Quick unlocking in one article: the whole process of the label system
Updated on: 41-0-0 0:0:0

The tagging system is the core of data analysis, which is related to business analysis, user portraits, recommendation strategies and other aspects. An accurate and efficient labeling system can provide rich materials and experience accumulation for subsequent analysis. In this article, we'll share how to avoid common mistakes when building a labeling system and provide a systematic approach to building and optimizing a labeling system.

The labeling system is definitely the most worthy of priority in data analysis work. Because it is related to all jobs, business analysis, placement analysis, user portraits, recommendation strategies, product operations, ...... It's all driven by labels.

If the labeling system is done well, there will be enough materials for follow-up analysis to accumulate experience. The labeling system is poorly done, not to mention the wasted effort, and it has not been relied on when I do in-depth analysis later.

So how do you do that? Let's share it briefly today.

1. Common mistakes in the labeling system

The most common mistake is that the label is just a basket of everything you put in.

  • When uploading articles, post a few: news, hot spots, products, ......
  • When the product is on the shelf, stick a few at hand: function, packaging, price......
  • When the event is released, post a few: name, gift, discount......

Not to mention, when tagging users, "high value", "potential", and "like XX" are casually pasted. Even "high value", "high quality" and "high quality" with similar names exist at the same time.

These vices can be seen everywhere when working on user portrait projects. People often show off to me proudly: "Mr. Chen, we are so good, we have more than 3000 ...... user tags"

At this time, you just have to ask him:

3000 tags, how many are there for business?

3000 tags, how much value is generated?

Ya was discouraged and dropped a sentence: I am still exploring how to apply it...... And then ran away.

Why? It's because these tags are just a bunch of dimensions lying in the database. If you want to use the business, you have to think first: what are the needs of the business, why should he use labels.

Second, to break the situation, start from understanding the needs

When building labels, there are at least 3 classes of completely different requirements.

1. Management

Quickly identify the need for value. Management is most afraid of seeing hundreds of pages of PPT reports. Labels can effectively distill the meaning of the business and identify the most critical factors.

Like what:

  • Mark users and distinguish between ABC users
  • Label channels to distinguish between stable/unstable channels
  • Mark your products to distinguish between before/during/after the life cycle

In this way, when the performance fluctuates, the management can see at a glance: Oh, it's the problem of XX place. Saves a lot of time.

2. Operations

Find what you need to be inspired to plan. Operations departments like to ask the most:

  • What do users like?
  • Where are the users?
  • Will users buy it?
  • If I don't buy it, will I send a coupon bank?
  • Coupons can't be used, but gifts can't be given?

In the final analysis, these problems revolve around the "5 elements of planning", and it would be nice to be able to clearly tell the answer to the business question through labels (rather than sparse data) (as shown in the figure below).

3. Frontline workers

The need for clear answers to questions. Frontline workers are much more aware of customer behavior and needs than data analysts who are thousands of miles away.

What frontline workers need is not: you teach me how to do it (and I can't teach it). But:

  • When a client comes to me, I can answer the question clearly
  • When the headquarters came to check, I was able to hand in the homework accurately
  • When there is a sales opportunity that is affordable, I can find the right tool

Like what:

When a customer consults a product, I can quickly find out the information

With the incentive policy, I can quickly find out how much I've achieved

There is an event going live, and I can quickly find out which guests can participate

Such clear guidance is the best tool.

After carefully understanding the business needs, you will find that there is no need for large-scale labeling. It is not necessary to use a large area of labels for business! The key to the success of the project is to provide fewer but more precise labels, cultivate business usage habits, and gradually build a complete system.

So, where to start?

3. Sorting, starting from the simplest

Note that the difficulty of implementing the above three types of requirements is different.

The easiest thing to achieve is the need for front-line personnel. Theoretically, as long as the activities, commodities, and articles that are frequently queried by the front line are labeled according to the standard format and put them into the library (as shown in the figure below).

