Operation personnel should understand user portraits

User portrait is a relatively new term. Initially, it was a fashionable concept that was always mentioned in the big data industry. Now when we talk about user portraits in operations , it is also directly linked to precision marketing and refined operations. This article mainly talks about user portraits from the product and operation perspectives. I hope that after reading this, all your questions about user portraits will be answered.

What is a User Portrait

User portraits are not mysterious at all. They are collected actively or passively based on the various data left by users on the Internet, and finally processed into a series of labels. For example, guess whether the user is male or female, where he is from, how much he earns, whether he is in a relationship, what he likes, and whether he is ready to shop?

We often equate user tags and user portraits. In any article about user portraits, pictures similar to the one above will appear, and there is a tendency for them to become overused. Labeling is the most intuitive explanation, but it is not equal to user portrait.

The official name of user portrait is User Profile, which is often confused with User Persona , the latter of which is more appropriately called user role. It is a method of product design and user research. When we discuss products, needs, scenarios, and user experience , we often need to focus on a certain group of people. User persona is an abstract method and a collection of target users .

User roles do not refer to specific people. "She is a 25-year-old white-collar worker who graduated from a 211 university. She is currently engaged in design work in the Internet industry and lives in Beijing. She is single and loves rock music." This sentence is often used to describe the typical user of a product.

The User Profile discussed in this article is more of a platform-level application that is closely related to operations and data. Its essence is that any user can be described with labels and data.

Application of User Portrait

It plays a vital role in the process of an enterprise's growth and development. The following are the main applications:

• Precision marketing : This is the most familiar way of operation. From extensive to refined, it divides the user group into finer granularity, supplemented by SMS, push, email, activities and other means, and drives strategies such as care, recovery and motivation.
• Data application : User portraits are the basis of many data products, such as the well-known recommendation system and advertising system. Those who have operated major advertising delivery systems must know that advertising is based on a series of demographic-related tags, including gender, age, education, interest preferences, mobile phones, etc.
• User analysis : Although different from Persona, user portrait is also a necessary supplement to understand users. In the early stages of a product, PMs get to know users through user surveys and interviews. As the number of product users increases, the effectiveness of the survey decreases, and user portraits will be used to assist in the research. What are the characteristics of new users, whether the attributes of core users have changed, etc.
• Data analysis : There is no need to say more about this. User portraits can be understood as data warehouses at the business level, and various labels are natural elements of multidimensional analysis. The data query platform will be connected with these data.

For most products, user portraits do not require a recommendation system, and personalized recommendations will not increase profits much, after all, it requires a large number of users and data to support it. Therefore, these products are more suitable for driving business based on user portraits.

So many benefits have been mentioned, but as far as I know, many companies have spent a lot of money and hired a lot of people to build user portrait systems, but in the end they cannot use them. Or you may make a report on user portrait, including gender, user location, user consumption amount. It looks very impressive and you will just stop reading it.

Ultimately, it's difficult to use well.

Many user portraits have good intentions, but have become formalism.

To give an example from my own experience, a friend of mine built user portraits in his company and divided them into hundreds of dimensions. It covers user consumption, attributes, and behaviors. This was originally a good idea, but after it went online, the operations team just stared at it helplessly.

The questions include but are not limited to: there are so many dimensions of users, how to choose labels reasonably? I want to define user levels. How much should the cumulative consumption of VIP users exceed? In what time window? Why choose these standards? How should it be maintained and monitored in the future? Should this label be changed if the business changes?

After setting up labels, how to verify the effectiveness of user portraits? How do I know that this system is successful? What to do if the effect is not good? Does it have more application scenarios?

The implementation of the strategy is also a tricky issue. From the perspective of job execution, operations are responsible for KPIs. When the KPI is not met at the end of the month, do you think they prefer to choose full-scale operation or refined operation?

I think many companies have similar situations: after using user portraits for a period of time, they find that they are just that, and gradually stop using them.

This is a long-standing problem encountered by user portraits at the business level. Although the company claims to have established user portraits, the application is still quite rough.

How to deeply understand user portraits

If you want to use it well, you must first have a deep understanding of user portraits.

Now operations have set up several labels according to the user life cycle , such as new users, active users, and churned users. These labels are certainly segmented enough. But is it really a good label? no.

Because these are all lagging. According to the general definition of lost users, it is often that users have not responded or taken action for a long time, but if there has been no response for several months, it will be of no use even if you know that they are lost users. It has value, but is too laggy.

