Practical case: How to build a user churn warning system!

Instead of waiting for users to churn and then using operational means to recall them, it is better to prepare a user churn warning system from the beginning, which can more effectively reduce the user churn rate. The author of this article will use the laundry channel as an example to teach you how to build a user churn warning system.

For a community o2o community product, user churn has always been one of the issues of concern to the market, operations, and products. In general user operations, the solution to lost users is to define users who have not consumed on the platform for a period of time as lost users, and to conduct activities to recall lost users.

This operation model is an intervention behavior taken after user loss, which often has a lag. The possibility of reaching or recalling users who have uninstalled the APP is very small.

Therefore, an effective user churn warning system is of great significance to preventing user churn. How to build a user churn warning model without the help of data operation platform and machine learning is the main topic discussed in this article.

In actual operations, we attribute user churn prediction to three questions:

• Question 1: Predict which users will churn?
• Question 2: What kind of data model is used for prediction?
• Question 3: What are the characteristics of users who may churn?

Let's use an example to discuss this issue. The platform's laundry channel has found that it has been experiencing serious customer churn recently, and plans to organize a user event to curb this trend. However, this customer retention event is one of many marketing activities in the channel, and the budget is limited. The user department needs to use data mining methods to find high-value users who may be lost, and characterize the characteristics of these users, so as to use the characteristics of the lost users to find other users who may be lost and conduct group operations. The following 5 steps will explore the specific operation methods of lost users:

1. How to define user churn?

First, we define the churn of sample users, which can help us predict the probability of churn of similar users based on the churn characteristics of sample users.

After communicating with the operators of the laundry channel, we first defined the laundry channel users into two types: churned users and non-churned users, and made a preliminary churn classification based on whether or not they had any consumption behavior within three months. At the data level, churned users are represented by 1 and non-churned users are represented by 0.

2. Which user data can characterize user churn characteristics?

What user data should be considered to influence user churn? This is a crucial step in building a user early warning model. From the data level, we need two dimensions: user portrait data and behavior data, namely:

• User profile information: ID, gender, age, region, member type, user source
• User behavior data: number of days logged in, online time, login frequency, number of days registered
• User consumption data: number of days since the last order, cumulative number of orders, cumulative consumption amount, average order value

3. How to define the prediction time window?

The purpose of churn analysis is to detect customers before they churn so that retention measures can be taken. So how long of a period of time should be used as the sample data volume during analysis?

If the time taken is too short, the user characteristics may not be representative; if the time taken is too long, the modeling operation time is too long and it is difficult to check errors, so it is very important to define an appropriate prediction time window.

After studying the historical consumption data of users with the laundry channel operators and considering the availability and effectiveness of the comprehensive data, we took the historical 3-month data as the prediction window and the user data of the current month as the verification basis to establish the churn prediction time window model:

4. How to build a user churn model?

The specific process of model building is too technical and is beyond the scope of this article. We mainly discuss the principles of building a churn model, hoping to learn from it.

First, we need to conduct an exploratory analysis between user feature data and churn fields to check whether there is a strong correlation between each feature dimension and churn. We retain highly correlated data and eliminate weakly correlated data.

The above is a laundry user exploration analysis model we built using big data analysis tools. Based on this model, we can get the association analysis results as follows:

The analysis results show that old users who have registered for a long time have a more serious loss, which means that the channel's marketing efforts for old users still need to be strengthened. Through association analysis, we retained 8 user data fields that are strongly correlated with churn, namely user type, source, membership type, city, gender, registration days, time since the last order, and average order value.

Secondly, we need to establish user churn rules to predict the churn of other users. A modeling approach is also needed, and a common approach is to use a decision tree algorithm to generate user churn rules:

After the model is established, the model needs to calculate the data to generate churn rules and rank the importance of churn in each data dimension. Since a small amount of data is used to quickly generate the model, the prediction results may not be very accurate and are only for reference:

According to the churn rules generated by the model, it can be found that the churn characteristics contain 4 rules.

Take Rule 1 as an example: If the number of registration days is greater than 53 days and the membership type is diamond membership, it is easy to churn. However, this group of users is high-value users at the operational level. The operators of the laundry channel need to pay special attention to retaining high-value old users. Similarly, the right figure shows that the number of registration days and membership type are two important influencing factors.

5. How to build a lost user operation strategy?

Once the churn rules and factors influencing churn are determined, it can help the laundry channel predict which users are likely to churn and retain them before they churn. This requires providing the channel with specific operating strategies.

• Segmentation strategy: By establishing a segmentation model, users with a tendency to churn are divided into several groups, the characteristics of the groups with a higher possibility of churn are analyzed, and targeted marketing measures are formulated to retain them.
• Scoring strategy: Separately group high-value customers of the channel, model and score the possibility of churn of high-value customers. The channel pays special attention to high-value users with high churn scores and designs appropriate retention activities to retain them.

The segmentation strategy is to operate all users in different groups. It requires the channel to design targeted marketing activities for each group. This strategy can be adopted if the channel has sufficient budget and energy. The scoring strategy only predicts and scores some high-value users, and takes retention marketing measures for high-value users with higher scores. Relatively speaking, it can achieve outstanding marketing results with low marketing costs.

Author: Zhao Wenbiao

Source: User Operation Observation (ID: yunyingguancha)