Editor’s Guide: Do a good job in data analysis, help clear the portrait of the user, find the core attention of each industrial user, to make refined operations to enhance the user’s repaction. But in the actual situation, many people don’t know how to do user portraits. The author of this article from project practice, combined with the case for the failure of user portraits, and shared with you.
The last case of sharing an algorithm model failed, see “Algorithm engineers who don’t understand data analysis”, there will be many students, I have aroused a lot of classmates, I have a classmate: “There is an example of the failure of the user portrait project, also shared Down.”
A: Failed user portrait project, simply óÀ bamboo book! One grabbing a big hand. Today, our system explains.
First, the user portrait failed sign
Everyone is not confused by himself: “Is the user portrait?”
If there is, congratulations, your project is doing a street, it is so intuitive!
Of course, the more general rolling mode is that when you start to make a user portrait, the business unit shook the brain: “We want to use user portraits, detailed understanding of users, such as user gender, age, geographical, prevail, consumption habits, …… This way we can refine the decision. ” Then the data department took a few months, and the 300,000 user label was played, and the leadership reported: “Our user portrait big data construction has taken considerable progress”.
At the first report of the project, the data department is proud of:
Our users have a proportion of men and women 6: 4 South China accounts for 30%, and 25% of East China purchase A product accounts for 50%.
The business unit has a white eye to:
I know! Our users are like this! Do you do this?
Of course, there is also a more miserable, that is, you posted a “loyal user” label, the business side said: Oh, since it is so loyal, you will not do it. The result is not logged in! You posted the label of the “A Products”, the business party pushed A products, but did not buy it! The business is angry and rushing to find the account: “This user portrait is not accurate!” So the project is completely cool.
Tucao, where is the problem?
Second, the user portrait project failed surface reasons
Reason 1: Confusing the past and the future
Q 1: A user bought apple yesterday, bought Apple before yesterday, bought Apple before, he didn’t buy Apple today?
Q 2: One user bought soy sauce, chicken wings, cola, did he still need to buy bamboo stick to barbecue?
Think about a second, don’t think about a second, everyone knows, the answer is: not necessarily, not necessarily, not necessarily. Buying apples in a row, it is possible to represent his favorite apple, and it may have already bought a lot, so don’t buy it. Buy soy sauce + chicken wings + cola, it may be to go to barbecue, or it may be a Coke chicken wings.
The past behavior is not equal to future behavior, and future behavior needs to be predicted. Regardless of the forecasting method is based on business logic, or based on the calculation of algorithm models, data analysis and test verification is required. Only a stable prediction method can be adopted.
However, when making user portraits, the business is often confused. It is often a big pile label to the past, and there is no concept for future forecasts, and there is no in the forecast analysis. When you look at the user portrait report, or when you set the push rule in the CDP, you want to be considered: I bought it later in the past. Finally, it is not predicted, but the pot is given to the user portrait system. The result is natural tragedy.
Cause 2: Confused behavior and motivation.
Ask a simple question: One user has been purchasing products in 1 day in the past 30 days, is the user is our product love? If there is 2 days, 3 days, 4 days … If there is 30 days? It’s time to buy every day in 30 days, it is definitely a lover!
A: Not necessarily. It’s time to buy every day in 30 days, you can manage him called “high frequency buyers” because the purchase frequency is indeed very high. But is it that people are very love to use our products, not necessarily, because you don’t know if he loves to love, and even don’t know if he uses it.
The purchase frequency cannot be directly equivalent to the user. Users love or don’t like it, need more dimensions of data to analyze, and the results have a certain probability of stability, in order to call it.
Similarly, many companies, businessfire and data analysts, treat this “love user” and other nouns are very casual, rough, basically use consumption amount, login frequency, etc., even if “like” “Love”, low, “the edge” “try”. The result of making it, naturally there is no accuracy. Needless to say, there is a problem, such as the recommended product, no one is bought, and it is on the user’s portrait head.