Smart operations will set up a new label, which shows the number of days since the last active user. If the user has not been active for six months, the number of days is 180. This is better than simply labeling lost users. It can be used to divide the time into different numbers from today and set time nodes of 30 days, 90 days, and 180 days.

The number of days since today is not the best either. Users are different. For the same two users, A and B, even if they have the same number of inactive days, I cannot assume that their possibility of churn is equal. This problem is more prominent in low-frequency scenarios. It is normal for a travel APP to have no activity for half a year, but it is not enough to do so today.

Looking back at lost users, we defined it not for the purpose of setting up a high-end system. Any business would certainly hope to lose as few users as possible at the outset, and then consider how to retain them. Under this business premise, preventive reduction of lost users is more important than labeling those that have already been lost.

So, the best label is the probability of user churn :

Churn probability > number of days since consumption > churn label

Don't take it for granted to summarize a complete system while ignoring the core value of portraiture. User portraits must first be a collection of user tags for commercial purposes.

Guess whether the user is male or female, where he is from, how much he earns, whether he is in a relationship, what he likes, and whether he is ready to shop? There is no point in discussing these. How does being a man or a woman affect consumption decisions, how salary affects consumption ability, whether or not one is in a relationship will bring about new marketing scenarios, and how to make accurate recommendations for shopping, these are the logic behind user portraits.

Just because I have user portraits doesn’t mean I can drive and improve my business. User portraits are needed to drive and improve business. This is an easy mistake to make.

User portrait labels are generally obtained in two forms: based on existing data or processed according to certain rules, including loss labels and the number of days since today. The other is to calculate the probability model based on existing data, which uses machine learning and data mining.

Probability is a value between 0 and 1. Take gender as an example. Unless the user's ID information can be directly obtained, the user will rarely fill in their gender. The gender filled in may not be accurate. There are a lot of female men in online games who are not good at playing.

Here we need to add an algorithm to infer the user’s true gender. There is a strong correlation between Chinese people’s gender and their names. For example, “Founding the Country”, “Founding the Army”, “Cuihua and Cuilan”, it is easy to judge. Bayesian is often used in the algorithm to predict the gender of newly added users through the existing name and gender database.

In special cases, many names are neutral and cannot distinguish between male and female. Like Xiaojing, who can be either male or female. In more special cases, a name that looks like a male's name may also be a female's. My junior high school teacher was named Jianjun, but she was a very amiable young lady.

Special circumstances mean special probabilities, so we cannot use an either-or dichotomy. The so-called probability is more accustomed to telling you that through model inference, Jianjun has a 95% chance of being a male name, expressed as 0.95; Xiaojing has a 55% chance of being a male, expressed as 0.55.

Although for convenience, the model will set a threshold, assuming that a probability above 50% is male and a probability below 50% is female. But students in the business department should be aware that the essence of user labels is often a probability between 0 and 1.

Probabilistic labels are difficult to verify. If a user is labeled as a student, it is difficult to know whether he is a real student, unless you really trick him into uploading his student status certificate. In this black box situation, the effectiveness of marketing activities targeting student users is affected by the accuracy of the labels. Advertising, recommendations, and precision marketing will all encounter this problem.

The probability is definitely higher or lower. Although users with a 90% churn probability and users with a 30% churn probability are predicted values ​​built by the model and are not real, we still believe that the former are more likely to leave and establish operating strategies based on this.

This brings up a new question: how to choose the probability threshold?

If we want to win back lost users, should we choose the group with a probability of over 80% or 60%? The answer has been given. We need to consider the business. Recovering lost users is a means, not an end. The actual goal is to increase profits by recovering lost users, so the choice of threshold is easily solved. Calculate the revenue and cost of recovering users under different thresholds and select the optimal solution.

By extension, whether it is a recommendation system or an advertising system, they have more complex dimensions, labels, and features, and their essence is to find out whether the user wants to buy a car or travel recently. Push the most appropriate information to users at the most appropriate time to obtain the greatest benefits.

The cases I listed are simplified. Like names, in the e-commerce and consumer industries, in addition to biological gender, gender labels are also established in consumption models. Although some people are male, their shopping behavior is that of women, which needs to be distinguished.

Don’t be afraid when you see this. Building a good user portrait is neither easy nor difficult.

How to create the right user persona

User portraits are first based on business models. The business department hasn’t even figured out the business model, so the data department can’t do anything. The data department should not close its doors and develop cars. This is the same as making a product. They don’t even understand user needs thoroughly, and they rush to launch an APP, but no one is interested in it.

