[人工智能]任务4 usercf

任务4

1. 推荐的指标的计算

下面实现了3中推荐指标
三种指标的具体定义见任务1

• 召回率
  $$\mathrm{Recall}=\frac{{\sum }_u|R(u)\cap T(u)|}{{\sum }_u|T(u)|}$$
• 准确率
  $$\mathrm{Precision}=\frac{{\sum }_u\mid R(u)\cap T(u)|}{{\sum }_u|R(u)|}$$
• 覆盖率
  $$\text{?Coverage?}=\frac{\left|{?}_{u\in U}R(u)\right|}{|I|}$$

1.1 召回率

def recall(y_pred, y_true):
   """计算召回率

   Args:
       y_pred (tensor): 每一行,都对应着给用户推荐的物品的序号(从0开始)
       y_true (tensor): 真实的测试集

   Returns:
       float: 所有用户的召回率的均值
   """
   y_pred = torch.scatter(input=torch.zeros_like(y_true), dim=1, index=y_pred, src=torch.ones_like(y_true))
   mask = (y_true > 0).float()
   common = (y_pred * mask).sum(dim=1)
   print(common)
   y_true_sum = mask.sum(dim=1)
   print(y_true_sum)
   return (common / y_true_sum).mean().item()

1.2 准确率

def precision(y_pred, y_true):
    """准确率

    Args:
        y_pred (tensor): 每一行都是给用户推荐的物品的序号(从0开始)
        y_true (tensor): 真实测试集

    Returns:
        float: 所有用户的准确率的均值
    """
    y_pred = torch.scatter(input=torch.zeros_like(y_true), dim=1, index=y_pred, src=torch.ones_like(y_true))
    mask = (y_true > 0).float()
    common = (y_pred * mask).sum(dim=1)
    print(common)
    y_pred_sum = y_pred.sum(dim=1)
    print(y_pred_sum)
    return common / y_pred_sum

1.3 覆盖率

def Coverage(y_pred, items):
    """覆盖率

    Args:
        y_pred (tensor): _description_
        items (int): 物品的数目

    Returns:
        float: 覆盖率
    """
    return len(y_pred.flatten().unique()) / items

2. 相关函数

2.1 LoadData

读取数据集

import numpy as np
import pandas as pd
import warnings
from tqdm import tqdm
from sklearn.model_selection import train_test_split
import torch
warnings.filterwarnings('ignore')


def loadData(filepath, name='ml-1m'):
    '''
    函数说明：
        读取打分文件的函数
    输入：
    filepath:
        string类型
        数据文件所在的路径，（注意最后必须是以\结尾的）
    name：
        string类型
        数据集的名称，目前只支持ml-1m和ml-100k
    输出：
    rating_data:
        pandas.DataFrame类型
        读取到的打分数据    
    '''
    print('开始读取数据')
    filepath = filepath + name
    if name == 'ml-1m':
        filepath = filepath + '/' + 'ratings.dat'
        rating_data = pd.read_csv(filepath, sep='::', names=['userid', 'movieid', 'rating', 'timestamp'])
    elif name == 'ml-100k':
        filepath = filepath + '/' + 'u.data'
        rating_data = pd.read_csv(filepath, sep='\t', names=['userid', 'movieid', 'rating', 'timestamp'])
    else:
        raise ValueError('输入的数据集不存在，可选的数据集只有ml-1m, ml-100k')
    
    user_num  = rating_data['userid'].max()
    movie_num = rating_data['movieid'].max()
    print(f'用户人数为: {user_num}')
    print(f'电影数目为: {movie_num}')
    
    return rating_data

2.2 rating2matrix

将评分的DataFrame转换成tensor的矩阵

def rating2matrix(ratings, shape=None):
    '''
    函数说明：
        将打分数据转换成一个打分矩阵, 需要注意userid和movieid在进行索引的时候需要减去1
    输入：
    ratings
        Pandas.DataFram
        movielens的打分数据，并且要求columns=['userid', 'movieid', 'rating', 'timestamp']
    输出：
    rating_matrix:
        numpy.ndarray
        输出的是打分矩阵
    '''
    print('开始将数据转换成矩阵')
    if shape is None:
        user_num  = ratings['userid'].max()
        movie_num = ratings['movieid'].max()
        shape = (user_num, movie_num)
    print(f'用户人数为: {shape[0]}')
    print(f'电影数目为: {shape[1]}')
    rating_matrix = np.zeros(shape=shape, dtype=np.float32)
    with tqdm(total=len(ratings)) as tq:
        for _, row in ratings.iterrows():
            row_id = int(row['userid']) - 1
            col_id = int(row['movieid']) - 1
            rating_matrix[row_id][col_id] = row['rating']
            tq.update()
    return rating_matrix

