class TrustSVD(MF):
"""
docstring for TrustSVD
implement the TrustSVD
Koren Y. Factor in the neighbors: Scalable and accurate collaborative filtering[J]. ACM Transactions on Knowledge Discovery from Data (TKDD), 2010, 4(1): 1.
"""
def __init__(self):
super(TrustSVD, self).__init__()
self.config.lr = 0.005
self.config.maxIter = 100
self.config.lambdaP = 1.2
self.config.lambdaQ = 1.2
self.config.lambdaB = 1.2
self.config.lambdaY = 1.2
self.config.lambdaW = 1.2
self.config.lambdaT = 0.9
self.tg = TrustGetter()
self.init_model()
def init_model(self):
super(TrustSVD, self).init_model()
self.Bu = np.random.rand(self.rg.get_train_size()[0]) / (
self.config.factor**0.5) # bias value of user
self.Bi = np.random.rand(self.rg.get_train_size()[1]) / (
self.config.factor**0.5) # bias value of item
self.Y = np.random.rand(self.rg.get_train_size()[1],
self.config.factor) / (self.config.factor**0.5
) # implicit preference
self.W = np.random.rand(self.rg.get_train_size()[0],
self.config.factor) / (self.config.factor**0.5
) # implicit preference
def train_model(self):
iteration = 0
while iteration < self.config.maxIter:
self.loss = 0
for index, line in enumerate(self.rg.trainSet()):
user, item, rating = line
u = self.rg.user[user]
i = self.rg.item[item]
error = rating - self.predict(user, item)
self.loss += error**2
p, q = self.P[u], self.Q[i]
nu, sum_y = self.get_sum_y(user)
nv, sum_w = self.get_sum_w(user)
frac = lambda x: 1.0 / math.sqrt(x)
# update latent vectors
self.Bu[u] += self.config.lr * (
error - self.config.lambdaB * frac(nu) * self.Bu[u])
self.Bi[i] += self.config.lr * (
error - self.config.lambdaB * frac(nv) * self.Bi[i])
self.Q[i] += self.config.lr * (
error *
(p + sum_y + sum_w) - self.config.lambdaQ * frac(nu) * q)
followees = self.tg.get_followees(user)
ws = np.zeros(self.config.factor)
for followee in followees:
if self.rg.containsUser(user) and self.rg.containsUser(
followee):
nw = len(self.tg.get_followers(followee))
vid = self.rg.user[followee]
w = self.W[vid]
weight = 1 # followees[followee]
err = w.dot(p) - weight
self.loss += err**2
ws += err * w
self.W[vid] += self.config.lr * (
err * frac(nv) * q - self.config.lambdaT * err * p
- self.config.lambdaW * frac(nw) * w) # 更新w
self.P[u] += self.config.lr * (
error * q - self.config.lambdaT * ws -
(self.config.lambdaP * frac(nu) +
self.config.lambdaT * frac(nv)) * p)
u_items = self.rg.user_rated_items(u) # 更新y
for j in u_items:
idj = self.rg.item[j]
self.Y[idj] += self.config.lr * (
error * frac(nu) * q -
self.config.lambdaY * frac(nv) * self.Y[idj])
self.loss += self.config.lambdaP * (self.P * self.P).sum() + self.config.lambdaQ * (self.Q * self.Q).sum() \
+ self.config.lambdaB * (
(self.Bu * self.Bu).sum() + (self.Bi * self.Bi).sum()) + self.config.lambdaY * (
self.Y * self.Y).sum() + self.config.lambdaW * (self.W * self.W).sum()
iteration += 1
if self.isConverged(iteration):
break
def predict(self, u, i):
if self.rg.containsUser(u) and self.rg.containsItem(i):
_, sum_y = self.get_sum_y(u)
_, sum_w = self.get_sum_w(u)
u = self.rg.user[u]
i = self.rg.item[i]
return self.Q[i].dot(
self.P[u] + sum_y +
sum_w) + self.rg.globalMean + self.Bi[i] + self.Bu[u]
else:
return self.rg.globalMean
def get_sum_y(self, u):
u_items = self.rg.user_rated_items(u)
nu = len(u_items)
sum_y = np.zeros(self.config.factor)
for j in u_items:
sum_y += self.Y[self.rg.item[j]]
sum_y /= (np.sqrt(nu))
return nu, sum_y
def get_sum_w(self, u):
followees = self.tg.get_followees(u)
nu = 1
sum_w = np.zeros(self.config.factor)
for v in followees.keys():
if self.rg.containsUser(v):
nu += 1
sum_w += self.W[self.rg.user[v]]
sum_w /= np.sqrt(nu)
return nu, sum_w
class SocialReg(MF):
"""
docstring for SocialReg
Ma H, Zhou D, Liu C, et al. Recommender systems with social regularization[C]//Proceedings of the fourth ACM international conference on Web search and data mining. ACM, 2011: 287-296.
