svmkit 0.2.6 → 0.2.7

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA1:
-  metadata.gz: b36f6b299c47d1107d587aafeb7bb66531f1208c
-  data.tar.gz: 1bd382f3339b8fb08454493a45a2338020791b6c
+  metadata.gz: 46878b59860b61bae7b522fb02af984208609f56
+  data.tar.gz: 6e889c6ad8382c654455a242d2f7f27de41de2d5
 SHA512:
-  metadata.gz: 0676f2e9b3ef4ac9786f10ca976721e73d2cd918a9c939900281e36267ab14a3413c3a719d9504415f527e5d6163d5640ea2af186e023c3980327cb7c476afba
-  data.tar.gz: e8dcf72f7d1641903a4625bb23399deabf9a19931a5c00bd2c5077b525a2d0361b194b8c43dee1b7a365b25eafdabb460e0a4ae21f17afd5b401379671379463
+  metadata.gz: cddb239bf0768e6d983ce942ed6a7bdda8b827fa2e73e51c1b4591e8af3c641339377417f844358159c3a2bdff51d2f5678ef07fe21fe86e51136289e69ea38c
+  data.tar.gz: 64c2029c729de580765ad9ee89fd57821a40773721eac291201cb9b9f4c72697f5945c8f5259ed8d6a879f0b35dac841bcdd6d5322014c9cc78b9a42046dc310

data/HISTORY.md

@@ -1,3 +1,6 @@
+# 0.2.7
+- Fixed to support multiclass classifiction into LinearSVC, LogisticRegression, KernelSVC, and FactorizationMachineClassifier
+
 # 0.2.6
 - Added class for Decision Tree classifier.
 - Added class for Random Forest classifier.

data/README.md

@@ -42,9 +42,7 @@ normalized = normalizer.fit_transform(samples)
 transformer = SVMKit::KernelApproximation::RBF.new(gamma: 2.0, n_components: 1024, random_seed: 1)
 transformed = transformer.fit_transform(normalized)
 
-base_classifier =
-  SVMKit::LinearModel::SVC.new(reg_param: 1.0, max_iter: 1000, batch_size: 20, random_seed: 1)
-classifier = SVMKit::Multiclass::OneVsRestClassifier.new(estimator: base_classifier)
+classifier = SVMKit::LinearModel::SVC.new(reg_param: 1.0, max_iter: 1000, batch_size: 20, random_seed: 1)
 classifier.fit(transformed, labels)
 
 File.open('trained_normalizer.dat', 'wb') { |f| f.write(Marshal.dump(normalizer)) }
@@ -76,12 +74,10 @@ require 'svmkit'
 
 samples, labels = SVMKit::Dataset.load_libsvm_file('pendigits')
 
-kernel_svc =
-  SVMKit::KernelMachine::KernelSVC.new(reg_param: 1.0, max_iter: 1000, random_seed: 1)
-ovr_kernel_svc = SVMKit::Multiclass::OneVsRestClassifier.new(estimator: kernel_svc)
+kernel_svc = SVMKit::KernelMachine::KernelSVC.new(reg_param: 1.0, max_iter: 1000, random_seed: 1)
 
 kf = SVMKit::ModelSelection::StratifiedKFold.new(n_splits: 5, shuffle: true, random_seed: 1)
-cv = SVMKit::ModelSelection::CrossValidation.new(estimator: ovr_kernel_svc, splitter: kf)
+cv = SVMKit::ModelSelection::CrossValidation.new(estimator: kernel_svc, splitter: kf)
 
 kernel_mat = SVMKit::PairwiseMetric::rbf_kernel(samples, nil, 0.005)
 report = cv.perform(kernel_mat, labels)

data/lib/svmkit/kernel_machine/kernel_svc.rb

@@ -6,7 +6,9 @@ require 'svmkit/base/classifier'
 module SVMKit
   # This module consists of the classes that implement kernel method-based estimator.
   module KernelMachine
-    # KernelSVC is a class that implements (Nonlinear) Kernel Support Vector Classifier with the Pegasos algorithm.
+    # KernelSVC is a class that implements (Nonlinear) Kernel Support Vector Classifier
+    # with stochastic gradient descent (SGD) optimization.
+    # For multiclass classification problem, it uses one-vs-the-rest strategy.
     #
     # @example
     #   training_kernel_matrix = SVMKit::PairwiseMetric::rbf_kernel(training_samples)
@@ -23,14 +25,18 @@ module SVMKit
       include Base::Classifier
 
       # Return the weight vector for Kernel SVC.
-      # @return [Numo::DFloat] (shape: [n_trainig_sample])
+      # @return [Numo::DFloat] (shape: [n_classes, n_trainig_sample])
       attr_reader :weight_vec
 
-      # Return the random generator for performing random sampling in the Pegasos algorithm.
+      # Return the class labels.
+      # @return [Numo::Int32] (shape: [n_classes])
+      attr_reader :classes
+
+      # Return the random generator for performing random sampling.
       # @return [Random]
       attr_reader :rng
 
-      # Create a new classifier with Kernel Support Vector Machine by the Pegasos algorithm.
+      # Create a new classifier with Kernel Support Vector Machine by the SGD optimization.
       #
       # @param reg_param [Float] The regularization parameter.
       # @param max_iter [Integer] The maximum number of iterations.
@@ -42,6 +48,7 @@ module SVMKit
         @params[:random_seed] = random_seed
         @params[:random_seed] ||= srand
         @weight_vec = nil
+        @classes
         @rng = Random.new(@params[:random_seed])
       end
 
