def __init__(self,
slop,
threshold,
init_theta=None,
score=None,
iter_method='newton'):
"""
不管是probit还是logit，都是用一样的参数估计算法，
基于牛顿迭代的极大似然算法和贝叶斯最大后验算法
:param slop: ndarray(float), 多维向量，斜率，区分度
:param threshold: ndarray(float), 单维向量，阈值，通俗度，难度
:param init_theta: ndarray(int|float), 特质向量初值
:param score: ndarray(0|1), 得分向量
"""
if not isinstance(slop, np.ndarray):
raise ItemParamError('item param must be ndarray')
if not isinstance(threshold, np.ndarray):
raise ItemParamError('item param must be ndarray')
if len(slop.shape) == 1:
slop.shape = 1, slop.shape[0]
if len(slop) != len(threshold):
raise ItemParamError('item param must be same length')
if score is not None:
if not isinstance(score, np.ndarray):
raise ScoreError('score must be ndarray')
if len(score) != len(slop):
raise ScoreError('score must be same length as item param')
if init_theta is not None and not isinstance(init_theta, np.ndarray):
raise ThetaError('init_theta must be ndarray')
if iter_method not in ('newton', 'gradient_ascent'):
raise IterMethodError(
'iter_method must be newton or gradient_ascent')
self._slop = slop
self._score = score
self._threshold = threshold
self._init_theta = init_theta if init_theta is not None else np.zeros(
len(self._slop[0]))
# 默认bayes先验正态分布标准差
# TODO 改为根据样本估计
self._inv_psi = np.identity(len(self._slop[0]))
self._iter_method = iter_method
def z(self, theta):
if not isinstance(theta, np.ndarray):
raise ThetaError('theta must be ndarray')
return self._z(theta)
def prob(self, theta):
# 回答为1的概率
if not isinstance(theta, np.ndarray):
raise ThetaError('theta must be ndarray')
return self._prob(theta)
def info(self, theta):
# 信息矩阵
if not isinstance(theta, np.ndarray):
raise ThetaError('theta must be ndarray')
_info = super(BayesProbitModel, self).info(theta)
return _info + self._inv_psi
def info(self, theta):
# 信息矩阵
if not isinstance(theta, np.ndarray):
raise ThetaError('theta must be ndarray')
h, prob_val, w = self._get_h_prob_val_w(theta)
return np.dot(self._slop.transpose() * w, self._slop)
def info(self, theta):
# 信息矩阵
if not isinstance(theta, np.ndarray):
raise ThetaError('theta must be ndarray')
prob_val = self._prob(theta)
return self._expect(prob_val)