rumale 0.22.2 → 0.23.1

data/lib/rumale/kernel_machine/kernel_ridge_classifier.rb

@@ -0,0 +1,92 @@
+# frozen_string_literal: true
+
+require 'rumale/base/base_estimator'
+require 'rumale/base/classifier'
+require 'rumale/preprocessing/label_binarizer'
+
+module Rumale
+  module KernelMachine
+    # KernelRidgeClassifier is a class that implements classifier based-on kernel ridge regression.
+    # It learns a classifier by converting labels to target values { -1, 1 } and performing kernel ridge regression.
+    #
+    # @example
+    #   require 'numo/linalg/autoloader'
+    #   require 'rumale'
+    #
+    #   kernel_mat_train = Rumale::PairwiseMetric::rbf_kernel(training_samples)
+    #   kridge = Rumale::KernelMachine::KernelRidgeClassifier.new(reg_param: 0.5)
+    #   kridge.fit(kernel_mat_train, traininig_values)
+    #
+    #   kernel_mat_test = Rumale::PairwiseMetric::rbf_kernel(test_samples, training_samples)
+    #   results = kridge.predict(kernel_mat_test)
+    class KernelRidgeClassifier
+      include Base::BaseEstimator
+      include Base::Classifier
+
+      # Return the class labels.
+      # @return [Numo::Int32] (size: n_classes)
+      attr_reader :classes
+
+      # Return the weight vector.
+      # @return [Numo::DFloat] (shape: [n_training_sample, n_classes])
+      attr_reader :weight_vec
+
+      # Create a new regressor with kernel ridge classifier.
+      #
+      # @param reg_param [Float/Numo::DFloat] The regularization parameter.
+      def initialize(reg_param: 1.0)
+        @params = {}
+        @params[:reg_param] = reg_param
+        @classes = nil
+        @weight_vec = nil
+      end
+
+      # Fit the model with given training data.
+      #
+      # @param x [Numo::DFloat] (shape: [n_training_samples, n_training_samples])
+      #   The kernel matrix of the training data to be used for fitting the model.
+      # @param y [Numo::Int32] (shape: [n_training_samples]) The labels to be used for fitting the model.
+      # @return [KernelRidgeClassifier] The learned classifier itself.
+      def fit(x, y)
+        x = check_convert_sample_array(x)
+        y = check_convert_label_array(y)
+        check_sample_label_size(x, y)
+        raise ArgumentError, 'Expect the kernel matrix of training data to be square.' unless x.shape[0] == x.shape[1]
+        raise 'KernelRidgeClassifier#fit requires Numo::Linalg but that is not loaded.' unless enable_linalg?
+
+        @encoder = Rumale::Preprocessing::LabelBinarizer.new
+        y_encoded = Numo::DFloat.cast(@encoder.fit_transform(y)) * 2 - 1
+        @classes = Numo::NArray[*@encoder.classes]
+
+        n_samples = x.shape[0]
+        reg_kernel_mat = x + Numo::DFloat.eye(n_samples) * @params[:reg_param]
+        @weight_vec = Numo::Linalg.solve(reg_kernel_mat, y_encoded, driver: 'sym')
+
+        self
+      end
+
+      # Calculate confidence scores for samples.
+      #
+      # @param x [Numo::DFloat] (shape: [n_testing_samples, n_training_samples])
+      #     The kernel matrix between testing samples and training samples to predict values.
+      # @return [Numo::DFloat] (shape: [n_samples, n_classes]) The confidence score per sample.
+      def decision_function(x)
+        x = check_convert_sample_array(x)
+        x.dot(@weight_vec)
+      end
+
+      # Predict class labels for samples.
+      #
+      # @param x [Numo::DFloat] (shape: [n_testing_samples, n_training_samples])
+      #     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)
+        x = check_convert_sample_array(x)
+        scores = decision_function(x)
+        n_samples, n_classes = scores.shape
+        label_ids = scores.max_index(axis: 1) - Numo::Int32.new(n_samples).seq * n_classes
+        @classes[label_ids].dup
+      end
+    end
+  end
+end

