The common concern of hundreds of thousands of Internet practitioners!
Author: Qin Lu, the author authorized reprint morning reading class.
Public Account: Qin Lu (ID: TracyKanc)
Editor: Dva
What are user operations?
It aims to maximize user value by increasing activity, retention, or payment metrics through a variety of operations. There is a classic framework in user operations called AARRR, which is Add, Retain, Active, Spread, Monetize (this has been covered in historical articles).
However, it’s not two simple steps to go from active to profitable. If users are active when they open the product, will the business model be profitable? Excellent user operation system, should be dynamic evolution.
Evolution is a pyramidal hierarchy of user groups, with dependencies at the top and bottom.
First, the state of the user base is constantly changing. Take e-commerce as an example. They register, download, use, recommend, review, buy, and pay for products, as well as cancel, uninstall, and lose them. From an operational perspective, we guide users to do what we want them to do (in this case, pay for it), which is called core goals.
Of course, the core goal is not achieved overnight. Users have to go through a series of processes.
Not all users will follow the steps we envision, and the steps will be funneled. We view the whole process as an evolution of the user community.
The figure above is a typical bottom-up presentation that Outlines the desired behavior of the user community.
Since the user group is no longer a simple whole, operators can no longer operate rudely at the same time, but need to operate according to different groups. This is called a refinement strategy, or user layering.
Its greatest value to operations is the use of different strategies through layering.
New users: I want them to download the product, and the common strategy is new user benefits;
Download user: I want them to be able to use the product, at this point should be a novice guide, let him familiar.
Active users: I hope to deepen the frequency of their use of the product, so the operation personnel should continue to operate, solidify users’ use habits, and be interested in the product content;
Interested users: I want them to make payment decisions, buy products, operate with different promotions and marketing tools;
Paying customers: This is my target audience, and I want them to stay that way.
Different levels of users take different means. Operations are also limited by resources. When we have limited resources to invest, we tend to choose the core group, namely the paying users mentioned above. Because according to the 80/20 rule, only the core group contributes the most value.
A typical example is in the game company, there will be a special human customer service or even telephone line to serve RMB players, sweet voice. Average players are probably auto-regen forever.
I think you know about layering, but how do you divide?
In fact, there is no fixed way of stratification, only according to the product form to set up the system according to local conditions. But it has a central idea:
According to the index, because the index is a clearly measurable standard, far better than the operational personnel’s empirical intuition.
The diagram above shows a simplified hierarchy of game users, each of which is quantifiable. In order to upper and lower users clear, groups should be as independent as possible, that is, calculation of RMB players, should be the local tyrants excluded, ordinary players, should be the results of the exclusion of the two layers, so that the operation of the targeted is strong.
Operators can then build layered reports from this, using data trends to develop ways to improve the data.
Next, let’s think about the form of user stratification of Zhihu. Its core is big V production content? Or do more people participate in Live to make money? It’s hard to decide. In fact, a lot of operating systems, user stratification is a two-tier structure.
It aims at two complementary cores, thus forming a double pyramid layer.
Under this structure, its core users are both big VS in content production and loyal fans in consumption, representing two types of business strategies.
Content production direction: early use of invitation system to obtain outstanding talents in various industries, maintain relationships through operation personnel, and encourage the production of content. The product mechanism will also motivate big V to create and produce better.
Content consumption direction: it is to find out the content interests of ordinary users, guide them, and cultivate their paying habits. Increase the exposure of Live, value and e-books, and design various coupons to promote the use of users.
This double pyramid structure brings content producers and consumers together in a virtuous circle across the platform: the Big Vs create content, attract ordinary people, ordinary people pay for it, and the big Vs reap the benefits.
Double pyramid user layering is not uncommon.
Take e-commerce, as we know it, where there are buyers and sellers.
The way the buyers operate is familiar, but what about the sellers? Shop opening tutorial, seller university, shop decoration, exposure display, shop background, all kinds of auxiliary products… The operation also needs to help the seller grow, so the seller can also be divided into ordinary sellers, senior sellers, big customers, super donors these levels.
Is O2O two-tier structure?
Is, of course.
Online refers to users, and offline refers to all kinds of offline or service entities. However, these sellers are mainly responsible for sales promotion and maintenance by marketing personnel, but we can still operate with a hierarchical idea. Others include Internet celebrities and people with live video streaming, big VS and grassroots weibo users, enterprises and employees with recruitment apps, and so on.