2.3 Split_Dataset_P

按照比例进行划分数据集(把每个用户的打分单独进行按照比例划分, 让后整合起来, 保证每个用户都能出现)

def Split_Dataset_P(rating_data, test_size=0.2, group_name='userid', random_state=42, shuffle=True):
    '''
    函数说明：
    把每个用户的打分数据单独挑出来，按照统一的比例进行划分数据集。
    这样可以保证每一个用户都出现在训练集当中，不会出现冷启动的问题

    输入：
    rating_data: pd.DataFrame类型的数据，表示输入的打分数据
    test_size  : float，测试集所占的比例
    userid : string，用户分组的标签名

    返回：
    train_ratings: pd.DataFrame，训练集数据
    test_ratings : pd.DataFrame, 测试集数据
    '''
    print('开始划分数据集')
    user_num = rating_data['userid'].max()
    train_ratings = pd.DataFrame(columns=rating_data.columns)
    test_ratings  = pd.DataFrame(columns=rating_data.columns)
    with tqdm(total=user_num) as tq:
        for i in rating_data.groupby(group_name):
            # print(i[1].values.shape)
            train_temp, test_temp, _, _ = train_test_split(i[1], i[1], test_size=test_size, shuffle=shuffle, random_state=random_state)
            train_ratings = train_ratings.append(train_temp)
            test_ratings = test_ratings.append(test_temp)
            tq.update()
    return train_ratings, test_ratings

3. 进行推荐

这个函数是将矩阵t按照行进行归一化

def normalize_e(t):
    return t / t.sum(dim=1).unsqueeze(dim=1)

利用用户打分的相似性对打分进行加权求和, 预测用户的打分.

def predict_score(scores, index, similarity):
    r1 = []
    similarity = normalize_e(similarly)
    for i,j in zip(index, similarity):
        # print(i)
        # print(j)
        selected_score = torch.index_select(input=scores, index=i, dim=0)
        # print(selected_score)
        final_score = torch.mm(j.unsqueeze(dim=0), selected_score)
        # print(final_score)
        r1.append(final_score)
    r1 = torch.cat(r1,dim=0)
    return r1

利用得到的预测的打分, 从高到低排列, 去除前k个

def recommend(score, k = 3):
    pred, index = torch.topk(input=score, k=k, dim=1)
    return pred, index

读取数据集

ratings = loadData('./data/', 'ml-1m')

开始读取数据
用户人数为: 6040
电影数目为: 3952

划分数据集

train_data, test_data = Split_Dataset_P(ratings)

开始划分数据集
100%|██████████| 6040/6040 [04:55<00:00, 20.44it/s]

把数据集转换成矩阵

train_data = rating2matrix(train_data)
# test_data = rating2matrix(test_data, shape=(943, 1682))
test_data = rating2matrix(test_data)

开始将数据转换成矩阵
用户人数为: 6040
电影数目为: 3952
100%|██████████| 797758/797758 [01:06<00:00, 12034.77it/s]
开始将数据转换成矩阵
用户人数为: 6040
电影数目为: 3952
100%|██████████| 202451/202451 [00:17<00:00, 11261.42it/s]

计算Jaccards相似性

train_data = torch.tensor(train_data)
simi = Similarity.SimJaccardS(train_data)

利用上面计算得到的相似性来预测打分

pred_score = predict_score(train_data, index=sim_user_index, similarity=sim_user_value)
pred_score.shape

torch.Size([6040, 3952])

把上面得到的打分进行过滤, 去掉用户之前看过的电影

mask = train_data > 0
mask = ~mask
filtered_score = mask * pred_score

利用之前预测得到的分数进行推荐

rec_value, rec_index = recommend(filtered_score, k=10)

对结果进行评估

recall_rate = Metirc.recall(rec_index, test_data)
recall_rate * 100

14.493964612483978

precision = Metirc.precision(rec_index, test_data)
print(precision)

0.3004801273345947

coverage = Metirc.Coverage(rec_index, 1682)
coverage

0.3055885850178359

4. itemCF

只需要将打分矩阵进行转置即可

5. 进阶

不使用矩阵乘法的代码我也实现了, 但是运行速度非常慢, 不推荐. 而且代码非常绕

# %%
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from sklearn.model_selection import train_test_split
import warnings
import os
from tqdm import tqdm
warnings.filterwarnings('ignore')

# %%
rnames = ['user_id','movie_id','rating','timestamp']
root_path = './data'
ratings = pd.read_csv(os.path.join(root_path, 'ratings.dat'), sep='::', engine='python', names=rnames)

