def initialize(data, lengths, factors):
mu = np.average(data[:, 2])
ulen, ilen = lengths[0], lengths[1]
p, q = randvec((ulen, factors)) - 0.5, randvec((ilen, factors)) - 0.5
nu, bu = np.zeros(ulen), np.zeros(ulen)
ni, bi = np.zeros(ilen), np.zeros(ilen)
for u, i, r in data:
nu[u] += 1
ni[i] += 1
bu[u] += r - mu
bi[i] += r - mu
for i in range(ilen):
bi[i] = bi[i] / ni[i] if ni[i] != 0 else 0
for u in range(ulen):
bu[u] = bu[u] / nu[u] if nu[u] != 0 else 0
return mu, bu, bi, p, q
def sgd(data, users, itesm, lam1, lam2, gamma1, gamma2):
print '[INFO] Start learning on λ1 = %.6f, λ2 = %.6f, ɣ1 = %.6f, ɣ2 = %.6f' % (lam1, lam2, gamma1, gamma2)
mu, bu, bi = get_normalized_info(data, users, itesm)
p, q = randvec((users, 50)) - 1/2., randvec((itesm, 50)) - 1/2.
steps = 10
for step in range(steps):
print '[INFO] Step %d/%d...' % (step + 1, steps)
for u, i, r in data:
e = r - (mu + bu[u] + bi[i] + np.dot(p[u], q[i]))
bu[u] += gamma1 * (e - lam1 * bu[u])
bi[i] += gamma1 * (e - lam1 * bi[i])
pu1 = p[u] + gamma2 * (e * q[i] - lam2 * p[u])
qi1 = q[i] + gamma2 * (e * p[u] - lam2 * q[i])
p[u], q[i] = pu1, qi1
print '[INFO] Learned.'
return lambda u, i: mu + bu[u] + bi[i] + np.dot(p[u], q[i])
def sgd(data, users, itesm, lam1, lam2, gamma1, gamma2):
print '[INFO] Start learning on λ1 = %.6f, λ2 = %.6f, ɣ1 = %.6f, ɣ2 = %.6f' % (
lam1, lam2, gamma1, gamma2)
mu, bu, bi = get_normalized_info(data, users, itesm)
p, q = randvec((users, 50)) - 1 / 2., randvec((itesm, 50)) - 1 / 2.
steps = 10
for step in range(steps):
print '[INFO] Step %d/%d...' % (step + 1, steps)
for u, i, r in data:
e = r - (mu + bu[u] + bi[i] + np.dot(p[u], q[i]))
bu[u] += gamma1 * (e - lam1 * bu[u])
bi[i] += gamma1 * (e - lam1 * bi[i])
pu1 = p[u] + gamma2 * (e * q[i] - lam2 * p[u])
qi1 = q[i] + gamma2 * (e * p[u] - lam2 * q[i])
p[u], q[i] = pu1, qi1
print '[INFO] Learned.'
return lambda u, i: mu + bu[u] + bi[i] + np.dot(p[u], q[i])