yelp数据集上识别潜在的热门商家

yelp数据集是研究B2C业态的一个很好的数据集，要识别潜在的热门商家是一个多维度的分析过程，涉及用户行为、商家特征和社区结构等多个因素。从yelp数据集里我们可以挖掘到下面信息有助于识别热门商家

用户评分和评论分析

评分均值: 商家的平均评分是反映其受欢迎程度的重要指标。较高的平均评分通常意味着顾客满意度高，从而可能成为热门商家。
评论数量: 评论数量可以反映商家的活跃度和用户的参与程度。评论数量多的商家更可能受到广泛关注。

用户活跃度

用户评分行为: 分析活跃用户（频繁评分的用户）对商家的评分，可以识别出哪些商家在用户群体中更受欢迎。
用户影响力: 一些用户的评分会对其他用户的选择产生较大影响（例如，社交媒体影响者）。识别这些高影响力用户对商家的评分可以帮助识别潜在热门商家。

社交网络分析

用户与商家的关系网络: 使用图神经网络等算法分析用户与商家之间的关系。商家与许多用户有互动，且用户在网络中有较高影响力的商家，可能会被视为热门商家。
社区发现: 通过分析用户和商家之间的关系网络，识别出相似用户群体，进而识别出在这些群体中受欢迎的商家。

多维度评价

综合评价: 结合多个指标（如评分、评论数、用户活跃度、地理位置等），使用加权方法或多指标决策模型来综合评估商家的受欢迎程度。

使用的文件

yelp_academic_dataset_business.json:
- 包含商家的基本信息，如商家 ID、名称、类别、位置等。
yelp_academic_dataset_review.json:
- 包含用户对商家的评论及评分，可以用来分析商家的受欢迎程度和用户的行为。
yelp_academic_dataset_user.json:
- 包含用户的基本信息，比如用户 ID、注册时间、评价数量等，可以用来分析用户的活跃度和影响力。

通过图神经网络（GNN）来识别商家的影响力：

先加载必要的库并读取数据文件：

import pandas as pd
import json

# 读取数据
with open('yelp_academic_dataset_business.json', 'r') as f:
    businesses = pd.DataFrame([json.loads(line) for line in f])

with open('yelp_academic_dataset_review.json', 'r') as f:
    reviews = pd.DataFrame([json.loads(line) for line in f])

with open('yelp_academic_dataset_user.json', 'r') as f:
    users = pd.DataFrame([json.loads(line) for line in f])

清洗数据以提取有用的信息：

# 过滤出需要的商家和用户数据
businesses = businesses[['business_id', 'name', 'categories', 'city', 'state', 'review_count', 'stars']]
reviews = reviews[['user_id', 'business_id', 'stars']]
users = users[['user_id', 'review_count', 'average_stars']]

# 处理类别数据
businesses['categories'] = businesses['categories'].str.split(', ').apply(lambda x: x[0] if x else None)

构建商家和用户之间的图，节点为商家和用户，边为用户对商家的评分。

    edges = []
    for _, row in reviews.iterrows():
        if row['user_id'] in node_mapping and row['business_id'] in node_mapping:
            edges.append([node_mapping[row['user_id']], node_mapping[row['business_id']]])

    edge_index = torch.tensor(edges, dtype=torch.long).t().contiguous()

    return node_mapping, edge_index, total_nodes

我们可以通过以下方式计算商家的影响力：

用户评分的平均值: 表示商家的受欢迎程度。
评论数: 提供商家影响力的直观指标。

business_reviews = reviews.groupby('business_id').agg({
    'stars': ['mean', 'count']
}).reset_index()
business_reviews.columns = ['business_id', 'average_rating', 'review_count']

# 合并商家信息和评论信息
merged_data = businesses.merge(business_reviews, on='business_id', how='left')

# 3. 目标变量定义
# 定义热门商家的标准
merged_data['is_popular'] = ((merged_data['average_rating'] > 4.0) &
                             (merged_data['review_count'] > 10)).astype(int)

