def generate_matrix(with_rating=False):
"""
:param with_rating: 训练集是否包括rating，True则rating范围为1~5，否则为0或1（只对部分算法有效）
"""
# UserCF.user_similarity_cosine(train, iif=False) # with/without rating
# UserCF.user_similarity_cosine(train, iif=True) # with/without rating
# UserCF.user_similarity_jaccard(train, iif=False) # with/without rating
# UserCF.user_similarity_jaccard(train, iif=True) # with/without rating
# UserCF.user_similarity_pearson(train, iif=False) # with rating
# UserCF.user_similarity_pearson(train, iif=True) # with rating
# UserCF.user_similarity_adjusted_cosine(train, iif=False) # with rating
# UserCF.user_similarity_adjusted_cosine(train, iif=True) # with rating
# UserCF.user_similarity_log_likelihood(train) # without rating
ItemCF.item_similarity_cosine(train, norm=False, iuf=False, with_rating=with_rating) # with/without rating
def PrecisionAndRecallAndCoverageAndPopularity(train, test, item_popularity, K, W, N, method):
""" 计算 准确率，召回率，覆盖率，流行度
Desc:
Args:
train
test
item_popularity
K: 取相似度前K个
W: 相似度矩阵
N: TopN
method: recommend 推荐方法
Returns:
precision: 准确率
recall: 召回率
coverage: 覆盖率
popularity: 流行度
"""
hit = 0
num_rank = 0
num_tu = 0
recommend_items = set()
all_items = set()
popularity = 0.0
for user in train: # test / train
if user not in test:
continue
all_items = all_items | set(train[user][0])
tu = test[user][0]
if method == 1:
rank = ItemCF.Recommend1(user, train, W, K, N)
elif method == 2:
rank = ItemCF.Recommend2(user, train, W, K, N)
for item, _ in rank:
if item in tu:
hit += 1
popularity += math.log(1 + item_popularity[item])
recommend_items.add(item)
num_rank += len(rank)
num_tu += len(tu)
print(K, ': ', hit / (num_rank * 1.0), hit / (num_tu * 1.0), len(recommend_items) / (len(all_items) * 1.0), popularity / (num_rank * 1.0))
return hit / (num_rank * 1.0), hit / (
num_tu * 1.0), len(recommend_items) / (len(all_items) * 1.0), popularity / (num_rank * 1.0)
def PrecisionRecall(train, test, W, N, K): # N为推荐物品的TopN，K为和用户兴趣相似的K个用户
hit = 0
p_all = 0
r_all = 0
for user in train.keys():
tu = test.get(user, {}) # 用户user在测试集上喜欢的物品集合，是一个字典
# if tu: # 由于|R(u)|=N是人为设定要推荐的物品数，所以此处就没必要为tu为空的user做推荐计算了
rank_all = ItemCF.ItemCF_IUFNormRecommend(train, user, W,
K) # K为指定的相似物品数
rank_TopN = sorted(rank_all.items(), key=itemgetter(1),
reverse=True)[0:N]
# 推荐给用户user的TopN物品及对应的预测出来的user对该物品的兴趣度
# for item, pui in rank_all.items():
for item, pui in rank_TopN:
if item in tu:
hit += 1
p_all += N
# p_all += len(rank_all)
r_all += len(tu)
return hit / p_all, hit / r_all
def evaluate(train_set, test_set):
print('Evaluating start ...')
