def nb_likelihood(x, m, r):
p = float(r) / (r + m)
return nbinom.pmf(x, r, p)
Poisson-NB *regression* mixture using concomitant variables (mixture of experts)
20-11-2014
"""
"""
Can use pois_likelihood from before
"""
"""
Negative binomial regression likelihood
"""
def nb_regression_likelihood(y, x, (a_0, a_1), r):
m = a_0 + a_1 * x
p = float(r) / (r + m)
return nbinom.pmf(y, r, p)
def mix_from_params(x, (w0, w1)): # just a sigmoid function
pi = np.zeros((x.size,2))
for i in range(x.size):
pi[i,0] = 1.0/(1 + np.exp(w0 + w1*x[i]))
pi[:,1] = 1 - pi[:,0]
return pi
def NB_regression_objective_function((a0, a1, r), y, x, gam):
value = 0
# convert to floats otherwise math.log(m / (m+r)) -> -inf
r = float(r)
m = a0 + a1 * x