使用 GNN 进一步分析商家的影响力，可以构建 GNN 模型并训练。以下是 GNN 模型的基本示例，使用 PyTorch Geometric：

class GNNModel(torch.nn.Module):
    def __init__(self, num_node_features):
        super(GNNModel, self).__init__()
        self.conv1 = GCNConv(num_node_features, 64)
        self.conv2 = GCNConv(64, 32)
        self.conv3 = GCNConv(32, 16)
        self.fc = torch.nn.Linear(16, 1)
        self.dropout = torch.nn.Dropout(0.3)

    def forward(self, x, edge_index):
        x = F.relu(self.conv1(x, edge_index))
        x = self.dropout(x)
        x = F.relu(self.conv2(x, edge_index))
        x = self.dropout(x)
        x = F.relu(self.conv3(x, edge_index))
        x = self.fc(x)
        return x

使用模型的输出嵌入来分析商家之间的相似度，识别潜在的热门商家。

print("Making predictions...")
    model.eval()
    with torch.no_grad():
        predictions = torch.sigmoid(model(data.x.to(device), data.edge_index.to(device))).cpu()

    # 将预测结果添加到数据框
    merged_data['predicted_popularity'] = 0.0
    for _, row in merged_data.iterrows():
        if row['business_id'] in node_mapping:
            idx = node_mapping[row['business_id']]
            merged_data.loc[row.name, 'predicted_popularity'] = predictions[idx].item()

    # 输出潜在热门商家
    potential_hot = merged_data[
        (merged_data['predicted_popularity'] > 0.5) &
        (merged_data['is_popular'] == 0)
        ].sort_values('predicted_popularity', ascending=False)

    print("\nPotential Hot Businesses:")
    print(potential_hot[['name', 'average_rating', 'review_count', 'predicted_popularity']].head())

使用上面定义流程跑一下训练, 报错了

Traceback (most recent call last):
File "/opt/miniconda3/envs/lora/lib/python3.10/site-packages/pandas/core/indexes/base.py", line 3805, in get_loc
return self._engine.get_loc(casted_key)
File "index.pyx", line 167, in pandas._libs.index.IndexEngine.get_loc
File "index.pyx", line 196, in pandas._libs.index.IndexEngine.get_loc
File "pandas/_libs/hashtable_class_helper.pxi", line 7081, in pandas._libs.hashtable.PyObjectHashTable.get_item
File "pandas/_libs/hashtable_class_helper.pxi", line 7089, in pandas._libs.hashtable.PyObjectHashTable.get_item
KeyError: 'review_count'

把print('merged_data', merged_data) 加上再试下

[150346 rows x 16 columns]
Index(['business_id', 'name', 'address', 'city', 'state', 'postal_code',
'latitude', 'longitude', 'stars', 'review_count_x', 'is_open',
'attributes', 'categories', 'hours', 'average_rating',
'review_count_y'],
dtype='object')

review_count 列被重命名为 review_count_x 和 review_count_y。这通常是因为在合并过程中，两个 DataFrame 中都存在 review_count 列。为了继续进行需要选择合适的列来作为评论数量的依据。选择 review_count_x 或 review_count_y: 通常，review_count_x 是从 businesses DataFrame 中来的，而 review_count_y 是从 business_reviews DataFrame 中来的。

代码修改下

import torch
import pandas as pd
import numpy as np
import torch.nn.functional as F
from torch_geometric.data import Data
from torch_geometric.nn import GCNConv
from sklearn.preprocessing import StandardScaler
from sklearn.model_selection import train_test_split


# 1. 数据加载
def load_data():
    businesses = pd.read_json('yelp_academic_dataset_business.json', lines=True)
    reviews = pd.read_json('yelp_academic_dataset_review.json', lines=True)
    users = pd.read_json('yelp_academic_dataset_user.json', lines=True)
    return businesses, reviews, users