kg_sim = {}
kg_sim_detail = {}
for line in lf.load_file_all("./data/kg_sim1.txt"):
temp = line.split("::")
details = temp[1].replace("{", "").replace("}", "").split(",")
kg_sim.setdefault(int(temp[0]), {})
for items in details:
kv = items.split(":")
kg_sim_detail[int(kv[0])] = kv[1]
kg_sim[temp[0]] = kg_sim_detail
user_sim_matrix = gg.calc_user_sim(train_set)
movie_sim_matrix = ic.calc_movie_sim(train_set)
all_rec_movies = set()
for i in sim_mix:
sim_score = i[0]
sim_kg = i[1]
for j in range(loop_count):
# group generate
group_mem = gg.generateGroup(user_sim_matrix, train_set,
group_count, group_sim)
group_rank_kg = ic.kg_ic_predict(movie_sim_matrix, train_set,
group_mem, kg_sim, sim_score,
sim_kg)
g_rating_kg_lm = gsp.least_strategy(group_rank_kg, N)
precision_kg_lm, recall_kg_lm, sim_score, sim_kg = ev_detail_kg(
group_mem, test_set, g_rating_kg_lm, all_rec_movies, sim_score,
sim_kg)
g_rating_kg_avg = gsp.avg_strategy(group_rank_kg, N)
precision_kg_avg, recall_kg_avg, sim_score, sim_kg = ev_detail_kg(
group_mem, test_set, g_rating_kg_avg, all_rec_movies,
sim_score, sim_kg)
print("执行完毕！")
def PrecisionAndRecallAndCoverageAndPopularity(train, test, item_popularity, K, W, N, method=1):
hit = 0
num_rank = 0
num_tu = 0
recommend_items = set()
all_items = set()
popularity = 0.0
if method == 1:
for user in train: # test / train
if user not in test:
continue
all_items = all_items | set(train[user][0])
tu = test[user][0]
rank = UserCF.Recommend(user, train, W, K, N)
recommend_items = recommend_items | set(rank)
# hit += len(np.intersect1d(rank, tu))
for item in rank:
if item in tu:
hit += 1
popularity += math.log(1 + item_popularity[item])
# for item, value in rank.items():
# if item in tu:
# hit += 1
num_rank += len(rank)
num_tu += len(tu)
# print('Hit: ', hit)
# print('Rank num: ', num_rank)
# print('Test user\'s item num:', num_tu)
# print(len(all_items), len(recommend_items))
return hit / (num_rank * 1.0), hit / (
num_tu * 1.0), len(recommend_items) / (len(all_items) * 1.0), popularity / (num_rank * 1.0)
elif method == 2:
for user in train: # test / train
if user not in test:
continue
all_items = all_items | set(train[user][0])
tu = test[user][0]
rank = ItemCF.Recommend(user, train, W, K, N)
for item, _ in rank:
if item in tu:
hit += 1
popularity += math.log(1 + item_popularity[item])
recommend_items.add(item)
num_rank += len(rank)
num_tu += len(tu)
# print('Hit: ', hit)
# print('Rank num: ', num_rank)
# print('Test user\'s item num:', num_tu)
# print(len(all_items), len(recommend_items))
return hit / (num_rank * 1.0), hit / (
num_tu * 1.0), len(recommend_items) / (len(all_items) * 1.0), popularity / (num_rank * 1.0)
def main():
# 读取数据集
file_path = '~/file/rs/dataset/ml-1m/ratings.dat'
d_file = eva.readData(file_path, '::')
M = 8 # 分组数
N = 10 # 推荐个数
K = [5, 10, 20, 40, 80, 120, 160]
# seeds = np.arange(M)
seeds = [0]
columns_list = [
'PrecisionUserCF', 'PrecisionItemCF', 'RecallUserCF', 'RecallItemCF',
'CoverageUserCF', 'CoverageItemCF', 'PopularityUserCF',
'PopularityItemCF'
]
userCF_columns = [
'PrecisionUserCF', 'RecallUserCF', 'CoverageUserCF', 'PopularityUserCF'
]
itemCF_columns = [
'PrecisionItemCF', 'RecallItemCF', 'CoverageItemCF', 'PopularityItemCF'
]
d = pd.DataFrame(np.zeros([len(K), len(columns_list)]),
index=K,
columns=columns_list)
for index, seed in enumerate(seeds):
train, test = eva.SplitData(d_file, M, seed)
item_popularity = eva.ItemsPopularity(d_file, M, seed)
W_user = UserCF.UserSimilarityVersion3(train)
W_item = ItemCF.ItemSimilarityVersion2(train)
for k in K:
precision, recall, coverage, popularity = eva.PrecisionAndRecallAndCoverageAndPopularity(
train, test, item_popularity, k, W_user, N, 1)
d.loc[k,
userCF_columns] += [precision, recall, coverage, popularity]
precision, recall, coverage, popularity = eva.PrecisionAndRecallAndCoverageAndPopularity(
