summaryrefslogtreecommitdiff
path: root/lib/regression.rb
blob: 8a52e7db4b9f218d8831728a9f9dcd1b5ad19ea3 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
# TODO install R packages kernlab, caret, doMC, class, e1071


        # log transform activities (create new dataset)
        # scale, normalize features, might not be necessary
        # http://stats.stackexchange.com/questions/19216/variables-are-often-adjusted-e-g-standardised-before-making-a-model-when-is
        # http://stats.stackexchange.com/questions/7112/when-and-how-to-use-standardized-explanatory-variables-in-linear-regression
        # zero-order correlation and the semi-partial correlation
        # seems to be necessary for svm
        #   http://stats.stackexchange.com/questions/77876/why-would-scaling-features-decrease-svm-performance?lq=1
        #   http://stackoverflow.com/questions/15436367/svm-scaling-input-values
        # use lasso or elastic net??
        # select relevant features
        #   remove features with a single value
        #   remove correlated features
        #   remove features not correlated with endpoint
module OpenTox
  module Algorithm
    
    class Regression

      def self.weighted_average neighbors
        weighted_sum = 0.0
        sim_sum = 0.0
        neighbors.each do |row|
          n,sim,acts = row
          acts.each do |act|
            weighted_sum += sim*Math.log10(act)
            sim_sum += sim
          end
        end
        confidence = sim_sum/neighbors.size.to_f
        sim_sum == 0 ? prediction = nil : prediction = 10**(weighted_sum/sim_sum)
        {:value => prediction,:confidence => confidence}
      end

      # Local support vector regression from neighbors 
      # @param [Hash] params Keys `:props, :activities, :sims, :min_train_performance` are required
      # @return [Numeric] A prediction value.
      def self.local_svm_regression neighbors, params={:min_train_performance => 0.1}

        confidence = 0.0
        prediction = nil

        $logger.debug "Local SVM."
        props = neighbors.collect{|row| row[3] }
        neighbors.shift
        activities = neighbors.collect{|n| n[2]}
        prediction = self.local_svm_prop( props, activities, params[:min_train_performance]) # params[:props].nil? signals non-prop setting
        prediction = nil if (!prediction.nil? && prediction.infinite?)
        $logger.debug "Prediction: '#{prediction}' ('#{prediction.class}')."
        if prediction
          confidence = get_confidence({:sims => neighbors.collect{|n| n[1]}, :activities => activities})
        else
          confidence = nil if prediction.nil?
        end
          [prediction, confidence]

      end


      # Local support vector prediction from neighbors. 
      # Uses propositionalized setting.
      # Not to be called directly (use local_svm_regression or local_svm_classification).
      # @param [Array] props, propositionalization of neighbors and query structure e.g. [ Array_for_q, two-nested-Arrays_for_n ]
      # @param [Array] activities, activities for neighbors.
      # @param [Float] min_train_performance, parameter to control censoring
      # @return [Numeric] A prediction value.
      def self.local_svm_prop(props, activities, min_train_performance)

        $logger.debug "Local SVM (Propositionalization / Kernlab Kernel)."
        n_prop = props[1..-1] # is a matrix, i.e. two nested Arrays.
        q_prop = props[0] # is an Array.

        prediction = nil
        if activities.uniq.size == 1
          prediction = activities[0]
        else
          t = Time.now
          #$logger.debug gram_matrix.to_yaml
          #@r = RinRuby.new(true,false) # global R instance leads to Socket errors after a large number of requests
          @r = Rserve::Connection.new#(true,false) # global R instance leads to Socket errors after a large number of requests
          rs = []
          ["caret", "doMC", "class"].each do |lib|
            #raise "failed to load R-package #{lib}" unless @r.void_eval "suppressPackageStartupMessages(library('#{lib}'))"
            rs << "suppressPackageStartupMessages(library('#{lib}'))"
          end
          #@r.eval "registerDoMC()" # switch on parallel processing
          rs << "registerDoMC()" # switch on parallel processing
          #@r.eval "set.seed(1)"
          rs << "set.seed(1)"
          $logger.debug "Loading R packages: #{Time.now-t}"
          t = Time.now
          p n_prop
          begin

