Rename output columns in ProjectionPursuit and LogRatio transforms (#248)

Author: eliascarv

src/transforms/logratio.jl

@@ -14,7 +14,6 @@ abstract type LogRatio <: StatelessFeatureTransform end
 # log-ratio transform interface
 function refvar end
 function newvars end
-function oldvars end
 function applymatrix end
 function revertmatrix end
 
@@ -24,19 +23,20 @@ assertions(::LogRatio) = [SciTypeAssertion{Continuous}()]
 
 function applyfeat(transform::LogRatio, feat, prep)
   cols = Tables.columns(feat)
-  names = Tables.columnnames(cols) |> collect
+  onames = Tables.columnnames(cols)
+  varnames = collect(onames)
 
   # reference variable
-  rvar = refvar(transform, names)
-  _assert(rvar ∈ names, "invalid reference variable")
-  rind = findfirst(==(rvar), names)
+  rvar = refvar(transform, varnames)
+  _assert(rvar ∈ varnames, "invalid reference variable")
+  rind = findfirst(==(rvar), varnames)
 
   # permute columns if necessary
-  perm = rind ≠ lastindex(names)
+  perm = rind ≠ lastindex(varnames)
   pfeat = if perm
-    popat!(names, rind)
-    push!(names, rvar)
-    feat |> Select(names)
+    popat!(varnames, rind)
+    push!(varnames, rvar)
+    feat |> Select(varnames)
   else
     feat
   end
@@ -46,35 +46,28 @@ function applyfeat(transform::LogRatio, feat, prep)
   Y = applymatrix(transform, X)
 
   # new variable names
-  newnames = newvars(transform, names)
+  newnames = newvars(transform, varnames)
 
   # return same table type
   𝒯 = (; zip(newnames, eachcol(Y))...)
   newfeat = 𝒯 |> Tables.materializer(feat)
 
-  newfeat, (rvar, rind, perm)
+  newfeat, (rind, perm, onames)
 end
 
 function revertfeat(transform::LogRatio, newfeat, fcache)
-  cols = Tables.columns(newfeat)
-  names = Tables.columnnames(cols)
-
   # revert transform
   Y = Tables.matrix(newfeat)
   X = revertmatrix(transform, Y)
 
   # retrieve cache
-  rvar, rind, perm = fcache
-
-  # original variable names
-  onames = oldvars(transform, names, rvar)
+  rind, perm, onames = fcache
 
   # revert the permutation if necessary
   if perm
     n = length(onames)
     inds = collect(1:(n - 1))
     insert!(inds, rind, n)
-    onames = onames[inds]
     X = X[:, inds]
   end
 

src/transforms/logratio/alr.jl

@@ -18,9 +18,7 @@ ALR() = ALR(nothing)
 
 refvar(transform::ALR, names) = isnothing(transform.refvar) ? last(names) : transform.refvar
 
-newvars(::ALR, names) = collect(names)[begin:(end - 1)]
-
-oldvars(::ALR, names, rvar) = [collect(names); rvar]
+newvars(::ALR, names) = Symbol.(:ARL, 1:(length(names) - 1))
 
 applymatrix(::ALR, X) = mapslices(alr ∘ Composition, X, dims=2)
 

src/transforms/logratio/clr.jl

@@ -11,9 +11,7 @@ struct CLR <: LogRatio end
 
 refvar(::CLR, names) = last(names)
 
-newvars(::CLR, names) = collect(names)
-
-oldvars(::CLR, names, rvar) = collect(names)
+newvars(::CLR, names) = Symbol.(:CLR, 1:length(names))
 
 applymatrix(::CLR, X) = mapslices(clr ∘ Composition, X, dims=2)
 

src/transforms/logratio/ilr.jl

@@ -18,9 +18,7 @@ ILR() = ILR(nothing)
 
 refvar(transform::ILR, names) = isnothing(transform.refvar) ? last(names) : transform.refvar
 