train, test, item_popularity, k, W_item, N, 2)
d.loc[k,
itemCF_columns] += [precision, recall, coverage, popularity]
d.loc[k] /= (index + 1)
d.to_excel('Result-UserCF-ItemCF-K.xlsx', 'UserCF-ItemCF-K')
fig, axes = plt.subplots(2, 2)
axes[0][0].set_title('Precision')
d.iloc[:, 0:2].plot(ax=axes[0][0], style=['o-', 'o-'])
axes[0][1].set_title('Recall')
d.iloc[:, 2:4].plot(ax=axes[0][1], style=['o-', 'o-'])
axes[1][0].set_title('Coverage')
d.iloc[:, 4:6].plot(ax=axes[1][0], style=['o-', 'o-'])
axes[1][1].set_title('Popularity')
d.iloc[:, 6:8].plot(ax=axes[1][1], style=['o-', 'o-'])
plt.show()
def TestItemCF():
""" 测试 ItemCF 算法
Desc:
Args:
Returns:
"""
start = time.time()
path = '~/file/rs/dataset/ml-1m/ratingsTest.dat'
d_file = pd.read_csv(path, sep='::', usecols=[0, 1, 2])
M = 8 # 分组数
N = 10 # 推荐个数
K = [5, 10, 20, 40, 80, 120, 160]
train, test, train_items_list, item_popularity, items_user = SplitData(d_file, M, 1) # 0: seed
w_item = ItemCF.ItemSimilarityVersion1(train, item_popularity, items_user)
columns_list = [
'precision', 'recall', 'coverage', 'popularity'
]
d = pd.DataFrame(
np.zeros([len(K), len(columns_list)]), index=K, columns=columns_list)
p = Pool(4)
result = dict()
for k in K:
result[k] = p.apply_async(Evaluation, args=(k, train, test, item_popularity, w_item, N, 2))
p.close()
p.join() # 等待所有子进程执行完毕
for k, v in result.items():
d.loc[k, columns_list] += v.get()
end = time.time()
print('total time: %.2fs' % (end - start))
d.to_excel('Result-ItemCF-K.xlsx', 'ItemCF-K')
fig, axes = plt.subplots(2, 2)
axes[0][0].set_title('Precision')
axes[0][0].plot(d.iloc[:, 0], 'o-', label='precision')
axes[0][1].set_title('Recall')
axes[0][1].plot(d.iloc[:, 1], 'o-', label='recall')
axes[1][0].set_title('Coverage')
axes[1][0].plot(d.iloc[:, 2], 'o-', label='coverage')
axes[1][1].set_title('Popularity')
axes[1][1].plot(d.iloc[:, 3], 'o-', label='popularity')
plt.legend()
plt.show()
def Coverge(train, W, N, K):
recommend_items = set()
all_items = set()
for user in train.keys():
for item in train[user].keys():
all_items.add(item)
rank_all = ItemCF.ItemCF_IUFNormRecommend(train, user, W, K)
rank_TopN = sorted(rank_all.items(), key=itemgetter(1),
reverse=True)[0:N]
# for item, pui in rank_all.items():
for item, pui in rank_TopN:
recommend_items.add(item)
return len(recommend_items) / len(all_items)
def Novelty(train, W, N, K):
item_popularity = PopularityNums(train)
ret = 0
n = 0 # n=用户数*TopN推荐列表中物品数N（推荐的物品总数）
for user in train.keys():
rank_all = ItemCF.ItemCF_IUFNormRecommend(train, user, W, K)
rank_TopN = sorted(rank_all.items(), key=itemgetter(1),
reverse=True)[0:N]
# for item, pui in rank_all.items():
for item, pui in rank_TopN:
ret += math.log(1 + item_popularity[item])
n += 1
return ret / n
def main(arg, f):
# 数据集按均匀分布划分为M份，M-1份均为训练集，剩下1份为测试集
M = 8
k = 0
seed = 42 # 随机数种子
data = [
tuple(line.split('::')[:2]) for line in open(
'G:/master/python/PycharmProjects/RecommendationSystem/ItemCF/MovieLens/data/ml-1m/ratings.dat'
).readlines()
] # win10上的ml-1m数据集
# data = [tuple(line.split(',')[:2]) for line in open('G:/Recommend/User-CF/MovieLens/ml-latest-small/ratings_test.csv').readlines()] # ml-latest-small数据集
train, test = SplitData(data, M, k, seed)
# 基于训练集计算物品相似度
W = ItemCF.ItemSimilarity(train)
# 离线指标计算
precision, recall = PrecisionRecall(train, test, W, arg[1], arg[0])
f.write(str(precision))
f.write(',')
f.write(str(recall))
f.write(',')
coverage = Coverge(train, W, arg[1], arg[0])
f.write(str(coverage))
f.write(',')
novelty = Novelty(train, W, arg[1], arg[0])
f.write(str(novelty))
f.write(',')
F1 = 2 * precision * recall / (precision + recall)
f.write(str(F1))
f.write(',')
# print(f'precision:{precision}\trecall:{recall}\tcoverage:{coverage}\tpopularity:{novelty}\tF1:{F1}')
'''
def get_recommendation_with_rating(user):
