def generate_conditions(data, attr, cls, default_conditions={}):
if not data[attr]:
return []
if attr in default_conditions:
return default_conditions[attr]
if not continuous(data[attr][0]):
return [Condition('=="%s"'%(value), partial((lambda v, x:x == v), value)) for value in set(data[attr])]
result = {}
for value in set(data[attr]):
result[value] = {
'<=': Counter(),
'>': Counter(),
}
count = 0.0
for i, value in enumerate(data[attr]):
cls_value = data[cls][i]
for base in result:
if value <= base:
result[base]['<='][cls_value] += 1
else:
result[base]['>'][cls_value] += 1
count += 1.0
for base in result:
sv1 = sum(result[base]['<='].values())
sv2 = sum(result[base]['>'].values())
result[base]['<='] =sv1*entropy_from_counter(result[base]['<=']) / count
result[base]['>'] = sv2*entropy_from_counter(result[base]['>']) / count
result[base] = result[base]['<='] + result[base]['>']
number = min(result, key=result.get)
return [
Condition('<=' + str(number), lambda x: x <= number),
Condition('>' + str(number), lambda x: x > number)
]
def extract_comp(tup):
result = []
for element in tup:
if not continuous(element):
result.append(lambda x, y: 0 if x == y else 1)
else:
result.append(lambda x, y: x - y)
return result
def extract_normalizer(data, cls):
result = {}
for attr, values in data.items():
if attr == cls:
continue
if not continuous(values[0]):
result[attr] = lambda x: x
else:
mi, ma = min(values), max(values)
mi, ma = float(mi), float(ma)
result[attr] = lambda x: (float(x) - mi) / ma
return result
def naive_bayes(data, cls, new_tup, adjust=False):
counter = Counter(data[cls])
result = Counter()
for ci, count in counter.items():
pci = float(count) / float(len(data[cls]))
mult = 1.0
base = {x:0.0 for x in new_tup}
p = {}
for x, x_value in new_tup.items():
countx_ci = count_x_and_y(data[x], data[cls], x_value, ci)
p[x] = p_continuous if continuous(x_value) else p_categorical
if adjust and not countx_ci:
base[x] = 1.0
for x, x_value in new_tup.items():
px_ci = p[x](data[x], data[cls], x_value, ci, count, base[x])
mult *= px_ci
result[ci] = mult * pci
return result.most_common(1)[0][0]