The form of different products will be different. Different stages of the same product can also be stratified by different users. In the early stage of a product, the target of user stratification is more users and Kols, and in the later stage, it will be closer to the commercial direction, which requires flexible stratification for operation.
User layer, generally four or five layers structure is ok, too many layers will become complex, not suitable for the implementation of business strategy.
Is there only user layer in the user operation system? Not exactly.
User hierarchy is a top-down structure, but the user community can not be completely summarized by structure. Think about it for a moment. We have defined the group of paying users based on whether they pay or not. However, there are also differences between these groups.
If you continue to add layers, conditions become complex and do not address business requirements.
So we use a horizontal hierarchy of users. The groups within the same layer continue to be segmented to meet the needs of higher refinement.
How to understand user clustering, we take the following case description.
There are significant differences between men and women in consumption-focused products, which are two distinct groups. The core goal of clustering is to improve the operation effect and maximize the value of the operation strategy. In e-commerce products, it is normal to distinguish between men and women, but in tool-based apps, it may not be necessary.
And that’s what I’ve been talking about, layering and clustering, building systems based on product and operational goals.
Next comes the practical application of clustering.
RFM model is a classic method in customer management. It is used to measure the value and profitability of consumer users. It is a typical cluster.
It builds consumption models based on three core metrics of charging: amount spent, frequency spent and last time spent.
Monetary consumption: The Monetary consumption is the golden indicator of marketing. According to the 80-20 rule, 80% of an enterprise’s revenue comes from 20% of its users, which directly reflects the contribution of users to the enterprise’s profits.
Frequency is the number of times a user makes a purchase within a limited period of time. Users who make the most purchases are more loyal.
Recency: measures the loss of users. The closer the consumption time is to the current user, the easier it is to maintain the relationship with the user. A user who spent a year ago is worth less than a user who spent a month ago.
Through these three indicators, it is easy to construct a coordinate system describing the user’s consumption level and form a data cube with three indicators:
In the coordinate system, the two ends of the three axes represent the consumption level from low to high, and users will fall into the coordinate system according to their consumption level. When we have enough user data, we can divide it into about eight user groups.
For example, if a user performs well in consumption amount, consumption frequency and last consumption time, then he is an important value user.
If a significant value user last spent a long time ago and has not spent again, he becomes a significant retention user. Because they used to be so valuable, we don’t want to lose them, so the operations and marketing people can call them back.
Different quadrants in the figure correspond to different consumer groups. Do we simply want to operate as a whole, or do we treat people differently?
This is the RFM model, which has been frequently used in traditional industries but can be transplanted to our use in a consumer-oriented operating system. It is not only the core of CRM system, but also the core of consumer user segmentation.
There are two main clustering methods of RFM model.
One is to set up indicators, which are used as the basis for classification, similar to user stratification.
The determination and establishment of indicators requires the experience of operational experts:
What counts as high consumption frequency, what counts as low, and how much consumption amount counts as valuable, these are all knowledge.
And it needs constant revision and improvement.
The diagram above is a simplified partition, which is more complicated in practice because the indicators are not necessarily representative. Most of the data related to fees will be distributed in a long tail. 80% of users are concentrated in the low-frequency and low-amount range, while 20% of users create most of the revenue. This is the difficulty of division.
Indicators are generally divided by the median, first quartile and third quartile of descriptive statistics.
The other is to use algorithms, through data mining to establish user groups, without manual division. The most common algorithm is called KMeans clustering algorithm, with the core idea of “birds of a feather flock together and people flock together”.
We do Python modeling using data from a company on the Internet. We first do z-Score processing and clean out the abnormal extreme values.
The three columns in the figure above are standardized user consumption data. The closer it gets to 0, the closer it gets to the average. Because r value is the last consumption time, the smaller the value is, the closer the time is; the larger the value is, the older the consumption is.
Through RFM three indicators (called features in machine learning), a visual scatter graph is first established. Below is the scatter diagram of the latest charge R and charge amount M. Each dot represents a user’s charging data
The scatterplot can not see the rule of user clustering temporarily, but can only be preliminarily judged, most of the data shows a central trend.
