Master the user portrait knowledge system from 0 to 1

I took a user portrait course some time ago. The instructor was Mr. Zhao Hongtian, the author of "User Portrait: Methodology and Engineering Solutions". I also read some articles about user portrait.

Based on the understanding of the above learning content and combined with work practice, this article will share with you the cognition, construction methods, productization and application of user portraits.

1. Getting to know the user portrait

1.1 User Profile

As all user behavior data can be tracked by enterprises, enterprises are increasingly focusing on how to use big data for business analysis and precision marketing services. In order to carry out refined operations, the first thing to do is to establish a user profile for the enterprise.

When it comes to the concept of user portrait, we distinguish between user role (Persona) and user portrait (Profile):

1.1.1 User Role

User personas are essentially a communication tool. When we discuss products, requirements, scenarios, and user experience, user personas come into being to avoid differences in understanding target users. User roles are based on a deep understanding of real users and a summary of highly accurate relevant data. They are fictional characters that contain typical user characteristics. The following is a typical user role:

1.1.2 User Profile

User portraits are more commonly used by operations and data analysts. Precision marketing, business analysis, and personalized recommendations are all applications based on user portraits. User portrait is a collection of variables that describe user data and can accurately describe any real user. The following is a simplified user portrait:

{ "ID": 123456,

“Name”: “Zhang Jianguo”,

“Gender”: “Male”,

"Date of Birth": 631123200,

“Place of Origin”: “Beijing”,

“Place of residence”: “Beijing”,

Educational Background:

{ "school": "Peking University",

“Major”: “CS”,

"Entrance Year":1220198400

}

}

1.2 User Tags and User Portraits

1.2.1 User Tags

User tags, which are descriptions of a certain dimension of user attributes, are independent and exhaustively enumerable. After collecting business, log, and tracking data, different statistical methods are used to calculate dimensional labels such as user attributes, user behavior, user consumption, risk control, and social interaction. For example: gender, age, number of visits in the past 30 days, purchase level, frequently active time periods, etc. For a detailed description of the construction of the user tag system, see the "2 Construction of tags and tag system" section.

1.2.2 User Profile

Building a user portrait means labeling users in various dimensions. In terms of business value, tags and portraits are system modules similar to the middle layer, which lay the foundation for data-driven operations. They can help big data "walk out" of the data warehouse and provide users with personalized recommendations, precision marketing and other diversified services. For detailed descriptions of the user portrait system and its practical applications, see the chapters “3 User Portrait Productization”, “4 User Portrait Application” and “5 User Portrait Practice Case Studies”.

1.3 User Groups and User Tags

User tags and user groups are two confusing concepts that are easy to confuse. Let's try to distinguish them:

1.3.1 User Groups

A combination of user attributes and behaviors is needed to select a comprehensive target group. With only behavioral data, we can only see what a person has done, but we don’t know whether this person is a man or a woman, how old he or she is, how long he or she has been registered, what his or her purchasing power is, etc. The user group selected in this way is flawed and is generally not directly used in precision marketing scenarios.

1.3.2 User Tags

When creating user tags, you don’t have to combine user attributes and behavioral events. You can use user attributes alone or behavioral events alone. User labels calculated based on user attributes and behavioral events are essentially user attributes, or user attributes themselves are labels.

1.3.3 Groups are a form of label application

As a middle-layer system module, tags are often not pushed with only one tag in precision marketing scenarios. In most cases, multiple tags need to be combined to meet the business definition of the crowd, as shown in the following figure:

Here we use a scenario to introduce an application that selects user groups based on user tags. During a women's clothing promotion event, channel operators need to screen out high-quality users on the platform and conduct marketing through channels such as SMS, email, and Push.

Step 1: Filter out potential users who may be interested in the women's clothing by circling the labels of "Browse", "Favorite", "Add to Cart", "Buy", and "Search" related to the women's clothing category

Step 2: Combine other tags (such as "gender", "consumption amount", "activity", etc.) to filter out the corresponding high-quality user groups and push them to the corresponding channels. Therefore, after abstracting user attributes and behavioral event data into labels, the target potential user group can be found by combining labels. From this perspective, user groups are a way of applying user tags.

2. Building labels and labeling systems

2.1 Classification of tags

There are many ways to classify labels themselves, but from the perspective of label implementation rules, they can be roughly divided into the following three types: ① statistical labels, ② rule labels, and ③ machine learning mining labels.

