N blog

# 分類器生成
n = NaiveBayes.new
# 学習
n.train('Nobody owns the water.','good')
n.train('the quick rabbit jumps fences','good')
n.train('buy pharmaceuticals now','bad')
n.train('make quick money at the online casino','bad')
n.train('the quick brown fox jumps','good')
# 推定
puts n.classify("quick rabbit",default="unknown")
puts n.classify("quick money",default="unknown")
# 閾値を与える
n.setthreshold("bad",3.0)
# 推定
puts n.classify("quick money",default="unknown")

# 特徴抽出クラス
class GetWordFeature
  def get_feature(doc)
    words = doc.split(' ')
    words.map{|w| w.downcase}.select{|w| w.length < 20 && w.length > 2 }.uniq
  end
end
 
# 分類器の基底クラス
class Classifier
  def initialize(f)
    @fc,@cc,@feature = {},{},f
  end
 
  def incf(f,cat)
    @fc[f] ||= {}
    @fc[f][cat] ||= 0.0
    @fc[f][cat] = @fc[f][cat] + 1.0
  end
 
  def incc(cat)
    @cc[cat] ||= 0.0
    @cc[cat] = @cc[cat] + 1.0
  end
 
  def fcount(f,cat)
    @fc[f] ||= {}
    @fc[f][cat] ||= 0.0
    @fc[f][cat]
  end
 
  def catcount(cat)
    @cc[cat] ? @cc[cat].to_f : 0.0
  end
 
  def totalcount
    sum = 0
    @cc.each_value {|value| sum += value}
    return sum
  end
 
  def categories
    @cc.keys
  end
 
  # 条件付き確率 Pr(A|B)
  def fprob(f,cat)
    return 0 if catcount(cat) == 0 || fcount(f,cat) == nil
    fcount(f,cat) / catcount(cat)
  end
 
  # 重み付き確率
  def weightedprob(f, cat, weight=1.0, ap=0.5)
    # 通常の条件付き確率
    basicprob = fprob(f,cat)
    # この特徴がすべてのカテゴリ中に存在する確率
    totals = 0
    categories.each do |c|
      totals = totals + fcount(f,c) if fcount(f,c)
    end
    # 重み付けした平均計算
    ((weight*ap)+(totals*basicprob)) / (weight+totals)
  end
 
  def train(item, cat)
    features =  @feature.get_feature(item);
    features.each do |f|
      incf(f, cat)
    end
    incc(cat)
  end
end
 
#　単純ベイズ分類器
class NaiveBayes < Classifier
  def initialize
    super(GetWordFeature.new)
    @thresholds = {}
  end
 
  # P(B|A):アイテム全体の与えられたカテゴリでの確率を求める（掛け合わせる）
  def docprob(item,cat)
    features = @feature.get_feature(item)
    p = 1.0
    features.each do |f|
      ep = weightedprob(f,cat)
      p = p * ep
    end
    return p
  end
 
  def prob(item,cat)
    catprob = catcount(cat) / totalcount
    docprob = docprob(item,cat)
    return docprob * catprob
  end
 
  def setthreshold(cat,t)
    @thresholds[cat] = t
  end
 
  def getthreshold(cat)
    @thresholds[cat] || 1.0
  end
 
  def classify(item, default=nil)
    probs,max,best = {},0.0,nil
    categories.each do |c|
      probs[c] = prob(item,c)
      if probs[c] > max
        max = probs[c]
        best = c
      end
    end
    probs.each_key do |c|
      # (2番目に確率の高いもの * 閾値)と比較
      second = probs[c].to_f * getthreshold(best)
      return default if second > probs[best]
    end
    return best
  end
end