# return UserCF.recommend_with_rating(user, train)
return ItemCF.recommend_with_rating(user, train)
def get_recommendation(user):
# return UserCF.recommend(user, train, _n, _user_k)
return ItemCF.recommend(user, train, _n, _item_k)
def test_recommend():
""" 对比原版推荐算法和改进后的推荐算法
Desc:
Args:
Returns:
"""
file_path = '~/file/rs/dataset/ml-1m/ratings.dat'
d_file = eva.readData(file_path, '::')
M = 8 # 分组数
N = 10 # 推荐个数
K = [5, 10, 20, 40, 80, 120, 160]
# K = [5, 10, 20, 40]
train, test, train_items_list, item_popularity, items_user = SplitData(d_file, M, 0) # 0: seed
# W = ItemCF.ItemSimilarityVersion1(train, item_popularity, items_user)
W = ItemCF.ItemSimilarityVersion2(train, item_popularity, items_user)
columns_list = [
'Precision-I', 'Precision-II', 'Recall-I', 'Recall-II',
'Coverage-I', 'Coverage-II', 'Popularity-I',
'Popularity-II'
]
I_columns = [
'Precision-I', 'Recall-I', 'Coverage-I', 'Popularity-I'
]
II_columns = [
'Precision-II', 'Recall-II', 'Coverage-II', 'Popularity-II'
]
d = pd.DataFrame(
np.zeros([len(K), len(columns_list)]), index=K, columns=columns_list)
# ItemCF
p = Pool(4)
resultI = dict()
resultII = dict()
for k in K:
resultI[k] = p.apply_async(Evaluation, args=(k, train, test, item_popularity, W, N, 1))
resultII[k] = p.apply_async(Evaluation, args=(k, train, test, item_popularity, W, N, 2))
p.close()
p.join() # 等待所有子进程执行完毕
for k, v in resultI.items():
d.loc[k, I_columns] += v.get()
for k, v in resultII.items():
d.loc[k, II_columns] += v.get()
d.to_excel('Result-ItemCF-Recommend-K.xlsx', 'ItemCF-K')
fig, axes = plt.subplots(2, 2)
axes[0][0].set_title('Precision')
d.iloc[:, 0:2].plot(ax=axes[0][0], style=['o-', 'o-'])
axes[0][1].set_title('Recall')
d.iloc[:, 2:4].plot(ax=axes[0][1], style=['o-', 'o-'])
axes[1][0].set_title('Coverage')
d.iloc[:, 4:6].plot(ax=axes[1][0], style=['o-', 'o-'])
axes[1][1].set_title('Popularity')
d.iloc[:, 6:8].plot(ax=axes[1][1], style=['o-', 'o-'])
plt.legend()
plt.show()
def TestItemCF_Norm():
""" 对比 ItemCF 和 ItemCF-Norm
Desc:
Args:
Returns:
"""
file_path = '~/file/rs/dataset/ml-1m/ratings.dat'
start = time.time()
d_file = eva.readData(file_path, '::')
M = 8 # 分组数
N = 10 # 推荐个数
K = [5, 10, 20, 40, 80, 120, 160]
train, test, train_items_list, item_popularity, items_user = SplitData(d_file, M, 0) # 0: seed
W_ItemCF = ItemCF.ItemSimilarityVersion1(train, item_popularity, items_user)
W_Norm = ItemCF.ItemSimilarityNorm(train, item_popularity, items_user)
columns_list = [
'Precision-ItemCF', 'Precision-Norm', 'Recall-ItemCF', 'Recall-Norm',
'Coverage-ItemCF', 'Coverage-Norm', 'Popularity-ItemCF',
'Popularity-Norm'
]
I_columns = [
'Precision-ItemCF', 'Recall-ItemCF', 'Coverage-ItemCF', 'Popularity-ItemCF'
]
II_columns = [
'Precision-Norm', 'Recall-Norm', 'Coverage-Norm', 'Popularity-Norm'
]
d = pd.DataFrame(
np.zeros([len(K), len(columns_list)]), index=K, columns=columns_list)
# ItemCF
p = Pool(4)
resultItemCF = dict()
resultNorm = dict()
for k in K:
resultItemCF[k] = p.apply_async(Evaluation, args=(k, train, test, item_popularity, W_ItemCF, N))
resultNorm[k] = p.apply_async(Evaluation, args=(k, train, test, item_popularity, W_Norm, N))
p.close()
p.join() # 等待所有子进程执行完毕
for k, v in resultItemCF.items():
d.loc[k, I_columns] += v.get()
for k, v in resultNorm.items():
d.loc[k, II_columns] += v.get()
end = time.time()
print('total time: %.2fs' % (end - start))
d.to_excel('Result-ItemCF-Norm-K.xlsx', 'ItemCF-K')
fig, axes = plt.subplots(2, 2)
axes[0][0].set_title('Precision')
d.iloc[:, 0:2].plot(ax=axes[0][0], style=['o-', 'o-'])
axes[0][1].set_title('Recall')
d.iloc[:, 2:4].plot(ax=axes[0][1], style=['o-', 'o-'])
axes[1][0].set_title('Coverage')
d.iloc[:, 4:6].plot(ax=axes[1][0], style=['o-', 'o-'])
axes[1][1].set_title('Popularity')
d.iloc[:, 6:8].plot(ax=axes[1][1], style=['o-', 'o-'])
plt.legend()
plt.show()