# %%
user_nums = ratings['user_id'].max()
print(user_nums)
movie_nums = ratings['movie_id'].max()
print(movie_nums)

# %%
query_ratings = np.zeros(shape=(user_nums + 1, movie_nums + 1))

# %%
np.save('ratings.npy', query_ratings)

# %%
for i in tqdm(ratings.iterrows()):
    query_ratings[i[1]['user_id']][i[1]['movie_id']] = i[1]['rating']

# %%
ratings.head()

# %%


# %%
trn_data, val_data, _, _ = train_test_split(ratings, ratings, test_size=0.2, random_state=42)

# %%
trn_data.groupby('user_id')['movie_id'].apply(list).reset_index()

# %%
trn_data = trn_data.groupby('user_id')['movie_id'].apply(list).reset_index()
val_data = val_data.groupby('user_id')['movie_id'].apply(list).reset_index()

# %%
trn_user_items = {}
val_user_items = {}

# %%
for user, movies in zip(*(list(trn_data['user_id']), list(trn_data['movie_id']))):
    trn_user_items[user] = set(movies)
for user, movies in zip(*(list(val_data['user_id']), list(val_data['movie_id']))):
    val_user_items[user] = set(movies)

# %%


# %%
from tqdm import tqdm

# %%
item_users = {}
for uid, items in tqdm(trn_user_items.items()):
    for item in items:
        if item not in  item_users:
            item_users[item] = set()
        item_users[item].add(uid)

# %%
sim = {}
num = {}

# %%
for item, users in tqdm(item_users.items()):
    for u in users:
        if u not in num:
            num[u] = 0
        num[u] += 1
        if u not in sim:
            sim[u] = {}
        for v in users:
            if u != v:
                if v not in sim[u]:
                    sim[u][v] = 0
                sim[u][v] += 1

# %%
import pickle

# %%
with open('sim_dict.pkl', 'wb') as fw:
    pickle.dump(sim, fw)

# %%
with open('sim_dict.pkl', 'rb') as fr:
   sim = pickle.load(fr)
print(type(sim))

# %%
with open('num.pkl', 'rb') as fr:
    num = pickle.load(fr)
print(type(num))

# %%
with open('num.pkl', 'wb') as fw:
    pickle.dump(num, fw)

# %%
trn_data.to_csv('traindata.csv')

# %%
trn_data = pd.read_csv('traindata.csv')
trn_data.head()

# %%
val_data.to_csv('valdata.csv')

# %%
val_data = pd.read_csv('valdata.csv')
val_data.head()

# %%
import math

# %%
for u, users in tqdm(sim.items()):
    for v, score in users.items():
       sim[u][v] = score / math.sqrt(num[u] * num[v])

# %%
K = 80
N = 10
item_rank = {}
item_scores = {}
for u, _ in tqdm(val_user_items.items()):
    item_rank[u] = {}
    item_scores[u] = {}
    for v, score in sorted(sim[u].items(), key=lambda x: x[1], reverse=True)[:K]:# v是与u相似的用户，score是u和v的相似度
        for item in trn_user_items[v]: #item是与与用户u相似的用户v打过分的item
            if item not in trn_user_items[u]:
                if item not in item_rank[u]:
                    item_rank[u][item] = 0
                    item_scores[u][item] = 0  
                user_rating = query_ratings[v][item]
                item_rank[u][item] += score * user_rating
                item_scores[u][item] += score

# %%
normal_sim = {}
for u, _ in tqdm(val_user_items.items()):
    temp = 0
    for I, temp2 in sorted(sim[u].items(), key=lambda x: x[1], reverse=True)[:K]:
        temp += temp2
    normal_sim[u] = temp

# %%
K = 80
N = 10
item_rank = {}
for u, _ in tqdm(val_user_items.items()):
    item_rank[u] = {}
    for v, score in sorted(sim[u].items(), key=lambda x: x[1], reverse=True)[:K]:# v是与u相似的用户，score是u和v的相似度
        for item in trn_user_items[v]: #item是与与用户u相似的用户v打过分的item
            if item not in trn_user_items[u]:
                if item not in item_rank[u]:
                    item_rank[u][item] = 0
                user_rating = query_ratings[v][item]
                item_rank[u][item] += (score / normal_sim[u]) * user_rating

# %%
sorted(sim[1].items(), key=lambda x: x[1], reverse=True)[:10]

# %%
item_rank[1]

# %%


# %%


# %%


# %%


# %%
for u, _ in tqdm(item_rank.items()):
    for item, temp in item_rank[u].items():
        item_rank[u][item] = temp / item_scores[u][item]

