Merge pull request #243 from JuliaStats/ast/fix_rnadindex

randindex(): fix int overflow

Author: alyst

src/confusion.jl

@@ -1,11 +1,14 @@
 """
-    confusion(a::Union{ClusteringResult, AbstractVector},
-              b::Union{ClusteringResult, AbstractVector}) -> Matrix{Int}
+    confusion([T = Int],
+              a::Union{ClusteringResult, AbstractVector},
+              b::Union{ClusteringResult, AbstractVector}) -> Matrix{T}
 
-Return 2×2 confusion matrix `C` that represents partition co-occurrence or
-similarity matrix between two clusterings by considering all pairs of samples
-and counting  pairs that are assigned into the same or into different clusters
-under the true and predicted clusterings.
+Calculate the *confusion matrix* of the two clusterings.
+
+Returns the 2×2 confusion matrix `C` of type `T` (`Int` by default) that
+represents partition co-occurrence or similarity matrix between two clusterings
+`a` and `b` by considering all pairs of samples and counting pairs that are
+assigned into the same or into different clusters.
 
 Considering a pair of samples that is in the same group as a **positive pair**,
 and a pair is in the different group as a **negative pair**, then the count of
@@ -17,23 +20,27 @@ true negatives is `C₂₂`:
 |Positive|C₁₁|C₁₂|
 |Negative|C₂₁|C₂₂|
 """
-function confusion(a::AbstractVector{<:Integer}, b::AbstractVector{<:Integer})
-    c = counts(a, b)
+function confusion(::Type{T}, a::AbstractVector{<:Integer}, b::AbstractVector{<:Integer}) where T<:Union{Integer, AbstractFloat}
+    cc = counts(a, b)
+    c = eltype(cc) === T ? cc : convert(Matrix{T}, cc)
 
     n = sum(c)
-    nis = sum(abs2, sum(c, dims=2))        # sum of squares of sums of rows
-    njs = sum(abs2, sum(c, dims=1))        # sum of squares of sums of columns
+    nis = sum(abs2, sum!(zeros(T, (size(c, 1), 1)), c))
+    (nis < 0) && OverflowError("sum of squares of sums of rows overflowed")
+    njs = sum(abs2, sum!(zeros(T, (1, size(c, 2))), c))
+    (njs < 0) && OverflowError("sum of squares of sums of columns overflowed")
 
-    t2 = sum(abs2, c)                      # sum over rows & columns of nij^2
+    t2 = sum(abs2, c)
+    (t2 < 0) && OverflowError("sum of squares of matrix elements overflowed")
     t3 = nis + njs
     C = [(t2 - n)÷2 (nis - t2)÷2; (njs - t2)÷2 (t2 + n^2 - t3)÷2]
     return C
 end
 
-confusion(a::ClusteringResult, b::ClusteringResult) =
-    confusion(assignments(a), assignments(b))
-confusion(a::AbstractVector{<:Integer}, b::ClusteringResult) =
-    confusion(a, assignments(b))
-confusion(a::ClusteringResult, b::AbstractVector{<:Integer}) =
-    confusion(assignments(a), b)
+confusion(T, a::ClusteringResultOrAssignments,
+             b::ClusteringResultOrAssignments) =
+    confusion(T, assignments(a), assignments(b))
 
+confusion(a::ClusteringResultOrAssignments,
+          b::ClusteringResultOrAssignments) =
+    confusion(Int, a, b)

src/randindex.jl

@@ -23,7 +23,7 @@ Returns a tuple of indices:
 > Rand Index.* Psychological Methods, Vol. 9, No. 3: 386-396
 """
 function randindex(a, b)
-    c11, c21, c12, c22 = confusion(a, b) # Table 2 from Steinley 2004
+    c11, c21, c12, c22 = confusion(Float64, a, b) # Table 2 from Steinley 2004
 
     t = c11 + c12 + c21 + c22   # total number of pairs of entities
     A = c11 + c22
@@ -32,7 +32,7 @@ function randindex(a, b)
     # expected index
     ERI = (c11+c12)*(c11+c21)+(c21+c22)*(c12+c22)
     # adjusted Rand - Hubert & Arabie 1985
-    ARI = D == 0 ? 1.0 : (t*A-ERI)/(t*t-ERI) # (9) from Steinley 2004
+    ARI = D == 0 ? 1.0 : (t*A - ERI)/(abs2(t) - ERI) # (9) from Steinley 2004
 
