class UserCF:
DataPath = "../dataset/ml-latest-small/ratings.csv" # 全部数据路径
Ratio = 0.8 # 训练集占数据集的比例
K = 10 # 最邻近的K个user
N = 10 # 为每个user推荐N个item
R_train = None # type:pd.DataFrame # 训练集评分矩阵
R_test = None # type:pd.DataFrame # 测试集评分矩阵
users = None # type:np.ndarray # 全部user列表
items = None # type:np.ndarray # 全部item列表
W = None # type:pd.DataFrame # user相似度矩阵
P = None # type:pd.DataFrame # 预测评分矩阵
logger = Logger.fileAndConsoleLogger('../result/UserCF/userCF.log')
def load_data(self):
"""
1. 读取原始数据。ratings数据格式为[Rating]
"""
ratings = FileUtil.readRatings(fileName=self.DataPath)
# 划分训练集和测试集
train, test = ratioSplitter(ratings, self.Ratio, fixed=True)
# 获取user和item列表,这里需要去重，并保持顺序
users = np.unique([r.uid for r in train])
items = np.unique([r.iid for r in train])
self.logger.debug(f"users长度:{len(users)}")
self.logger.debug(f"items长度:{len(items)}")
# 构建训练集评分矩阵
R_train = pd.DataFrame(np.zeros((len(users), len(items))), index=users, columns=items)
for r in train:
R_train.at[r.uid, r.iid] = 1 # 隐式反馈，这里就设置为1，忽略具体的评分数据
self.logger.debug(f"训练集评分矩阵:\n{R_train}")
# 构建测试集评分矩阵
test_users = np.unique([r.uid for r in test])
test_items = np.unique([r.iid for r in test])
R_test = pd.DataFrame(np.zeros((len(test_users), len(test_items))), index=test_users, columns=test_items)
for r in test:
R_test.at[r.uid, r.iid] = 1 # 隐式反馈，这里就设置为1，忽略具体的评分数据
self.logger.debug(f"测试集评分矩阵:\n{R_test}")
self.users = users
self.items = items
self.R_train = R_train
self.R_test = R_test
return users, items, R_train, R_test
def calc_user_sim(self):
"""
2. 计算user相似度。直接使用sklearn的pairwise_distances函数来计算，速度大大提高
"""
W = 1 - pairwise_distances(self.R_train.to_numpy(), metric="cosine")
W = pd.DataFrame(W, index=self.users, columns=self.users)
self.logger.debug(f"user相似度矩阵：\n{W}")
self.W = W
return W
def rec(self):
"""
3. 推荐。计算推荐(兴趣)矩阵P，P[u][i]表示u对i的兴趣值(预测的u对i的评分)
"""
self.logger.info(f"开始推荐,K={self.K}")
P = pd.DataFrame(np.zeros((len(self.users), len(self.items))), index=self.users, columns=self.items)
for u in self.users:
K_Wu = self.W[u].nlargest(self.K + 1).iloc[1:]
for v, wuv in K_Wu.items():
Rv = self.R_train.loc[v]
vis = Rv[Rv != 0.0].index.values
for i in vis:
# 如果i已经在u的评分列表中，则跳过
if self.R_train.at[u, i] != 0.0:
continue
P.at[u, i] += self.W.at[u, v] * self.R_train.at[v, i]
self.logger.debug(f"预测矩阵：\n{P}")
self.P = P
return P
def evaluate(self):
"""
4. 评估效果
"""
Ru_Dict = {}
Tu_Dict = {}
for u in self.users:
# 为u推荐的N个item
N_Pu = self.P.loc[u].nlargest(self.N).index.values
# 测试集中u的items
try: # 测试集中可能没有训练集中全部的user
Tu = self.R_test.loc[u]
except KeyError:
continue
N_Tu = Tu[Tu != 0.0].index.values
Ru_Dict[u] = set(N_Pu)
Tu_Dict[u] = set(N_Tu)
precs = Evaluate.precision(Ru_Dict, Tu_Dict)
recl = Evaluate.recall(Ru_Dict, Tu_Dict)
covrg = Evaluate.coverage(Ru_Dict, len(self.items))
popu = Evaluate.popularity(self.R_train, Ru_Dict)
self.logger.info(
f"K={self.K},N={self.N},准确率{precs * 100:.2f}%"
f",召回率{recl * 100:.2f}%,覆盖率{covrg * 100:.2f}%,流行度{popu}")
return precs, recl, covrg, popu
def run(self):
"""
执行上面4个步骤,也可以在外部调用，分步调试
"""
self.load_data()
self.calc_user_sim()
self.rec()
self.evaluate()