# 2. 数据预处理
def preprocess_data(businesses, reviews):
    # 聚合评论数据
    business_reviews = reviews.groupby('business_id').agg({
        'stars': ['mean', 'count'],
        'useful': 'sum',
        'funny': 'sum',
        'cool': 'sum'
    }).reset_index()

    # 修复列名
    business_reviews.columns = ['business_id', 'average_rating', 'review_count',
                                'total_useful', 'total_funny', 'total_cool']

    # 合并商家信息
    # 删除businesses中的review_count列（如果存在）
    if 'review_count' in businesses.columns:
        businesses = businesses.drop('review_count', axis=1)

    # 合并商家信息
    merged_data = businesses.merge(business_reviews, on='business_id', how='left')

    # 填充缺失值
    merged_data = merged_data.fillna(0)

    return merged_data


# 3. 特征工程
def engineer_features(merged_data):
    # 确保使用正确的列名创建特征
    merged_data['engagement_score'] = (merged_data['total_useful'] +
                                       merged_data['total_funny'] +
                                       merged_data['total_cool']) / (merged_data['review_count'] + 1)  # 加1避免除零

    # 定义热门商家
    merged_data['is_popular'] = ((merged_data['average_rating'] >= 4.0) &
                                 (merged_data['review_count'] >= merged_data['review_count'].quantile(0.75))).astype(
        int)

    return merged_data


# 4. 图构建
def build_graph(merged_data, reviews):
    # 创建节点映射
    business_ids = merged_data['business_id'].unique()
    user_ids = reviews['user_id'].unique()

    # 修改索引映射，确保从0开始
    node_mapping = {user_id: i for i, user_id in enumerate(user_ids)}
    # 商家节点的索引接续用户节点的索引
    business_start_idx = len(user_ids)
    node_mapping.update({business_id: i + business_start_idx for i, business_id in enumerate(business_ids)})

    # 获取节点总数
    total_nodes = len(user_ids) + len(business_ids)

    # 创建边
    edges = []
    for _, row in reviews.iterrows():
        if row['user_id'] in node_mapping and row['business_id'] in node_mapping:
            edges.append([node_mapping[row['user_id']], node_mapping[row['business_id']]])

    edge_index = torch.tensor(edges, dtype=torch.long).t().contiguous()

    return node_mapping, edge_index, total_nodes


def prepare_node_features(merged_data, node_mapping, num_user_nodes, total_nodes):
    feature_cols = ['average_rating', 'review_count', 'engagement_score']

    # 确保所有特征列都是数值类型
    for col in feature_cols:
        merged_data[col] = merged_data[col].astype(float)

    # 标准化特征
    scaler = StandardScaler()
    merged_data[feature_cols] = scaler.fit_transform(merged_data[feature_cols])

    # 创建特征矩阵，使用总节点数
    num_features = len(feature_cols)
    x = torch.zeros(total_nodes, num_features, dtype=torch.float)

    # 用户节点特征（使用平均值）
    mean_values = merged_data[feature_cols].mean().values.astype(np.float32)
    x[:num_user_nodes] = torch.tensor(mean_values, dtype=torch.float)

    # 商家节点特征
    for _, row in merged_data.iterrows():
        if row['business_id'] in node_mapping:
            idx = node_mapping[row['business_id']]
            feature_values = row[feature_cols].values.astype(np.float32)
            if not np.isfinite(feature_values).all():
                print(f"警告: 发现无效值 {feature_values}")
                feature_values = np.nan_to_num(feature_values, 0)
            x[idx] = torch.tensor(feature_values, dtype=torch.float)

    return x


def main():
    print("Starting the program...")