Cause 3: Confused causes and results.
Q: The user who spends more than 10,000 yuan, she has purchased more than 5 times, so the user will purchase 5 times, the user will spend 10000, not right … Of course not right. However, the business is often so dry! Holding a high-consumption user behavior, a low consumption, thinking that as long as the consumption is low, it can become a high consumption. Also beautiful, name: “Magic Numbers”.
It is very likely to be from the source, high consumption and low consumption are two types of people. It should be made through in-depth analysis.
From the surface, the user’s portrait failed reason is: Heavy data, light analysis. Over-input energy refinement has occurred, label too much of factorial labels. In enough for predictive investment, it is not enough to analyze causality. Finally, it is judged that the business is full of business. You ask him to push products / activities based on these labels, he is:
I feel that he bought so many times. He surely bought I felt that he bought related products. He sure I felt that he bought A. How can I buy B?
After reading the user portrait, I took my head and read the report to take my head. It is a head, and there is no essential difference, thank you. Based on past data, it is easy; the precipitation has predicted, and it is difficult for user labels. Not only requires depth data analysis and modeling, but also need repeated, multi-wheel, contrast testing. It’s not a chance.
Therefore, in the business department, I have been very understanding that the data department is happy to announce the “30,000 labels”, the root has been buried.
However, it is also a scene of the business expectation value too high + data preparation, which is a user portrait class project, which is more likely to flop the street than the data model item?
Third, the deep reason for the failure of the user portrait
Deeply see that because the data modeling is difficult, the business parties cannot participate in the intermediate process, and they can only be arrested in the results. To model the model, as long as you don’t die, don’t close the door, take the initiative to reduce your business expectation, it is to avoid problems. Therefore, the modeling project failed, basically the result of the blind horse riding.
But the user portrait project is just the opposite: the business partner thinks it is very understandable! Data’s little brother is also very understanding! Almost all the business parties, the user portrait, will say this: “For example, I know that the user is 24 years old, women, I will push an XX product to her.”
Everyone thought: I understand myself, it is a few! Give me me. So the business kept the data to make the past data, then fine, fine, and the data rushed all the way on the label road. The most important prediction, analysis, and experimental three-piece set no one.
Of course, this label based on past data is useful for some departments – for customer service, supply chain, and logistics. For example, customers receive a customer complaint. “After the sale, the master has not gone yet!!!” If you don’t play the label, the customer has to turn several tables, confirm: The customer is buying products, is a product, the product body, The details of the master of the time. The light confirmation process will give the customer half a dead. With labels, the brush can be positioned to the problem, which can greatly enhance the customer experience.
But the tragedy is that this kind of usefulness is only to make the operation, marketing, planning, design, etc. need to make brains, creative, and want to strategically. Strong them “I am really powerful, it is a few!”
So the tragedy is constantly sourced. If the modeling is a blind horse, the user portrait project is riding an electric car. The mobile phone is retrograde the red light – the electric car thinks that he is a car, and the motor vehicle is self-riding 666.
In order to avoid this problem, Teacher Chen often uses this trick. When the business party is open, “If I know 24 years old, women, I will push a product”, I found it directly from the database, 24 years old women, have you bought A, then go to the business side: ” No need for user portrait, I now tell you now, the purchase rate is 12%. You also on the user portrait, let your little brother run the number according to the rules. “
At this time, the business parties who relieve a little, they will immediately wake up here, indicating that this simple splicing is not based on the factual label. This will take a stable after the project.
However, students should use this trick, your corporate environment is not suitable for this hard-style Ha, in short, everyone understands the key to the problem. The key to the problem is that the simple fact label prediction ability is too bad, and the insight is too much. Not enough to meet the needs of operation, planning, sales, and marketing. A large amount of data + in-depth analysis is the way to solve the problem.
This article is over, it is estimated that many classmates will be curious: CDP is awkward. Teacher who is interested in paying attention to the ground, we will share it. Please look forward to it.