"""
def __init__(self):
super(SocialReg, self).__init__()
# self.config.lambdaP = 0.001
# self.config.lambdaQ = 0.001
self.config.alpha = 0.1
self.tg = TrustGetter()
# self.init_model()
def init_model(self, k):
super(SocialReg, self).init_model(k)
from collections import defaultdict
self.user_sim = SimMatrix()
print('constructing user-user similarity matrix...')
# self.user_sim = util.load_data('../data/sim/ft_cf_soreg08_cv1.pkl')
for u in self.rg.user:
for f in self.tg.get_followees(u):
if self.user_sim.contains(u, f):
continue
sim = self.get_sim(u, f)
self.user_sim.set(u, f, sim)
# util.save_data(self.user_sim,'../data/sim/ft_cf_soreg08.pkl')
def get_sim(self, u, k):
sim = (pearson_sp(self.rg.get_row(u), self.rg.get_row(k)) +
1.0) / 2.0 # fit the value into range [0.0,1.0]
return sim
def train_model(self, k):
super(SocialReg, self).train_model(k)
iteration = 0
while iteration < self.config.maxIter:
self.loss = 0
for index, line in enumerate(self.rg.trainSet()):
user, item, rating = line
u = self.rg.user[user]
i = self.rg.item[item]
error = rating - self.predict(user, item)
self.loss += 0.5 * error**2
p, q = self.P[u], self.Q[i]
social_term_p, social_term_loss = np.zeros(
(self.config.factor)), 0.0
followees = self.tg.get_followees(user)
for followee in followees:
if self.rg.containsUser(followee):
s = self.user_sim[user][followee]
uf = self.P[self.rg.user[followee]]
social_term_p += s * (p - uf)
social_term_loss += s * ((p - uf).dot(p - uf))
social_term_m = np.zeros((self.config.factor))
followers = self.tg.get_followers(user)
for follower in followers:
if self.rg.containsUser(follower):
s = self.user_sim[user][follower]
ug = self.P[self.rg.user[follower]]
social_term_m += s * (p - ug)
# update latent vectors
self.P[u] += self.config.lr * (
error * q - self.config.alpha *
(social_term_p + social_term_m) - self.config.lambdaP * p)
self.Q[i] += self.config.lr * (error * p -
self.config.lambdaQ * q)
self.loss += 0.5 * self.config.alpha * social_term_loss
self.loss += 0.5 * self.config.lambdaP * (self.P * self.P).sum(
) + 0.5 * self.config.lambdaQ * (self.Q * self.Q).sum()
iteration += 1
if self.isConverged(iteration):
break
class TrustWalker(MF):
"""
docstring for TrustWalker
Jamali M, Ester M. Trustwalker: a random walk model for combining trust-based and item-based recommendation[C]//Proceedings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 2009: 397-406.
