def fit(self, x, y, sample_weight, eps=1e-8):
self._x, self._y = np.atleast_2d(x), np.array(y)
self.sample_weight = sample_weight
# 满足第一种停止准则，退出函数体
if self.stop1(eps):
return
# 用该 Node 数据实例化 Cluster 计算各种信息量
_cluster = Cluster(self._x, self._y, sample_weight, self.base)
# 对于根节点，需要额外计算数据的不确定性
if self.is_root:
if self.criterion == "gini":
self.chaos = _cluster.gini()
else:
self.chaos = _cluster.ent()
_max_gain, _chaos_lst = 0, []
_max_feature = _max_tar = None
# 遍历还能选择的特征
for feat in self.feats:
# 如果连续型或是CART算法，需要额外计算二分标准的取值集合
if self.wc[feat]:
_samples = np.sort(self._x.T[feat])
_set = (_samples[:-1] + _samples[1:]) * 0.5
elif self.is_cart:
_set = self.tree.feature_sets[feat]
# 遍历二分标准并调用二类问题相关的计算信息量的方法
if self.is_cart or self.wc[feat]:
for tar in _set:
_tmp_gain, _tmp_chaos_lst = _cluster.b
def get_threshold(self):
if self.category is None:
rs = 0
for leaf in self.leafs.values():
_cluster = Cluster(None, leaf["y"], None, self.base)
rs += len(leaf["y"]) * _cluster.ent()
return Cluster(self._x, self._y, None, self.base).ent() - rs / (len(self.leafs) - 1)
return 0
def fit(self, x, y, sample_weight, feature_bound=None, eps=1e-8):
self._x, self._y = np.atleast_2d(x), np.asarray(y)
self.sample_weight = sample_weight
if self.stop1(eps):
return
cluster = Cluster(self._x, self._y, sample_weight, self.base)
if self.is_root:
if self.criterion == "gini":
self.chaos = cluster.gini()
else:
self.chaos = cluster.ent()
max_gain, chaos_lst = 0, []
max_feature = max_tar = None
feat_len = len(self.feats)
if feature_bound is None:
indices = range(0, feat_len)
elif feature_bound == "log":
indices = np.random.permutation(feat_len)[:max(1, int(log2(feat_len)))]
else:
indices = np.random.permutation(feat_len)[:feature_bound]
tmp_feats = [self.feats[i] for i in indices]
xt, feat_sets = self._x.T, self.tree.feature_sets
bin_ig, ig = cluster.bin_info_gain, cluster.info_gain
for feat in tmp_feats:
if self.wc[feat]:
samples = np.sort(xt[feat])
feat_set = (samples[:-1] + samples[1:]) * 0.5
else:
if self.is_cart:
feat_set = feat_sets[feat]
else:
feat_set = None
if self.is_cart or self.wc[feat]:
for tar in feat_set:
tmp_gain, tmp_chaos_lst = bin_ig(
feat, tar, criterion=self.criterion, get_chaos_lst=True, continuous=self.wc[feat])
if tmp_gain > max_gain:
(max_gain, chaos_lst), max_feature, max_tar = (tmp_gain, tmp_chaos_lst), feat, tar
else:
tmp_gain, tmp_chaos_lst = ig(
feat, self.criterion, True, self.tree.feature_sets[feat])
if tmp_gain > max_gain:
(max_gain, chaos_lst), max_feature = (tmp_gain, tmp_chaos_lst), feat
if self.stop2(max_gain, eps):
return
self.feature_dim = max_feature
if self.is_cart or self.wc[max_feature]:
self.tar = max_tar
self._gen_children(chaos_lst, feature_bound)
if (self.left_child.category is not None and
self.left_child.category == self.right_child.category):
self.prune()
self.tree.reduce_nodes()
else:
self._gen_children(chaos_lst, feature_bound)
def fit(self, x, y, sample_weights, eps=1e-8):
self.feed_data(x, y)
self.sample_weights = sample_weights
if self.stop1(eps):
return
_cluster = Cluster(self._x, self._y, sample_weights, self.base)
_max_gain, _chaos_lst = 0, []
_max_feature = _max_tar = None
for feat in self.feats:
if self.wc[feat]:
_samples = np.sort(self._x.T[feat])