# %%
item_rank[1]

# %%
item_scores[1][3712]

# %%


# %%


# %%


# %%
K = 80
N = 10
sl = []
item_rank = {}
for u, _ in tqdm(val_user_items.items()):
    item_rank[u] = {}
    for v, score in sorted(sim[u].items(), key=lambda x: x[1], reverse=True)[:K]:
        for item in trn_user_items[v]:
            if item not in trn_user_items[u]:
                if item not in item_rank[u]:
                    item_rank[u][item] = 0
                item_rank[u][item] += score
                sl.append(score)

    

# %%
for u, _ in tqdm(val_user_items.items()):
    for v, score in sorted(sim[u].items(), key=lambda x: x[1], reverse=True)[:K]:
        print(u, v)
        print(score)
        break
    break

# %%
item_rank2 = {k: sorted(v.items(), key=lambda x: x[1], reverse=True)[:N] for k, v in item_rank.items()}

# %%
item_rank2 = {k: set([x[0] for x in v]) for k, v in item_rank2.items()}

# %%
item_rank2

# %%
def Recall(Rec_dict, Val_dict):
    '''
    Rec_dict: 推荐算法返回的推荐列表, 形式:{uid: {item1, item2,...}, uid: {item1, item2,...}, ...} 
    Val_dict: 用户实际点击的商品列表, 形式:{uid: {item1, item2,...}, uid: {item1, item2,...}, ...}
    '''
    hit_items = 0
    all_items = 0
    for uid, items in Val_dict.items():
        rel_set = items
        rec_set = Rec_dict[uid]
        for item in rec_set:
            if item in rel_set:
                hit_items += 1
        all_items += len(rel_set)

    return round(hit_items / all_items * 100, 2)

# 推荐系统推荐正确的商品数量占给用户实际推荐的商品数
def Precision(Rec_dict, Val_dict):
    '''
    Rec_dict: 推荐算法返回的推荐列表, 形式:{uid: {item1, item2,...}, uid: {item1, item2,...}, ...} 
    Val_dict: 用户实际点击的商品列表, 形式:{uid: {item1, item2,...}, uid: {item1, item2,...}, ...}
    '''
    hit_items = 0
    all_items = 0
    for uid, items in Val_dict.items():
        rel_set = items
        rec_set = Rec_dict[uid]
        for item in rec_set:
            if item in rel_set:
                hit_items += 1
        all_items += len(rec_set)

    return round(hit_items / all_items * 100, 2)

# 所有被推荐的用户中,推荐的商品数量占这些用户实际被点击的商品数量
def Coverage(Rec_dict, Trn_dict):
    '''
    Rec_dict: 推荐算法返回的推荐列表, 形式:{uid: {item1, item2,...}, uid: {item1, item2,...}, ...} 
    Trn_dict: 训练集用户实际点击的商品列表, 形式:{uid: {item1, item2,...}, uid: {item1, item2,...}, ...}
    '''
    rec_items = set()
    all_items = set()
    for uid in Rec_dict:
        for item in Trn_dict[uid]:
            all_items.add(item)
        for item in Rec_dict[uid]:
            rec_items.add(item)
    return round(len(rec_items) / len(all_items) * 100, 2)

# 使用平均流行度度量新颖度,如果平均流行度很高(即推荐的商品比较热门),说明推荐的新颖度比较低
def Popularity(Rec_dict, Trn_dict):
    '''
    Rec_dict: 推荐算法返回的推荐列表, 形式:{uid: {item1, item2,...}, uid: {item1, item2,...}, ...} 
    Trn_dict: 训练集用户实际点击的商品列表, 形式:{uid: {item1, item2,...}, uid: {item1, item2,...}, ...}
    '''
    pop_items = {}
    for uid in Trn_dict:
        for item in Trn_dict[uid]:
            if item not in pop_items:
                pop_items[item] = 0
            pop_items[item] += 1
    
    pop, num = 0, 0
    for uid in Rec_dict:
        for item in Rec_dict[uid]:
            pop += math.log(pop_items[item] + 1) # 物品流行度分布满足长尾分布,取对数可以使得平均值更稳定
            num += 1  
    return round(pop / num, 3)

# 将几个评价指标指标函数一起调用
def rec_eval(val_rec_items, val_user_items, trn_user_items):
    print('recall:',Recall(val_rec_items, val_user_items))
    print('precision',Precision(val_rec_items, val_user_items))
    print('coverage',Coverage(val_rec_items, trn_user_items))
    print('Popularity',Popularity(val_rec_items, trn_user_items))

# %%
rec_eval(item_rank2, val_user_items, trn_user_items)

# %%
item_rank2[1]

# %%
item_rank2[1]

# %%
val_user_items[1]

# %%
sorted(item_rank[1].items(), key=lambda x: x[1], reverse=True)

# %%

创作打卡挑战赛

赢取流量/现金/CSDN周边激励大奖