How to build a user portrait that can be implemented?

User portrait refers to a labeled user model abstracted from information such as the user's basic attributes, user preferences, living habits, and user behavior. Each label and label weight is a vector that represents user preference. A user can be understood as the sum of multiple preference vectors (labels). – An incomprehensible explanation from a certain website

Speak human language! ! ! User portraits are based on the various data you leave on the Internet. These data are processed with or without your knowledge to generate a group of labels that depict your interests and preferences. This is the user portrait.

So the next question is: What are the usage scenarios of user portraits?

1. User grouping operations

Group operation is the most commonly used scenario for user portraits. Different user groups can be screened out by filtering user portrait labels. The platform can be configured through push or pop-up windows to implement refined operations for different user groups.

For example, we can define users whose purchase amount in the past 30 days is greater than 500, whose active days are greater than 10 days, and whose last active interval is within 5 days as a high-paying potential user group, and then carry out refined operations for these high-potential users.

2. Automated Reach

As the stratification becomes more and more refined, the granularity of the user group will become smaller and smaller, and finally it will be as small as an individual. We push or pop-up messages to each individual, but the operational efficiency is still too low. Automated outreach based on user portraits comes into play. For example, for the high paying potential user group we created in the previous step, we send a coupon worth 40 yuan off for purchases over 200 yuan through App Push.

3. Personalized recommendations

Building recommendation systems, search engines, and advertising systems based on user portraits can effectively improve conversion rates. Let's first look at the user preference table storage table for user portraits (user portraits have many such tags, which we will explain in detail below):

When we purchase a product with labels 1, 2, and 3, there will usually be a cross-selling scenario on the purchase completion page. We use the labels and weights in the user preference table and the collaborative filtering algorithm based on user similarity or product similarity to recommend other products that the user may like.

4. User statistics & industry research

After classifying users according to their attributes and behavioral characteristics, statistics are collected on usage and distribution under different characteristics, and the distribution characteristics and trends of different user portrait groups are analyzed.

In addition, user portrait analysis can also be used to understand industry dynamics, such as people’s consumption habits, consumption preference analysis, and consumption difference analysis of product categories in different regions. For example, we often hear Jack Ma use some labels to say what people in this place like to buy? Why do you like to buy? What models do people in that place buy?

When I just graduated from college, I started working in an Internet company as a data operator. One day, my boss suddenly gave me a task: to make an APP user portrait report. I was a little confused at the time. First, I had never come across user portraits before, so the concept was not very clear to me. Second, after receiving the task from my leader, I only knew that I had to make a user portrait, but I had no idea what this user portrait would be used for.

In this situation, I consulted various materials, asked classmates and friends, and also found some advanced drawing tools. Finally, I completed the task and got the company's recognition. However, it is not clear what value user portraits have brought to the company and what role they have played in operational decisions. This is actually a typical case of following the crowd and doing it for the sake of doing it.

This kind of face-saving project is very common. You can look at the situation of your own company and find out which situation applies to you. Because everyone is talking about user portraits, many leaders have also begun to require user portraits. Others have them, so we cannot do without them. However, it is not clear what problems user portraits are intended to solve, so naturally there is no possibility of implementation, let alone generating value.

So in order to solve this problem, we must think clearly at the very beginning: do we really need user portraits? Just think about these three questions and you will get the answer.

1. Are there scenarios in our business where user portraits can be used? For example, user stratification, intelligent reach and personalized recommendations.

2. If there is a usage scenario, is it necessary to implement it through user portraits? User profiling itself is a very huge project, which is very time-consuming and labor-intensive. Do we have any other alternative solutions?

3. If we must do this, after we label and stratify users , do we have corresponding product services or operation plans? Without corresponding actions after stratification, it is still difficult to implement a closed loop for user portraits.

If you don't think through the above issues and just blindly create user portraits, the result will most likely be that the user portraits are far from the business, have no practical value, and contain no substance, and are ridiculed by the business department as a "big and useless" chicken-rib product.

Therefore, when planning a user portrait platform, you must have purpose and a sense of scenario. You cannot just do superficial work without paying attention to the actual application value.

