Many students are excited when they hear about user segmentation. There are countless articles on user segmentation on the Internet. However, in actual work, after completing the segmentation, they are often asked a question: "So how much can it improve performance? If it can't improve performance, why do you segment it?" The questions were so sharp that many students began to doubt their own lives: What is the point of grouping? How can we break the impasse? 01 Pain points of user segmentationThe biggest pain point is that clustering itself cannot directly generate value. No matter what model is used for clustering, the final result is just a data label. The problem is that no user will pay for this label. The reasons why users pay are always:
And it is the result of the combined influence of these factors. It is 100% impossible to expect a model to predict all of the above. That is equivalent to asking the algorithm engineer to be accurate: Zhang San will open the mobile mall at 10:19 on July 29, 2021, and he just wants to buy a Dafeng brand electric fan with a regular price of 200 yuan and a discount of 20 yuan... In this way, it is better to hire a fortune teller. Knowing the key to the problem, the idea of breaking the deadlock becomes clear: give up the naive idea of using a powerful and invincible general model to solve all problems in one shot. Start from the factors that affect user consumption, divide them into modules, and break them down one by one. 02 The key to improving performanceThere is an inherent logical relationship between the factors that affect user consumption (as shown in the following figure): Users have needs and happen to be exposed to a certain brand, and then think about whether the solutions provided by this brand can meet their needs. Therefore, from the big category, it can be broken down into three modules:
Most companies are not like Tencent, Amei, or Didi, and do not have a monopoly position. Therefore, the biggest difficulty lies in the middle layer: contact channels. If users cannot be effectively contacted, front-end demand insights and back-end solution matching are all empty talk. Therefore, if you really want the user segmentation model to be effective, you must first break through the contact channel. Segmenting user contact behaviors and distinguishing the contact channels suitable for different users can directly improve the implementation effect. (As shown in the figure below) Most companies are not able to spend a lot of money, so resources should be used with caution. Therefore, the second biggest difficulty is to distinguish the user's ability to pay and calculate how much resources are worth investing. Note: Many students are lazy and like to calculate the cumulative consumption. They think that high cumulative consumption means high ability to pay, and it is worth investing resources. This is lazy politics. Because historical consumption does not represent future consumption, some categories may have user loyalty and will be repurchased. Some categories are simply one-time business. Therefore, it is recommended to split several tags and classify them separately (as shown below): Most companies are not platform companies. As long as there is a transaction, it doesn’t matter whose product it is. Therefore, when matching solutions, some products are destined to be uncompetitive and can only be matched to customers with high loyalty; some products are just for the purpose of exchanging price for sales and attracting traffic. Therefore, the solution is closely combined with the user's value, needs, and touchpoints (as shown below): With the above preparations, you can start building blocks and assembling a layered model that can improve performance. 03 User grouping assembly methodAfter the analysis in the previous step, prepare the basic modules (as shown below) and start assembling. Step 1: Based on user payment + brand stickiness, confirm the basic segmentation strategy to determine how much investment each user is worth and the expected goals to be achieved. If you want to further improve the efficiency, you can add two models here: Model 1: Natural consumption prediction. Through modeling/analysis, we can find users with a high probability (70% or more) of natural consumption, and then reduce the investment in these users, which can further improve the ROI of resource allocation in the segmentation strategy. Model 2: Price sensitivity analysis. Through modeling/analysis, we can find users who are highly price sensitive. Then, depending on the current business development, if there is a lack of sales volume, we will invest in these users; if there is a lack of profit, we will temporarily give up, which can further improve ROI. Step 2: Determine the contact strategy based on user preferences and contact channels. Note that for key users, you can use carpet bombing tactics, with store clerks, APP information, and text messages taking turns until 100% coverage. For marginal users, you need to look at the contact probability and evaluate whether it is worth investing in contact. Step 3: User grouping + user preference, determine products and promotion plans. User preference determines which products to match, and the grouping strategy determines whether to offer discounts and how much discount to offer. Step 4: Confirm the content. After the above three items are confirmed, the content will be a natural outcome. Because the products, discount plans, and contact channels pushed to each group of people have been determined, the content only needs to create corresponding pictures, texts, and videos based on the hot spots of different contact channels. The above combination is a complete operational implementation plan. Compared with simply giving a prediction result of whether a customer will consume or churn, this combination has several obvious advantages:
Because resources and contact channels have been locked in advance, the only control variables are content and products. Later optimization is to see how to change the content to increase clicks and how to change the products to increase conversions. This will also make the post-analysis process clear: through multiple rounds of product and content testing, we can find out the upper limit of user operations under the existing product structure and which method is more effective. This will not only improve short-term benefits, but also lay the foundation for promoting product upgrades in the long term. 04 SummaryEssentially, the reason why this combined clustering model is needed is that when the analysis goal is to "improve performance", there are too many factors that affect the results, and if they are not carefully considered, it is very likely that all will fail. If you want data to be effective, you can never expect to "one model to rule the world", and the right way is to mobilize resources from multiple parties. This grouping idea and modeling process is actually the overall idea of CDP operation. Author: Down-to-earth Teacher Chen WeChat public account: Down-to-earth Teacher Chen (ID: gh_abf29df6ada8) |
<<: The "Kuang Biao" is not only about acting, but also about the lines and copywriting
>>: From the fact that Keep medals sold for 500 million, I summarized two cognitive gains:
Digital marketing has the advantages of high accur...
In order to optimize the product structure and enh...
Amazon is the world's largest cross-border e-c...
The author of this article analyzes the new consum...
Shoptop announced that in order to address the pai...
Shopee sellers who want to get explosive sales in ...
The author of this article uses the classic framew...
In the 99 promotion event held yesterday, in the f...
Shopee is one of the largest e-commerce platforms ...
In addition to some stores, Alibaba International ...
This article deeply analyzes the business strategi...
Amazon is one of the world's largest e-commerc...
How should self-media realize cash? The most direc...
The current Internet business model has spawned a ...
Nowadays, e-commerce merchants generally have some...