feat: support for probabilistic outputs of one-class SVM

lib/rumale/svm/one_class_svm.rb

@@ -24,15 +24,16 @@ class OneClassSVM
       # @param gamma [Float] The gamma parameter in rbf/poly/sigmoid kernel function.
       # @param coef0 [Float] The coefficient in poly/sigmoid kernel function.
       # @param shrinking [Boolean] The flag indicating whether to use the shrinking heuristics.
+      # @param probability [Boolean] The flag indicating whether to train the parameter for probability estimation.
       # @param cache_size [Float] The cache memory size in MB.
       # @param tol [Float] The tolerance of termination criterion.
       # @param verbose [Boolean] The flag indicating whether to output learning process message
       # @param random_seed [Integer/Nil] The seed value using to initialize the random generator.
       def initialize(nu: 1.0, kernel: 'rbf', degree: 3, gamma: 1.0, coef0: 0.0,
-                     shrinking: true, cache_size: 200.0, tol: 1e-3, verbose: false, random_seed: nil)
+                     shrinking: true, probability: true, cache_size: 200.0, tol: 1e-3, verbose: false, random_seed: nil)
         check_params_numeric(nu: nu, degree: degree, gamma: gamma, coef0: coef0, cache_size: cache_size, tol: tol)
         check_params_string(kernel: kernel)
-        check_params_boolean(shrinking: shrinking, verbose: verbose)
+        check_params_boolean(shrinking: shrinking, probability: probability, verbose: verbose)
         check_params_numeric_or_nil(random_seed: random_seed)
         @params = {}
         @params[:nu] = nu.to_f
@@ -41,6 +42,7 @@ def initialize(nu: 1.0, kernel: 'rbf', degree: 3, gamma: 1.0, coef0: 0.0,
         @params[:gamma] = gamma.to_f
         @params[:coef0] = coef0.to_f
         @params[:shrinking] = shrinking
+        @params[:probability] = probability
         @params[:cache_size] = cache_size.to_f
         @params[:tol] = tol.to_f
         @params[:verbose] = verbose
@@ -82,6 +84,19 @@ def predict(x)
         Numo::Int32.cast(Numo::Libsvm.predict(x, libsvm_params, @model))
       end
 
+      # Predict class probability for samples.
+      # This method works correctly only if the probability parameter is true.
+      #
+      # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to predict the probailities.
+      #   If the kernel is 'precomputed', the shape of x must be [n_samples, n_training_samples].
+      # @return [Numo::DFloat] (shape: [n_samples, n_classes]) Predicted probability of each class per sample.
+      def predict_proba(x)
+        raise "#{self.class.name}\##{__method__} expects to be called after training the model with the fit method." unless trained?
+        raise "#{self.class.name}\##{__method__} expects to be called after training the probablity parameters." unless trained_probs?
+        x = check_convert_sample_array(x)
+        Numo::Libsvm.predict_proba(x, libsvm_params, @model)
+      end
+
       # Dump marshal data.
       # @return [Hash] The marshal data about SVC.
       def marshal_dump
@@ -150,6 +165,10 @@ def libsvm_params
       def trained?
         !@model.nil?
       end
+
+      def trained_probs?
+        @model[:prob_density_marks].is_a?(Numo::NArray)
+      end
     end
   end
 end

rumale-svm.gemspec

@@ -36,6 +36,6 @@ Gem::Specification.new do |spec|
   spec.require_paths = ['lib']
 
   spec.add_dependency 'numo-liblinear', '~> 2.0'
-  spec.add_dependency 'numo-libsvm', '~> 2.0'
+  spec.add_dependency 'numo-libsvm', '~> 2.1'
   spec.add_dependency 'rumale', '~> 0.14', '< 0.24'
 end

spec/rumale/svm/one_class_svm_spec.rb

@@ -24,6 +24,7 @@
   shared_examples 'distribution estimation task' do
     let(:dfs) { ocsvm.decision_function(x) }
     let(:predicted) { ocsvm.predict(x) }
+    let(:probs) { ocsvm.predict_proba(x) }
     let(:n_sv) { ocsvm.n_support }
 
     before { ocsvm.fit(x_pos) }
@@ -58,6 +59,13 @@
       expect(dfs.ge(0).count).to eq(predicted.eq(1).count)
     end
 
+    it 'estimates probabilities', :aggregate_failures do
+      expect(probs.class).to eq(Numo::DFloat)
+      expect(probs.ndim).to eq(2)
+      expect(probs.shape[0]).to eq(n_samples)
+      expect(probs.shape[1]).to eq(2)
+    end
+
     it 'dumps and restores itself using Marshal module', :aggregate_failures do
       expect(copied.instance_variable_get(:@model)).to eq(ocsvm.instance_variable_get(:@model))
       expect(copied.params).to eq(ocsvm.params)