为什么要做分层
-
去平均化
-
精细化运营
-
举例：
- 用户运营
- 库存管理
Inventory management is one of the key elements of supply chain management for any organization. Controlling inventory helps integrate upstream processes such as manufacturing and purchasing with downstream activities such as demand and sales, thereby decreasing stock-out incidents and inevitably increasing customer satisfaction.
In today’s competitive market, companies are increasing their portfolio of SKUs to attract more customers, increase sales, and capture more market share, by offering buyers more choice. However, the large portfolio size also affects their inventory, since adding a greater number of items to the offering leads to companies holding obsolete or slow-moving inventory.
方法
分类维度+分层标准
商品常见的分类维度：
- 生命周期（新品引入、发展成长、稳定成熟、衰退淘汰）
- 毛利结构：高毛利、普通毛利商品
- 基于商品综合得分的分层分级
平台流量分配基础规则是基于商品综合得分，尤其是前面我们介绍的召回排序规则，总体原则是越受用户欢迎。商品得分来源于：
- 品牌得分；
- 人气得分；
- 商品主数据得分；
- 服务得分；
- 商品所在的店铺得分。
通过得分将商品分为：畅销品、普通品和滞销品，通过流量向畅销品分配完成规模，通过普通品的提升促进竞争，淘汰滞销品改善流量分发效率。对应商品生命周期，畅销品是商品进入成长后期和成熟期努力进化的方向。
与标签、分类classification：
-
描述标 vs 效果标
-
分类 vs 排序
举例 – 1688
实践 – Heyday
Transforming eCommerce Product Segmentation with Machine Learning
背景
Heyday is an aggregator of brands that are native to digital marketplaces such as Amazon FBA and Shopify. It acquires and incubates small but successful deliver-to-consumer (D2C) brands and helps them grow organically by providing them with capabilities such as brand management, analytics, marketing, and operations.
方案调研
1. 2-8
问题：维度单一，商品必须同质 heterogeneous
2. Multi-Criteria Inventory Classification
2.1 Optimization of inventory classification（线性 加权 ）
These models involve subjective assignment of weights to the different criteria done manually by a decision maker and then optimization is performed using those assigned weights.
问题：
they require a lot of effort from decision makers,
fail to include categorical data points and therefore miss out on many of the characteristics relevant in inventory management.
2.2 Multi-criteria decision making （打分表）
AHP uses both qualitative and quantitative criteria for classification and requires subjective judgement to rank relative importance of elements into numeric value.
问题：
These methods, while efficient,require management to spend a huge amount of time on developing information about each SKU
2.3 Machine learning
The k-nearest neighbor approach requires manually choosing appropriate distance metrics between data samples, which can be tedious for a dataset with large set of features
Both SVM and ANN can be employed to find a non-linear function approximation to estimate the class of the input inventory, which is significantly more powerful than heuristic or linear methods
2.4 Hybrid models using multi-criteria decision making and machine learning****
The authors initially conduct the ABC analysis using three different multi-criteria decision-making methods, simple additive weighting (SAW), Analytic Hierarchy Process (AHP), and VIKOR (Opricovic, 1998).
The curated ABC classes using these methods are then used as the target labels for machine learning methods such as Naïve Bayes, Bayesian networks, support vector machine, and artificial neural networks (ANN). They show that support vector machines and ANN outperform all other techniques.
3. Inventory Segmentation in eCommerce and CPG industries 电商增加了难度
E-commerce has raised consumers’ expectations and supply chains are becoming more reliant on effective inventory management strategies to avoid impacting customer satisfaction
Velocity, volatility, and profit margin factors to determine and improve customer service levels
ARIMA gmv 预测
4. Conclusion
具体模型
-
特征
-
Profit Margin
-
Unit Price
-
Demand: forecasted demand calculated in number of units
-
Demand Fluctuation: coefficient of variation (stddev/avg) of demand
-
Inventory Turnover
-
Priority:
-
The final attribute was a special request by Heyday, which we named Brand Priority. This is a subjective ranging from 1-5 that Heyday can use to elevate the priority of any SKU due to a management decision at any time.
-
-
-
Analytic Hierarchy Process (AHP)
- 问卷：两两特征对比：
- 根据问卷产出：特征权重，l1 归一
-
对每个商品，特征 min max norm，线性加权，2-8排序
-
结果：
- 对比单维度：
-
Running Machine Learning Models
- 特征：同上，min max norm
- 样本：AHP 结果
- 预测目标：新 sku
- 模型：svm，ann，超参，oversample A B，k-fold cross-validation
实践 – 腾讯lookalike
背景：
问题：马太效应，冷启动
While the head contents get more and more popular, many competitive long-tail contents are difficult to achieve timely exposure because of lacking behavior features.
解决方案：lookalike 扩人群
Because of less features for long-tail contents in recommendation campaign, the look-alike model is a good choice which only depends on the seed users as inputs instead of specific features of contents
启发：
- 从 User – item 排序，转为 User – User – item：对于非联盟商品，是否可以通过这个商品用户达人的关联关系，建设联盟相关特征
- 种子用户聚类：增加表征能力
- attention：关注个别弱特征减少马太效应
模型
特点：
-
Improve the effectiveness of user representation learning
-
Improve the robustness and adaptivity of seeds representation learning
-
Realize a real-time and high-performance look-alike model （实时+准实时异步+离线）
pipeline ：
模型：
User embedding learning
- 模型：Youtube DNN，ctr
- 样本：负样本采样
- 优化：concatenation 改为 attention
Look-alike Learning：
-
模型，双塔 cosin
-
优化：avg pooling 改 attention
- 优化：seed 用户 avg pool 改成聚类
实验
- Baseline 对比，avg 不如 dnn
- Attention 优于 concat
- Seed 聚簇，k 越大效果越好，结合性能k=20
- 线上多样性有改善
Reference
Transforming eCommerce Product Segmentation with Machine Learning.pdf