@@ -52,25 +59,22 @@ module SVMKit
       # @param y [Numo::Int32] (shape: [n_training_samples]) The labels to be used for fitting the model.
       # @return [KernelSVC] The learned classifier itself.
       def fit(x, y)
-        # Generate binary labels
-        negative_label = y.to_a.uniq.sort.shift
-        bin_y = y.to_a.map { |l| l != negative_label ? 1 : -1 }
-        # Initialize some variables.
-        n_training_samples = x.shape[0]
-        rand_ids = []
-        weight_vec = Numo::DFloat.zeros(n_training_samples)
-        # Start optimization.
-        @params[:max_iter].times do |t|
-          # random sampling
-          rand_ids = [*0...n_training_samples].shuffle(random: @rng) if rand_ids.empty?
-          target_id = rand_ids.shift
-          # update the weight vector
-          func = (weight_vec * bin_y[target_id]).dot(x[target_id, true].transpose).to_f
-          func *= bin_y[target_id] / (@params[:reg_param] * (t + 1))
-          weight_vec[target_id] += 1.0 if func < 1.0
+        @classes = Numo::Int32[*y.to_a.uniq.sort]
+        n_classes = @classes.size
+        _n_samples, n_features = x.shape
+
+        if n_classes > 2
+          @weight_vec = Numo::DFloat.zeros(n_classes, n_features)
+          n_classes.times do |n|
+            bin_y = Numo::Int32.cast(y.eq(@classes[n])) * 2 - 1
+            @weight_vec[n, true] = binary_fit(x, bin_y)
+          end
+        else
+          negative_label = y.to_a.uniq.sort.first
+          bin_y = Numo::Int32.cast(y.ne(negative_label)) * 2 - 1
+          @weight_vec = binary_fit(x, bin_y)
         end
-        # Store the learned model.
-        @weight_vec = weight_vec * Numo::DFloat[*bin_y]
+
         self
       end
 
@@ -78,9 +82,9 @@ module SVMKit
       #
       # @param x [Numo::DFloat] (shape: [n_testing_samples, n_training_samples])
       #     The kernel matrix between testing samples and training samples to compute the scores.
-      # @return [Numo::DFloat] (shape: [n_testing_samples]) Confidence score per sample.
+      # @return [Numo::DFloat] (shape: [n_testing_samples, n_classes]) Confidence score per sample.
       def decision_function(x)
-        x.dot(@weight_vec)
+        x.dot(@weight_vec.transpose)
       end
 
       # Predict class labels for samples.
@@ -89,7 +93,11 @@ module SVMKit
       #     The kernel matrix between testing samples and training samples to predict the labels.
       # @return [Numo::Int32] (shape: [n_testing_samples]) Predicted class label per sample.
       def predict(x)
-        Numo::Int32.cast(decision_function(x).map { |v| v >= 0 ? 1 : -1 })
+        return Numo::Int32.cast(decision_function(x).ge(0.0)) * 2 - 1 if @classes.size <= 2
+
+        n_samples, = x.shape
+        decision_values = decision_function(x)
+        Numo::Int32.asarray(Array.new(n_samples) { |n| @classes[decision_values[n, true].max_index] })
       end
 
       # Claculate the mean accuracy of the given testing data.
@@ -105,7 +113,10 @@ module SVMKit
       # Dump marshal data.
       # @return [Hash] The marshal data about KernelSVC.
       def marshal_dump
-        { params: @params, weight_vec: @weight_vec, rng: @rng }
+        { params: @params,
+          weight_vec: @weight_vec,
+          classes: @classes,
+          rng: @rng }
       end
 
       # Load marshal data.
@@ -113,9 +124,30 @@ module SVMKit
       def marshal_load(obj)
         @params = obj[:params]
         @weight_vec = obj[:weight_vec]
+        @classes = obj[:classes]
         @rng = obj[:rng]
         nil
       end
+
+      private
+
+      def binary_fit(x, bin_y)
+        # Initialize some variables.
+        n_training_samples = x.shape[0]
+        rand_ids = []
+        weight_vec = Numo::DFloat.zeros(n_training_samples)
+        # Start optimization.
+        @params[:max_iter].times do |t|
+          # random sampling
+          rand_ids = [*0...n_training_samples].shuffle(random: @rng) if rand_ids.empty?
+          target_id = rand_ids.shift
+          # update the weight vector
+          func = (weight_vec * bin_y[target_id]).dot(x[target_id, true].transpose).to_f
+          func *= bin_y[target_id] / (@params[:reg_param] * (t + 1))
+          weight_vec[target_id] += 1.0 if func < 1.0
+        end
+        weight_vec * Numo::DFloat[*bin_y]
+      end
     end
   end
 end

data/lib/svmkit/linear_model/logistic_regression.rb

@@ -8,7 +8,7 @@ module SVMKit
   module LinearModel
     # LogisticRegression is a class that implements Logistic Regression
     # with stochastic gradient descent (SGD) optimization.
-    # Note that the class performs as a binary classifier.
+    # For multiclass classification problem, it uses one-vs-the-rest strategy.
     #
     # @example
     #   estimator =
@@ -23,14 +23,18 @@ module SVMKit
       include Base::Classifier
 
       # Return the weight vector for Logistic Regression.
-      # @return [Numo::DFloat] (shape: [n_features])
+      # @return [Numo::DFloat] (shape: [n_classes, n_features])
       attr_reader :weight_vec
 
       # Return the bias term (a.k.a. intercept) for Logistic Regression.
-      # @return [Float]
+      # @return [Numo::DFloat] (shape: [n_classes])
       attr_reader :bias_term
 
