Sharing guest: Qi Jie, Ali senior algorithm expert
Article finishing: Ling Ming
Content source: DataFunTalk
Production platform: DataFun
Note: Welcome to reprint, please leave a message for reprinting.
Guide: In the process of online shopping on Tmall and Taobao, the front-end recommendations on the App are the ones that people come into contact with. However, the work and effort behind it may not be clear to everyone. Today, I would like to introduce to you the intelligent operation of the ToB scene behind the scenes of 1688.
The ToB industry cares about efficiency, function, and process, and the algorithm in the ToB operating mechanism can generally be summarized as: simplification of complex things, proceduralization of simple things, automation of procedural things, and intelligence of automated things. It is hoped that the machine can solve the simple problems of vertical labor, and can exert horizontal coordination and complex ability for people.
Today's introduction will focus on the following three points:
❶ Intelligent supply and demand matching
❷ Smart assistant
❸ Smart content operation
▌ background
The operation scenarios are very general and very large. Here are a few scenarios to introduce:
From left to right, as shown in the figure, they are ToC supermarket, Tmall e-commerce, Yiwu Small Commodity Wholesale Store, and finally 1688 Merchant's Day. It can be seen that the gameplay of the operation scene is diversified, and hundreds of millions of products need to be operated. Regarding the traditional industry operation model, can it handle thousands of products and see the overall operation situation? Obviously it is difficult to deal with it manually. At the same time, the cost of manual creation of product content copywriting operations is also very high. For each product, venue, and e-commerce activities that require manual editing and creation, the cost is unimaginable, so the face is very huge.
▌Smart supply and demand matching
Supply and demand matching is mainly divided into two parts, the supply side and the marketing side. Everyone is more concerned about the marketing side, how to sell products to users, then the supply side needs to find the sold products and provide them to users. The supply side involves which goods should be sold and which factories or merchants should be found.
先提出一个问题，该运营什么货呢？我们可以分成市场洞察和买家洞察。首先我们需要定义市场 ( 市场发现层 )，然后分析市场 ( 市场分析层 )，接着为评估市场 ( 市场评估层 )，最后赋能运营对市场进行决策 ( 市场运营层 )。这些都会涉及到子的模块包括指标预测、销量预测、定价能力、新兴市场的发现和不同市场的运营，这些问题都是小二运营需要面对的。
❶ 市场定义和生成
以饮料为例，可以划分为碳酸、水和茶饮料等。然后水中还会划分为包装水和纯净水，可以看到包装水和纯净水市场都是逐年递增的，所以我们认为水的市场非常好。
那么我们怎么定义市场呢？我们可以根据在线的商品挂载到市场上，CPV 市场作为承接还有一些类目、修饰和面料等属性。
市场数据中，我们会把对齐上的商品图谱化和用户输入 query 等属性挂载到 CPV 上，可以做特征聚类和标签聚合，最后生成一个市场的模型。对于未登陆的就可以用这模型预测属于哪个市场。
❷ 市场分析与评估
有了市场的定义，那么可对于市场做分析与评估。
如图根据消费者需求的维度进行分析，比如可看买家、成交、流量、供给等指标。我们可用 gbdt，W&D 模型进行预测。
有了这些数据指标，那么可根据波士顿矩阵去分析市场的类型 ( 成熟型、优质型、风险型和新兴型 )。对于不同的市场类型做不同运营的方案，比如新兴市场则可做促销和流量扶持的操作。
分析的指标则可以帮助小二做运营决策，比如选择羽绒服时，可看到销量的决策热度属性，中等厚薄在10月开始是卖的最好的。根据这些全面的数据指标，刚入门的小二也能变成类目运营的专家。
❸ 整体框架图
可从图中看到，我们通过智能供需匹配系统计算的结果，最终输出到运营平台中，例如补品任务发给商家，营销方案发给小二。可看到供需匹配的市场是一片蓝海。
▌智能助手
❶ 流程架构
如图为运营过程智能化的流程示意，包括：
洞察与精耕客户，挖掘多源数据
衔接市场需求，市场定义和分析数据、商家/商品图谱
打通商品和商家的成长路
协同策略举例：
经营策略，如何经营商品，增加营收
供货策略，洞察下游市场，结合商家供货能力，生成供货策略
权益策略，1688会员优惠价，或者流量分析工具
引擎触达方式：有商家工作台、钉钉、千机助手和阿里通信。
❷ 场景示例
下面我们来看一下千机助理，以上新品场景为例：
首先我们需要定义市场是需要哪些供货策略，那些需求单。再通过策略中心做智能分发。小二收到分发信息定向做 SOP 生成包括可解释的模型。
从整条链路看，可分为：
发现问题 -> 找出问题，通过大数据分析洞察商家上新能力及意愿。
问题归因 -> 发现解决方案，根据洞察结果将合适的策略分发给合适的运营，然后运营再交付给商家。
最后生成 SOP，通过可解释的模型把合适的因子拿出来生成一套 SOP。
▌智能内容运营
通过人机协作方式，降低人工创作的成本。内容运营分成三部分：
智能文案生成吸引人的描述和推荐理由
智能短标题生成千人千面的标题
智能标签解决买家的标签和卖家的标签不同域的问题，如卖家上架只写货品类目属性，而买家更关注货品使用场景、功效等
智能内容运营整体上线后，效果转化提升10%。
我们在2019年做了一个智能化风格文案 [ Automatic Generation of Pattern-controlled Product Description in E-commerce WWW19 ]，我们希望生成的文案能满足买家需求同时长度字数可控的。同一个标题输入不同风格类型则输出对应的文案，比如输入简约则生成 "连帽棉服，简约不单调的穿搭" 的文案。
The diversification of the model is achieved by coordinate encoder, and the deduplication is performed by attention based decoder, and style Classification is added to the loss functions.
▌ summary
The above shared three parts of intelligent supply and demand matching, intelligent assistants, and intelligent content operations. Our work also includes research on knowledge graphs (commodity knowledge graphs, venue knowledge graphs, factory knowledge graphs), human-machine collaboration interpretability, etc.
Sharing guests▬
Odd diameter
Alibaba | Advanced Algorithm Expert
——END——