Bump LIBSVM version to v0.7

The svmtrain and svmpredict functions slightly changed in 0.7 for custom
kernels, therefore some code had to be changed.

Author: simonschoelly

Project.toml

@@ -18,7 +18,7 @@ ThreadsX = "ac1d9e8a-700a-412c-b207-f0111f4b6c0d"
 [compat]
 DataStructures = "0.18.9"
 KernelFunctions = "0.10.5"
-LIBSVM = "0.6"
+LIBSVM = "0.7"
 LightGraphs = "1.3"
 SimpleValueGraphs = "0.3"
 ThreadsX = "0.1"

src/GraphKernels.jl

@@ -55,7 +55,7 @@ include("integrations/LIBSVM.jl")
 
 Simple k-fold cross validation implementation for quick testing during development.
 """
-function k_fold_cross_validation(kernel::AbstractGraphKernel, graphs; k_folds=5, class_key=1, kwargs...)
+function k_fold_cross_validation(kernel::KernelFunctions.Kernel, graphs::AbstractVector{<:AbstractGraph}; k_folds=5, class_key=1, kwargs...)
 
     n = length(graphs)
     indices = randperm(MersenneTwister(123), n)

src/integrations/LIBSVM.jl

@@ -1,17 +1,14 @@
 
 struct GraphSVMModel
     svm::LIBSVM.SVM
-    kernel::AbstractGraphKernel
+    kernel::KernelFunctions.Kernel
     graphs::AbstractVector
 end
 
 
-function svmtrain(graphs::AbstractVector{<:AbstractGraph}, labels, kernel::AbstractGraphKernel; kwargs...)
-
-    n = length(graphs)
-
-    X = vcat(transpose(1:n), kernelmatrix(kernel, graphs))
+function svmtrain(graphs::AbstractVector{<:AbstractGraph}, labels, kernel::KernelFunctions.Kernel; kwargs...)
 
+    X = kernelmatrix(kernel, graphs)
     svm = svmtrain(X, labels, kernel=Kernel.Precomputed; kwargs...)
 
     return GraphSVMModel(svm, kernel, graphs)
@@ -22,12 +19,7 @@ function svmpredict(model::GraphSVMModel, unpredicted_graphs::AbstractVector{<:A
     graphs = model.graphs
     kernel = model.kernel
 
-    m = length(graphs)
-    n = length(unpredicted_graphs)
-
-    X = Matrix{Float64}(undef, m + 1, n)
-    X[1, :] = 1:n
-    X[2:end, :] = kernelmatrix(kernel, graphs, unpredicted_graphs)
-
+    # TODO might only be necessary to do the calculations for support vectors
+    X = kernelmatrix(kernel, graphs, unpredicted_graphs)
     return svmpredict(model.svm, X)[1] # for simplicity return only the labels for now
 end