-      # Return the random generator for transformation.
+      # Return the class labels.
+      # @return [Numo::Int32] (shape: [n_classes])
+      attr_reader :classes
+
+      # Return the random generator for performing random sampling.
       # @return [Random]
       attr_reader :rng
 
@@ -42,79 +46,59 @@ module SVMKit
       #   If fit_bias is true, the feature vector v becoms [v; bias_scale].
       # @param max_iter [Integer] The maximum number of iterations.
       # @param batch_size [Integer] The size of the mini batches.
+      # @param normalize [Boolean] The flag indicating whether to normalize the weight vector.
       # @param random_seed [Integer] The seed value using to initialize the random generator.
-      def initialize(reg_param: 1.0, fit_bias: false, bias_scale: 1.0, max_iter: 100, batch_size: 50, random_seed: nil)
+      def initialize(reg_param: 1.0, fit_bias: false, bias_scale: 1.0,
+                     max_iter: 100, batch_size: 50, normalize: true, random_seed: nil)
         @params = {}
         @params[:reg_param] = reg_param
         @params[:fit_bias] = fit_bias
         @params[:bias_scale] = bias_scale
         @params[:max_iter] = max_iter
         @params[:batch_size] = batch_size
+        @params[:normalize] = normalize
         @params[:random_seed] = random_seed
         @params[:random_seed] ||= srand
         @weight_vec = nil
-        @bias_term = 0.0
+        @bias_term = nil
+        @classes = nil
         @rng = Random.new(@params[:random_seed])
       end
 
       # Fit the model with given training data.
       #
       # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The training data to be used for fitting the model.
-      # @param y [Numo::Int32] (shape: [n_samples]) The categorical variables (e.g. labels)
-      #   to be used for fitting the model.
+      # @param y [Numo::Int32] (shape: [n_samples]) The labels to be used for fitting the model.
       # @return [LogisticRegression] The learned classifier itself.
       def fit(x, y)
-        # Generate binary labels.
-        negative_label = y.to_a.uniq.sort.shift
-        bin_y = y.to_a.map { |l| l != negative_label ? 1 : 0 }
-        # Expand feature vectors for bias term.
-        samples = x
-        if @params[:fit_bias]
-          samples = Numo::NArray.hstack(
-            [samples, Numo::DFloat.ones([x.shape[0], 1]) * @params[:bias_scale]]
-          )
-        end
-        # Initialize some variables.
-        n_samples, n_features = samples.shape
-        rand_ids = [*0...n_samples].shuffle(random: @rng)
-        weight_vec = Numo::DFloat.zeros(n_features)
-        # Start optimization.
-        @params[:max_iter].times do |t|
-          # random sampling
-          subset_ids = rand_ids.shift(@params[:batch_size])
-          rand_ids.concat(subset_ids)
-          # update the weight vector.
-          eta = 1.0 / (@params[:reg_param] * (t + 1))
-          mean_vec = Numo::DFloat.zeros(n_features)
-          subset_ids.each do |n|
-            z = weight_vec.dot(samples[n, true])
-            coef = bin_y[n] / (1.0 + Math.exp(bin_y[n] * z))
-            mean_vec += samples[n, true] * coef
+        @classes = Numo::Int32[*y.to_a.uniq.sort]
+        n_classes = @classes.size
+        _n_samples, n_features = x.shape
+
+        if n_classes > 2
+          @weight_vec = Numo::DFloat.zeros(n_classes, n_features)
+          @bias_term = Numo::DFloat.zeros(n_classes)
+          n_classes.times do |n|
+            bin_y = Numo::Int32.cast(y.eq(@classes[n]))
+            weight, bias = binary_fit(x, bin_y)
+            @weight_vec[n, true] = weight
+            @bias_term[n] = bias
           end
-          mean_vec *= eta / @params[:batch_size]
-          weight_vec = weight_vec * (1.0 - eta * @params[:reg_param]) + mean_vec
-          # scale the weight vector.
-          norm = Math.sqrt(weight_vec.dot(weight_vec))
-          scaler = (1.0 / @params[:reg_param]**0.5) / (norm + 1.0e-12)
-          weight_vec *= [1.0, scaler].min
-        end
-        # Store the learned model.
-        if @params[:fit_bias]
-          @weight_vec = weight_vec[0...n_features - 1]
-          @bias_term = weight_vec[n_features - 1]
         else
-          @weight_vec = weight_vec[0...n_features]
-          @bias_term = 0.0
+          negative_label = y.to_a.uniq.sort.first
+          bin_y = Numo::Int32.cast(y.ne(negative_label))
+          @weight_vec, @bias_term = binary_fit(x, bin_y)
         end
+
         self
       end
 
       # Calculate confidence scores for samples.
       #
       # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to compute the scores.
-      # @return [Numo::DFloat] (shape: [n_samples]) Confidence score per sample.
+      # @return [Numo::DFloat] (shape: [n_samples, n_classes]) Confidence score per sample.
       def decision_function(x)
-        @weight_vec.dot(x.transpose) + @bias_term
+        x.dot(@weight_vec.transpose) + @bias_term
       end
 
       # Predict class labels for samples.
@@ -122,7 +106,11 @@ module SVMKit
       # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to predict the labels.
       # @return [Numo::Int32] (shape: [n_samples]) Predicted class label per sample.
       def predict(x)
-        Numo::Int32.cast(sigmoid(decision_function(x)).map { |v| v >= 0.5 ? 1 : -1 })
+        return Numo::Int32.cast(decision_function(x).ge(0.5)) * 2 - 1 if @classes.size <= 2
+
+        n_samples, = x.shape
+        decision_values = decision_function(x)
+        Numo::Int32.asarray(Array.new(n_samples) { |n| @classes[decision_values[n, true].max_index] })
       end
 
