base MF

exp3/src/MatrixFactorization.py

@@ -0,0 +1,170 @@
+import numpy as np
+import cmath
+
+
+class Matrix_Factorization(object):
+
+ def __init__(self, K=10, alpha=0.05, beta=0.02, epoch=500, regularization=True, random_state=100):
+
+ self.R = None
+ self.K = K
+ self.P = None
+ self.Q = None
+ self.r_index = None
+ self.r = None
+ self.length = None
+ self.init_alpha=alpha
+ self.alpha = alpha
+ self.beta = beta
+ self.epoch = epoch
+ self.regularization = regularization
+ self.random_state = random_state
+
+
+ def fit(self, R):
+ #将矩阵R分解为M*K和K*N
+
+ np.random.seed(self.random_state)
+ self.R = R
+ M, N = self.R.shape
+ #P是M*K的随机矩阵，Q是K*N的随机矩阵
+ self.P = np.random.rand(M, self.K)
+ self.Q = np.random.rand(N, self.K)
+
+ #r_index是R中不为0的元素的数组下标
+ self.r_index = self.R.nonzero()
+ #r是R中不为0的元素值
+ self.r = self.R[self.r_index[0], self.r_index[1]]
+ self.length = len(self.r)
+
+
+ def _comp_descent(self, index):
+ #r中第index个元素
+ r_i = self.r_index[0][index]
+ r_j = self.r_index[1][index]
+
+ p_i = self.P[r_i]
+ q_j = self.Q[r_j]
+
+ r_ij_hat = p_i.dot(q_j)
+ e_ij = self.R[r_i, r_j] - r_ij_hat
+
+ #正则化？？？？？？？？？
+ if self.regularization == True:
+ descent_p_i = -2 * e_ij * q_j + self.beta * p_i
+ descent_q_j = -2 * e_ij * p_i + self.beta * q_j
+ else:
+ descent_p_i = -2 * e_ij * q_j
+ descent_q_j = -2 * e_ij * p_i
+
+ return r_i, r_j, p_i, q_j, descent_p_i, descent_q_j
+
+
+ def _update(self, p_i, q_j, descent_p_i, descent_q_j):
+
+ p_i_new = p_i - self.alpha * descent_p_i
+ q_j_new = q_j - self.alpha * descent_q_j
+
+ return p_i_new, q_j_new
+
+
+ def _estimate_r_hat(self):
+
+ r_hat = self.P.dot(self.Q.T)[self.r_index[0], self.r_index[1]]
+
+ return r_hat
+
+
+ def start(self):
+
+ epoch_num = 1
+ #epoch_cnt = 0
+ while epoch_num <= self.epoch:
+ for index in range(0, self.length):
+
+ r_i, r_j, p_i, q_j, descent_p_i, descent_q_j = self._comp_descent(index)
+ p_i_new, q_j_new = self._update(p_i, q_j, descent_p_i, descent_q_j)
+
+ self.P[r_i] = p_i_new
+ self.Q[r_j] = q_j_new
+
+ self.alpha=max(0.01,self.init_alpha*pow(0.98,epoch_num))
+ #if epoch_num%10==0:
+ # epoch_cnt+=1
+ # self.alpha=self.init_alpha/cmath.sqrt(epoch_cnt)
+ r_hat = self._estimate_r_hat()
+ e = r_hat - self.r
+ error = e.dot(e)
+ #if epoch_num%10==0:
+ # print ('The error is %s=================Epoch:%s' %(error, epoch_num))
+ print ('The error is %s=================Epoch:%s' %(error, epoch_num))
+ epoch_num += 1
+
+ R_hat = self.P.dot(self.Q.T)
+ return R_hat
+
+
+if __name__ == '__main__':
+
+ read_data_path=".//data//DoubanMusic.txt"
+ write_data_path=".//result//"
+ f=open(read_data_path,"r")
+ lines=f.readlines()
+ f.close()
+
+ user_total=23599
+ item_total=21602
+ sep='\t'
+ comma=','
+
+ #R=np.zeros((user_total,item_total),dtype=np.int32)
+ R=np.zeros((user_total,item_total))
+ for line in lines:
+ temp=line.split()
+ UserID=int(temp[0])
+ pos_cnt=0
+ pos_total=0
+ music_list=[]
+ for pair in temp[1:]:
+ pair=pair.split(',')
+ MusicID=int(pair[0])
+ MusicRating=int(pair[1])
+ R[UserID,MusicID]=MusicRating
+ if MusicRating>0:
+ pos_cnt+=1
+ pos_total+=MusicRating
+ music_list.append(MusicID)
+ if pos_cnt:
+ MeanRating=pos_total/pos_cnt
+ else:
+ MeanRating=2.5
+ for musicID in music_list:
+ #if R[UserID,musicID]>0:
+ # R[UserID,musicID]-=MeanRating
+ # #########+1
+ # R[UserID,musicID]+=1
+ #else:
+ # if R[UserID,musicID]<0:
+ # ########+1
+ # R[UserID,musicID]=MeanRating+1
+ if R[UserID,musicID]<0:
+ R[UserID,musicID]=(R[UserID,musicID]+1)/5
+ if R[UserID,musicID]<0:
+ R[UserID,musicID]=(MeanRating+1)/5
+ np.save("Rarray",R)
+
+ aa = Matrix_Factorization(K = 5)
+ aa.fit(R)
+ R_hat=aa.start()
+
+ RatingRank=np.argsort(R_hat)
+
+ with open(write_data_path+"res.txt","w") as f:
+ for userID in range(0,user_total):
+ output_str=str(userID)+sep
+ for music_index in range(item_total-100,item_total):
+ output_str+=str(RatingRank[userID][music_index])+comma
+ f.write(output_str[:-1]+'\n')
+
+
+