Add new distance functions for Multiclass and OrderedFactor. (#5)

* Add new distance functions for Multiclass and OrderedFactor. * Update src/TableDistances.jl * Move CategoricalArrays to before CoDa. Remove unecessary levelcode and type conversion. * Add MulticlassDistance and OrderedFactorDistance to tests. * Update result_type to return Bool instead of Float64 for MulticlassDistance. * Add docstring and move from the top of file the new defined distances. * Quick update. * Update src/distances.jl * Update src/distances.jl Co-authored-by: Júlio Hoffimann <[email protected]>
JuliaML · Oct 5, 2021 · 43eacfc · 43eacfc
1 parent 0700289
commit 43eacfc
Show file tree

Hide file tree

Showing 5 changed files with 40 additions and 3 deletions.
diff --git a/Project.toml b/Project.toml
@@ -4,6 +4,7 @@ authors = ["Júlio Hoffimann <[email protected]>", "José Augusto <jose.
 version = "0.1.0"
 
 [deps]
+CategoricalArrays = "324d7699-5711-5eae-9e2f-1d82baa6b597"
 CoDa = "5900dafe-f573-5c72-b367-76665857777b"
 Distances = "b4f34e82-e78d-54a5-968a-f98e89d6e8f7"
 ScientificTypes = "321657f4-b219-11e9-178b-2701a2544e81"
@@ -13,6 +14,7 @@ TableOperations = "ab02a1b2-a7df-11e8-156e-fb1833f50b87"
 Tables = "bd369af6-aec1-5ad0-b16a-f7cc5008161c"
 
 [compat]
+CategoricalArrays = "0.10"
 CoDa = "0.6"
 Distances = "0.10"
 ScientificTypes = "2.3"

diff --git a/src/TableDistances.jl b/src/TableDistances.jl
@@ -9,10 +9,11 @@ using TableOperations
 using ScientificTypes
 using Distances
 using StringDistances
+using CategoricalArrays
 using CoDa
 using Statistics
 
-import Distances: pairwise
+import Distances: pairwise, result_type
 
 include("distances.jl")
 include("normalizations.jl")

diff --git a/src/distances.jl b/src/distances.jl
@@ -2,10 +2,32 @@
 # Licensed under the MIT License. See LICENCE in the project root.
 # ------------------------------------------------------------------
 
+"""
+    MulticlassDistance()(x, y)
+
+Return `true` if x and y are different classes, else return `false`.
+"""
+struct MulticlassDistance <: Metric end
+
+(::MulticlassDistance)(x, y) = x != y
+
+result_type(::MulticlassDistance, x, y) = Bool
+
+"""
+    OrderedFactorDistance()(x, y)
+
+Return the absolute value of the difference between the categorical codes of x and y.
+"""
+struct OrderedFactorDistance <: Metric end
+
+(::OrderedFactorDistance)(x, y) = abs(levelcode(x) - levelcode(y))
+
+result_type(::OrderedFactorDistance, x, y) = Float64
+
 default_distance(::Type{Continuous})      = Distances.Euclidean()
 default_distance(::Type{Count})           = Distances.Cityblock()
-default_distance(::Type{<:Multiclass})    = Distances.Hamming()
-default_distance(::Type{<:OrderedFactor}) = Distances.Chebyshev()
+default_distance(::Type{<:Multiclass})    = MulticlassDistance()
+default_distance(::Type{<:OrderedFactor}) = OrderedFactorDistance()
 default_distance(::Type{Textual})         = StringDistances.Levenshtein()
 default_distance(::Type{<:Compositional}) = CoDa.CoDaDistance()
 

diff --git a/test/Project.toml b/test/Project.toml
@@ -1,4 +1,5 @@
 [deps]
+CategoricalArrays = "324d7699-5711-5eae-9e2f-1d82baa6b597"
 CoDa = "5900dafe-f573-5c72-b367-76665857777b"
 Distances = "b4f34e82-e78d-54a5-968a-f98e89d6e8f7"
 ScientificTypes = "321657f4-b219-11e9-178b-2701a2544e81"

diff --git a/test/runtests.jl b/test/runtests.jl
@@ -1,6 +1,7 @@
 using TableDistances
 using Tables
 using Test
+using CategoricalArrays
 using CoDa
 using Distances
 using ScientificTypes
@@ -10,12 +11,15 @@ using ScientificTypes
     # test data
     table₁ = (a = rand(4), b = rand(Composition{5}, 4))
     table₂ = (a = rand(6), b = rand(Composition{5}, 6))
+    table₃ = (a = categorical(["a", "b", "a", "c"]), b = categorical([1, 4, 1, 5]))
 
     # specific columns
     euclidcol₁ = Tables.getcolumn(table₁, :a)
     euclidcol₂ = Tables.getcolumn(table₂, :a)
     codacol₁   = Tables.getcolumn(table₁, :b)
     codacol₂   = Tables.getcolumn(table₂, :b)
+    multiclass = Tables.getcolumn(table₃, :a)
+    ordered    = Tables.getcolumn(table₃, :b)
 
     # column normalization
     D₁ = pairwise(TableDistance(normalize=true), table₁, table₂)
@@ -36,5 +40,12 @@ using ScientificTypes
     D₂ = 0.5*pairwise(Euclidean(), euclidcol₁) +
          0.5*pairwise(CoDaDistance(), codacol₁)
     @test D₁ ≈ D₂
+
+    # pairwise with multiclass and ordered factor
+    table₃ = coerce(table₃, :a => Multiclass, :b => OrderedFactor)
+    D₁ = pairwise(TableDistance(normalize=false), table₃)
+    D₂ = 0.5*pairwise(TableDistances.MulticlassDistance(), multiclass) +
+         0.5*pairwise(TableDistances.OrderedFactorDistance(), ordered)
+    @test D₁ ≈ D₂
   end
 end