One point that needs to be emphasized here is: It is not because I have user portraits that I can drive and improve my business. It is only in order to drive and improve business that we need to build user portraits. It is an easy mistake to make by reversing cause and effect!

If you haven’t thought through the above issues, don’t start development. If you have thought through the above issues and still want to do it, then we must plan carefully and think about how to make user portraits valuable.

Now that we have decided to create user portraits, the next thing to think about is: how can we implement the business? How can we create value?

When we first started to create user portraits, the business department shook their heads and said: "We need to have a detailed and in-depth understanding of users based on user portraits, such as their gender, age, region, preferences, consumption habits, ... so that we can make refined decisions." Then the data department worked hard for several months and added 30,000 user tags. They proudly reported to their leaders: "We have made great progress in building our user portrait big data."

Then at the first project report meeting, the data department proudly said:

The male-female ratio of our users is 6:4. 30% are from South China, 25% from East China, and 50% purchase product A. Blah blah blah…

The business department rolled their eyes.

I knew it! Our users are all like this! Then what? What's the point of you doing this? …

Of course there are worse ones.

It means that you put a label of "loyal user" on them, and the business side says: Oh, since they are so loyal, we won't do anything about it. As a result, he didn’t make any purchases or log in next month! You put a label on yourself as a "lover of product A", and the business side promoted product A, but you didn't buy it! The business side came to settle the score angrily: "This user portrait is not accurate at all!"

As a result, the project was completely shelved. What exactly is the problem? You don't understand the business needs at all. The business wants to eat simple spicy and sour potato shreds, but you are talking to them about the 108 processes of making Buddha Jumps. Or even though you understand the business needs, you don't label it correctly. What I want is spicy and sour potato shreds, not green pepper potato shreds...

Therefore, in order to implement user portraits, the following steps are particularly critical.

1. Clarify business requirements

When planning user portraits, you must have purpose and a sense of context. You cannot just do superficial work without paying attention to the actual application value. Let me emphasize again: It is not because I have user portraits that I can drive and improve my business. Instead, it is necessary to build user portraits only when there is business demand!

The first step is also the most critical step. You must figure out what the business needs are. What is the problem to be solved?

For example, a content-based community is preparing to launch a knowledge payment module in the near future, and use this model for commercial monetization. It wants to recommend precise content to precise people through user portraits, thereby promoting paid monetization. Based on this, the business goals and problems to be solved can be sorted out as follows:

2. Tag selection

Which tags should I select? Why were these labels chosen? Why not other tags? What are the misunderstandings about choosing labels?

We all know that there are labels for basic user attributes, as well as various user behavior labels. We need to continuously subdivide and improve the label system within this large framework. But why do we choose these labels? For example, in the user's purchasing behavior label, why do we choose the label of new or old users? Because our store will give red envelopes to new users who have not made any purchases to guide them to make purchases, we need to distinguish between new and old users. For old users, why do we choose the most recent purchase time, purchase frequency and transaction amount as labels? Because RFM can be used to stratify user value and then carry out refined operations.

We will introduce the specific steps of building the label system in the third part, so we will not expand on it here.

3. Develop operational strategies for users with different profiles

Different labels divide users into different user groups, but only by formulating targeted operation strategies for different user groups can user portraits be implemented and valuable.

For example, in the RFM model, some students are not clear about how to determine the threshold of the user transaction amount M and then divide it into high and low? Some students are struggling with the question of whether spending 1,000 yuan is high or 10,000 yuan is high. Why don't they think about it the other way around? If the threshold is set to 1000 or 2000, what is my corresponding strategy? What’s the point of dividing things into high and low if there is no corresponding strategy? At this time, you may want to ask your colleagues in product and operation. You will find that in order to promote price-sensitive low-unit-price users, operations have a 200 yuan discount activity for purchases over 1,000 yuan. For sponsors with strong purchasing power, there is an exclusive customized diamond membership card for purchases over 2,000 yuan, and they can enjoy super VIP treatment. Starting from the end in mind, isn’t it easier to implement user portraits at this time?

Now that we have identified the business needs and strategies for different user profiles, we can start creating user profiles.