data/lib/rumale/kernel_machine/kernel_svc.rb

@@ -152,7 +152,7 @@ module Rumale
 
         if @classes.size > 2
           probs = 1.0 / (Numo::NMath.exp(@prob_param[true, 0] * decision_function(x) + @prob_param[true, 1]) + 1.0)
-          return (probs.transpose / probs.sum(axis: 1)).transpose
+          return (probs.transpose / probs.sum(axis: 1)).transpose.dup
         end
 
         n_samples, = x.shape

data/lib/rumale/linear_model/elastic_net.rb

@@ -81,7 +81,7 @@ module Rumale
       # 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, n_outputs]) The target values to be used for fitting the model.
+      # @param y [Numo::DFloat] (shape: [n_samples, n_outputs]) The target values to be used for fitting the model.
       # @return [ElasticNet] The learned regressor itself.
       def fit(x, y)
         x = check_convert_sample_array(x)

data/lib/rumale/linear_model/lasso.rb

@@ -77,7 +77,7 @@ module Rumale
       # 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, n_outputs]) The target values to be used for fitting the model.
+      # @param y [Numo::DFloat] (shape: [n_samples, n_outputs]) The target values to be used for fitting the model.
       # @return [Lasso] The learned regressor itself.
       def fit(x, y)
         x = check_convert_sample_array(x)

data/lib/rumale/linear_model/linear_regression.rb

@@ -1,12 +1,15 @@
 # frozen_string_literal: true
 
+require 'lbfgsb'
+
 require 'rumale/linear_model/base_sgd'
 require 'rumale/base/regressor'
 
 module Rumale
   module LinearModel
     # LinearRegression is a class that implements ordinary least square linear regression
-    # with stochastic gradient descent (SGD) optimization or singular value decomposition (SVD).
+    # with stochastic gradient descent (SGD) optimization,
+    # singular value decomposition (SVD), or L-BFGS optimization.
     #
     # @example
     #   estimator =
@@ -41,31 +44,32 @@ module Rumale
       #
       # @param learning_rate [Float] The initial value of learning rate.
       #   The learning rate decreases as the iteration proceeds according to the equation: learning_rate / (1 + decay * t).
-      #   If solver = 'svd', this parameter is ignored.
+      #   If solver is not 'sgd', this parameter is ignored.
       # @param decay [Float] The smoothing parameter for decreasing learning rate as the iteration proceeds.
       #   If nil is given, the decay sets to 'learning_rate'.
-      #   If solver = 'svd', this parameter is ignored.
+      #   If solver is not 'sgd', this parameter is ignored.
       # @param momentum [Float] The momentum factor.
-      #   If solver = 'svd', this parameter is ignored.
+      #   If solver is not 'sgd', this parameter is ignored.
       # @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 epochs that indicates
       #   how many times the whole data is given to the training process.
-      #   If solver = 'svd', this parameter is ignored.
+      #   If solver is 'svd', this parameter is ignored.
       # @param batch_size [Integer] The size of the mini batches.
-      #   If solver = 'svd', this parameter is ignored.
+      #   If solver is not 'sgd', this parameter is ignored.
       # @param tol [Float] The tolerance of loss for terminating optimization.
-      #   If solver = 'svd', this parameter is ignored.
-      # @param solver [String] The algorithm to calculate weights. ('auto', 'sgd' or 'svd').
-      #   'auto' chooses the 'svd' solver if Numo::Linalg is loaded. Otherwise, it chooses the 'sgd' solver.
+      #   If solver is 'svd', this parameter is ignored.
+      # @param solver [String] The algorithm to calculate weights. ('auto', 'sgd', 'svd' or 'lbfgs').
+      #   'auto' chooses the 'svd' solver if Numo::Linalg is loaded. Otherwise, it chooses the 'lbfgs' solver.
       #   'sgd' uses the stochastic gradient descent optimization.
       #   'svd' performs singular value decomposition of samples.
+      #   'lbfgs' uses the L-BFGS method for optimization.
       # @param n_jobs [Integer] The number of jobs for running the fit method in parallel.
       #   If nil is given, the method does not execute in parallel.
       #   If zero or less is given, it becomes equal to the number of processors.
-      #   This parameter is ignored if the Parallel gem is not loaded.
+      #   This parameter is ignored if the Parallel gem is not loaded or solver is not 'sgd'.
       # @param verbose [Boolean] The flag indicating whether to output loss during iteration.
-      #   If solver = 'svd', this parameter is ignored.
+      #   If solver is 'svd', this parameter is ignored.
       # @param random_seed [Integer] The seed value using to initialize the random generator.
       def initialize(learning_rate: 0.01, decay: nil, momentum: 0.9,
                      fit_bias: true, bias_scale: 1.0, max_iter: 1000, batch_size: 50, tol: 1e-4,
@@ -80,9 +84,9 @@ module Rumale
         super()
         @params.merge!(method(:initialize).parameters.map { |_t, arg| [arg, binding.local_variable_get(arg)] }.to_h)
         @params[:solver] = if solver == 'auto'
-                             load_linalg? ? 'svd' : 'sgd'
+                             enable_linalg?(warning: false) ? 'svd' : 'lbfgs'
                            else
-                             solver != 'svd' ? 'sgd' : 'svd' # rubocop:disable Style/NegatedIfElseCondition
+                             solver.match?(/^svd$|^sgd$|^lbfgs$/) ? solver : 'lbfgs'
                            end
         @params[:decay] ||= @params[:learning_rate]
         @params[:random_seed] ||= srand
@@ -95,15 +99,17 @@ module Rumale
       # 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, n_outputs]) The target values to be used for fitting the model.
+      # @param y [Numo::DFloat] (shape: [n_samples, n_outputs]) The target values to be used for fitting the model.
       # @return [LinearRegression] The learned regressor itself.
       def fit(x, y)
         x = check_convert_sample_array(x)
         y = check_convert_tvalue_array(y)
         check_sample_tvalue_size(x, y)
 
