How to segment users? 6 models and 5 dimensions

Nowadays, with the continuous development of the Internet and the advancement of technology, users have been labeled and classified in today's environment. They are classified according to different needs and other aspects to achieve refined operations. The author of this article shares the model and dimensions of user segmentation. Let's take a look at it together.

Classification is the recognition pattern of the human brain, and classification is one of the ways to simplify things.

——Zhang Xiaolong

With the development of commodity economy, social division of labor has become increasingly refined, and product production and manufacturing has gradually evolved from unified supply to personalized supply.

Internet products are even more diverse for each individual, and users can be divided into different types and provided with different services, which can be said to be adapted to local conditions, take advantage of the situation, teach students in accordance with their aptitude, and tailor-make services to their needs.

However, some common problems may occur in user segmentation and refined operations:

• Without tailor-made services and differentiation, it is difficult to attract new customers.
• "Tailoring" is unreasonable, needs are difficult to meet, and retention is difficult

So, how do we conduct reasonable user segmentation?

Two thousand years ago, books were made of bamboo slips, and only the nobility could read books. There were very few people who could read. How did Buddhism spread the Dharma to every household?

People from different regions, tribes, rich or poor, and cultures have different spiritual needs and levels of acceptance, so different approaches are needed; Buddhism in China has evolved into eight major sects, each with a very different missionary style. Each sect adopts different methods for people with different mindsets: with people of the highest aptitude, they speculate on philosophy and explore the truth of the world; with people of the highest aptitude, they keep their hearts pure and desires few, and practice cultivation in every aspect of life; with people of average aptitude, they discuss cases and extend the principles; with people of the lowest aptitude, they lure them with the thought of heaven and scare them with the thought of hell, and guide them to do good.

This is even more true for enterprises to users.

User segmentation is the behavior of enterprises in order to realize the heterogeneity of user needs and concentrate limited resources for effective market competition; in a clear strategic business model and a specific market, enterprises classify users according to user attributes, behaviors and other factors, and provide targeted products, services, sales and operation models to maximize user value and product goals.

At the system implementation level, under the guidance of abstract theories, algorithms are used to conduct labeling statistics and classification, and presented in the form of user portraits. Finally, the strategy, interface, and operation methods are tailored.

From what angles and dimensions can we segment users?

How to use algorithms for labeled statistics and classification?

How to verify the rationality of user segmentation and make adjustments?

1. 6 common models and 5 dimensions of user segmentation

First of all, what is a user?

We often hear conversations like this:

• How many student users does your company have? 5 million.
• There are three types of users: those who post videos, those who watch videos, and those who place advertisements.
• In this scenario, different users have different cognitions and needs!

If you think about it, the meaning of users in spoken conversations is ambiguous:

• In the first sentence, users actually refer to the number of registered IDs of car owners.
• In the second sentence, users actually refer to the roles that exist in the product.
• In the third sentence, users actually refer to a collection of a certain type of needs.

Obviously, the first two sentences are lazy and simplified expressions in spoken language, and the meaning contained in the third sentence "users are a collection of needs" is more accurate.

Now that we have a clear definition of users, let’s look at the six classification models commonly used by Internet companies:

1. Liang Ning: Product Thinking

According to the main roles in the business model and the user classification under a certain role:

For example: Xindian’s free meal on Dianping.com. Dianping cooperates with newly opened merchants to launch free meals, most of which are allocated to the top users with high user levels, high activity levels and many classic reviews. After enjoying the meals for free, the top users will make reviews, attracting Daming Sheep, Xiaoxian Sheep and Benben Sheep to consume.

2. UCPM-Product Management Knowledge System

In a certain scenario, users have feelings and needs, and then they look for solutions, select products, buy products, use products, and finally carry out after-sales service. Users can be divided into five categories:

• Buyers: perform purchase actions and are primarily concerned with price
• User: Uses the product and considers its performance first
• Influencers: Groups that provide guidance on how decisions are made
• Information managers: control the flow of information and communicate with others
• Decision makers: Those who formally approve the purchase decision and are concerned with the internal policy aspects of the decision.