2.1.1 Statistical Tags

It is the most basic and common tag type. For example, for a certain user, fields such as gender, age, city, zodiac sign, active time in the past 7 days, active days in the past 7 days, and active times in the past 7 days can be statistically obtained from user registration data, user access, and consumption data. This type of label forms the basis of user portraits.

2.1.2 Rule-based tags

This type of tag is generated based on user behavior, user attributes and certain rules. For example, the definition of "active consumer" users on the platform is "number of transactions ≥ 2 in the past 30 days." In the actual process of portrait development, since operations personnel are more familiar with the business, while data personnel are more familiar with the structure, distribution, and characteristics of the data, the rules of rule-based labels are determined by operations personnel and data personnel through consultation.

2.1.3 Machine Learning to Mining Class Labels

This type of label is generated through machine learning mining and is used to predict and judge certain attributes or behaviors of users. For example, judging whether a user is male or female based on his/her behavioral habits, and judging the user's preference for a product based on his/her consumption habits. This type of label needs to be generated through algorithm mining.

In project engineering practice, general statistical and rule-based tags are sufficient to meet application requirements and account for a large proportion in development. Machine learning mining labels are mostly used in prediction scenarios, such as determining user gender, user purchase preferences, user churn intention, etc. Generally, the development cycle of machine learning labels is long and the development cost is high, so its development proportion is relatively small.

In fact, the final labeling system is defined from the user's perspective and needs to be combined with specific business. For example, the label classification of an e-commerce business includes user attribute dimension labels, user behavior dimension labels, user consumption dimension labels, risk control dimension labels, and social attribute dimension labels.

2.2 Label construction process

The following figure is a label construction process, which focuses on the product manager's perspective, mainly describing the analysis process and output documents of requirements, and briefly summarizing the development principles of labels.

2.2.1 Requirements Collection and Analysis

In the demand collection and analysis phase, you can proceed according to the steps of restoring the business process - clarifying the business purpose - deriving labels from the strategy - aggregating labels.

A clothing retailer expanded its business by setting up an online shopping mall and offline physical stores. Online, it is mainly through WeChat official accounts that traffic is directed to the mini program, and then the transaction is completed in the mini program. The following is a detailed description of how to collect and analyze label requirements through the clothing retail case:

1. Identify and analyze business processes and business scenario touchpoints

User portraits are based on business, so the first step in building labels is to identify and analyze users' decision-making processes and business scenarios in order to quickly become familiar with the business. Refer to the restoration of the business process of the case below: First, WeChat users who are attracted through various scenarios follow the official account and become fans. Then the official account operator will push picture and text messages to WeChat fans to conduct fan operations, and at the same time guide fans to the mini program mall. The official account fans will eventually convert transactions in the mini program mall.

During the entire process, the official account operators will continue to carry out operations such as maintaining WeChat fans and recovering lost fans.

Recommended here:

a. The detailed requirements chapter in "Effective Requirements Analysis" supports the main line requirements analysis method of business functions

2. Clarify the business purpose of each business scenario touchpoint. This step is based on the previous review of business processes, insight into business problems, clarification of what business purpose you want to achieve, and the breakdown of business purposes. Refer to the following case study for the process from clarifying the overall business goals to breaking down and quantifying the business goals:

O: Assuming that the clothing retailer's online layout is relatively complete, the primary business goal at this stage is to increase sales. Therefore, "increasing sales" is the North Star indicator of the retail e-commerce. Then increasing traffic, increasing conversion rate, increasing average order value, and increasing repurchase rate are the core indicators after decomposition.

S: Here we assume that you want to increase the traffic into the mini program mall, and there are many strategies you can adopt. For example, by scanning the QR code and then following the app and sending it coupons, you can attract more WeChat users to follow and become fans. Another example is to produce higher quality WeChat pictures and texts to better operate WeChat private domain traffic.

M: Following the previous step, for the strategy of pushing coupons to attract users to follow the official account, we can focus on the ratio of following the official account through scanning the code, the ratio of unfollowing, and the ratio of new and old fans.