-        if @params[:solver] == 'svd' && enable_linalg?
+        if @params[:solver] == 'svd' && enable_linalg?(warning: false)
           fit_svd(x, y)
+        elsif @params[:solver] == 'lbfgs'
+          fit_lbfgs(x, y)
         else
           fit_sgd(x, y)
         end
@@ -124,24 +130,46 @@ module Rumale
 
       def fit_svd(x, y)
         x = expand_feature(x) if fit_bias?
-
         w = Numo::Linalg.pinv(x, driver: 'svd').dot(y)
+        @weight_vec, @bias_term = single_target?(y) ? split_weight(w) : split_weight_mult(w)
+      end
 
-        is_single_target_vals = y.shape[1].nil?
-        if @params[:fit_bias]
-          @weight_vec = is_single_target_vals ? w[0...-1].dup : w[0...-1, true].dup
-          @bias_term = is_single_target_vals ? w[-1] : w[-1, true].dup
-        else
-          @weight_vec = w.dup
-          @bias_term = is_single_target_vals ? 0 : Numo::DFloat.zeros(y.shape[1])
+      def fit_lbfgs(x, y)
+        fnc = proc do |w, x, y| # rubocop:disable Lint/ShadowingOuterLocalVariable
+          n_samples, n_features = x.shape
+          w = w.reshape(y.shape[1], n_features) unless y.shape[1].nil?
+          z = x.dot(w.transpose)
+          d = z - y
+          loss = (d**2).sum.fdiv(n_samples)
+          gradient = 2.fdiv(n_samples) * d.transpose.dot(x)
+          [loss, gradient.flatten.dup]
         end
-      end
 
-      def fit_sgd(x, y)
-        n_outputs = y.shape[1].nil? ? 1 : y.shape[1]
+        x = expand_feature(x) if fit_bias?
+
         n_features = x.shape[1]
+        n_outputs = single_target?(y) ? 1 : y.shape[1]
+
+        res = Lbfgsb.minimize(
+          fnc: fnc, jcb: true, x_init: init_weight(n_features, n_outputs), args: [x, y],
+          maxiter: @params[:max_iter], factr: @params[:tol] / Lbfgsb::DBL_EPSILON,
+          verbose: @params[:verbose] ? 1 : -1
+        )
 