For example:

• Early childhood education products: the decision makers, buyers, information managers, and influencers are parents, and the users are students;
• Primary school education products: decision makers and buyers are parents, users are students, and information managers and influencers are parents and students together;
• Enterprise service products: The purchase of an office software requires comparison of products from different suppliers by the purchasing department, budget by the finance department, opinions of employees, management and maintenance by dedicated personnel, and decision-making by senior management.

3. User experience elements

According to the degree of familiarity with the product, they are divided into:

• Novice users: those who have just started using the product or use it infrequently and are not familiar with the product.
• Ordinary users: They account for the majority and have a good use of the product.
• Expert users: They are highly dependent on the product, can put forward many opinions and views, and are strong supporters of the product.

This is how it is divided for easy analysis:

• What problems do the same group of people encounter in scenarios A/B/C?
• In the same scenario, three groups of people, a/b/c, encounter different problems.

4. New generation of business models

According to the scope of user needs and the relationship between users, they are divided into:

• Mass market: serving a large user group with broad similar needs, such as Taobao, Pinduoduo, Douyin, etc.
• Niche market: Serving a specific, specialized user group with tailored needs, such as Huaban.com serving designers.
• User groups that seek common ground while reserving differences: serve multiple segmented user groups with similar but different needs, such as stock trading apps that serve "leeks", independent economic people, and traders.
• Diversified user groups: Serving user groups with different needs; just like a CRM product, it can be deployed privately, on the public cloud, or on a hybrid cloud.
• Multi-sided platform (multi-sided market): serves two or more independent yet interdependent user groups; such as O2O food delivery, serving people who order food, people who deliver food, merchants, and advertising buyers.

5. The Essence of Interaction Design

According to the design goals of the interface, it is divided into:

• Main character: A product can only have one main character model per interface.
• Secondary Personas: There are additional specific needs that can be met without weakening the product's ability to serve the primary personas.
• Supplementary characters: The main characters and secondary characters combined can fully represent the needs of supplementary characters. An interface can be connected with any number of supplementary characters.
• Customer persona: It is the needs of customers rather than end users. The interfaces of some customer personas may become the main characters of their own unique management interfaces.
• Service recipients: They are not the users of the product but are directly affected by the use of the product.
• Negative characters: used to communicate that the product will not serve a specific type of user, that is, they are not the actual users of the product.

For example:

In 2B products, there are actually not many data that a certain type of users really care about frequently and many functions that they use frequently in daily use. Placing these key data and functions on the system homepage, module homepage, and function homepage, allowing users to complete most of their daily work using only a few functions, is the ultimate user experience.

6. RFM Model

The RFM model is an important tool and means to measure customer value and customer profitability. It is widely used in user analysis of many CRM products. It mainly uses three indicators - "the number of days from the last consumption to the current time, the cumulative number of consumptions, and the cumulative consumption amount" to describe the value of customers. It can be divided into 8 categories:

These six models are suitable for different work scenarios of PMs. They are established, conventional, and rough classification models. However, due to the trend of increasingly vertical products on the market, the effectiveness of the public segmentation models has been compromised.

Therefore, more and more products are gradually classified from more detailed dimensions. The author has read most articles in the industry and summarized them based on work practice, which can be summarized into five dimensions:

After refining the dimensions, how to quantify it?

How is the technology implemented? How do users behave?

2. Label the data and visualize the users

What are tags? Tags are used to mark your product goals, categories, and content. They are keywords for your goals and are tools for easy searching and positioning.

User tags can be divided into static tags and dynamic tags based on update frequency; they can be divided into statistical tags, rule tags, and algorithm tags (also known as basic tags, model tags, and prediction tags) based on development method; and they can be divided into tags automatically added by the system, added by developers, and added by users themselves based on the source of tags.

What is a user portrait? From the classification models in the previous section, we can see that some models are more emotional, as if they can see a real person, and some models are more rational, as if they can see a bunch of labeled data. Yes, user portraits are currently divided into two categories: User Persona and User Profile.