       # Predict probability for samples.
@@ -130,17 +118,24 @@ module SVMKit
       # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to predict the probailities.
       # @return [Numo::DFloat] (shape: [n_samples, n_classes]) Predicted probability of each class per sample.
       def predict_proba(x)
+        proba = 1.0 / (Numo::NMath.exp(-decision_function(x)) + 1.0)
+        return (proba.transpose / proba.sum(axis: 1)).transpose if @classes.size > 2
+
         n_samples, = x.shape
-        proba = Numo::DFloat.zeros(n_samples, 2)
-        proba[true, 1] = sigmoid(decision_function(x))
-        proba[true, 0] = 1.0 - proba[true, 1]
-        proba
+        probs = Numo::DFloat.zeros(n_samples, 2)
+        probs[true, 1] = proba
+        probs[true, 0] = 1.0 - proba
+        probs
       end
 
       # Dump marshal data.
       # @return [Hash] The marshal data about LogisticRegression.
       def marshal_dump
-        { params: @params, weight_vec: @weight_vec, bias_term: @bias_term, rng: @rng }
+        { params: @params,
+          weight_vec: @weight_vec,
+          bias_term: @bias_term,
+          classes: @classes,
+          rng: @rng }
       end
 
       # Load marshal data.
@@ -149,14 +144,53 @@ module SVMKit
         @params = obj[:params]
         @weight_vec = obj[:weight_vec]
         @bias_term = obj[:bias_term]
+        @classes = obj[:classes]
         @rng = obj[:rng]
         nil
       end
 
       private
 
-      def sigmoid(x)
-        1.0 / (Numo::NMath.exp(-x) + 1.0)
+      def binary_fit(x, bin_y)
+        # Expand feature vectors for bias term.
+        samples = @params[:fit_bias] ? expand_feature(x) : x
+        # Initialize some variables.
+        n_samples, n_features = samples.shape
+        rand_ids = [*0...n_samples].shuffle(random: @rng)
+        weight_vec = Numo::DFloat.zeros(n_features)
+        # Start optimization.
+        @params[:max_iter].times do |t|
+          # random sampling
+          subset_ids = rand_ids.shift(@params[:batch_size])
+          rand_ids.concat(subset_ids)
+          # update the weight vector.
+          df = samples[subset_ids, true].dot(weight_vec.transpose)
+          coef = bin_y[subset_ids] / (Numo::NMath.exp(-bin_y[subset_ids] * df) + 1.0)
+          mean_vec = samples[subset_ids, true].transpose.dot(coef) / @params[:batch_size]
+          weight_vec -= learning_rate(t) * (@params[:reg_param] * weight_vec - mean_vec)
+          # scale the weight vector.
+          normalize_weight_vec(weight_vec) if @params[:normalize]
+        end
+        split_weight_vec_bias(weight_vec)
+      end
+
+      def expand_feature(x)
+        Numo::NArray.hstack([x, Numo::DFloat.ones([x.shape[0], 1]) * @params[:bias_scale]])
+      end
+
+      def learning_rate(iter)
+        1.0 / (@params[:reg_param] * (iter + 1))
+      end
+
+      def normalize_weight_vec(weight_vec)
+        norm = Math.sqrt(weight_vec.dot(weight_vec))
+        weight_vec * [1.0, (1.0 / @params[:reg_param]**0.5) / (norm + 1.0e-12)].min
+      end
+
+      def split_weight_vec_bias(weight_vec)
+        weights = @params[:fit_bias] ? weight_vec[0...-1] : weight_vec
+        bias = @params[:fit_bias] ? weight_vec[-1] : 0.0
+        [weights, bias]
       end
     end
   end

data/lib/svmkit/linear_model/svc.rb

@@ -6,7 +6,9 @@ require 'svmkit/base/classifier'
 module SVMKit
   # This module consists of the classes that implement generalized linear models.
   module LinearModel
-    # SVC is a class that implements Support Vector Classifier with the Pegasos algorithm.
+    # SVC is a class that implements Support Vector Classifier
+    # with stochastic gradient descent (SGD) optimization.
+    # For multiclass classification problem, it uses one-vs-the-rest strategy.
     #
     # @example
     #   estimator =
@@ -21,36 +23,44 @@ module SVMKit
       include Base::Classifier
 
       # Return the weight vector for SVC.
-      # @return [Numo::DFloat] (shape: [n_features])
+      # @return [Numo::DFloat] (shape: [n_classes, n_features])
       attr_reader :weight_vec
 
       # Return the bias term (a.k.a. intercept) for SVC.
-      # @return [Float]
+      # @return [Numo::DFloat] (shape: [n_classes])
       attr_reader :bias_term
 
-      # Return the random generator for performing random sampling in the Pegasos algorithm.
+      # Return the class labels.
+      # @return [Numo::Int32] (shape: [n_classes])
+      attr_reader :classes
+
+      # Return the random generator for performing random sampling.
       # @return [Random]
       attr_reader :rng
 