"""
def __init__(self):
super(TrustWalker, self).__init__()
np.random.seed(0)
self.tg = TrustGetter()
self.init_model()
def init_model(self):
self.p = 1.0
pass
def single_random_walk(self, user=5, item=3, k=0):
print(user, item, k)
print('%s%d' % ('k=', k))
if self.rg.containsUserItem(
user, item
): # judge whether user u rate on item i, if so, return the rating.
return self.p, self.rg.trainSet_u[user][item]
else:
rand_num = np.random.rand(1) # get random number
# compute the stop probability
stop_prob, max_item, p_j = self.get_stop_prob(user, item, k)
print('stop probability:' + str(stop_prob))
# print('%s%d'%('stop probbability:',stop_prob))
# the probability of stopping walk
print(rand_num, stop_prob)
if rand_num < stop_prob or k >= 6: # no more than six steps
# get the most similar item j, and return r(u,j)
rating = self.rg.trainSet_u[user][max_item]
self.p = self.p * stop_prob * p_j
return (self.p, rating)
# compute the probability of next random walk
else:
# get next user for random walk
next_user, tu_prob = self.get_followee_user(
user) # if user don't have friends in trust network
print('next step user is:' + str(next_user))
if next_user == None: # if no next user
_, max_item, p_j = self.get_stop_prob(
user, item, -1) # no sense if k=-1
if max_item == 0: # if no next user and no similar users
return self.p, 0
rating = self.rg.trainSet_u[user][max_item]
self.p = self.p * p_j
return (self.p, rating)
self.p = self.p * (1 - stop_prob) * tu_prob
k += 1
return self.single_random_walk(user=next_user, item=item,
k=k) # 特么忘了return 当然返回None啦
def get_followee_user(self, user):
p = 0
followees = list(self.tg.get_followees(user))
num_foll = len(followees)
if num_foll == 0:
return None, 0
# pick one randomly
ind = np.random.randint(num_foll)
p = 1.0 / num_foll
return followees[ind], p
# def get_max_item(self,user,item):
# u_items=self.rg.user_rated_items(user)
# sum_sim=0.0
# max_sim=0
# max_prob_item=0
# if len(u_items)==0:
# return 0,0
# print(u_items)
# for i,u_item in enumerate(u_items):
# sim=self.get_sim(item,u_item)
# sum_sim+=sim
# if sim>max_sim:
# max_sim=sim
# max_prob_item=u_item
# return max_prob_item,max_sim/sum_sim
def get_stop_prob(self, user, item, k):
p = 1.0
sum_sim = 0.0
max_sim = 0
max_prob = 0.0
max_prob_item = 0
if k == 0: # if k==0，the stop probability=0
self.p = 1.0
return 0, 0, 0
param = sigmoid_2(k)
u_items = self.rg.user_rated_items(user)
print(u_items)
if len(u_items) == 0:
return 0, 0, 0
for u_item in u_items:
sim = self.get_sim(item, u_item)
sum_sim += sim
prob = sim * param
if prob > max_prob:
max_sim = sim
max_prob = prob
max_prob_item = u_item
return max_prob, max_prob_item, max_sim / sum_sim # 返回停止概率，最相似item，选择item j的概率
def get_sim(self, item1, item2):
return cosine_improved_sp(self.rg.get_col(item1),
self.rg.get_col(item2))
class SocialRec(MF):
"""
docstring for SocialRec
Ma H, Yang H, Lyu M R, et al. Sorec: social recommendation using probabilistic matrix factorization[C]//Proceedings of the 17th ACM conference on Information and knowledge management. ACM, 2008: 931-940.