    # 设置设备
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    print(f"Using device: {device}")

    # 加载数据
    print("Loading data...")
    businesses, reviews, users = load_data()

    # 预处理数据
    print("Preprocessing data...")
    merged_data = preprocess_data(businesses, reviews)
    merged_data = engineer_features(merged_data)

    # 构建图
    print("Building graph...")
    node_mapping, edge_index, total_nodes = build_graph(merged_data, reviews)
    num_user_nodes = len(reviews['user_id'].unique())

    # 打印节点信息
    print(f"Total nodes: {total_nodes}")
    print(f"User nodes: {num_user_nodes}")
    print(f"Business nodes: {total_nodes - num_user_nodes}")
    print(f"Max node index in mapping: {max(node_mapping.values())}")

    # 准备特征
    print("Preparing node features...")
    x = prepare_node_features(merged_data, node_mapping, num_user_nodes, total_nodes)

    # 准备标签
    print("Preparing labels...")
    labels = torch.zeros(total_nodes)
    business_mask = torch.zeros(total_nodes, dtype=torch.bool)

    for _, row in merged_data.iterrows():
        if row['business_id'] in node_mapping:
            idx = node_mapping[row['business_id']]
            labels[idx] = row['is_popular']
            business_mask[idx] = True

    # 创建图数据对象
    data = Data(x=x, edge_index=edge_index)

    # 初始化模型
    print("Initializing model...")
    model = GNNModel(num_node_features=x.size(1)).to(device)

    # 训练模型
    print("Training model...")
    train_model(model, data, labels, business_mask, device)

    # 预测
    print("Making predictions...")
    model.eval()
    with torch.no_grad():
        predictions = torch.sigmoid(model(data.x.to(device), data.edge_index.to(device))).cpu()

    # 将预测结果添加到数据框
    merged_data['predicted_popularity'] = 0.0
    for _, row in merged_data.iterrows():
        if row['business_id'] in node_mapping:
            idx = node_mapping[row['business_id']]
            merged_data.loc[row.name, 'predicted_popularity'] = predictions[idx].item()

    # 输出潜在热门商家
    potential_hot = merged_data[
        (merged_data['predicted_popularity'] > 0.5) &
        (merged_data['is_popular'] == 0)
        ].sort_values('predicted_popularity', ascending=False)

    print("\nPotential Hot Businesses:")
    print(potential_hot[['name', 'average_rating', 'review_count', 'predicted_popularity']].head())

# 6. GNN模型定义
class GNNModel(torch.nn.Module):
    def __init__(self, num_node_features):
        super(GNNModel, self).__init__()
        self.conv1 = GCNConv(num_node_features, 64)
        self.conv2 = GCNConv(64, 32)
        self.conv3 = GCNConv(32, 16)
        self.fc = torch.nn.Linear(16, 1)
        self.dropout = torch.nn.Dropout(0.3)

    def forward(self, x, edge_index):
        x = F.relu(self.conv1(x, edge_index))
        x = self.dropout(x)
        x = F.relu(self.conv2(x, edge_index))
        x = self.dropout(x)
        x = F.relu(self.conv3(x, edge_index))
        x = self.fc(x)
        return x


# 7. 训练函数
def train_model(model, data, labels, business_mask, device, epochs=100):
    optimizer = torch.optim.Adam(model.parameters(), lr=0.01, weight_decay=5e-4)
    criterion = torch.nn.BCEWithLogitsLoss()

    model.train()
    for epoch in range(epochs):
        optimizer.zero_grad()
        out = model(data.x.to(device), data.edge_index.to(device))
        loss = criterion(out[business_mask], labels[business_mask].unsqueeze(1).to(device))
        loss.backward()
        optimizer.step()
        print(f'Epoch [{epoch + 1}/{epochs}], Loss: {loss.item():.4f}')



if __name__ == "__main__":
    main()

开始正式训练，先按照epoch=100做迭代训练测试，loss向收敛方向滑动

识别出热门店铺

Potential Hot Businesses:
name average_rating review_count predicted_popularity
100024 Mother's Restaurant -0.154731 41.821089 0.999941
31033 Royal House 0.207003 40.953749 0.999933
113983 Pat's King of Steaks -0.361171 34.103369 0.999805
64541 Felix's Restaurant & Oyster Bar 0.389155 32.023360 0.999725
42331 Gumbo Shop 0.340872 31.517411 0.999701