-      # Create a new classifier with Support Vector Machine by the Pegasos algorithm.
+      # Create a new classifier with Support Vector Machine by the SGD optimization.
       #
       # @param reg_param [Float] The regularization parameter.
       # @param fit_bias [Boolean] The flag indicating whether to fit the bias term.
       # @param bias_scale [Float] The scale of the bias term.
       # @param max_iter [Integer] The maximum number of iterations.
       # @param batch_size [Integer] The size of the mini batches.
+      # @param normalize [Boolean] The flag indicating whether to normalize the weight vector.
       # @param random_seed [Integer] The seed value using to initialize the random generator.
-      def initialize(reg_param: 1.0, fit_bias: false, bias_scale: 1.0, max_iter: 100, batch_size: 50, random_seed: nil)
+      def initialize(reg_param: 1.0, fit_bias: false, bias_scale: 1.0,
+                     max_iter: 100, batch_size: 50, normalize: true, random_seed: nil)
         @params = {}
         @params[:reg_param] = reg_param
         @params[:fit_bias] = fit_bias
         @params[:bias_scale] = bias_scale
         @params[:max_iter] = max_iter
         @params[:batch_size] = batch_size
+        @params[:normalize] = normalize
         @params[:random_seed] = random_seed
         @params[:random_seed] ||= srand
         @weight_vec = nil
-        @bias_term = 0.0
+        @bias_term = nil
+        @classes = nil
         @rng = Random.new(@params[:random_seed])
       end
 
@@ -60,56 +70,34 @@ module SVMKit
       # @param y [Numo::Int32] (shape: [n_samples]) The labels to be used for fitting the model.
       # @return [SVC] The learned classifier itself.
       def fit(x, y)
-        # Generate binary labels
-        negative_label = y.to_a.uniq.sort.shift
-        bin_y = y.to_a.map { |l| l != negative_label ? 1 : -1 }
-        # Expand feature vectors for bias term.
-        samples = x
-        if @params[:fit_bias]
-          samples = Numo::NArray.hstack(
-            [samples, Numo::DFloat.ones([x.shape[0], 1]) * @params[:bias_scale]]
-          )
-        end
-        # Initialize some variables.
-        n_samples, n_features = samples.shape
-        rand_ids = [*0...n_samples].shuffle(random: @rng)
-        weight_vec = Numo::DFloat.zeros(n_features)
-        # Start optimization.
-        @params[:max_iter].times do |t|
-          # random sampling
-          subset_ids = rand_ids.shift(@params[:batch_size])
-          rand_ids.concat(subset_ids)
-          target_ids = subset_ids.map { |n| n if weight_vec.dot(samples[n, true]) * bin_y[n] < 1 }.compact
-          n_subsamples = target_ids.size
-          next if n_subsamples.zero?
-          # update the weight vector.
-          eta = 1.0 / (@params[:reg_param] * (t + 1))
-          mean_vec = Numo::DFloat.zeros(n_features)
-          target_ids.each { |n| mean_vec += samples[n, true] * bin_y[n] }
-          mean_vec *= eta / n_subsamples
-          weight_vec = weight_vec * (1.0 - eta * @params[:reg_param]) + mean_vec
-          # scale the weight vector.
-          norm = Math.sqrt(weight_vec.dot(weight_vec))
-          scaler = (1.0 / @params[:reg_param]**0.5) / (norm + 1.0e-12)
-          weight_vec *= [1.0, scaler].min
-        end
-        # Store the learned model.
-        if @params[:fit_bias]
-          @weight_vec = weight_vec[0...n_features - 1]
-          @bias_term = weight_vec[n_features - 1]
+        @classes = Numo::Int32[*y.to_a.uniq.sort]
+        n_classes = @classes.size
+        _n_samples, n_features = x.shape
+
+        if n_classes > 2
+          @weight_vec = Numo::DFloat.zeros(n_classes, n_features)
+          @bias_term = Numo::DFloat.zeros(n_classes)
+          n_classes.times do |n|
+            bin_y = Numo::Int32.cast(y.eq(@classes[n])) * 2 - 1
+            weight, bias = binary_fit(x, bin_y)
+            @weight_vec[n, true] = weight
+            @bias_term[n] = bias
+          end
         else
-          @weight_vec = weight_vec[0...n_features]
-          @bias_term = 0.0
+          negative_label = y.to_a.uniq.sort.first
+          bin_y = Numo::Int32.cast(y.ne(negative_label)) * 2 - 1
+          @weight_vec, @bias_term = binary_fit(x, bin_y)
         end
+
         self
       end
 
       # Calculate confidence scores for samples.
       #
       # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to compute the scores.
-      # @return [Numo::DFloat] (shape: [n_samples]) Confidence score per sample.
+      # @return [Numo::DFloat] (shape: [n_samples, n_classes]) Confidence score per sample.
       def decision_function(x)
-        @weight_vec.dot(x.transpose) + @bias_term
+        x.dot(@weight_vec.transpose) + @bias_term
       end
 
       # Predict class labels for samples.
@@ -117,13 +105,21 @@ module SVMKit
       # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to predict the labels.
       # @return [Numo::Int32] (shape: [n_samples]) Predicted class label per sample.
       def predict(x)
-        Numo::Int32.cast(decision_function(x).map { |v| v >= 0 ? 1 : -1 })
+        return Numo::Int32.cast(decision_function(x).ge(0.0)) * 2 - 1 if @classes.size <= 2
+
+        n_samples, = x.shape
+        decision_values = decision_function(x)
+        Numo::Int32.asarray(Array.new(n_samples) { |n| @classes[decision_values[n, true].max_index] })
       end
 
       # Dump marshal data.
       # @return [Hash] The marshal data about SVC.
       def marshal_dump
-        { params: @params, weight_vec: @weight_vec, bias_term: @bias_term, rng: @rng }
+        { params: @params,
+          weight_vec: @weight_vec,
+          bias_term: @bias_term,
+          classes: @classes,
+          rng: @rng }
       end
 