"""
def __init__(self):
super(SocialRec, self).__init__()
# self.config.lr=0.0001
self.config.alpha = 0.1
self.config.lambdaZ = 0.01
self.tg = TrustGetter()
# self.init_model()
def init_model(self, k):
super(SocialRec, self).init_model(k)
self.Z = np.random.rand(
self.rg.get_train_size()[0], self.config.factor) / (
self.config.factor**0.5) # latent user social matrix
def train_model(self, k):
super(SocialRec, self).train_model(k)
iteration = 0
while iteration < self.config.maxIter:
# tempP=np.zeros((self.rg.get_train_size()[0], self.config.factor))
self.loss = 0
for index, line in enumerate(self.rg.trainSet()):
user, item, rating = line
u = self.rg.user[user]
i = self.rg.item[item]
error = rating - self.predict(user, item)
self.loss += error**2
p, q = self.P[u], self.Q[i]
followees = self.tg.get_followees(user)
zs = np.zeros(self.config.factor)
for followee in followees:
if self.rg.containsUser(user) and self.rg.containsUser(
followee):
vminus = len(
self.tg.get_followers(followee)) # ~ d - (k)
uplus = len(self.tg.get_followees(user)) # ~ d + (i)
import math
try:
weight = math.sqrt(vminus / (uplus + vminus + 0.0))
except ZeroDivisionError:
weight = 1
zid = self.rg.user[followee]
z = self.Z[zid]
err = weight - z.dot(p)
self.loss += err**2
zs += -1.0 * err * p
self.Z[zid] += self.config.lr * (
self.config.alpha * err * p -
self.config.lambdaZ * z)
self.P[u] += self.config.lr * (error * q - self.config.alpha *
zs - self.config.lambdaP * p)
self.Q[i] += self.config.lr * (error * p -
self.config.lambdaQ * q)
self.loss += self.config.lambdaP * (self.P * self.P).sum() + self.config.lambdaQ * (self.Q * self.Q).sum() \
+ self.config.lambdaZ * (self.Z * self.Z).sum()
iteration += 1
if self.isConverged(iteration):
break
class RSTE(MF):
"""
docstring for RSTE
Ma H, King I, Lyu M R. Learning to recommend with social trust ensemble[C]//Proceedings of the 32nd international ACM SIGIR conference on Research and development in information retrieval. ACM, 2009: 203-210.
"""
def __init__(self):
super(RSTE, self).__init__()
# self.maxIter=700
self.config.alpha = 0.5
# self.config.lambdaH=0.01
self.tg = TrustGetter()
# self.init_model()
def init_model(self, k):
super(RSTE, self).init_model(k)
# from collections import defaultdict
# self.Sim = defaultdict(dict)
# print('constructing similarity matrix...')
# for user in self.rg.user:
# for k in self.tg.get_followees(user):
# if user in self.Sim and k in self.Sim[user]:
# pass
# else:
# self.Sim[user][k]=self.get_sim(user,k)
def train_model(self, k):
super(RSTE, self).train_model(k)
iteration = 0
while iteration < self.config.maxIter:
self.loss = 0
for index, line in enumerate(self.rg.trainSet()):
user, item, rating = line
error = rating - self.predict(user, item)
self.loss += error**2
social_term, _ = self.get_social_term_Q(user, item)
u = self.rg.user[user]
i = self.rg.item[item]
p, q = self.P[u], self.Q[i]
# update latent vectors
self.P[u] += self.config.lr * (self.config.alpha * error * q + \
(1 - self.config.alpha) * self.get_social_term_P(user,
item) - self.config.lambdaP * p)
self.Q[i] += self.config.lr * (error * (self.config.alpha * p + (1 - self.config.alpha) * social_term) \
- self.config.lambdaQ * q)
self.loss += self.config.lambdaP * (self.P * self.P).sum(
) + self.config.lambdaQ * (self.Q * self.Q).sum()
iteration += 1
if self.isConverged(iteration):
break
def get_social_term_Q(self, user, item):
if self.rg.containsUser(user) and self.rg.containsItem(item):
i = self.rg.item[item]
u = self.rg.user[user]
social_term_loss = 0
social_term = np.zeros(self.config.factor)
followees = self.tg.get_followees(user)
weights = []
indexes = []
for followee in followees:
if self.rg.containsUser(followee): # followee is in rating set
indexes.append(self.rg.user[followee])
weights.append(followees[followee])
weights = np.array(weights)