-        if n_outputs > 1
+        @weight_vec, @bias_term =
+          if single_target?(y)
+            split_weight(res[:x])
+          else
+            split_weight_mult(res[:x].reshape(n_outputs, n_features).transpose)
+          end
+      end
+
+      def fit_sgd(x, y)
+        if single_target?(y)
+          @weight_vec, @bias_term = partial_fit(x, y)
+        else
+          n_outputs = y.shape[1]
+          n_features = x.shape[1]
           @weight_vec = Numo::DFloat.zeros(n_outputs, n_features)
           @bias_term = Numo::DFloat.zeros(n_outputs)
           if enable_parallel?
@@ -150,20 +178,23 @@ module Rumale
           else
             n_outputs.times { |n| @weight_vec[n, true], @bias_term[n] = partial_fit(x, y[true, n]) }
           end
-        else
-          @weight_vec, @bias_term = partial_fit(x, y)
         end
       end
 
-      def fit_bias?
-        @params[:fit_bias] == true
+      def single_target?(y)
+        y.ndim == 1
       end
 
-      def load_linalg?
-        return false if defined?(Numo::Linalg).nil?
-        return false if Numo::Linalg::VERSION < '0.1.4'
+      def init_weight(n_features, n_outputs)
+        Rumale::Utils.rand_normal([n_outputs, n_features], @rng.dup).flatten.dup
+      end
 
-        true
+      def split_weight_mult(w)
+        if fit_bias?
+          [w[0...-1, true].dup, w[-1, true].dup]
+        else
+          [w.dup, Numo::DFloat.zeros(w.shape[1])]
+        end
       end
     end
   end