Recommended here:

a. OSM model (Objective, Strategy, Measurement)

b. Sales formula = traffic * conversion rate * average order value * repurchase rate

3. Guide operational strategy design and user labeling needs from business purposes. The construction of the labeling system is different for different business purposes, so labels should be derived from operational strategies. For example, if the business department wants to make personalized recommendations, it would be more valuable to label objects or people’s interests and preferences; but if it wants to perform refined operations, labels about user retention and activity would be more valuable. Refer to the following example of user tag selection:

Taking increasing the rate of attention through scanning code as a quantitative goal, the selected operation strategy is to attract WeChat users to scan the code by pushing coupons. After new fans scan the code and follow, a 100-yuan coupon will be pushed, and after old fans scan the code, a 50-yuan coupon will be pushed. Therefore, the label "Is it a new fan?" needs to be used in the execution of the operation strategy.

At this stage, you can prepare a simple Excel template to record communication content. The column header should include label name, label rules, usage scenarios, etc., and record the communication content together with the business party.

4. Organizational tags Regarding organizational tags, classification management needs to be carried out from the user's perspective based on an understanding of the business and strategy. Here is a frame of reference:

(1) User attribute tags: gender, age, province, city, registration date, mobile phone number, etc.

(2) User behavior tags: number of visits in the past 30 days, average order value in the past 30 days, number of active days in the past 30 days, visit duration in the past 30 days, average visit depth, etc.

(3) User consumption tags: income status, purchasing power level, purchased products, purchase channel preference, last purchase time, purchase frequency, etc. (4) Product category tags: high heels, boots, shirts, French dresses, jeans, etc.

(5) Social attribute tags: frequently active time periods, active locations, single, number of reviews, positive reviews, etc.

2.2.2 Output label requirement document

After the previous demand collection and analysis, the business side's label requirements have been clarified. In order to smoothly deliver R&D, the following steps are required: writing label system documentation - determining tracking points based on label rules - writing data requirement documentation.

1. Write label system documentation

In this phase, the data product manager needs to produce specific label system documents based on the previous communication with the business side:

(1) Tag ID: For example, ATTRITUBE_U_01_001, where "ATTRITUBE" is the demographic attribute theme, the "U" after the "_" is the userid dimension, the "01" after the "_" is the first-level classification, and the last "001" is the tag details under the first-level tag.

(2) Label name: English format name, for example, famale

(3) Chinese label: female

(4) Tag topic: describes the topic to which the tag belongs, for example, user attribute dimension tag, user behavior dimension tag, user consumption dimension tag (5) Tag level ID: the level to which the tag belongs, generally divided into 2 levels (6) Name: the name corresponding to the ID

(7) Label type: statistical labels, rule labels, machine learning algorithm labels

(8) Update frequency: real-time update, offline T+1 update, single calculation (9) Label algorithm rules:

a. You need to describe which specific field in which data table to select. If you need to associate multiple tables, you also need to explain which field to join.

b. Specific algorithm logic and statistical cycle, such as "number of payments in the past 7 days", which means counting the total number of payments in the past 7 days.

(10) Description of usage scenarios

(11) Scheduling

(12) Developer

(13) Demand side

(14) Priority

2. Determine the embedding point according to the label rules

The algorithm rules of the label have been clarified above. Next, we need to further determine which points should be buried to collect the required data. The following is a specific

Examples:

For the label "Preference for Purchased Product Categories", the event data of clicking the order button, as well as event attribute data such as product name and product category, will be used. Therefore, it is necessary to embed the event of clicking the order button.

3. Write data requirements document

Once the data to be collected by the tracking point has been determined, a specific data requirement document needs to be produced and handed over to the development colleague responsible for the tracking point to collect the data. In the data requirements document, the following should be clearly stated:

(1) Click order name: click_order

(2) Display name of the tracking point: Click the order button

(3) Timing of reporting: Choose when to report based on actual conditions. For example, for a click-to-order event, you can choose to report when the order button is clicked.

(4) Tracking method: Choose between client-side tracking or server-side tracking based on actual conditions. For example, the order button click event of the "Purchase Product Category Preference" label, because we only want to determine the user's preference for purchasing products, the user can indicate whether he has a preference after clicking the button, and there is no need to wait for the server to return a reminder of whether it is successful. Therefore, it is suitable to use the client-side embedding method.

(5) Attribute name: the name of the event attribute, such as the product name attribute of the order button click event

(6) Attribute value: such as a shirt

(7) Notes

In actual work, writing label system documents, determining tracking points based on label rules, and writing data requirement documents will be a process of mutual improvement and complementation.