Understanding consumer decisions, considering business scenarios, considering business models, considering the needs of business departments... These concepts sound abstract, but a good user portrait cannot be separated from them. This article does not mention data, models and algorithms because I believe they are more important than the technical level.

Let’s start by building a user persona with a story.

Lao Wang is a core member of an Internet startup company. Its main products are green and healthy salads. Lao Wang and green go well together. The company has launched an app that sells a variety of salads and now needs to build user portraits to guide operations.

At the current business level, the company is more focused on marketing and sales: how to sell salads better. The picture below is a simple summary of the operation process by Lao Wang.

Lao Wang divides customers into potential users and new customers based on whether they have purchased salads before. Potential users are those who have registered for the APP but have not placed an order yet. New customers are those who have only purchased salad once. In addition, there are old customers, that is, people who have consumed twice or more.

To make it easier for everyone to understand, I use JSON format to represent a simple user portrait.

Why create a separate new customer tag? Because Lao Wang’s salad will give new red envelopes to non-consumers to guide them to consume, everything is difficult at the beginning. This also brings about the problem that new customers may not consume again after the first visit, so there is a need to divide customers into potential, new and old customers.

As an aspiring operator, it is not enough to classify old customers, so we will continue with user stratification.

Traditional stratification is measured using the three dimensions of RFM. The average customer spending on salad is relatively fixed, so one of F and M is enough. Lao Wang is now calculating the difference in user retention rates for different consumption levels. For example, whether a user who spends XX yuan in a certain period of time will continue to consume in the future.

Salads and other foods are high-frequency consumption. XX should choose a narrower time window. Statistics on consumption within 365 days are not very meaningful. Another point to note is that the sales volume of salads varies in different seasons. Salads definitely do not sell as well in winter as in summer, so the consumption distribution needs to be considered comprehensively.

Here we define that those who spend more than 200 yuan within 30 days are VIP users. If Lao Wang’s business is particularly good, he can continue to divide it into super VIPs. This kind of label is often used in conjunction with business, for example, VIPs have the right to free drinks and priority delivery. Non-VIP people also need incentives to develop into VIPs.

Lao Wang relies on the user to fill in the recipient's name on the order to determine the demographic attributes of the portrait. Place of origin and age are not particularly helpful for the salad business. Should we increase the number of spicy salads for Sichuan users?

The user address can be determined by setting rules for the delivery address. For example, if a certain address appears X times, it can be considered a frequently used address. Based on whether the delivery location is an office building or a school, we can infer whether the user is a white-collar worker or a student.

Lao Wang has adopted special operating strategies for people with different attributes. For example, for the student group, since July and August are summer vacation, Lao Wang predicted in advance that sales in campus areas would decline. When the school starts in September, students can be recalled to return to school.

White-collar workers are more concerned about consumption experience, and price sensitivity is secondary. If the consumption proportion of female users on the platform is high, Lao Wang will focus on salads with weight loss function and increase sales in the form of monthly packages.

For a salad shop, Lao Wang’s user portrait was already good, but he was still worried because the user churn rate began to increase. There are various reasons for user churn: competition from rival Lao Li Salad, the taste of the salad, users feel that the price-performance ratio is not high, Lao Wang is not handsome enough, etc.

Churn is a difficult problem to predict. Lao Wang defines a lost user as one who has not made any purchase for 30 days. To make accurate predictions, we need to try machine learning modeling. We will skip the technical aspects here. The so-called modeling is to find the key factors before the time when users start to stop consuming, which can be behaviors or attributes.

If the amount of money spent in the user's history window is small, there is a possibility of loss; if the frequency of consumption in the user's history window is low, there is a possibility of loss; if the number of times the user opens the APP in the history window is small, there is a possibility of loss; if the user has given a bad review, there is a possibility of loss; if the user waits for a long time for the meal, there is a possibility of loss; if there is a gender difference in the user, there is a possibility of loss; if there is a seasonal factor in the restaurant, there is a possibility of loss...

Based on the business, Lao Wang selected features that may affect the business and submitted them to the data group to try to predict churn. It should be noted that these user behaviors cannot reflect the real situation. You might want to think about this: Is the behavior of lost users a dynamic process of change?

I used to consume there many times, but suddenly got tired of it, so I reduced my consumption, and then consumed less, and eventually left. Consumer loyalty within a unit time period decreases in a gradient manner. In order to better describe the change process, the time window is subdivided into multiple equidistant segments. This segmentation, within the first 30 to 20 days, the first 20 to 10 days, and the first 10 days, can better express the downward trend and better predict churn than within the first 30 days.

