Alibaba intends to make the customer base matrix a weather vane of the platform, allowing merchants to choose products with goals, levels and differences, and efficiently carry out scene operations and business operations, providing power for user growth and algorithm model optimization, and providing a foundation for digital operations. We mainly build around four dimensions: people, goods, market and business, and the customer base matrix is shown in figure 1.
The superposition of customer base matrix and scenario matrix can not only build the target users of scenarios and measure the differences of scenarios, but also improve the performance of scenarios and effectively guide the target traffic, thus providing the underlying data foundation for algorithmic modeling of various business scenarios.
Unlike C-type buyers, B-type buyers have no clear basic coordinate dimensions such as age and gender, and B-type users are mostly enterprises or wholesalers. How to describe the customer base matrix with B-type characteristics is very important for B-type e-commerce, and it is also a problem that B-type e-commerce "Xiaoer" has been thinking about.
Since B user groups are mainly enterprises and wholesalers, how to accurately describe the customer group matrix? Purchasing power is an outstanding performance, which includes purchase quantity and purchase frequency. From the purchasing power, we can see the user's business scale and consumption power. Therefore, we take purchasing power as the basic coordinate dimension and provide accurate and differentiated services at different levels.
The purchase amount is mainly the amount purchased by the user in a certain period of time. In order to avoid the layered interference caused by the large price difference between different categories, the purchase amount is first classified by category, and then viewed by amount file regardless of category, and the amount file with the largest proportion is the purchase amount file layer of the user.
Purchase frequency is the purchase frequency of users in a certain period of time. Rank users according to the purchase time, and then calculate the purchase frequency of users in a certain period of time. According to the proportion of Gaussian distribution, all users are divided into three grades: high, medium and low, which are used as the grading grades of purchase frequency.
Including new clothes, new users, low activity, medium activity, medium activity, high activity, deep sleep, loss and other stages, this life cycle is mainly divided according to the user's activity in the e-commerce platform, which also incorporates some business knowledge. For example, newly installed users refer to newly installed users, new users refer to users with less than 2 orders, and low-income users refer to users who visit less than 2 days a month.
From the transaction cycle, the user life cycle is analyzed, as shown in Figure 2, including new users, login users, first-time users, active buyers (buyers with high purchasing power and potential), sleeper buyers, sleeper buyers and other stages, and the transformation relationship between each life cycle stage is also intuitively presented in the figure. Accurate user operation adjusts the target according to different stages of the buyer's life cycle, and the adopted strategy will be adjusted accordingly.
Knowing the user's life cycle, we can do a good job of user innovation, promotion and retention, and improve the user's stickiness: for new installations and new users, it is mainly to improve their user experience, cultivate users' consumption habits, and carry out retention transformation; For middle and low-level users, it is mainly to promote and retain; For middle and high-level users, it is mainly to maintain user habits and strengthen stickiness; For users who have lost sleep, they mainly promote their lives through the rights and interests of red envelopes. The maintenance of user life cycle plays a vital role in the continuous growth of e-commerce users.
As a typical representative of B2B e-commerce platform, CBU has been committed to serving hundreds of millions of Class B buyers around the world. As one of the core attributes of Class B user portraits, the user's identity verification and main categories (such as the owner of imported maternal and child stores, the owner of boutique women's stores, the part-time WeChat merchants, the owner of small supermarkets, etc.). ) not only represents the offline entity identity of users, but also directly affects users' behavior preferences, procurement cycle and demands for merchant service capabilities on e-commerce platforms, so it has always been one of the core user portrait attributes of class B e-commerce platforms.
Most C-level user portraits can be modeled directly based on the historical behavior of users on the website, while B-level user portraits are different. In general, the B-type e-commerce platform needs to verify the user's identity, and the main categories require accuracy, so it mainly determines the user's verification identity in the form of self-filling forms. This user filling method has high accuracy, but the location is hidden, the link is lengthy, and there is no guidance of interest points. Not only the user filling rate is low, but also the combination with the scene is insufficient.