       # Load marshal data.
@@ -132,9 +128,55 @@ module SVMKit
         @params = obj[:params]
         @weight_vec = obj[:weight_vec]
         @bias_term = obj[:bias_term]
+        @classes = obj[:classes]
         @rng = obj[:rng]
         nil
       end
+
+      private
+
+      def binary_fit(x, bin_y)
+        # Expand feature vectors for bias term.
+        samples = @params[:fit_bias] ? expand_feature(x) : x
+        # Initialize some variables.
+        n_samples, n_features = samples.shape
+        rand_ids = [*0...n_samples].shuffle(random: @rng)
+        weight_vec = Numo::DFloat.zeros(n_features)
+        # Start optimization.
+        @params[:max_iter].times do |t|
+          # random sampling
+          subset_ids = rand_ids.shift(@params[:batch_size])
+          rand_ids.concat(subset_ids)
+          target_ids = subset_ids.map { |n| n if weight_vec.dot(samples[n, true]) * bin_y[n] < 1 }.compact
+          n_subsamples = target_ids.size
+          next if n_subsamples.zero?
+          # update the weight vector.
+          mean_vec = samples[target_ids, true].transpose.dot(bin_y[target_ids]) / n_subsamples
+          weight_vec -= learning_rate(t) * (@params[:reg_param] * weight_vec - mean_vec)
+          # scale the weight vector.
+          normalize_weight_vec(weight_vec) if @params[:normalize]
+        end
+        split_weight_vec_bias(weight_vec)
+      end
+
+      def expand_feature(x)
+        Numo::NArray.hstack([x, Numo::DFloat.ones([x.shape[0], 1]) * @params[:bias_scale]])
+      end
+
+      def learning_rate(iter)
+        1.0 / (@params[:reg_param] * (iter + 1))
+      end
+
+      def normalize_weight_vec(weight_vec)
+        norm = Math.sqrt(weight_vec.dot(weight_vec))
+        weight_vec * [1.0, (1.0 / @params[:reg_param]**0.5) / (norm + 1.0e-12)].min
+      end
+
+      def split_weight_vec_bias(weight_vec)
+        weights = @params[:fit_bias] ? weight_vec[0...-1] : weight_vec
+        bias = @params[:fit_bias] ? weight_vec[-1] : 0.0
+        [weights, bias]
+      end
     end
   end
 end

data/lib/svmkit/multiclass/one_vs_rest_classifier.rb

@@ -4,13 +4,16 @@ require 'svmkit/base/base_estimator.rb'
 require 'svmkit/base/classifier.rb'
 
 module SVMKit
-  # This module consists of the classes that implement multi-label classification strategy.
+  # This module consists of the classes that implement multi-class classification strategy.
   module Multiclass
-    # OneVsRestClassifier is a class that implements One-vs-Rest (OvR) strategy for multi-label classification.
+    # @note
+    #   All classifier in SVMKit support multi-class classifiction since version 0.2.7.
+    #   There is no need to explicitly use this class for multiclass classifiction.
+    #
+    # OneVsRestClassifier is a class that implements One-vs-Rest (OvR) strategy for multi-class classification.
     #
     # @example
-    #   base_estimator =
-    #    SVMKit::LinearModel::PegasosSVC.new(penalty: 1.0, max_iter: 100, batch_size: 20, random_seed: 1)
+    #   base_estimator = SVMKit::LinearModel::LogisticRegression.new
     #   estimator = SVMKit::Multiclass::OneVsRestClassifier.new(estimator: base_estimator)
     #   estimator.fit(training_samples, training_labels)
     #   results = estimator.predict(testing_samples)
@@ -26,9 +29,9 @@ module SVMKit
       # @return [Numo::Int32] (shape: [n_classes])
       attr_reader :classes
 
-      # Create a new multi-label classifier with the one-vs-rest startegy.
+      # Create a new multi-class classifier with the one-vs-rest startegy.
       #
-      # @param estimator [Classifier] The (binary) classifier for construction a multi-label classifier.
+      # @param estimator [Classifier] The (binary) classifier for construction a multi-class classifier.
       def initialize(estimator: nil)
         @params = {}
         @params[:estimator] = estimator

data/lib/svmkit/polynomial_model/factorization_machine_classifier.rb

@@ -6,9 +6,9 @@ require 'svmkit/base/classifier'
 module SVMKit
   # This module consists of the classes that implement polynomial models.
   module PolynomialModel
-    # FactorizationMachineClassifier is a class that
-    # implements Fatorization Machine for binary classification
-    # with (mini-batch) stochastic gradient descent optimization.
+    # FactorizationMachineClassifier is a class that implements Factorization Machine
+    # with stochastic gradient descent (SGD) optimization.
+    # For multiclass classification problem, it uses one-vs-the-rest strategy.
     #
     # @example
     #   estimator =
@@ -26,22 +26,26 @@ module SVMKit
       include Base::Classifier
 
       # Return the factor matrix for Factorization Machine.
-      # @return [Numo::DFloat] (shape: [n_factors, n_features])
+      # @return [Numo::DFloat] (shape: [n_classes, n_factors, n_features])
       attr_reader :factor_mat
 
       # Return the weight vector for Factorization Machine.
-      # @return [Numo::DFloat] (shape: [n_features])
+      # @return [Numo::DFloat] (shape: [n_classes, n_features])
       attr_reader :weight_vec
 