2.2.3 Label Development

In the entire engineering solution, the system relies on infrastructure including Spark, Hive, HBase, Airflow, MySQL, Redis, and Elasticsearch. In addition to the infrastructure, the main body of the system also includes three important components: ETL operations, user portrait topic modeling, and storage of label result data on the application side. The figure shows the user portrait data warehouse architecture diagram, which is briefly introduced below.

1. Hive data warehouse ETL job

The dotted box below is the common data warehouse ETL processing flow, which is to process daily business data, log data, embedded point data, etc. through the ETL process and process them into the ODS layer, DW layer, and DM layer corresponding to the data warehouse.

2. Hive Data Warehouse User Portrait Topic Modeling

The dotted box in the middle is the main link of user portrait modeling, which will perform secondary modeling and processing on user-related data in the data warehouse ODS layer, DW layer, and DM layer.

3. Storage of label result data on the application side

During the user portrait topic modeling process, the user tag calculation results will be written into Hive. Since different databases have different application scenarios, they are described below:

(1) MySQL

As a relational database, it can be used in applications such as metadata management, monitoring and early warning data, and result set storage in user profiling. The following are the three application scenarios:

a. Metadata management: MySQL has faster reading and writing speeds. The tag metadata in the platform tag view (Web-based products) can be maintained in the MySQL relational database, which facilitates tag editing, querying, and management.

b. Monitoring and warning data: In the data monitoring of the portrait, the monitoring data of each module is inserted into MySQL after the scheduling flow finishes running the corresponding module. When the verification task determines that the triggering alarm threshold is reached, an alarm is triggered.

c. Result set storage: stores tags used for multi-dimensional perspective analysis, user tags used for tagging services, and the number of tags recorded on the day.

(2) HBase

Unlike Hive, HBase can run in real time on the database instead of running MapReduce tasks, making it suitable for real-time queries of big data.

The following example introduces the application scenario and engineering implementation of HBase in the portrait system:

In order to encourage unregistered new users to register and place orders, a channel operator plans to guide them by distributing red envelopes or coupons through pop-up windows on the App homepage. After the ETL scheduling of the portrait system is completed every day, the corresponding population data will be pushed to the advertising system (stored in the HBase database).

When a new user who meets the conditions visits the App, the online interface reads the HBase database and pushes the pop-up window to the user when the user is found.

(3) Elasticsearch

It is an open source distributed full-text search engine that can store and retrieve data in near real time. For scenarios that require high response time, such as user tag query, user population calculation, and multi-dimensional perspective analysis of user groups, you can also consider using Elasticsearch for storage.

2.2.4 Label Release and Effect Tracking

Through development and testing, after going online, it is necessary to continuously track the label application effect and business feedback, and adjust and optimize the model and related weight configuration.

3. Productization of User Portraits

In terms of business value, tags and portraits are like a middle-layer system module that provides data support for front-end services. After developing portrait label data, if it just "lies" in the data warehouse, it cannot play a greater business value. Only after the portrait data is productized can the efficiency of each link in the data processing chain be improved in a standardized manner, and it will also be more convenient for business parties to use. The following is a summary from the perspectives of label production architecture and functional modules covered after productization:

3.1 User Portrait Product System Architecture

The figure below is a structural diagram of a user portrait product system. The data is from left to right, mainly including four levels: data collection, data access, data integration/label calculation, and label application. Here is an attempt to briefly describe it:

3.1.1 Data Collection

In the data collection module, log data, business data, and third-party data are collected mainly through three methods: client/server SDK, import, and docking with third-party applications.

1. SDK

(1) Client SDK: Through the client SDK tracking, you can collect user behavior data and user attribute information from various clients such as iOS, Android, mini-programs, and websites.

(2) Server SDK: If the data already exists in a database or data warehouse, such as order information, you can use the server SDK of the corresponding development language to collect the data.

2. Importer

You can choose different large import methods based on influencing factors such as the operating environment, source data format, and the amount of imported data to import historical file data into the user portrait product system.

3. In view of the characteristics of OpenAPI of different third-party products, Link adopts the method of receiving event message push or active polling to collect users' personal attributes and behavioral event data in different third-party application systems.

3.1.2 Data Access

The buried data first enters Kafka in large quantities, and then is slowly consumed and connected to the subsequent data integration storage system.

3.1.3 Data Integration/Label Calculation

In the user portrait system, Hive is mainly used as the data warehouse for ETL processing, development of corresponding user attribute tables and user behavior tables, and label calculation.