     RI = A/t            # Rand 1971      # Probability of agreement
     MI = D/t            # Mirkin 1970    # p(disagreement)

src/utils.jl

@@ -9,6 +9,11 @@ Base type for the output of clustering algorithm.
 """
 abstract type ClusteringResult end
 
+# vector of cluster indices for each clustered point
+ClusterAssignments = AbstractVector{<:Integer}
+
+ClusteringResultOrAssignments = Union{ClusteringResult, ClusterAssignments}
+
 # generic functions
 
 """
@@ -48,6 +53,7 @@ Get the vector of cluster indices for each point.
 is assigned.
 """
 assignments(R::ClusteringResult) = R.assignments
+assignments(A::ClusterAssignments) = A
 
 
 ##### convert display symbol to disp level

test/confusion.jl

@@ -10,21 +10,27 @@ using Clustering
         @test confusion([0,0,1], [0,0,0]) == [1 0; 2 0]
         @test confusion([0,1,1], [0,0,0]) == [1 0; 2 0]
         @test confusion([1,1,1], [0,0,0]) == [3 0; 0 0]
-        
+
         @test confusion([0,0,0], [0,0,1]) == [1 2; 0 0]
         @test confusion([0,0,1], [0,0,1]) == [1 0; 0 2]
         @test confusion([0,1,1], [0,0,1]) == [0 1; 1 1]
         @test confusion([1,1,1], [0,0,1]) == [1 2; 0 0]
-        
+
         @test confusion([0,0,0], [0,1,1]) == [1 2; 0 0]
         @test confusion([0,0,1], [0,1,1]) == [0 1; 1 1]
         @test confusion([0,1,1], [0,1,1]) == [1 0; 0 2]
         @test confusion([1,1,1], [0,1,1]) == [1 2; 0 0]
-        
+
         @test confusion([0,0,0], [1,1,1]) == [3 0; 0 0]
         @test confusion([0,0,1], [1,1,1]) == [1 0; 2 0]
         @test confusion([0,1,1], [1,1,1]) == [1 0; 2 0]
         @test confusion([1,1,1], [1,1,1]) == [3 0; 0 0]
+
+    end
+
+    @testset "specifying element type" begin
+        @test @inferred(confusion(Int, [1,1,1], [1,1,1])) isa Matrix{Int}
+        @test @inferred(confusion(Float64, [1,1,1], [1,1,1])) isa Matrix{Float64}
     end
 
     @testset "comparing 2 k-means clusterings" begin
@@ -38,6 +44,9 @@ using Clustering
         r2 = kmeans(x, k; maxiter=5)
         C = confusion(r1, r2)
         @test C == [n*(n-1)/2 0; 0 0]
+
+        C = confusion(Float64, r1, r2)
+        @test C == [n*(n-1)/2 0; 0 0]
     end
 
 end

test/randindex.jl

@@ -36,7 +36,15 @@ a3 = [3, 3, 3, 2, 2, 2, 1, 1, 1, 1]
 
 @test randindex(ones(Int, 3), ones(Int, 3)) == (1, 1, 0, 1)
 
-a, b = rand(1:5, 10_000), rand(1:5, 10_000)
-@test randindex(a, b)[1] < 1.0e-2
+@testset "large independent clusterings (#225)" begin
+    rng = MersenneTwister(123)
+
+    n = 10_000_000
+    k = 5 # number of clusters
+    a = rand(rng, 1:k, n)
+    b = rand(rng, 1:k, n)
+
+    @test collect(randindex(a, b)) ≈ [0.0, ((k-1)^2 + 1)/k^2, 2*(k-1)/k^2, ((k-2)/k)^2] atol=1e-5
+end
 
 end