品牌商家库分析数据

基于品牌、门店的名称、原子词、类目词等基础特征，确定商户与品牌的挂载关系，在品牌商户的精细化运营场景中，结合商户经营情况，圈选平台头部品牌，针对品牌的经营状况、菜品结构、营业时长等其他条件，给品牌商户输出诊断建议，在同一品牌的不同商户间进行经营策略的复制，辅助平台品牌商户成长。商家挂载品牌算法大体分为两步：召回与排序打分。对于召回部分：运用自然语言处理的NER命名实体识别算法分别对店铺名以及品牌名识别出原子词和类目词（原子词：即名称中的核心词，能够精准描述店铺和品牌的最短文本，如“全家便利店”的“全家”，“瑞幸咖啡”的“瑞幸”；类目词：即描述店铺或品牌的品类词，如“全家便利店”的“便利店”，“瑞幸咖啡”的“咖啡”）；通过原子词和类目词的文本相似度召回店铺的潜在挂载品牌。对于排序打分部分：使用深度学习二分类模型对召回部分的结果进行相似度打分，其中模型输入特征选取文本特征和类目信息特征（对于品牌类目，通过已挂载品牌的商户清洗得到top3类目）；然后将特征输入到基于bert+transformer的双塔二分类模型；选取模型阈值大于0.8，且匹配分数最高的品牌作为最终商户与品牌的挂载结果。

This dataset aims to determine the merchant-brand binding relationship based on basic features including brand names, store names, atomic words, category words, etc. In the refined operation scenario for brand merchants, it selects top-tier brands on the platform by combining merchants' business conditions, outputs diagnostic suggestions for brand merchants according to other conditions such as the brand's business status, dish structure and business hours, replicates business strategies across different merchants under the same brand, and assists the growth of platform brand merchants.

The merchant-brand binding algorithm is roughly divided into two stages: recall and ranking scoring.
For the recall stage: The NER (Named Entity Recognition) algorithm of natural language processing is used to identify atomic words and category words from store names and brand names respectively. (Atomic words refer to the core words in a name, which are the shortest text that can accurately describe a store and brand, such as "Quanjia" in "Quanjia Convenience Store", "Ruixing" in "Ruixing Coffee"; Category words refer to category terms describing a store or brand, such as "convenience store" in "Quanjia Convenience Store", "coffee" in "Ruixing Coffee"). Potential binding brands for the store are recalled through the text similarity between atomic words and category words.
For the ranking scoring stage: A deep learning binary classification model is used to conduct similarity scoring on the results from the recall stage. The input features of the model include text features and category information features (for brand categories, the top 3 categories are obtained by cleaning merchants with bound brands). Then the features are input into the twin-tower binary classification model based on BERT+Transformer. Finally, the brand with a model threshold greater than 0.8 and the highest matching score is selected as the final merchant-brand binding result.

搜集汇总

数据集介绍

特点

品牌商家库分析数据是一个包含1001条记录的企业数据集，每日更新，主要用于品牌商户的精细化运营。数据集包含门店和品牌的详细信息，如ID、名称、匹配分等，并通过自然语言处理和深度学习算法确定商户与品牌的挂载关系。