But! This does not meet the needs of the frontline. Because searching for information on the front line is inherently difficult. For example, there are 30 activities launched at the same time this month, and there may be two or three of the most popular ones that are interested in the front line. And the two or three most popular ones, front-line personnel and customers, often give them nicknames, resulting in bizarre search keywords. If you directly open a tag library query, the usage rate is often low and the search accuracy is low.

Therefore, the tools provided to the front line can be further optimized:

  • Actively collect first-line opinions and optimize keyword search
  • Commonly used and popular tags are actively pushed to the front line to understand
  • Tools such as sales assistants are optimized to highlight key tags

In this way, the frequency of label use can be increased, and there is an opportunity to drive front-line efficiency.

Fourth, classification, the most important value of identification

The second type of easy to push is the label that identifies the value.

First, the definition of value is relatively simple, clear, and easy to do.

Second, the management of the value tag often looks at it and can brush the sense of existence.

Third, the fluctuation of daily diagnostic indicators can be used, and the appearance rate is high.

Even if you don't do anything else, you have to prioritize the value of these signs.

Common ones, such as:

  • Product value tags: explosive models, drainage models, profit models, matching models, and defensive models
  • Pipeline value tag: past input-output ratio, output volume + stability
  • User value tag: generated consumption + expected future consumption

(as shown below)

  • The value of the commodity, as long as the cost and pricing are calculated, can be calculated clearly, and it is the simplest.
  • The value of the pipeline can be seen clearly as long as the input and output of the channel are calculated and the past trend is observed
  • The user value, which has been directly counted, may only be the expected output, which requires some workload

(as shown below)

The only challenge here is to popularize the concept in management. It is very likely that the company has not done similar labels before, and there is no consensus within the management on "what is a high-value user" and "what is a high-quality channel", so it may be difficult to mention it for the first time. However, as long as it is not so ignorant that it doesn't even know what its own products, channels, and users look like, they can gradually promote the application of labels. After all, reducing the pressure of reading reports and focusing on core issues is the common appeal of everyone.

Fifth, explore, and gradually achieve accuracy

Of the three types of requirements, the most difficult to meet is the needs of the operations department. "Likes" and "Preferences" tags are very difficult to make.

  • 數據少:不是頭騰阿,採集不到足夠用戶決策路徑
  • Difficult to define: what does it mean to like? Does buying count as a liking? Does it count if you just look at it and don't buy it?
  • Unstable: I didn't like it, but as soon as you lowered the price, he liked it......
  • Difficult to see results: he likes it, but the promotion channel can't contact the customer......

Not to mention, even if it works, what proportion of advertising copy, promotional offers, and user needs will be accounted for......

Therefore, if you want to do this clearly, you must need to iterate many times.

The way of iteration is to move from more data to less data. Like what:

  • From the point of view of promotion, the people who like wool the most are sorted out first
  • From the perspective of consumption, the people who repurchase most frequently should be studied first
  • From a behavioral point of view, the people who interact with the most frequently should be understood first

These extreme groups are generally large contributors to performance, and there is a lot of data, so it is easy to summarize the rules. And when high consumption does not consume and high activity does not convert, the business department will be in a hurry to find a way, and it can further verify the accuracy of prediction in combination with business actions.

As for users with little data, they can first fry fish according to the fixed recommended route (as shown in the figure below) combined with business actions to test user needs and gradually improve the accuracy of prediction.

VI. Xiaoyu

Labeling is crucial, it is an important tool to quantify qualitative factors and provide value judgments, and it is a very basic construction. However, to do a label project, it must be combined with business analysis (for management), activity support (for operation), and system tools (for the front line), and you can't dedicate yourself in obscurity. Otherwise, everyone thinks that you can refine a furnace of elixir in a muffled voice, and if you don't participate in the process and don't use it, you will definitely be disappointed at the end.

This article was published in Operation School by the author of the operation faction [Down-to-earth Teacher Chen], WeChat public account: [Down-to-earth Teacher Chen], original/authorized Published in Operation School, without permission, it is forbidden to reprint.

The title image is from Unsplash and is licensed under CC0.