电商数据建模

一、分析背景与目的

1.1 背景介绍

电商平台数据分析是最为典型的一个数据分析赛道，且电商数据分析有着比较成熟的数据分析模型，比如：人货场模型。此文中我将通过分析国内最大的电商平台——淘宝的用户行为，来巩固数据分析技能以及思维。通过分析用户行为，以此来实现精准营销，总结现有问题，获得业务增长

1.2 数据说明

该数据包含了20230523用户一天购物行为的所有数据，主要包括人货场三个维度信息。用户维度、商品维度、地区维度构成一个用户下单的行为事实表。

1.3数据分析流程

提出业务问题—确认粒度—数据处理和清洗—构建模型—数据可视化——根据数据可视化分析解决问题

业务问题：

1.如何提高品牌销售力度？

2.如何刺激地区市场消费潜力？

3.如何刺激用户消费？

4.如何减少产品成本？

确认粒度：

用户信息表、订单表、订单明细表

商品信息表、品牌信息表、一二三级分类信息表

省份信息表、地区信息表

核心算法代码分享如下：

import sys

import numpy as np
from torch.utils.data import DataLoader
from torch import nn
import torch.nn.functional as F
import torch
from sklearn.metrics import precision_score,recall_score,accuracy_score

import dataloader


class ALS_MLP (nn.Module):
    def __init__(self, n_users, n_items, dim):
        super(ALS_MLP, self).__init__()
        '''
        :param n_users: 用户数量
        :param n_items: 物品数量
        :param dim: 向量维度
        '''
        # 随机初始化用户的向量,
        self.users = nn.Embedding( n_users, dim, max_norm=1 )
        # 随机初始化物品的向量
        self.items = nn.Embedding( n_items, dim, max_norm=1 )

        #初始化用户向量的隐层
        self.u_hidden_layer1 = self.dense_layer(dim, dim // 2)
        self.u_hidden_layer2 = self.dense_layer(dim//2, dim // 4)

        #初始化物品向量的隐层
        self.i_hidden_layer1 = self.dense_layer(dim, dim // 2)
        self.i_hidden_layer2 = self.dense_layer(dim//2, dim // 4)

        self.sigmoid = nn.Sigmoid()

    def dense_layer(self,in_features,out_features):
        #每一个mlp单元包含一个线性层和激活层，当前代码中激活层采取Tanh双曲正切函数。
        return nn.Sequential(
            nn.Linear(in_features, out_features),
            nn.Tanh()
        )

    def forward(self, u, v, isTrain=True):
        '''
         :param u: 用户索引id shape:[batch_size]
         :param i: 用户索引id shape:[batch_size]
         :return: 用户向量与物品向量的内积 shape:[batch_size]
         '''
        u = self.users(u)
        v = self.items(v)

        u = self.u_hidden_layer1(u)
        u = self.u_hidden_layer2(u)

        v = self.i_hidden_layer1(v)
        v = self.i_hidden_layer2(v)

        #训练时采取dropout来防止过拟合
        if isTrain:
            u = F.dropout(u)
            v = F.dropout(v)

        uv = torch.sum( u*v, axis = 1)
        logit = self.sigmoid(uv*3)
        return logit

def doEva(net, d):
    d = torch.LongTensor(d)
    u, i, r = d[:, 0], d[:, 1], d[:, 2]
    with torch.no_grad():
        out = net(u,i,False)
    y_pred = np.array([1 if i >= 0.5 else 0 for i in out])
    y_true = r.detach().numpy()
    p = precision_score(y_true, y_pred)
    r = recall_score(y_true, y_pred)
    acc = accuracy_score(y_true,y_pred)
    return p,r,acc

def train( epochs = 10, batchSize = 1024, lr = 0.001, dim = 256, eva_per_epochs = 1):
    '''
    :param epochs: 迭代次数
    :param batchSize: 一批次的数量
    :param lr: 学习率
    :param dim: 用户物品向量的维度
    :param eva_per_epochs: 设定每几次进行一次验证
    '''
    #读取数据
    user_set, item_set, train_set, test_set = \
        dataloader.readRecData(test_ratio = 0.1)
    #初始化ALS模型
    net = ALS_MLP(len(user_set), len(item_set), dim)
    #定义优化器
    optimizer = torch.optim.AdamW( net.parameters(), lr = lr, weight_decay=0.2)
    #定义损失函数
    criterion = torch.nn.BCELoss()
    #开始迭代
    for e in range(epochs):
        all_lose = 0
        #每一批次地读取数据
        for u, i, r in DataLoader(train_set,batch_size = batchSize, shuffle = True):
            optimizer.zero_grad()
            r = torch.FloatTensor(r.detach().numpy())
            result = net(u,i)
            loss = criterion(result,r)
            all_lose += loss
            loss.backward()
            optimizer.step()
        print('epoch {}, avg_loss = {:.4f}'.format(e,all_lose/(len(train_set)//batchSize)))

        #评估模型
        if e % eva_per_epochs==0:
            p, r, acc = doEva(net, train_set)
            print('train: Precision {:.4f} | Recall {:.4f} | accuracy {:.4f}'.format(p, r, acc))
            p, r, acc = doEva(net, test_set)
            print('test: Precision {:.4f} | Recall {:.4f} | accuracy {:.4f}'.format(p, r, acc))

def als_mlp_predict(userId=1, itemSize=100, count=4, dim=64):
    # 读取数据
    user_set, item_set, train_set, test_set = \
        dataloader.readRecData( test_ratio=0.1)

    # 预测一个用户的所有的评分形成一个元祖
    train_set = []
    for i in range(1, itemSize):
        train_set.append((userId, i, 0))
    # print(train_set)
    # print(train_set)
    # 初始化ALS模型
    net = ALS_MLP(len(user_set), len(item_set), dim)

    d = torch.LongTensor(train_set)
    u, i, r = d[:, 0], d[:, 1], d[:, 2]
    with torch.no_grad():
        out = net(u, i)

    predict = []
    preds = out.tolist()
    # print(len(preds))
    # 找出最大值,通过这种方式可以求出多个
    for i in range(0, count):
        m = max(preds)
        idx = preds.index(m)
        predict.append(dict(iid=idx, score=m))
        del preds[idx]

    # print(predict)
    return predict

def test(dim = 64):
    result = als_mlp_predict(1, 2000, 5)
    print(result)

if __name__ == '__main__':
    # train()
    # test()
    param1 = sys.argv[1]
    # param1 = "1"
    result = als_mlp_predict(int(param1), 55, 4)
    list = []
    # print(result)
    for r in result:
        list.append(dict(iid=r['iid'], rate=r['score']))
    print(list)