The construction of the network is mainly divided into label system construction and label weight calculation. In simple terms, it is about which labels are used to represent users and how much users prefer each label.

1. Label system construction

The most direct way to understand user portraits is to label the portrait information. Information labeling is a method used by user portrait systems to depict the full picture of users using information technology, and it is also one of the most core links in the user portrait system. We can characterize and understand users' interests and preferences by building rich tags and obtaining relevant tag data.

1) How to build a label system

There are two main approaches to building a tagging system:

Method 1: Systematic construction of structure (commonly used). By dividing label categories and dimensions, a dimensional system is constructed from the perspective of describing the complete dimensions of users. This division method has a clear structure and strong logic, and can comprehensively sort out all information dimensions. However, it is difficult to implement due to limitations on actual data.

Method 2: Build scene effects. Combined with the actual needs of population targeting, the user's psychological needs and tendencies are described through the user's behavior records in different industry fields or platforms. This method is highly purposeful and has a relatively accurate crowd targeting. It is easy to integrate with actual application scenarios and land on specific business goals. It combines well with actual data, but labels will fluctuate with changes in consumption trends, entertainment hotspots, etc.

2) Design ideas for label dimension

The label dimension design not only needs to be clear and intuitive, but also needs to consider multiple scenario-based usage requirements, while taking into account multiple product operation requirements and commercial launch requirements.

Label design requires a forward-looking design of label usage based on an understanding of the business side’s planning.

The sources of portrait data generally include: user surveys, user behavior data acquisition, client/server data content reporting, third-party data platforms, basic data, and crawled third-party data.

ps: Here we need to pay special attention to the following two aspects:

Label granularity: Too coarse granularity is not conducive to operational use and promotion. It is easy to deviate from the business itself and over-refine behavioral data, resulting in information loss. Too fine granularity will lead to low label coverage and coupled operational business promotion.

Label data: The acquisition of label data is directly related to its actual use value and needs to be carried out within the scope of available data. It is difficult to expand the data source, and it is usually necessary to prioritize data feasibility statistics.

3) Basic framework of the label system

The portrait labeling systems for different businesses are not the same, and we need to refine them specifically. There is a relatively simple way: we can first find some general portrait tags, and then add business portrait tags based on actual scenarios and needs. The label system obtained in this way will be relatively complete and can be adjusted and optimized in time with business changes.

General portrait label system (reference):

Business portrait label system (taking an e-commerce company as an example for reference):

Labeling user portrait information according to the above framework can better obtain relevant user portrait data based on actual needs. However, it should be noted that the analysis of product user portraits does not require the use of all label data, and the more complete the label system is, the more difficult it is to implement. Moreover, the greater difficulty lies in how to accurately describe user characteristics. Because only when the description of user characteristics is more accurate, the user portrait we get will be clearer, and the more help it will be in the actual application process. Therefore, how to accurately calculate the weight of the user's label becomes a top priority.

2. Label weight calculation

The user's preference on different tags is reflected by weights. The higher the weight, the stronger the user's preference on the tag, and vice versa. Moreover, this weight will change over time, and the calculation of label weight is mainly through the TF-IDF algorithm.

1) TF-IDF algorithm idea

The weight of a user tag is determined by the importance of the tag to the user (TF-IDF weight) and the importance of the tag to the user in terms of business (business weight).

That is: user tag weight = business weight * TF-IDF weight.

The TF-IDF weight is calculated by TF-IDF, and the business weight is determined by the user's behavior on the tag.

That is: business weight = behavior type weight * behavior times * time decay

2) Simple understanding

The importance of a tag to users is expressed through different behaviors. Different behaviors have different levels of difficulty. For example, for e-commerce users, the behavioral difficulty is payment>collection and purchase>sharing>browsing>clicking. Different behaviors will have different weights. The more difficult the behavior is, the more you like it and the higher the weight. Similarly, the more times you perform the behavior, the more you like it.

The rarer the tag is for this user, the more he likes it. The degree of liking will gradually decrease with time. The tag weight is calculated using this formula.

3) Behavior type weight

Different behaviors of users, such as browsing, clicking, searching, collecting, sharing, ordering, and purchasing, have different importance to users. Generally, a basic behavior weight is defined based on business experience or using the hierarchical analysis method.