            # set data
            rs << "n_prop <- c(#{n_prop.flatten.join(',')})"
            rs << "n_prop <- c(#{n_prop.flatten.join(',')})"
            rs << "n_prop_x_size <- c(#{n_prop.size})"
            rs << "n_prop_y_size <- c(#{n_prop[0].size})"
            rs << "y <- c(#{activities.join(',')})"
            rs << "q_prop <- c(#{q_prop.join(',')})"
            rs << "y = matrix(y)"
            rs << "prop_matrix = matrix(n_prop, n_prop_x_size, n_prop_y_size, byrow=T)"
            rs << "q_prop = matrix(q_prop, 1, n_prop_y_size, byrow=T)"

            $logger.debug "Setting R data: #{Time.now-t}"
            t = Time.now
            # prepare data
            rs << "
              weights=NULL
              if (!(class(y) == 'numeric')) { 
                y = factor(y)
                weights=unlist(as.list(prop.table(table(y))))
                weights=(weights-1)^2
              }
            "

            rs << "
              rem = nearZeroVar(prop_matrix)
              if (length(rem) > 0) {
                prop_matrix = prop_matrix[,-rem,drop=F]
                q_prop = q_prop[,-rem,drop=F]
              }
              rem = findCorrelation(cor(prop_matrix))
              if (length(rem) > 0) {
                prop_matrix = prop_matrix[,-rem,drop=F]
                q_prop = q_prop[,-rem,drop=F]
              }
            "

            #p @r.eval("y").to_ruby
            #p "weights"
            #p @r.eval("weights").to_ruby
            $logger.debug "Preparing R data: #{Time.now-t}"
            t = Time.now
            # model + support vectors
            #train_success = @r.eval <<-EOR
            rs << '
              model = train(prop_matrix,y,
                             method="svmRadial",
                             preProcess=c("center", "scale"),
                             class.weights=weights,
                             trControl=trainControl(method="LGOCV",number=10),
                             tuneLength=8
                           )
              perf = ifelse ( class(y)!="numeric", max(model$results$Accuracy), model$results[which.min(model$results$RMSE),]$Rsquared )
            '
            File.open("/tmp/r.r","w+"){|f| f.puts rs.join("\n")}
            p rs.join("\n")
            p `Rscript /tmp/r.r`
=begin
            @r.void_eval <<-EOR
              model = train(prop_matrix,y,
                             method="svmRadial",
                             #preProcess=c("center", "scale"),
                             #class.weights=weights,
                             #trControl=trainControl(method="LGOCV",number=10),
                             #tuneLength=8
                           )
              perf = ifelse ( class(y)!='numeric', max(model$results$Accuracy), model$results[which.min(model$results$RMSE),]$Rsquared )
            EOR
=end

            $logger.debug "Creating R SVM model: #{Time.now-t}"
            t = Time.now
            if train_success
              # prediction
              @r.eval "predict(model,q_prop); p = predict(model,q_prop)" # kernlab bug: predict twice
              #@r.eval "p = predict(model,q_prop)" # kernlab bug: predict twice
              @r.eval "if (class(y)!='numeric') p = as.character(p)"
              prediction = @r.p

              # censoring
              prediction = nil if ( @r.perf.nan? || @r.perf < min_train_performance.to_f )
              prediction = nil if prediction =~ /NA/
              $logger.debug "Performance: '#{sprintf("%.2f", @r.perf)}'"
            else
              $logger.debug "Model creation failed."
              prediction = nil 
            end
            $logger.debug "R Prediction: #{Time.now-t}"
          rescue Exception => e
            $logger.debug "#{e.class}: #{e.message}"
            $logger.debug "Backtrace:\n\t#{e.backtrace.join("\n\t")}"
          ensure
            #puts @r.inspect
            #TODO: broken pipe
            #@r.quit # free R
          end
        end
        prediction
      end
    end

  end
end