In order to solve the problem of high operating cost of the original form-based core users, Alibaba CBU e-commerce platform predicts the user core through the user core component borrowing algorithm model, and puts forward the Top K option for users to choose according to the confidence ranking. The overall algorithm solution is as follows.
The behavior of users in the station is the first feedback basis of users' needs and preferences, and it is also the data source that the algorithm needs to focus on. Compared with other preferred portrait attributes, user kernel is a relatively stable and long-term user attribute, so in the application of the algorithm, we select the global behavior of users in the last six months as the underlying data. There are two main considerations in choosing a long window of half a year: First, there are abundant and high-quality products on the website, the search and recommendation algorithms are increasingly refined, the cost for users to browse all kinds of products is low, it is difficult for B-class users to keep their attention on the website, the needs and behaviors of B-class /C-class users are mixed, and the data is dirty. A long time window is conducive to filtering out interference and capturing users' longer-term and stable needs; Second, the data of user behavior, especially purchase behavior, is relatively sparse. The purchase behavior of class B users is one of the core characteristics reflecting the user's identity, and the user's purchase behavior has certain periodicity, so a long time window is helpful for the algorithm to understand users more comprehensively.
Different from many preferred user portrait attributes, the user's identity can have a real mapping relationship with the user's identity in reality, such as the owner of milk tea shop-milk tea shop owner, baking shop owner-Baodao Jindian store owner, boutique women's clothing store owner-Taobao women's clothing store owner and so on. Therefore, the identity mapping relationship between upstream and downstream outside the user station can help us to further improve the prediction of user identity and improve coverage and accuracy.
In view of the mixed and noisy B /C behaviors of users on the website, the core preferences of users B are easily disturbed by popular categories and commodities on the website, so we also introduce a lot of industry knowledge as a guide to help predict the core identity of users B, and on this basis, precipitate a core preference category data.
Using the data of user's behavior inside the station, upstream and downstream identities outside the station and industry knowledge, the algorithm can realize the prediction of user's identity through the following steps, and the prediction process is shown in Figure 3.
Fig. 3 user kernel prediction flow chart
Seed users are mainly defined as users who have been authenticated in the station and users whose authentication information has upstream and downstream mapping relationship outside the station.
Based on the in-station behavior data of seed users in recent period, the salient features are mined and identified, which are provided to operation colleagues, and seed users are allocated for another round, so as to exclude users whose daily core behaviors are obviously inconsistent with industry preferences and optimize the selection of seed users.
Taking the industry preference category as the threshold, the commodities purchased by seed users in the last 6 months below the threshold are screened out as seed commodities.
Based on the I2I table of existing commodities stored by the team, seed commodities are used as trigger keys to expand the seed commodities, and the preference score of the expanded seed commodities is equal to the product of the I2I similarity score of the commodities and the preference score of the trigger seed commodities.
For a user's nuclear prediction, we select his behavior data in the last six months to model and score. Then, based on the scored user behavior goods, the user's preference confidence for each possible identity is calculated, which is used to distinguish the user's personal purchase behavior from the B-type purchase behavior, reduce the influence of the user's personal purchase behavior on the prediction result, and increase the weight of the user's B-type purchase behavior.
This article is taken from "Alibaba B2B E-commerce Algorithm in Actual Combat" and published with the authorization of the publisher.
This book is a summary of the experience of Alibaba CBU Technology Department (1688. com) 15 years of deep cultivation of B2B e-commerce. Alibaba B2B has experienced the upgrade iteration of information platform, trading platform and marketing platform in strategic form. This book focuses on the algorithm and technical ability behind the business form of marketing platform, trying to explain how technology empowers business from the perspective of mutual drive between technology and business, and combining the precipitation of Alibaba Group in basic field setting and algorithm innovation to create an intelligent B2B business operation system.
Recommended reading
User Portrait: Methodology and Engineering Solution
This is a book that explains how to build a user portrait system from 0 to 1 from the perspectives of technology, products and operation. At the same time, it also gives a solution for how to use the user portrait system to drive the income growth of enterprises.
Pay attention to "real-time stream calculation" and reply "08 14" in the background to participate in the lucky draw.
Draw prizes at noon on August 15 12:00. * * * Send two books and welcome everyone to participate ~