4) Number of behaviors

The number of behaviors here indicates the number of times each behavior occurs.

5) Time decay

Time decay means that user behavior will gradually weaken over time, and user preferences will gradually weaken. When building functions related to time decay, we can apply the mathematical model of Newton's law of cooling.

Newton's law of cooling: The temperature F(t) of a hot object decays exponentially with the increase of time t. The temperature decay formula is:

F(t)=T*exp(-α*t). T: Initial temperature α: Attenuation constant, also known as cooling coefficient, is a self-defined value that can generally be calculated through regression t: Time interval

How to calculate the cooling coefficient?

The cooling coefficient is a self-defined value, which can generally be calculated through regression. For example: the initial preference is set to 1, and the preference after 1 day is 0.85, that is, 0.85=1*exp(-α*1), and α=0.16 is obtained, which is the cooling coefficient of the label.

Here we use Python language to simulate this cooling curve:

import numpy as np import matplotlib.pyplot as plt import matht = np.arange(0,100)

plt.plot(t,1*np.ma.exp(-0.16*t))

plt.title('Preference Cooling Trend')

plt.ylabel('Preference')

plt.xlabel('Time (days)')

plt.show()

6) Take an example

For user "Xiaomei", weight calculation for the label "lipstick": Assume that we previously defined the cooling coefficient α = 0.16, and based on business experience or through the hierarchical analysis method, assume the behavior type weights: click (0.1), browse (0.2), share (0.5), add to favorites (0.6), and pay (0.9).

Xiaomei's daily behavior chart:

2021-05-01:

2021-05-02:

2021-05-03:

The daily weight of user "Xiaomei" for the tag "lipstick":

2021-05-01: 2*0.1+2*0.2+3*0.6+1*0.5+1*0.9=3.8

2021-05-02: 3.8 * exp (-α * 1) + 1 * 0.1 + 1 * 0.2 + 2 * 0.6 + 1 * 0.5 + 0 = 5.06

2021-05-03: 5.067718*exp(-α*1)= 4.32

This makes the calculation clear.

After the user portrait is initially formed, it cannot be directly handed over to operations and business personnel for direct use. It is also necessary to evaluate the accuracy of the user portrait and continuously iterate the user portrait after delivery to obtain a more accurate user portrait.

There are three main evaluation methods: logic verification, A/B test, and user feedback.

Logical verification: also called cross-validation. In a complete user portrait labeling system, some labels often have some correlation. For example, the longer the user's cumulative online time, the higher the order volume will usually be. For example, among the user group that purchases 3C products, the number of male users is usually greater than the number of female users. In addition, if the company purchases data from a third-party agency, it can also be used for cross-validation.

A/B test: also called grayscale test. Taking the above-mentioned loyalty as an example, ensure that the traffic of the control group and the experimental group is the same. For users in the experimental group, implement operational strategies to improve loyalty (promotional activities, point rewards, etc.). If the loyalty of users in the experimental group is improved compared to that of users in the control group, it can be considered that the user portrait is more accurate.

User revisit: The simplest evaluation method, such as the user portrait system, defines 100,000 users as low-loyalty users; at this time, 1,000 people are randomly selected from them and handed over to customer service for revisit. Based on the follow-up results, you can determine whether the user portrait results are accurate; you can even perform text mining on the follow-up results to form a word cloud and view the proportion of negative words.

This article is very long, so I have to make a summary at the end. Unlike other articles that simply explain how to build user portraits, this article systematically and comprehensively analyzes what user portraits are? Is it necessary to do it? How to do it? It explains step by step: What is user portrait? → In what scenarios will user portraits be used? → Do we really need user portraits? → How can user portraits be implemented? → Steps to build a feasible user portrait → User portrait effectiveness evaluation and iteration.

I hope we can ask more why questions when making user portraits? Knowing why you should do something is far more important than knowing how to do it. Only in this way can you create a practical and valuable user portrait.

The above is the content of the data analysis thinking - user portrait part. More articles on data analysis thinking are being updated continuously. Please stay tuned. If you think it is good, you are welcome to share it~