-newvars(::ILR, names) = collect(names)[begin:(end - 1)]
-
-oldvars(::ILR, names, rvar) = [collect(names); rvar]
+newvars(::ILR, names) = Symbol.(:ILR, 1:(length(names) - 1))
 
 applymatrix(::ILR, X) = mapslices(ilr ∘ Composition, X, dims=2)
 

src/transforms/projectionpursuit.jl

@@ -153,13 +153,13 @@ end
 
 sphering() = Quantile() → EigenAnalysis(:VDV)
 
-function applyfeat(transform::ProjectionPursuit, table, prep)
-  # retrieve column names
-  cols = Tables.columns(table)
-  names = Tables.columnnames(cols)
+function applyfeat(transform::ProjectionPursuit, feat, prep)
+  # original columns names
+  cols = Tables.columns(feat)
+  onames = Tables.columnnames(cols)
 
   # preprocess the data to approximately spherical shape
-  ptable, pcache = apply(sphering(), table)
+  ptable, pcache = apply(sphering(), feat)
 
   # initialize scores and Gaussian quantiles
   Z = Tables.matrix(ptable)
@@ -183,20 +183,19 @@ function applyfeat(transform::ProjectionPursuit, table, prep)
     iter += 1
   end
 
+  # new column names
+  names = Symbol.(:PP, 1:size(Z, 2))
+
   𝒯 = (; zip(names, eachcol(Z))...)
-  newtable = 𝒯 |> Tables.materializer(table)
-  newtable, (pcache, caches)
+  newtable = 𝒯 |> Tables.materializer(feat)
+  newtable, (pcache, caches, onames)
 end
 
-function revertfeat(::ProjectionPursuit, newtable, fcache)
-  # retrieve column names
-  cols = Tables.columns(newtable)
-  names = Tables.columnnames(cols)
-
+function revertfeat(::ProjectionPursuit, newfeat, fcache)
   # caches to retrieve transform steps
-  pcache, caches = fcache
+  pcache, caches, onames = fcache
 
-  Z = Tables.matrix(newtable)
+  Z = Tables.matrix(newfeat)
   for (Q, qcache) in reverse(caches)
     table = revert(Quantile(1), Tables.table(Z * Q), qcache)
     Z = Tables.matrix(table) * Q'
@@ -205,6 +204,6 @@ function revertfeat(::ProjectionPursuit, newtable, fcache)
   table = revert(sphering(), Tables.table(Z), pcache)
   Z = Tables.matrix(table)
 
-  𝒯 = (; zip(names, eachcol(Z))...)
-  newtable = 𝒯 |> Tables.materializer(newtable)
+  𝒯 = (; zip(onames, eachcol(Z))...)
+  𝒯 |> Tables.materializer(newfeat)
 end

test/data/projectionpursuit-1.png

@@ -10,13 +10,16 @@
 
   T = ALR()
   n, c = apply(T, t)
+  @test Tables.schema(n).names == (:ARL1, :ARL2)
   @test n == t |> ALR(:c)
   talr = revert(T, n, c)
   T = CLR()
   n, c = apply(T, t)
+  @test Tables.schema(n).names == (:CLR1, :CLR2, :CLR3)
   tclr = revert(T, n, c)
   T = ILR()
   n, c = apply(T, t)
+  @test Tables.schema(n).names == (:ILR1, :ILR2)
   @test n == t |> ILR(:c)
   tilr = revert(T, n, c)
   @test Tables.matrix(talr) ≈ Tables.matrix(tclr)

test/data/projectionpursuit-3.png

@@ -10,7 +10,7 @@
   T = ProjectionPursuit(rng=MersenneTwister(2))
   n, c = apply(T, t)
 
-  @test Tables.columnnames(n) == (:a, :b, :c, :d)
+  @test Tables.columnnames(n) == (:PP1, :PP2, :PP3, :PP4)
 
   μ = mean(Tables.matrix(n), dims=1)
   Σ = cov(Tables.matrix(n))