data/lib/rumale/linear_model/nnls.rb

@@ -0,0 +1,137 @@
+# frozen_string_literal: true
+
+require 'lbfgsb'
+
+require 'rumale/base/base_estimator'
+require 'rumale/base/regressor'
+
+module Rumale
+  module LinearModel
+    # NNLS is a class that implements non-negative least squares regression.
+    # NNLS solves least squares problem under non-negative constraints on the coefficient using L-BFGS-B method.
+    #
+    # @example
+    #   estimator = Rumale::LinearModel::NNLS.new(reg_param: 0.01, random_seed: 1)
+    #   estimator.fit(training_samples, traininig_values)
+    #   results = estimator.predict(testing_samples)
+    #
+    class NNLS
+      include Base::BaseEstimator
+      include Base::Regressor
+
+      # Return the weight vector.
+      # @return [Numo::DFloat] (shape: [n_outputs, n_features])
+      attr_reader :weight_vec
+
+      # Return the bias term (a.k.a. intercept).
+      # @return [Numo::DFloat] (shape: [n_outputs])
+      attr_reader :bias_term
+
+      # Returns the number of iterations when converged.
+      # @return [Integer]
+      attr_reader :n_iter
+
+      # Return the random generator for initializing weight.
+      # @return [Random]
+      attr_reader :rng
+
+      # Create a new regressor with non-negative least squares method.
+      #
+      # @param reg_param [Float] The regularization parameter for L2 regularization term.
+      # @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 epochs that indicates
+      #   how many times the whole data is given to the training process.
+      # @param tol [Float] The tolerance of loss for terminating optimization.
+      #   If solver = 'svd', this parameter is ignored.
+      # @param verbose [Boolean] The flag indicating whether to output loss during iteration.
+      # @param random_seed [Integer] The seed value using to initialize the random generator.
+      def initialize(reg_param: 1.0, fit_bias: true, bias_scale: 1.0,
+                     max_iter: 1000, tol: 1e-4, verbose: false, random_seed: nil)
+        check_params_numeric(reg_param: reg_param, bias_scale: bias_scale, max_iter: max_iter, tol: tol)
+        check_params_boolean(fit_bias: fit_bias, verbose: verbose)
+        check_params_numeric_or_nil(random_seed: random_seed)
+        check_params_positive(reg_param: reg_param, max_iter: max_iter)
+        @params = method(:initialize).parameters.each_with_object({}) { |(_, prm), obj| obj[prm] = binding.local_variable_get(prm) }
+        @params[:random_seed] ||= srand
+        @n_iter = nil
+        @weight_vec = nil
+        @bias_term = 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::DFloat] (shape: [n_samples, n_outputs]) The target values to be used for fitting the model.
+      # @return [NonneagtiveLeastSquare] The learned regressor itself.
+      def fit(x, y)
+        x = check_convert_sample_array(x)
+        y = check_convert_tvalue_array(y)
+        check_sample_tvalue_size(x, y)
+
+        x = expand_feature(x) if fit_bias?
+
+        n_features = x.shape[1]
+        n_outputs = single_target?(y) ? 1 : y.shape[1]
+
+        w_init = Rumale::Utils.rand_normal([n_outputs, n_features], @rng.dup).flatten.dup
+        w_init[w_init.lt(0)] = 0
+        bounds = Numo::DFloat.zeros(n_outputs * n_features, 2)
+        bounds.shape[0].times { |n| bounds[n, 1] = Float::INFINITY }
+
+        res = Lbfgsb.minimize(
+          fnc: method(:nnls_fnc), jcb: true, x_init: w_init, args: [x, y, @params[:reg_param]], bounds: bounds,
+          maxiter: @params[:max_iter], factr: @params[:tol] / Lbfgsb::DBL_EPSILON, verbose: @params[:verbose] ? 1 : -1
+        )
+
+        @n_iter = res[:n_iter]
+        w = single_target?(y) ? res[:x] : res[:x].reshape(n_outputs, n_features).transpose
+
+        if fit_bias?
+          @weight_vec = single_target?(y) ? w[0...-1].dup : w[0...-1, true].dup
+          @bias_term = single_target?(y) ? w[-1] : w[-1, true].dup
+        else
+          @weight_vec = w.dup
+          @bias_term = single_target?(y) ? 0 : Numo::DFloat.zeros(y.shape[1])
+        end
+
+        self
+      end
+
+      # Predict values for samples.
+      #
+      # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to predict the values.
+      # @return [Numo::DFloat] (shape: [n_samples, n_outputs]) Predicted values per sample.
+      def predict(x)
+        x = check_convert_sample_array(x)
+        x.dot(@weight_vec.transpose) + @bias_term
+      end
+
+      private
+
+      def nnls_fnc(w, x, y, alpha)
+        n_samples, n_features = x.shape
+        w = w.reshape(y.shape[1], n_features) unless y.shape[1].nil?
+        z = x.dot(w.transpose)
+        d = z - y
+        loss = (d**2).sum.fdiv(n_samples) + alpha * (w * w).sum
+        gradient = 2.fdiv(n_samples) * d.transpose.dot(x) + 2.0 * alpha * w
+        [loss, gradient.flatten.dup]
+      end
+
+      def expand_feature(x)
+        n_samples = x.shape[0]
+        Numo::NArray.hstack([x, Numo::DFloat.ones([n_samples, 1]) * @params[:bias_scale]])
+      end
+
+      def fit_bias?
+        @params[:fit_bias] == true
+      end
+
+      def single_target?(y)
+        y.ndim == 1
+      end
+    end
+  end
+end

data/lib/rumale/linear_model/ridge.rb

@@ -1,12 +1,15 @@
 # frozen_string_literal: true
 
+require 'lbfgsb'
+
 require 'rumale/linear_model/base_sgd'
 require 'rumale/base/regressor'
 
