朴素贝叶斯

代码实现：

  1. # naive Bayes algorithm + add-k smoothing
  2. import math, collections
  4. ADDK = 1
  7. # D: documents
  8. # C: classes
  9. # return log P(c) and log P(w|c)
  10. def train_naive_bayes(D, C, inlog=False):
  11.     V = vocabulary(D) #vocabulary of D
  12.     print("|V| =", len(V))
  13.     prior = collections.defaultdict()
  14.     likelihood = collections.defaultdict(dict)
  15.     Ndoc = len(D) # number of documents in D
  16.     for c in C:
  17.         # calculate P(c)
  18.         Dc = [d for d in D if d.inclass(c)]
  19.         Nc = len(Dc)# number of documents from D in class c
  20.         if inlog:
  21.             prior[c] = math.log(Nc/Ndoc) # log P(c)
  22.         else:
  23.             prior[c] = Nc/Ndoc
  24.         # bigdoc contains all the documents from D in class c
  25.         bigdoc = " ".join(d.text for d in Dc)
  27.         # calculate P(w|c)
  28.         for w in V: # w == word
  29.             # occurrences() calculate the occurrences of w in bigdoc[c]
  30.             count_wc = occurrences(bigdoc, w)
  31.             if inlog:
  32.                 likelihood[w][c] = math.log((count_wc + ADDK)/(sum(occurrences(bigdoc, wd) + ADDK for wd in V))) # log P(w|c)
  33.             else:
  34.                 likelihood[w][c] = (count_wc + ADDK)/(sum(occurrences(bigdoc, wd) + ADDK for wd in V)) # P(w|c)
  36.     return prior, likelihood, V
  38. # return best c
  39. def test_naive_bayes(testdoc, logprior, loglikelihood, C, V, inlog=False):
  40.     sum_ = {}
  41.     for c in C:
  42.         sum_[c] = logprior[c]
  43.         words = [word.strip() for word in testdoc.split(" ")]
  44.         for word in words:
  45.             if word in V: # remove unknown
  46.                 if inlog:
  47.                     sum_[c] += loglikelihood[word][c]
  48.                 else:
  49.                     sum_[c] *= loglikelihood[word][c]
  50.     #print(sum_)
  51.     return argmaxc(sum_)
  53. ## utils
  54. def vocabulary(D, sep=" "):
  55.     return {ele.strip() for d in D for ele in d.text.split(" ") if ele.strip()}
  57. def occurrences(doc, w):
  58.     return [ele.strip() for ele in doc.split(" ")].count(w)
  60. def argmaxc(ps):
  61.     return max(ps, default=None, key=lambda c: ps[c])
  63. ##### 测试
  64. class Document():
  65.     def __init__(self, text, c):
  66.         self.text = text
  67.         self.c = c
  68.     def inclass(self, c):
  69.         return self.c == c
  71. ## 例子来源于 《speech and language processing》4.3
  72. def testcase1():
  73.     Ds = [
  74.         Document("just plain boring", "-"),
  75.         Document("entirely predictable and lacks energy", "-"),
  76.         Document("no surprises and very few laughs", "-"),
  77.         Document("very powerful", "+"),
  78.         Document("the most fun film of the summer", "+"),
  79.     ]
  80.     Cs = ["+", "-"]
  81.     prior, likelihood, V = train_naive_bayes(Ds, Cs)
  82.     # print
  83.     for c in Cs:
  84.         print("P('%s') = %f" % (c, prior[c]))
  85.         for w in likelihood:
  86.             print("P('%s'|'%s') = %f" % (w, c, likelihood[w][c]))
  87.     # test doc
  88.     testdoc = "predictable with no fun"
  89.     predict = test_naive_bayes(testdoc, prior, likelihood, Cs, V)
  90.     print("predict class:", predict)
  92. if __name__ == "__main__":
  93.     testcase1()