Question

I have to often transpose a “rectangular” collection-of-collections in Scala, e.g.: a list of maps, a map of lists, a map of maps, a set of lists, a map of sets etc. Since collections can be uniformly viewed as a mapping from a specific domain to a co-domain (e.g.: a List[A]/Array[A] is a mapping from the Int domain to the A co-domain, Set[A]is a mapping from the A domain to the Boolean co-domain etc.), I’d like to write a clean, generic function to do a transpose operation (e.g.: turn a map of lists to the transposed list of maps). However, I’m having trouble because other than the () operator, Scala doesn’t seem to have a unified API to view collections abstractly as mappings ?

So I end up writing a separate transpose for each type of collection-of-collections as follows:

def transposeMapOfLists[A,B]( mapOfLists: Map[A,List[B]] ) : List[Map[A,B]] = {
  val k = ( mapOfLists keys ) toList
  val l = ( k map { mapOfLists(_) } ) transpose;
  l map {  v => ( k zip v ) toMap }
}

def transposeListOfMaps[A,B]( listOfMaps: List[Map[A,B]]) : Map[A,List[B]] = {
  val k = ( listOfMaps(0) keys ) toList
  val l = ( listOfMaps map { m => k map { m(_) } } ) transpose;
  ( k zip l ) toMap
}

def transposeMapOfMaps[A,B,C]( mapOfMaps: Map[A,Map[B,C]] ) : Map[B,Map[A,C]] = {
  val k = ( mapOfMaps keys ) toList
  val listOfMaps = k map { mapOfMaps(_) }
  val mapOfLists = transposeListOfMaps( listOfMaps )
  mapOfLists map { p => ( p._1, ( k zip p._2 ) toMap ) }
}

Can someone help me unify these methods into one generic collection-of-collections transpose ? It will also help me (and I am sure others) learn some useful Scala features in the process.

ps: I have ignored exception handling and have assumed the input collection-of-collections is rectangular, i.e., all of the inner collections’ domain elements constitute the same set.

Answer 1

I’m sure the following messy version using type classes could be cleaned up a lot, but it works as a quick proof-of-concept. I don’t see an easy way to get the return types right without dependent method types (I’m sure it’s possible), so you’ll have to use -Xexperimental:

trait Mapping[A, B, C] {
  type M[D] <: PartialFunction[A, D]
  def domain(c: C): Seq[A]
  def fromPairs[D](ps: Seq[(A, D)]): M[D]
  def codomain(c: C)(implicit ev: C <:< PartialFunction[A, B]) =
    domain(c).map(c)
  def toPairs(c: C)(implicit ev: C <:< PartialFunction[A, B]) =
    domain(c).map(a => (a, c(a)))
}

implicit def seqMapping[A, B <: Seq[A]] = new Mapping[Int, A, B] {
  type M[C] = Seq[C]
  def domain(c: B) = 0 until c.size
  def fromPairs[C](ps: Seq[(Int, C)]) = ps.sortBy(_._1).map(_._2)
}

implicit def mapMapping[A, B, C <: Map[A, B]] = new Mapping[A, B, C] {
  type M[D] = Map[A, D]
  def domain(c: C) = c.keys.toSeq
  def fromPairs[D](ps: Seq[(A, D)]) = ps.toMap
}

def transpose[A, B, C, M, N](m: M)(implicit
  pev: M <:< PartialFunction[A, N],
  qev: N <:< PartialFunction[B, C],
  mev: Mapping[A, N, M],
  nev: Mapping[B, C, N]
) = nev.fromPairs(nev.domain(mev.codomain(m).head).map(b =>
    b -> mev.fromPairs(mev.toPairs(m).map { case (a, c) => a -> c(b) })
))

And now for some tests:

scala> println(transpose(List(Map("a" -> 1, "b" -> 13), Map("b" -> 99, "a" -> 14))))
Map(a -> Vector(1, 14), b -> Vector(13, 99))

scala> println(transpose(Map('a' -> List(1, 2, 3), 'z' -> List(4, 5, 6))))
Vector(Map(a -> 1, z -> 4), Map(a -> 2, z -> 5), Map(a -> 3, z -> 6))

scala> println(transpose(Map("x" -> Map(4 -> 'a, 99 -> 'z), "y" -> Map(4 -> 'b, 99 -> 's))))
Map(4 -> Map(x -> 'a, y -> 'b), 99 -> Map(x -> 'z, y -> 's))

So it’s working as desired.