 module Rumale
   module LinearModel
     # Ridge is a class that implements Ridge Regression
-    # with stochastic gradient descent (SGD) optimization or singular value decomposition (SVD).
+    # with stochastic gradient descent (SGD) optimization,
+    # singular value decomposition (SVD), or L-BFGS optimization.
     #
     # @example
     #   estimator =
@@ -41,32 +44,33 @@ module Rumale
       #
       # @param learning_rate [Float] The initial value of learning rate.
       #   The learning rate decreases as the iteration proceeds according to the equation: learning_rate / (1 + decay * t).
-      #   If solver = 'svd', this parameter is ignored.
+      #   If solver is not 'sgd', this parameter is ignored.
       # @param decay [Float] The smoothing parameter for decreasing learning rate as the iteration proceeds.
       #   If nil is given, the decay sets to 'reg_param * learning_rate'.
-      #   If solver = 'svd', this parameter is ignored.
+      #   If solver is not 'sgd', this parameter is ignored.
       # @param momentum [Float] The momentum factor.
-      #   If solver = 'svd', this parameter is ignored.
+      #   If solver is not 'sgd', this parameter is ignored.
       # @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 epochs that indicates
       #   how many times the whole data is given to the training process.
-      #   If solver = 'svd', this parameter is ignored.
+      #   If solver is 'svd', this parameter is ignored.
       # @param batch_size [Integer] The size of the mini batches.
-      #   If solver = 'svd', this parameter is ignored.
+      #   If solver is not 'sgd', this parameter is ignored.
       # @param tol [Float] The tolerance of loss for terminating optimization.
-      #   If solver = 'svd', this parameter is ignored.
-      # @param solver [String] The algorithm to calculate weights. ('auto', 'sgd' or 'svd').
-      #   'auto' chooses the 'svd' solver if Numo::Linalg is loaded. Otherwise, it chooses the 'sgd' solver.
+      #   If solver is 'svd', this parameter is ignored.
+      # @param solver [String] The algorithm to calculate weights. ('auto', 'sgd', 'svd', or 'lbfgs').
+      #   'auto' chooses the 'svd' solver if Numo::Linalg is loaded. Otherwise, it chooses the 'lbfgs' solver.
       #   'sgd' uses the stochastic gradient descent optimization.
       #   'svd' performs singular value decomposition of samples.
+      #   'lbfgs' uses the L-BFGS method for optimization.
       # @param n_jobs [Integer] The number of jobs for running the fit method in parallel.
       #   If nil is given, the method does not execute in parallel.
       #   If zero or less is given, it becomes equal to the number of processors.
-      #   This parameter is ignored if the Parallel gem is not loaded or the solver is 'svd'.
+      #   This parameter is ignored if the Parallel gem is not loaded or solver is not 'sgd'.
       # @param verbose [Boolean] The flag indicating whether to output loss during iteration.
-      #   If solver = 'svd', this parameter is ignored.
+      #   If solver is 'svd', this parameter is ignored.
       # @param random_seed [Integer] The seed value using to initialize the random generator.
       def initialize(learning_rate: 0.01, decay: nil, momentum: 0.9,
                      reg_param: 1.0, fit_bias: true, bias_scale: 1.0,
@@ -83,9 +87,9 @@ module Rumale
         super()
         @params.merge!(method(:initialize).parameters.map { |_t, arg| [arg, binding.local_variable_get(arg)] }.to_h)
         @params[:solver] = if solver == 'auto'
-                             load_linalg? ? 'svd' : 'sgd'
+                             enable_linalg?(warning: false) ? 'svd' : 'lbfgs'
                            else
-                             solver != 'svd' ? 'sgd' : 'svd' # rubocop:disable Style/NegatedIfElseCondition
+                             solver.match?(/^svd$|^sgd$|^lbfgs$/) ? solver : 'lbfgs'
                            end
         @params[:decay] ||= @params[:reg_param] * @params[:learning_rate]
         @params[:random_seed] ||= srand
@@ -99,15 +103,17 @@ module Rumale
       # 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, n_outputs]) The target values to be used for fitting the model.
+      # @param y [Numo::DFloat] (shape: [n_samples, n_outputs]) The target values to be used for fitting the model.
       # @return [Ridge] The learned regressor itself.
       def fit(x, y)
         x = check_convert_sample_array(x)
         y = check_convert_tvalue_array(y)
         check_sample_tvalue_size(x, y)
 
-        if @params[:solver] == 'svd' && enable_linalg?
+        if @params[:solver] == 'svd' && enable_linalg?(warning: false)
           fit_svd(x, y)
+        elsif @params[:solver] == 'lbfgs'
+          fit_lbfgs(x, y)
         else
           fit_sgd(x, y)
         end
@@ -127,27 +133,51 @@ module Rumale
       private
 
       def fit_svd(x, y)
-        samples = @params[:fit_bias] ? expand_feature(x) : x
+        x = expand_feature(x) if fit_bias?
 
