => s"${x._2}:${x._1}").toList
+        val label    = math.round(query.labels(rowid)).toString
+        val line     = (List(label) ++ features).mkString("", " ", "\n")
         data.write(line.getBytes())
       }
     }

src/test/scala/io/github/metarank/ltrlib/input/LibsvmInputFormatTest.scala

@@ -11,6 +11,10 @@ class LibsvmInputFormatTest extends AnyFlatSpec with Matchers {
     parse1("1 qid:1 1:1", LabeledItem(1, 1, Array(1.0)))
   }
 
+  it should "load 1 feature with double label and qid" in {
+    parse1("1.0 qid:1 1:1", LabeledItem(1, 1, Array(1.0)))
+  }
+
   it should "fail on nan" in {
     Try(LibsvmInputFormat.parseLine(1, "1 qid:1 1:NaN")).isFailure shouldBe true
   }

src/test/scala/io/github/metarank/ltrlib/output/LibSVMOutputFormatTest.scala

@@ -23,9 +23,9 @@ class LibSVMOutputFormatTest extends AnyFlatSpec with Matchers {
     LibSVMOutputFormat.write(out, ds)
     val str = new String(out.toByteArray)
     str shouldBe
-      """1.0 qid:1 0:1.0 1:2.0
-        |0.0 qid:1 
-        |1.0 qid:2 0:1.0 1:NaN
+      """1 qid:1 0:1.0 1:2.0
+        |0 qid:1
+        |1 qid:2 0:1.0 1:NaN
         |""".stripMargin
   }
 
@@ -36,9 +36,9 @@ class LibSVMOutputFormatTest extends AnyFlatSpec with Matchers {
     val str  = new String(data.toByteArray)
     val gstr = new String(groups.toByteArray)
     str shouldBe
-      """1.0 0:1.0 1:2.0
-        |0.0 
-        |1.0 0:1.0 1:NaN
+      """1 0:1.0 1:2.0
+        |0
+        |1 0:1.0 1:NaN
         |""".stripMargin
     gstr shouldBe "2\n1\n"
   }