       # Return the bias term for Factoriazation Machine.
-      # @return [Float]
+      # @return [Numo::DFloat] (shape: [n_classes])
       attr_reader :bias_term
 
-      # Return the random generator for transformation.
+      # Return the class labels.
+      # @return [Numo::Int32] (shape: [n_classes])
+      attr_reader :classes
+
+      # Return the random generator for random sampling.
       # @return [Random]
       attr_reader :rng
 
-      # Create a new classifier with Support Vector Machine by the Pegasos algorithm.
+      # Create a new classifier with Factorization Machine.
       #
       # @param n_factors [Integer] The maximum number of iterations.
       # @param loss [String] The loss function ('hinge' or 'logistic').
@@ -67,7 +71,8 @@ module SVMKit
         @params[:random_seed] ||= srand
         @factor_mat = nil
         @weight_vec = nil
-        @bias_term = 0.0
+        @bias_term = nil
+        @classes = nil
         @rng = Random.new(@params[:random_seed])
       end
 
@@ -77,33 +82,27 @@ module SVMKit
       # @param y [Numo::Int32] (shape: [n_samples]) The labels to be used for fitting the model.
       # @return [FactorizationMachineClassifier] The learned classifier itself.
       def fit(x, y)
-        # Generate binary labels.
-        negative_label = y.to_a.uniq.sort.shift
-        bin_y = y.map { |l| l != negative_label ? 1.0 : -1.0 }
-        # Initialize some variables.
-        n_samples, n_features = x.shape
-        rand_ids = [*0...n_samples].shuffle(random: @rng)
-        @factor_mat = rand_normal([@params[:n_factors], n_features], 0, @params[:init_std])
-        @weight_vec = Numo::DFloat.zeros(n_features)
-        @bias_term = 0.0
-        # Start optimization.
-        @params[:max_iter].times do |t|
-          # Random sampling.
-          subset_ids = rand_ids.shift(@params[:batch_size])
-          rand_ids.concat(subset_ids)
-          data = x[subset_ids, true]
-          label = bin_y[subset_ids]
-          # Calculate gradients for loss function.
-          loss_grad = loss_gradient(data, label)
-          next if loss_grad.ne(0.0).count.zero?
-          # Update each parameter.
-          @bias_term -= learning_rate(@params[:reg_param_bias], t) * bias_gradient(loss_grad)
-          @weight_vec -= learning_rate(@params[:reg_param_weight], t) * weight_gradient(loss_grad, data)
-          @params[:n_factors].times do |n|
-            @factor_mat[n, true] -= learning_rate(@params[:reg_param_factor], t) *
-                                    factor_gradient(loss_grad, data, @factor_mat[n, true])
+        @classes = Numo::Int32[*y.to_a.uniq.sort]
+        n_classes = @classes.size
+        _n_samples, n_features = x.shape
+
+        if n_classes > 2
+          @factor_mat = Numo::DFloat.zeros(n_classes, @params[:n_factors], n_features)
+          @weight_vec = Numo::DFloat.zeros(n_classes, n_features)
+          @bias_term = Numo::DFloat.zeros(n_classes)
+          n_classes.times do |n|
+            bin_y = Numo::Int32.cast(y.eq(@classes[n])) * 2 - 1
+            factor, weight, bias = binary_fit(x, bin_y)
+            @factor_mat[n, true, true] = factor
+            @weight_vec[n, true] = weight
+            @bias_term[n] = bias
           end
+        else
+          negative_label = y.to_a.uniq.sort.first
+          bin_y = Numo::Int32.cast(y.ne(negative_label)) * 2 - 1
+          @factor_mat, @weight_vec, @bias_term = binary_fit(x, bin_y)
         end
+
         self
       end
 
@@ -112,8 +111,12 @@ module SVMKit
       # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to compute the scores.
       # @return [Numo::DFloat] (shape: [n_samples]) Confidence score per sample.
       def decision_function(x)
-        linear_term = @bias_term + x.dot(@weight_vec)
-        factor_term = 0.5 * (@factor_mat.dot(x.transpose)**2 - (@factor_mat**2).dot(x.transpose**2)).sum
+        linear_term = @bias_term + x.dot(@weight_vec.transpose)
+        factor_term = if @classes.size <= 2
+                        0.5 * (@factor_mat.dot(x.transpose)**2 - (@factor_mat**2).dot(x.transpose**2)).sum
+                      else
+                        0.5 * (@factor_mat.dot(x.transpose)**2 - (@factor_mat**2).dot(x.transpose**2)).sum(1).transpose
+                      end
         linear_term + factor_term
       end
 
@@ -122,26 +125,37 @@ module SVMKit
       # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to predict the labels.
       # @return [Numo::Int32] (shape: [n_samples]) Predicted class label per sample.
       def predict(x)
-        Numo::Int32.cast(decision_function(x).map { |v| v >= 0.0 ? 1 : -1 })
+        return Numo::Int32.cast(decision_function(x).ge(0.0)) * 2 - 1 if @classes.size <= 2
+
+        n_samples, = x.shape
+        decision_values = decision_function(x)
+        Numo::Int32.asarray(Array.new(n_samples) { |n| @classes[decision_values[n, true].max_index] })
       end
 