1. Data Integration

The data received from various channels have data quality problems such as isolation, null values, format mismatch, and exceeding the limit range. Therefore, dirty data cleaning, format conversion, user identification and merging and other integration work are required:

(1) Clean/Transform

a.Clean: For example, if a user's date of birth is a future date, then this dirty data needs to be filtered out.

b.Transform: For example, the regional information of all users obtained through a third-party application API is in the IPB standard encoding format. In order to analyze it together with information from other channels, it needs to be converted into the standard province and city format according to the IPB standard encoding.

(2) Id Mapping

a. The user attribute data, behavioral event data, etc. received from various channels are all isolated. In order to calculate the user's comprehensive labels, it is necessary to identify and merge the users. For example, through unionID, the information of the same user in the public accounts, mini-programs, and websites bound to the same WeChat open platform can be identified and merged.

After data integration, the data will enter the following data model:

2. Tag calculation

In the user portrait system, a batch offline label processing engine will be built, which relies on a relatively stable underlying data structure. This tag engine reads event data and user attribute data at the same time, and then combines it with specific tag rules to perform a batch calculation and finally generate user tags.

3.1.4 Label Application

The application of tags is mainly divided into two categories: front-end portrait display and access to other systems through APIs. They are described in detail in the following "3.2 User Portrait Productization Function Module" section.

3.2 User portrait product function module

3.2.1 System Dashboard

Usually, the data dashboard of the user portrait system displays the core user data assets of the enterprise or the data of key focus groups in a visual form. Aims to establish and unify users' basic understanding of enterprise data assets or core population data, which are mainly divided into the following categories:

1. User volume and change trend: the volume of IDs of different device types, the volume of different types of users (such as registered and unregistered users, paying and non-paying users, etc.);

2. Tag assets: count the number of tags by major categories, etc.

3. Core user tags: display key tag portrait data of inherent or customized groups;

3.2.2 Tag Management

It allows business personnel to add, delete, modify, and query tags, including tag classification, new tag creation, tag review, tag listing and delisting, and tag coverage number monitoring.

Based on user behavior data and user attribute data, create tags by setting tag rules:

3.2.3 Single User Portrait

The main capabilities include viewing detailed data of a single user portrait, such as user attribute information, user behavior, and other data, by entering the user ID.

3.2.4 User Grouping and User Group Profiling

1. User Segmentation

The user grouping function is mainly used by business personnel. When applying tags, product managers, operations, customer service and other business personnel may not only look at the population corresponding to a certain tag, but may need to combine multiple tags to meet their business definition of the population. For example: combine the three tags "number of coupons received in the past 7 days is greater than 1", "activity level is high and extremely high", and "female" users to define the target population and check the number of users covered by this group of people.

2. User group portrait

Similar to the user grouping function, the user group portrait function also needs to combine tags to define the user group. The difference is that the user group portrait function supports analyzing the characteristics of the user group from multiple dimensions, while the user grouping function focuses on pushing the screened user groups to various business systems to provide service support.

3.2.5 BI Analysis

After the BI platform is connected with these data, the data dimensions can be enriched, supporting richer and deeper analysis and comparison through a variety of analysis models.

3.2.6 OpenAPI

OpenAPI can ensure that the portrait system data is connected with various systems, such as push systems, marketing systems, advertising systems, recommendation systems, BI platforms, and ensure real-time updating of data in each system to avoid the problem of different numbers from the same source.

4. User portrait application

As mentioned earlier, user portraits mainly have three applications: business analysis, precision marketing, and personalized recommendations and services. Specifically, it can be divided into:

4.1 Business Analysis

After the label data of the user portrait system enters the analysis system through the API, the dimensions of the analysis data can be enriched to support operational analysis of various business objects. The following is a summary of some of the indicators that marketing, operations, and product personnel will pay attention to during analysis:

4.1.1 Traffic Analysis

1. Traffic Sources

2. Flow quantity: UV, PV

3. Traffic quality: browsing depth (UV, PV), dwell time, source conversion, ROI (return on investment)

4.1.2 User Analysis

1. Number of users: number of new users, number of old users, ratio of new to old users

2. User quality: number of new users (App launch), number of active users (App launch), user retention (App launch-App launch), user engagement, dormancy, and average order value