From Lao Wang’s point of view, the so-called churn can be predicted through the details of user behavior. Although machine learning modeling relies on statistical methods, it is also inseparable from business insights. This proves again that user portraits are built on business models.

The churn probability solves Lao Wang’s worries by reducing churn users through early detection. After implementing the policy of recovering lost users for a period of time, Lao Wang found that although the number of lost users decreased, the cost increased, because it costs money to recover users. He couldn't afford to lose money, so Lao Wang came up with another plan. He only wanted to retain the valuable users. He didn't want those users who only made purchases after receiving red envelopes! What Lao Wang wants is true fans. So he treated customers differently according to their consumption levels. Although the number of lost users was not controlled well, profits increased.

In the above user portraits, no label is separated from the business. Based on business scenarios, we can also imagine many ways to play with user portraits. Salads come in different flavors, including vegetables, fruits, chicken and seafood. Users' taste preferences can be calculated using matrix decomposition, fuzzy clustering or multi-classification problems, and the degree of preference is also represented by a number between 0 and 1. Similarly, there are price preferences, namely price sensitivity.

Let's think more deeply about the business scenario. If a certain office location has five or six orders every day, from different customers at different time periods, the delivery guy has to make five or six deliveries. What a waste of labor costs. Operations can analyze relevant data in the background and facilitate order consolidation in the form of group buying or group buying. Although sales profits may decrease, the labor cost of food delivery is also saved. This is also the basis for using portraits as data analysis.

Now that Lao Wang’s operation story is over, you should have some ideas about how to build user portraits.

User portrait architecture

The portrait labeling systems for different businesses are not consistent, which requires the refinement of data and operational purposes.

User portraits are generally divided into multiple category modules according to business attributes. In addition to the common demographic and social attributes. There are also user consumption portraits, user behavior portraits, user interest portraits, etc. The specific portrait depends on the product form. In the financial field, there will also be risk portraits, including credit reporting, default, money laundering, repayment ability, insurance blacklist, etc. In the e-commerce field, there will be product category preferences, product category preferences, brand preferences, and so on.

The picture above is an example I drew casually. It is not difficult to draw an architecture, but it is difficult to understand the business logic and implementation methods behind each label. As for the algorithm, it can be discussed in many separate articles.

From the perspective of data flow and processing, user portraits contain a progressive relationship between superiors and subordinates.

Take the churn coefficient above as an example. It is modeled based on the user’s earlier historical behavior. The user's early historical behavior, namely the consumption amount, consumption times, login times, etc. within 10 days, is itself a label, which is obtained through the original detailed data.

The above figure lists the process of label processing and calculation, which is easy to understand. The top-level policy tags are for business implementation. Operations personnel form a user group through the combination of multiple tags for easy execution.

The larger the company, the more complex the user profile. A company focusing on content distribution has entered the new video field and now has two apps, so the structure of user portraits also needs to change. There are both content-related tags and video-related tags, which are parallel and related.

For example, user A is a heavy user under the content tag, but a light user under the video tag. For example, user B has not opened the content app for a long time and is at risk of churn, but is very loyal based on the length of time he uses the video app. All these things depend on flexible application. Of course, demographic labels such as name and gender are universal.

User portraits are platform-level applications, and many operational strategies and tools are built on their basis.

Based on marketing and consumption-related labels, new customers, old customers, user churn and loyalty, user consumption levels and frequency, etc., are all the basis of CRM (customer relationship management). Perhaps everyone is more accustomed to calling it a user/member management and operation platform.

Its function is to convert data-based labels into product operation strategies. Different labels correspond to different user groups and different marketing methods. The CRM structure will include various common channels for reaching users such as SMS, email, push, etc. It also includes CMS (content management system), through which executives can quickly configure activity pages, activity channels, coupons, etc., and drive data through marketing activities.

If Lao Wang's salad business grows, the operating platform will be built according to the structure shown in the figure. Lao Wang combines tags in CRM, monitors the data of new, old and lost customers with the help of BI, and then configures red envelopes, coupons, etc. through the CMS system, and then reaches out through short messages or Push.

A good user portrait system is not only a data ecosystem, but also a business and operation ecosystem. It is a complex cross-field. Due to limited space, algorithms and data products are not discussed in detail. I will talk about them again when I have the chance. I hope everyone can understand the core idea. If you have any comments or questions, please leave a message.

Thousands of uses, all in one mind.