_set = (_samples[:-1] + _samples[1:]) * 0.5
else:
if self.is_cart:
_set = self.tree.feature_sets[feat]
else:
_set = None
if self.is_cart or self.wc[feat]:
for tar in _set:
_tmp_gain, _tmp_chaos_lst = _cluster.bin_info_gain(
feat, tar, criterion=self.criterion, get_chaos_lst=True, continuous=self.wc[feat])
if _tmp_gain > _max_gain:
(_max_gain, _chaos_lst), _max_feature, _max_tar = (_tmp_gain, _tmp_chaos_lst), feat, tar
else:
_tmp_gain, _tmp_chaos_lst = _cluster.info_gain(
feat, self.criterion, True, self.tree.feature_sets[feat])
if _tmp_gain > _max_gain:
(_max_gain, _chaos_lst), _max_feature = (_tmp_gain, _tmp_chaos_lst), feat
if self.stop2(_max_gain, eps):
return
self.feature_dim = _max_feature
if self.is_cart or self.wc[_max_feature]:
self.tar = _max_tar
self._gen_children(_chaos_lst)
if (self.left_child.category is not None and
self.left_child.category == self.right_child.category):
self.prune()
self.tree.reduce_nodes()
else:
self._gen_children(_chaos_lst)
def fit(self, x, y, sample_weight, feature_bound=None, eps=1e-8):
"""
1.根据划分标准将数据划分为若干份
2.依次用这若干份数据实例化新Node(新Node即是当前Node的子节点),同时将当前Node的相关信息传递新的Node
这里需要注意的是，如果划分标准是离散型的特征的话：
.若算法是ID3或C4.5,需将该特征对应的维度从新的Node的self.feats属性中除去
.若算法是CART，需要将二分标准从新的Node的二分标准取值集合中除去。
最后对新的Node调用fit方法，完成递归
"""
self._x, self._y = np.atleast_2d(x), np.asarray(y)
self.sample_weight = sample_weight
if self.stop1(eps):
return
cluster = Cluster(self._x, self._y, sample_weight, self.base)
if self.is_root:
if self.criterion =="gini":
self.chaos = cluster.gini()
else:
self.chaos = cluster.ent()
max_gain, chaos_lst = 0, []
max_feature = max_tar = None
feat_len = len(self.feats)
if feature_bound is None:
indices = range(0, feat_len)
elif feature_bound == "log":
indices = np.random.permutation(feat_len)[:max(1, int(log2(feat_len)))]
else:
indices = np.random.permutation(feat_len)[:feature_bound]
tmp_feats = [self.feats[i] for i in indices]
xt, feat_sets = self._x.T, self.tree.feature_sets
bin_ig, ig = cluster.bin_info_gain, cluster.info_gain
for feat in tmp_feats:#[0.1,2,3]遍历每个维度,通过遍历这些维度特征，选取使得不确定性最小的特征
if self.wc[feat]: #是否连续 wc=whether_continuous
samples = np.sort(xt[feat])
feat_set = (samples[:-1] + samples[1:]) *5
else:#非连续
if self.is_cart:
feat_set = feat_sets[feat] #取第feat维的特征类别 #{'y', 'p'}
else:
feat_set = None
if self.is_cart or self.wc[feat]:
for tar in feat_set:
tmp_gain, tmp_chaos_lst = bin_ig(feat,tar,criterion = self.criterion, get_chaos_lst=True,continuous=self.wc[feat])
if tmp_gain >max_gain:
(max_gain, chaos_lst), max_feature,max_tar = (tmp_gain, tmp_chaos_lst), feat, tar
else:
tmp_gain, tmp_chaos_lst = ig(feat, self.criterion, True, self.tree.feature_sets[feat])
if tmp_gain > max_gain:
(max_gain, chaos_lst), max_feature = (tmp_gain, tmp_chaos_lst), feat
if self.stop2(max_gain, eps):
return
self.feature_dim = max_feature #max_feature记录划分后获取最大信息增益的维度
if self.is_cart or self.wc[max_feature]:
self.tar = max_tar
#调用根据划分标准进行生成的方法
self._gen_children(chaos_lst, feature_bound)
#如果该Node的左子节点和右子节点都是叶节点且所属类别一样，那么就将他们合并，亦进行局部剪枝
if (self.left_child.category is not None and self.left_child.category == self.right_child.category):
self.prune()
#调用Tree的相关方法，将被剪掉的该Node的左右子节点从tree的记录所有Node的列表nodes中除去
self.tree.reduce_nodes()
else:
#调用根据划分标准进行生成的方法
self._gen_children(chaos_lst,feature_bound)