4.1.3 Product Analysis

1. Product sales: GMV, average order value, number of people placing orders, number of people canceling purchases, number of people returning goods, repurchase rate on each end, purchase frequency distribution, and purchase conversion of operation positions

2. Product categories: payment order status (number of times, number of people, trend, repeat purchase), purchase status, return application status, order cancellation status, attention status

4.1.4 Order Analysis

1. Order indicators: total order volume, refunded order volume, order payable amount, order actual payment amount, number of people placing orders

2. Conversion rate indicators: new orders/visit UV, effective orders/visit UV

4.1.5 Channel Analysis

1. Active users

(1) Active users: UV, PV

(2) New users: number of registrations and year-on-year growth

2. User quality

(1) Retention: Next-day/7-day/30-day retention rate

3. Channel revenue

(1) Orders: order volume, average daily order volume, and year-on-year and month-on-month changes in orders

(2) Revenue: amount of payments, average daily amount of payments, amount compared to the same period last year (3) Users: number of orders per person, amount of orders per person

4.1.6 Product Analysis

1. Search function: number of searches/number of searches, search function penetration rate, search keywords

2. Product function design analysis such as critical path funnel

4.2 Precision Marketing

4.2.1 SMS/email/push marketing

In our daily lives, we often receive marketing information from many channels. A text message push about the arrival of a red envelope may prompt a user to open an app that they have not visited for a long time. An email message push about a price reduction of a book on their wish list may stimulate a user to open the push link and place an order directly. What types of marketing methods are there? It can be roughly divided into the following 4 categories:

1. Behavior-based marketing: product browsing, adding to shopping cart, store code scanning, order cancellation, order return, etc.

2. Location-based marketing: nearby stores, nearby activities, frequently visited areas, etc. 3. Festival-based marketing: birthdays, Spring Festival, Double 11, Double 12, Christmas, etc.

4. Membership marketing: welcome to membership, card and coupon reminders, points changes, level changes, member benefits, etc.

4.2.2 Customer Service Talk

When we complain, consult or give feedback to the customer service department of a platform, the customer service staff can accurately tell us our purchase situation on the platform, the results of the last consultation issue and other information, propose targeted solutions, and provide special services such as VIP customer service channels for high-value users.

4.3 Personalized recommendations and services

The operator of the application can recommend different content to users based on labels such as gender, age group, interests and hobbies, browsing and purchasing behavior in the user portrait. Such as the personalized article content recommendations on Toutiao, the personalized video content recommendations based on user portraits on Douyin, and the personalized product recommendations based on user browsing behavior and other portrait data on Taobao.

5. User portrait practice case

Based on the portrait system, we can conduct multi-faceted data analysis and reach user operation plans, quickly apply label data to the service layer (T+1, real-time application), and obtain user feedback through effect analysis to help iterate marketing strategies or product designs. The following uses some practical cases to reproduce the application points and methods of user portraits in a scenario-based manner.

5.1 A/B crowd effect test

5.1.1 Case Background

In order to achieve good sales during the big promotion, a fast-moving consumer goods company of snacks planned to promote a series of articles on new products, health functions of products, etc. through message push, so as to build momentum for the big promotion and stimulate sales conversion. In order to accurately locate the target population flow, channel operators now plan to conduct two A/B population effect tests:

1. The impact of different content titles on traffic;

2. Precision push brings more traffic than ordinary push.

5.1.2 User Profile Entry Point

The entire project needs to clarify how to divide the traffic into group A and group B, and how to design the crowd rules and effect monitoring for group A and group B. The following is a step-by-step introduction on how to use the portrait system in AB population testing.

1. Divide users into groups A and B. In order to conduct A/B group testing, you first need to divide the traffic. You can use A/B allocation random diversion to divide users into A/B groups.

2. A plan to test the impact of copy titles on traffic. In order to recall more users to visit the App during a big promotion, a platform channel operator plans to select a small number of users during the event warm-up period to conduct an AB effect test on a version of the copy title.

In this test plan, control group A selected the user group that followed path A and had visited the app in the past x days and had browsed/collected/added to purchase the snack in the past x days, and pushed retail copy A to this group of users; control group B selected the user group that followed path B and had visited the app in the past x days and had browsed/collected/added to purchase the snack in the past x days, and pushed snack copy B to this group of users. The control group and the comparison group have the same number of users, but different copywriting. The click-through rates of the two groups are subsequently monitored to analyze the impact of different copywriting on user clicks. For example, circle the users in group A through the user group function, as shown in the figure below:

3. Test solution for traffic increase brought by precise push compared with ordinary push

Before using the portrait system to refine the push groups, a certain platform pushed messages to users indiscriminately. In order to test the increase in traffic brought by refined operations compared to indiscriminate operations, channel operators decided to conduct an AB effect test at the snack marketing venue that has been the focus of recent operations.