qw = weights.sum()
indexes = np.array(indexes)
if qw != 0:
social_term = weights.dot(self.P[indexes])
social_term /= qw
social_term_loss += weights.dot(
(self.P[indexes].dot(self.Q[i]))) / qw
return social_term, social_term_loss
def get_social_term_P(self, user, item):
i = self.rg.item[item]
# social_term_loss = 0
social_term = np.zeros(self.config.factor)
followers = self.tg.get_followers(user)
weights = []
indexes = []
errs = []
for follower in followers:
if self.rg.containsUser(follower) and self.rg.containsItem(
item) and self.rg.containsUserItem(
follower, item): # followee is in rating set
indexes.append(self.rg.user[follower])
weights.append(followers[follower])
errs.append(self.rg.trainSet_u[follower][item] -
self.predict(follower, item))
weights = np.array(weights)
indexes = np.array(indexes)
errs = np.array(errs)
qw = weights.sum()
if qw != 0:
for es in errs * weights:
social_term += es * self.Q[i]
social_term /= qw
# social_term_loss += weights.dot((self.P[indexes].dot(self.Q[i])))
return social_term
def predict(self, u, i):
if self.rg.containsUser(u) and self.rg.containsItem(i):
_, social_term_loss = self.get_social_term_Q(u, i)
i = self.rg.item[i]
u = self.rg.user[u]
if social_term_loss != 0:
return self.config.alpha * self.P[u].dot(
self.Q[i]) + (1 - self.config.alpha) * social_term_loss
else:
return self.P[u].dot(self.Q[i])
else:
return self.rg.globalMean
class SocialMF(MF):
"""
docstring for SocialMF
Jamali M, Ester M. A matrix factorization technique with trust propagation for recommendation in social networks[C]//Proceedings of the fourth ACM conference on Recommender systems. ACM, 2010: 135-142.
"""
def __init__(self):
super(SocialMF, self).__init__()
# self.config.lr=0.0001
self.config.alpha = 1 # 0.8 rmse=0.87605
self.tg = TrustGetter() # loading trust data
self.init_model()
def train_model(self):
iteration = 0
while iteration < self.config.maxIter:
self.loss = 0
for index, line in enumerate(self.rg.trainSet()):
user, item, rating = line
u = self.rg.user[user]
i = self.rg.item[item]
error = rating - self.predict(user, item)
self.loss += error**2
p, q = self.P[u], self.Q[i]
total_weight = 0.0
social_term = np.zeros(self.config.factor)
followees = self.tg.get_followees(user) # 获得u所关注的用户列表
for followee in followees:
weight = followees[followee]
if self.rg.containsUser(followee):
uk = self.P[self.rg.user[followee]]
social_term += weight * uk
total_weight += weight
if total_weight != 0:
social_term = p - social_term / total_weight
social_term_a = np.zeros(self.config.factor)
total_count = 0
followers = self.tg.get_followers(user)
for follower in followers:
if self.rg.containsUser(follower):
total_count += 1
uv = self.P[self.rg.user[follower]]
social_term_m = np.zeros(self.config.factor)
total_weight = 0.0
followees = self.tg.get_followees(
follower) # 获得u所关注的用户列表
for followee in followees:
weight = followees[followee]
if self.rg.containsUser(followee):
uw = self.P[self.rg.user[followee]]
social_term_m += weight * uw
total_weight += weight
if total_weight != 0:
social_term_a += uv - social_term_m / total_weight
if total_count != 0:
social_term_a /= total_count
# update latent vectors
self.P[u] += self.config.lr * (
error * q - self.config.alpha * social_term +
self.config.alpha * social_term_a - self.config.lambdaP * p
) #
self.Q[i] += self.config.lr * (error * p -
self.config.lambdaQ * q)
self.loss += self.config.alpha * social_term.dot(
social_term).sum()
self.loss += self.config.lambdaP * (self.P * self.P).sum(
) + self.config.lambdaQ * (self.Q * self.Q).sum()
iteration += 1
if self.isConverged(iteration):
break
class TrustWalker(MF):
"""
docstring for TrustWalker
Jamali M, Ester M. Trustwalker: a random walk model for combining trust-based and item-based recommendation[C]//Proceedings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 2009: 397-406.