-        s, u, vt = Numo::Linalg.svd(samples, driver: 'sdd', job: 'S')
+        s, u, vt = Numo::Linalg.svd(x, driver: 'sdd', job: 'S')
         d = (s / (s**2 + @params[:reg_param])).diag
         w = vt.transpose.dot(d).dot(u.transpose).dot(y)
 
-        is_single_target_vals = y.shape[1].nil?
-        if @params[:fit_bias]
-          @weight_vec = is_single_target_vals ? w[0...-1].dup : w[0...-1, true].dup
-          @bias_term = is_single_target_vals ? w[-1] : w[-1, true].dup
-        else
-          @weight_vec = w.dup
-          @bias_term = is_single_target_vals ? 0 : Numo::DFloat.zeros(y.shape[1])
-        end
+        @weight_vec, @bias_term = single_target?(y) ? split_weight(w) : split_weight_mult(w)
       end
 
-      def fit_sgd(x, y)
-        n_outputs = y.shape[1].nil? ? 1 : y.shape[1]
+      def fit_lbfgs(x, y)
+        fnc = proc do |w, x, y, a| # rubocop:disable Lint/ShadowingOuterLocalVariable
+          n_samples, n_features = x.shape
+          w = w.reshape(y.shape[1], n_features) unless y.shape[1].nil?
+          z = x.dot(w.transpose)
+          d = z - y
+          loss = (d**2).sum.fdiv(n_samples) + a * (w * w).sum
+          gradient = 2.fdiv(n_samples) * d.transpose.dot(x) + 2.0 * a * w
+          [loss, gradient.flatten.dup]
+        end
+
+        x = expand_feature(x) if fit_bias?
+
         n_features = x.shape[1]
+        n_outputs = single_target?(y) ? 1 : y.shape[1]
+
+        res = Lbfgsb.minimize(
+          fnc: fnc, jcb: true, x_init: init_weight(n_features, n_outputs), args: [x, y, @params[:reg_param]],
+          maxiter: @params[:max_iter], factr: @params[:tol] / Lbfgsb::DBL_EPSILON,
+          verbose: @params[:verbose] ? 1 : -1
+        )
+
+        @weight_vec, @bias_term =
+          if single_target?(y)
+            split_weight(res[:x])
+          else
+            split_weight_mult(res[:x].reshape(n_outputs, n_features).transpose)
+          end
+      end
 
-        if n_outputs > 1
+      def fit_sgd(x, y)
+        if single_target?(y)
+          @weight_vec, @bias_term = partial_fit(x, y)
+        else
+          n_outputs = y.shape[1]
+          n_features = x.shape[1]
           @weight_vec = Numo::DFloat.zeros(n_outputs, n_features)
           @bias_term = Numo::DFloat.zeros(n_outputs)
           if enable_parallel?
@@ -156,16 +186,23 @@ module Rumale
           else
             n_outputs.times { |n| @weight_vec[n, true], @bias_term[n] = partial_fit(x, y[true, n]) }
           end
-        else
-          @weight_vec, @bias_term = partial_fit(x, y)
         end
       end
 
-      def load_linalg?
-        return false if defined?(Numo::Linalg).nil?
-        return false if Numo::Linalg::VERSION < '0.1.4'
+      def single_target?(y)
+        y.ndim == 1
+      end
+
+      def init_weight(n_features, n_outputs)
+        Rumale::Utils.rand_normal([n_outputs, n_features], @rng.dup).flatten.dup
+      end
 
-        true
+      def split_weight_mult(w)
+        if fit_bias?
+          [w[0...-1, true].dup, w[-1, true].dup]
+        else
+          [w.dup, Numo::DFloat.zeros(w.shape[1])]
+        end
       end
     end
   end