User Persona is a typical user abstracted from the user group by product designers and operators. It usually comes from user interviews and user research, helping us to intuitively understand what types of users the current product mainly serves.

User Profile is a collection of labels that describe users based on their real data in the product. It is a relatively rational data representation. It is generally used in specific product design, decision-making basis, operational marketing, risk prediction, credit assessment, personalized recommendation and other processes, such as the five-dimensional table presented at the end of the first section.

User Persona and User Profile are two sides of the same coin, and they are identical and unified. In actual applications, they should be used in conjunction with business and scenarios.

This section mainly summarizes the implementation of User Profile. The product structure is as follows:

The implementation steps can be divided into three steps:

1. Determine the dimensions of the portrait

• Mining virtual representations of real users based on business scenarios: User Persona.
• Based on User Persona, determine the label dimensions, hierarchical relationships, label types, label values, and preliminary label rules of the User Profile in the system.

2. Establish data processing model

• Tag weight: the weight of a tag in a business indicator
• Update frequency: real-time update, offline T+1 update, single calculation
• Tag statistics rules: Ordering takeout > 2 times in 7 days is considered moderately active
• Label algorithm: TF-IDF weight classification algorithm, similarity matrix algorithm, LAP propagation algorithm...

User tag weight = behavior type weight × time decay coefficient × number of user behaviors × TF-IDF calculated tag weight

Behavior type weight: Different behaviors such as browsing, searching, collecting, ordering, and purchasing have different importance to users. Generally speaking, the more complex the operation, the greater the behavior weight. The weight value is usually given subjectively by operators or data analysts.

Time decay coefficient: The influence of time on some user behaviors is gradually weakened. The farther the behavior time is from now, the less significance the behavior has to the user at present. The mathematical model of Newton's cooling law is applied here, which means that when a hotter object is in an environment with a lower temperature than this object, the temperature of this hotter object will decrease, and the temperature of the surrounding objects will rise. Finally, the temperature of the object and the surrounding temperature reach equilibrium. In this equilibrium process, the temperature F(t) of the hotter object decays exponentially with the increase of time t. The temperature decay formula is: F(t) = initial temperature × exp (-cooling coefficient × interval time).

Corresponding to the influence of user tags over time, the cooling coefficient is equivalent to the coefficient of tag weight decay over time. The formula is as follows:

λ=-ln(dN/dt)/T=-ln(current value/initial value)/interval time

Number of user behaviors: User tag weights are calculated periodically. The more behaviors a user has with the tag within a period, the greater the impact of the tag on the user.

TF-IDF calculates tag weights: the importance of a tag increases in direct proportion to the number of times it is tagged by a user, but decreases in inverse proportion to the frequency of its appearance in the tag library; w(P, T) represents the number of times a tag T is used to tag user P. TF (P, T) represents the proportion of this tag count in all tags of user P. The larger the TF, the more important the tag. The corresponding IDF (P, T) represents the scarcity of tag T among all tags, that is, the probability of occurrence of this tag. The larger the IDF, the less important the tag.

Then, the weight value of the tag for the user can be obtained based on TF * IDF. The formula is as follows:

(The denominator +1 prevents the denominator from being 0)

For example:

• TF(user P1, tag T1) = 5/(5+2+1+0) = 0.625
• IDF(user, tag T1) = log(5/(2+1)) = 0.222
• TD-IDF=0.625*0.222=0.139

3. Data collection, data processing, and generation of User Profile classification

Bring business data, log data, embedded data, and third-party data into the data processing model to generate a User Profile.

(The picture comes from the Internet)

For example:

During the epidemic, everyone was bored staying at home. Many social platforms saw cases of criminals seducing ignorant and lonely men to chat naked, and then recording the screen to blackmail them for money. The number of reports to the police increased, and districts, counties and communities sent text messages to remind residents.

So, as a social networking platform for strangers, how can we avoid the rampant criminals?

