def best_features():
ham_train = bayes.HAM + bayes.TRAIN
spam_train = bayes.SPAM + bayes.TRAIN
ham_words = map(lambda s: '/' + s + '/',
set.union(*map(toolkit.word_bag, toolkit.get_files(ham_train))))
spam_words = map(lambda s: '/' + s + '/',
set.union(*map(toolkit.word_bag, toolkit.get_files(spam_train))))
all_words = set(ham_words + spam_words)
# p(w_i | c)
p_w_ham = defaultdict(lambda: toolkit.ZERO, zip(ham_words,
toolkit.countre(ham_train, ham_words, smoothing=toolkit.ONE)[0]))
p_w_spam = defaultdict(lambda: toolkit.ZERO, zip(spam_words,
toolkit.countre(spam_train, spam_words, smoothing=toolkit.ONE)[0]))
#for word in all_words:
# if type(p_w_ham[word]) != toolkit.NUM or type(p_w_spam[word]) != toolkit.NUM:
# print p_w_ham[word], p_w_spam[word], word
#return range(0, 100)
mut_inf = dict()
maxlog = math.log(toolkit.MAX)
def no_error(x, y):
try:
return math.log(x / y)
except Exception:
return maxlog
for word in all_words:
p_w_h = p_w_ham[word]
#if type(p_w_h)() == list() or type(p_w_s)() == list:
# print "'" + word + "'"
#else:
# print "*",
p_nw_h = toolkit.ONE - p_w_h
p_w_s = p_w_spam[word]
p_nw_s = toolkit.ONE - p_w_s
p_w = p_w_h * toolkit.PRIOR_HAM + p_w_s * toolkit.PRIOR_SPAM
p_nw = toolkit.ONE - p_w
log_w_h = no_error(p_w, p_w_h)
log_nw_h = no_error(p_nw, p_nw_h)
log_w_s = no_error(p_w, p_w_s)
log_nw_s = no_error(p_nw, p_nw_s)
mut_inf[word] = p_w_h * toolkit.PRIOR_HAM * log_w_h
mut_inf[word] += p_nw_h * toolkit.PRIOR_HAM * log_nw_h
mut_inf[word] += p_w_s * toolkit.PRIOR_SPAM * log_w_s
mut_inf[word] += p_nw_s * toolkit.PRIOR_SPAM * log_nw_s
return sorted(mut_inf.iteritems(), key=operator.itemgetter(1), reverse=True)
def instance_feature_prob(instance, features, clss, train=True, smoothing=toolkit.ZERO):
"""
Corresponds to [p(x_i | C_k) for x_i in x] from the assignment.
The instance is a message, which is either of clss HAM or clss SPAM.
features is a collection of (compiled) regular expressions. If applied to
the instance, we compute for every feature, if it occurs in the instance.
We divide that number by how many training/testing instances of the class
match against the feature.
``The probability of an observation (i.e., an email) given the class (i.e.,
ham or spam), p(x|Ck ) is then modelled as the probability of seeing
specific keywords in the email.''
"""
folder = clss + TRAIN if train else clss + TEST
countre_result = toolkit.countre(folder, features, smoothing=smoothing)[0]
feature_presence = toolkit.presentre(instance, features)
return map(lambda c, p: 1 if c == toolkit.ZERO and p == toolkit.ZERO else pow(c, p) * pow(1 - c, 1 - p), countre_result, feature_presence)