       # Predict probability for samples.
-      # Note that this method works normally only if the 'loss' parameter is set to 'logistic'.
       #
       # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to predict the probailities.
       # @return [Numo::DFloat] (shape: [n_samples, n_classes]) Predicted probability of each class per sample.
       def predict_proba(x)
+        proba = 1.0 / (Numo::NMath.exp(-decision_function(x)) + 1.0)
+        return (proba.transpose / proba.sum(axis: 1)).transpose if @classes.size > 2
+
         n_samples, = x.shape
-        proba = Numo::DFloat.zeros(n_samples, 2)
-        proba[true, 1] = 1.0 / (Numo::NMath.exp(-decision_function(x)) + 1.0)
-        proba[true, 0] = 1.0 - proba[true, 1]
-        proba
+        probs = Numo::DFloat.zeros(n_samples, 2)
+        probs[true, 1] = proba
+        probs[true, 0] = 1.0 - proba
+        probs
       end
 
       # Dump marshal data.
       # @return [Hash] The marshal data about FactorizationMachineClassifier
       def marshal_dump
-        { params: @params, factor_mat: @factor_mat, weight_vec: @weight_vec, bias_term: @bias_term, rng: @rng }
+        { params: @params,
+          factor_mat: @factor_mat,
+          weight_vec: @weight_vec,
+          bias_term: @bias_term,
+          classes: @classes,
+          rng: @rng }
       end
 
       # Load marshal data.
@@ -151,39 +165,76 @@ module SVMKit
         @factor_mat = obj[:factor_mat]
         @weight_vec = obj[:weight_vec]
         @bias_term = obj[:bias_term]
+        @classes = obj[:classes]
         @rng = obj[:rng]
         nil
       end
 
       private
 
-      def hinge_loss_gradient(x, y)
-        evaluated = y * decision_function(x)
+      def binary_fit(x, bin_y)
+        # Initialize some variables.
+        n_samples, n_features = x.shape
+        rand_ids = [*0...n_samples].shuffle(random: @rng)
+        factor_mat = rand_normal([@params[:n_factors], n_features], 0, @params[:init_std])
+        weight_vec = Numo::DFloat.zeros(n_features)
+        bias_term = 0.0
+        # Start optimization.
+        @params[:max_iter].times do |t|
+          # Random sampling.
+          subset_ids = rand_ids.shift(@params[:batch_size])
+          rand_ids.concat(subset_ids)
+          data = x[subset_ids, true]
+          label = bin_y[subset_ids]
+          # Calculate gradients for loss function.
+          loss_grad = loss_gradient(data, label, factor_mat, weight_vec, bias_term)
+          next if loss_grad.ne(0.0).count.zero?
+          # Update each parameter.
+          bias_term -= learning_rate(@params[:reg_param_bias], t) * bias_gradient(loss_grad, bias_term)
+          weight_vec -= learning_rate(@params[:reg_param_weight], t) * weight_gradient(loss_grad, data, weight_vec)
+          @params[:n_factors].times do |n|
+            factor_mat[n, true] -= learning_rate(@params[:reg_param_factor], t) *
+                                   factor_gradient(loss_grad, data, factor_mat[n, true])
+          end
+        end
+        [factor_mat, weight_vec, bias_term]
+      end
+
+      def bin_decision_function(x, factor, weight, bias)
+        bias + x.dot(weight) + 0.5 * (factor.dot(x.transpose)**2 - (factor**2).dot(x.transpose**2)).sum
+      end
+
+      def hinge_loss_gradient(x, y, factor, weight, bias)
+        evaluated = y * bin_decision_function(x, factor, weight, bias)
         gradient = Numo::DFloat.zeros(evaluated.size)
         gradient[evaluated < 1.0] = -y[evaluated < 1.0]
         gradient
       end
 
-      def logistic_loss_gradient(x, y)
-        evaluated = y * decision_function(x)
+      def logistic_loss_gradient(x, y, factor, weight, bias)
+        evaluated = y * bin_decision_function(x, factor, weight, bias)
         sigmoid_func = 1.0 / (Numo::NMath.exp(-evaluated) + 1.0)
         (sigmoid_func - 1.0) * y
       end
 
-      def loss_gradient(x, y)
-        @params[:loss] == 'hinge' ? hinge_loss_gradient(x, y) : logistic_loss_gradient(x, y)
+      def loss_gradient(x, y, factor, weight, bias)
+        if @params[:loss] == 'hinge'
+          hinge_loss_gradient(x, y, factor, weight, bias)
+        else
+          logistic_loss_gradient(x, y, factor, weight, bias)
+        end
       end
 
       def learning_rate(reg_param, iter)
         1.0 / (reg_param * (iter + 1))
       end
 
-      def bias_gradient(loss_grad)
-        loss_grad.mean + @params[:reg_param_bias] * @bias_term
+      def bias_gradient(loss_grad, bias)
+        loss_grad.mean + @params[:reg_param_bias] * bias
       end
 
-      def weight_gradient(loss_grad, data)
-        (loss_grad.expand_dims(1) * data).mean(0) + @params[:reg_param_weight] * @weight_vec
+      def weight_gradient(loss_grad, data, weight)
+        (loss_grad.expand_dims(1) * data).mean(0) + @params[:reg_param_weight] * weight
       end
 
       def factor_gradient(loss_grad, data, factor)

data/lib/svmkit/version.rb

@@ -3,5 +3,5 @@
 # SVMKit is a machine learning library in Ruby.
 module SVMKit
   # @!visibility private
-  VERSION = '0.2.6'
+  VERSION = '0.2.7'
 end

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: svmkit
 version: !ruby/object:Gem::Version
-  version: 0.2.6
+  version: 0.2.7
 platform: ruby
 authors:
 - yoshoku
 autorequire: 
 bindir: exe
 cert_chain: []
-date: 2018-03-11 00:00:00.000000000 Z
+date: 2018-04-01 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: numo-narray