In this test plan, control group A selected the user group that followed path A, visited the app in the past x days, and browsed/collected/added the snack to purchase in the past x days; control group B selected the user group that followed path B, visited the app in the past x days, and had no category preference. The same message is pushed to user groups A and B, and the click-through rates of the two groups are subsequently monitored to analyze the growth points brought about by precision marketing push.

5.1.3 Effect Analysis

After the AB group crowd message push is launched, you need to build a monitoring report to monitor the traffic and conversion of the control group and the test group, focusing on the indicators in the following list:

For example, the GMV comparison report of population A and population B built using the event analysis model is shown in the figure below:

5.2 Targeted Marketing for Women’s Day

5.2.1 Case Background

A brand that focuses on women's products plans to carry out targeted marketing to goddesses with preferences for different categories on Women's Day. Marketing information will be pushed twice. The first time is at 10:00 a.m. on the same day, and the second time is at 10:00 p.m. on the same day. Finally, the marketing effectiveness is evaluated by tracking the target audience’s same-day payment order completion rate.

5.2.2 Implementation Logic

First, based on the user's gender label and age label, female users aged 18 to 40 are selected. Then, the time will be uniformly extended to 10:00 am on March 8, 2020, and different marketing content will be pushed according to the user's category preference tags. For example, marketing information about the Spring Beauty Festival will be pushed to people whose category preference = makeup and skin care. The second wave of push notifications will be delayed until 10:00 pm on March 8, 2020, and the push information will be a unified promotion reminder.

5.3 Real-time marketing for newly installed unregistered users

5.3.1 Case Background

In order to promote the registration and ordering of unregistered newly installed users, the operator of a snack mall App has formulated operating rules: when a newly installed unregistered user opens the App, coupons are pushed to them through App pop-ups for marketing purposes. For example, if a user installs the App but does not register, the App will immediately send him a pop-up coupon when he opens the App the next day to better guide him to complete the registration and place an order.

5.3.2 User Profile Entry Point

Channel operators filter out the corresponding user groups by combining user tags (such as "unregistered users" and "installation date less than ×× days ago"), and then choose to push the corresponding groups to the "advertising system". In this way, after the ETL scheduling of the daily portrait system is completed, the corresponding population data will be pushed to the HBase database for storage. When a new user who meets the conditions visits the App, the online interface reads the HBase database and pushes the pop-up window to the user when the user is found.

5.4 Remarketing Advertisement of an E-commerce Company

5.4.1 Case Background

The product operation team of an e-commerce app wanted to increase the repurchase rate of old customers and the order rate of new customers for electronic products, so they chose to cooperate with Toutiao to launch remarketing advertisements. For example, a user saw a vivo phone on the e-commerce app, and when he was browsing Toutiao the next day, he saw advertising information for the corresponding phone.

5.4.2 Implementation Logic

First, it is necessary to ensure that the API of the e-commerce app and Toutiao are connected, and then algorithm mining is performed based on the user's behavior in the app (browsing, collecting, adding to cart, searching, etc.) to generate labels for the user's product preferences.

Once Toutiao captures the user's device information, it will send a request to the e-commerce company, asking whether it needs to display ads to this user.

At this time, the e-commerce platform will determine whether the user is its own user. If it is its own user, it will return a recommendation result to Toutiao. Then the user will see the product information that he has browsed before on Toutiao. After clicking, he can jump to the product details page in the e-commerce app.

VI. Conclusion

1. First, the cognitive concepts of user portraits, user tags, and user groups are described;

2. Then, the classification of label system, the process and method of label construction are explained;

3. In order to explain how to make the portrait label data "lying in" the data warehouse have greater business value, the construction of the user portrait system is briefly summarized from the perspectives of system architecture and application layer functions;

4. Finally, the application of user portraits is summarized from three perspectives: business analysis, precise marketing, and personalized recommendation, and several cases of practical application of user portraits are listed in the practical case section.

Author: Linkflow

Source: Linkflow