Since the core idea of KMeans algorithm is “birds of a feather flock together and people flock together”, it takes distance as the objective function. In short, the closer the distance between two users is, the more likely they are to be similar. Therefore, KMeans will find out the similar groups and call them clusters. The larger the distance between clusters, the more independent the user groups are, which is called clustering. The tighter the distance within the cluster, the more similar the users are, which is called clustering.
Speaking through charts:
The users circled in red are more likely to be similar and belong to the same user group. Because they have similar data in R and M, both of them belong to the group with low consumption amount and recent consumption.
As for whether or not, let the algorithm to solve it, the specific algorithm principle and process will not demonstrate. Let’s say we can identify five user groups and see what those groups look like.
The different colors in the figure above are the user groups calculated by the algorithm.
The red user group: represents the high amount of money spent, because the number is small, there is no discernable difference in the date of the last purchase, but it is not long ago. These are the dads and funders of the product.
Green user group: represents the users who have the tendency to lose. These users do not consume too much, and the operation can take appropriate recovery strategies.
Purple user group: represents the recent consumption of users with less consumption amount. The operation needs to explore their value to develop and cultivate.
Cyan and blue seem indistinguishable. What if we change the dimensions of the scatter diagram?
With indicators R and F, there is another perspective. The cyan user group has more consumption times than the blue user group, while the blue user group has a lower consumption frequency and needs more incentives. Purple users have a high consumption frequency.
At this point, the user groups are clearly differentiated. Can you accurately outline the characteristics of these users? Although the long tail pattern affects readability to some extent in terms of data distribution, the operation can still make corresponding operation means for different groups.
Look at the final result with a scatter matrix (the image may not be very sharp) :
That’s the CONTENT of the RFM model. It can provide users with dynamic consumption profile, to market, sales, product and operation personnel to provide the basis of fine operation.
This is also one of the applications of data mining in user operation, you should understand.
How to divide groups is a science. If there are fewer groups, the distinction is not obvious. Divided more, there is no business value, more than 20 groups you how to operate? Population is a balance between data and business.
In a word, the clustering method, one is through indicators and attributes to manually divide the user groups. The other is to give business meaning to results through data mining. Ultimately, the goal is to increase operational effectiveness and value.
We can use the RFM model and try to think more broadly. Can we do something new? You can try it.
Finance: amount of investment, frequency of investment, time of last investment;
Live streaming: duration of watching live streaming, last watching time and amount of reward;
Content: number of comments, number of comments, number of likes;
Website: login times, login duration, last login time;
Game: level, game duration, game recharge amount;
These are my simple list of references, may not be accurate, as you refer to other rocks. Different products have different clustering strategies, such as hotel products, accommodation is not a solid demand, whether to add the dimension of time? Maybe the accommodation will be better grouped.
It should be noted that the number of groups is not fixed, it can be two or four, depending on the business needs, mainly to include most users. Just not too much. On the one hand, it is complicated, and on the other hand, the performance of KMeans clustering in multiple features is not good.
Through user stratification and user clustering, you already know the cornerstones of user operations.
User stratification is a division based on the general direction, what core goals you want users to work towards;
User clustering, then, is to slice them into finer granularity to improve the effect. The two go hand in hand.
If the number of users reaches a certain level, stratification and clustering may not be a good method, because with the further expansion of the product, no matter how to subdivide it, it is difficult to meet the complexity of users, which is common in all kinds of platform products. At this time, the UserProfile system needs to be introduced. At this time, user stratification and clustering are only part of the portrait.
— — —
This is an article to answer readers’ questions. I originally wanted to split stratification and clustering into two pieces, but it was more coherent, so I combined them together, which took a little more time. Slightly involved in a little data mining content, according to the progress of the direction of data, the second half of the year to see this will be better, as a preview.
But then again, this is probably the first article I’ve written that combines high-level operations with high-level data analysis.
Email: [email protected]
This article is published by the author authorized by morning Reading class. Please contact the author for reprinting.
There is no shortage of job opportunities for good people, only good opportunities for themselves. But they often don’t have the energy to sift through a sea of opportunities to find the best fit.
100Offer will strictly screen talents and enterprises on the platform, so that “the best talents” and “the best companies” meet.
Scan the QR code below to register 100Offer and talk about your expectations for your next job. Within a week, you’ll receive 5-10 great opportunities to meet your needs!