"""
def __init__(self):
super(TrustWalker, self).__init__()
np.random.seed(0)
self.tg = TrustGetter()
self.init_model()
def init_model(self):
self.p = 1.0
pass
def single_random_walk(self, user=5, item=3, k=0):
print(user, item, k)
print('%s%d' % ('k=', k))
if self.rg.containsUserItem(user, item): # 判断用户user是否对item评分，若评分则返回r
return self.p, self.rg.trainSet_u[user][item]
else:
rand_num = np.random.rand(1) # 获取随机数
# 计算停止概率
stop_prob, max_item, p_j = self.get_stop_prob(user, item, k)
print('stop probability:' + str(stop_prob))
# print('%s%d'%('stop probbability:',stop_prob))
# 停止游走的概率
print(rand_num, stop_prob)
if rand_num < stop_prob or k >= 6: # （不超过6步）
# 获取相似项目j，然后返回r(u,j)
rating = self.rg.trainSet_u[user][max_item]
self.p = self.p * stop_prob * p_j
return (self.p, rating)
# 获取继续游走的概率
else:
# 得到下一个游走的用户v
next_user, tu_prob = self.get_followee_user(
user) # 对于user在信任网络中没有朋友的情况
print('next step user is:' + str(next_user))
if next_user == None: # 不存在下一个用户的情况
_, max_item, p_j = self.get_stop_prob(user, item,
-1) # k=-1无意义
if max_item == 0: # 没有一个用户并且不存在相似的用户的情况
return self.p, 0
rating = self.rg.trainSet_u[user][max_item]
self.p = self.p * p_j
return (self.p, rating)
self.p = self.p * (1 - stop_prob) * tu_prob
k += 1
return self.single_random_walk(user=next_user, item=item,
k=k) # 特么忘了return 当然返回None啦
def get_followee_user(self, user):
p = 0
followees = list(self.tg.get_followees(user))
num_foll = len(followees)
if num_foll == 0:
return None, 0
# 随机选一个
ind = np.random.randint(num_foll)
p = 1.0 / num_foll
return followees[ind], p
# def get_max_item(self,user,item):
# u_items=self.rg.user_rated_items(user)
# sum_sim=0.0
# max_sim=0
# max_prob_item=0
# if len(u_items)==0:
# return 0,0
# print(u_items)
# for i,u_item in enumerate(u_items):
# sim=self.get_sim(item,u_item)
# sum_sim+=sim
# if sim>max_sim:
# max_sim=sim
# max_prob_item=u_item
# return max_prob_item,max_sim/sum_sim
def get_stop_prob(self, user, item, k):
p = 1.0
sum_sim = 0.0 # 用于计算P(Y=j)-分母
max_sim = 0 # 用于计算P(Y=j)-分子
max_prob = 0.0 # 用于计算停止概率
max_prob_item = 0 # 用于方便选择相似item-j
if k == 0: # k==0时，停留概率为0
self.p = 1.0
return 0, 0, 0
param = sigmoid_2(k)
u_items = self.rg.user_rated_items(user)
print(u_items)
if len(u_items) == 0:
return 0, 0, 0
for u_item in u_items:
sim = self.get_sim(item, u_item)
sum_sim += sim
prob = sim * param
if prob > max_prob:
max_sim = sim
max_prob = prob
max_prob_item = u_item
return max_prob, max_prob_item, max_sim / sum_sim # 返回停止概率，最相似item，选择item j的概率
def get_sim(self, item1, item2):
return cosine_improved_sp(self.rg.get_col(item1),
self.rg.get_col(item2))