Step 1: Confirm the image dimensions

1) Mining User Persona Based on Scenarios

2) Determine the label of the User Profile in the system

Step 2: Establish a data processing model

1) Label weight:

• Behavior type weight: manual judgment, sorting, and weighting
• Time decay coefficient: calculated according to the formula
• User behavior count: cycle count
• TF-IDF calculation weight: calculated according to the formula

2) Update frequency: real-time update

3) Tag statistics rules: the rules described in the tag value

4) Labeling algorithm: TF-IDF weighted classification algorithm, decision tree classification algorithm, neural network, KNN classification, SVM...

Step 3: Data collection, data processing, and classification generation

1) After the model is built, you can import sample data and perform simulation to find all the accounts of Guoliao fraudsters.

2) Multiple algorithms can be used for classification at the same time, and their respective results can be observed, compared comprehensively, and finally the best one can be used.

Once the User Profile is determined, what is the gap between the User Profile and the User Persona? What is the difference between User Persona and real user situations? What is the difference between the User Profile and the real user situation?

3. Use regression analysis, A/B testing, and user research

Verify the accuracy of user segmentation:

1. Regression Analysis

Continuing with the example in the previous section, it can be divided into two steps:

• By comparing the judgment situation with the actual situation, four types of results are obtained.
• Use recall, accuracy, and precision to evaluate, the larger the value, the better.

Recall: R = TP / (TP + FN); that is, (number of correctly identified criminals) / (number of correctly identified criminals and normal users)

Accuracy: ACC = (TP + TN) / (TP + TN + FP + FN); that is, (number of correct judgments) / (number of all judgments)

Precision: P = TP / (TP + FP); that is, (number of criminals correctly identified) / (number of criminals identified by the system)

Basically, only recall (R) and accuracy (ACC) can be used to evaluate the quality of the strategy and make optimization adjustments.

2. A/B Testing

Multiple schemes are tested in parallel, and the single variable method is used to observe the effect of the schemes, and finally the best one is selected; the principle at the implementation level is as follows:

(The picture comes from the Internet)

From left to right, the four thick vertical lines represent four key roles: client, server, data layer, and data warehouse.

From top to bottom, the three parts represent three types of tests: no A/B test group, backend-based A/B test group, and front-end-based A/B test group.

3. User research

The purpose of user research is to get close to users, understand them, and also to make it easier for you to become a user and experience the user.

There are many ways to conduct research: user interviews, focus groups, participatory design, questionnaires, observing user behavior, entering scenarios, analyzing user data, the 10-100-1000 rule... These are used to continuously obtain accurate user portraits. The specific steps can be found by searching online. In your work, you only need to choose one or two methods that are most convenient for you and use them to the extreme, to perfection, and until you can perceive the users.

So, how to reduce the gap between survey results and real users? Sony once conducted a face-to-face interview survey on users' preferences for the color of Boomboxes speakers. The speakers were available in two colors: yellow and black. Most people said that the yellow color looked better and they were more willing to buy yellow speakers. Interestingly, after the survey, the organizers allowed everyone to take a speaker away when they left as a thank you, and the vast majority of people took away the black speakers.

Therefore, no matter which research method is adopted, there are still many things to pay attention to when interacting with others. Many articles have elaborated on this content. To make it easier to understand, let’s look at the story of Tang Monk who went to the West to obtain Buddhist scriptures while walking and begging for food. Why did he have to walk? Don’t let Wukong carry you and fly away? Why do we have to beg for food? Because we need to get close to all living beings, understand all living beings, and integrate into all living beings, what should we pay attention to when begging for food?

In summary, regression analysis and A/B testing verify the accuracy of user segmentation from a quantitative perspective, while user research confirms the accuracy from a qualitative perspective.

User segmentation is the behavior of enterprises in order to realize the heterogeneity of user needs and concentrate limited resources for effective market competition; at the system implementation level, it is under the guidance of abstract theories, using algorithms to conduct labeling statistics and classification, and presenting them in the form of user portraits, and finally "tailoring" in terms of strategy, interface, and operation methods.

Author: Qi Niu

Source: Qiniu