How to build user churn warning?

Before a user becomes a churned user, the risk of user churn is identified based on his own attributes, behavior and other characteristics, and timely measures are taken to retain the user. This is user churn warning. Building a user churn warning system mainly consists of three steps: definition - analysis - construction. This article provides a detailed analysis and shares it with you.

1. What is user churn warning?

We all know that for a relatively mature product with a relatively saturated market, the cost of acquiring a new user is much higher than retaining an old user. The loss of old users means a decrease in revenue. Therefore, I believe that many people will build a system to recall lost users. They will first define lost users, and then use various contact methods, such as SMS, push, etc. to recall lost users.

However, in many cases the recall rate of such recall work is not ideal. On the one hand, users who have truly lost their app may have uninstalled the app and turned off push notifications, making it impossible to reach them effectively. On the other hand, users have given up on the app for some reason and are likely to ignore or even feel disgusted when they receive a recall message. The difficulty of recalling a user may not be any lower than acquiring a new user.

Therefore, once the user has left, it is very difficult to get him back. Therefore, we hope to be able to identify the user's churn risk based on his own attributes and behavior characteristics before a user becomes a churned user, and take timely measures to retain the user. This is the user churn warning. Loss warning can firstly advance the time of user recall. Secondly, compared with loss recall, it has low cost and low difficulty. Thirdly, it can be used to conduct recall and activation within the app, with more diverse gameplay forms.

2. How to build a user churn warning?

1. Define churned users

User churn actually refers to users who no longer use the product within a period of time. In fact, different products have different dimensional rules for measuring user churn, and there is no universal definition. The definition of churn is usually a combination of two dimensions, namely behavior plus cycle. For example, some products define not logging in for one week as churn, and some products define non-payment for half a year as churn.

In addition, the definition of churn can also be stratified based on user attributes. For example, different churn thresholds can be set for users of different genders or different levels.

There are many user behaviors, and we need to combine the product type and the overall operational goals of this stage to find out the core behaviors that can define users. For example, e-commerce products can be defined by purchasing behavior, and a user is considered to have churned if they have not made a purchase for a long time; consumers of content products can be defined by user browsing, and a user is considered to have churned if they have not browsed for a long time; creators can be defined by user publications, that is, a creator is considered to have churned if they have not published a work for a long time.

The cycle can be defined by combining the inflection point theory with business characteristics as a reference, and ultimately the lost users can be defined by behavior + cycle.

2. Analyze the reasons for churn

Why do we need to analyze the reasons for user churn? This is because after building the churn warning model, we need to know the reasons why different users have the idea and behavior of leaving, and carry out targeted user recovery. And find the key behaviors for user retention and guide user behavior.

According to different lost users, we can conduct targeted loss reason analysis, mainly in the following four ways:

3. Construction of the churn warning model

The churn warning model needs to adopt different models to make predictions for users in different life cycles. Users can be divided into the acquisition period, promotion period, maturity period, and decline period. The purpose of dividing the user into cycles is to incorporate the user life stages into the early warning and recall strategy of refined operations in the future. Churn warning is to extract historical user data, observe relevant data in a certain window time, and then evaluate the user churn situation within the performance window according to the above-mentioned churn user definition, so as to predict the current user's churn probability in the future.

So what user data can affect user churn? It can be roughly divided into three dimensions, namely user portrait data, user behavior data, and user consumption data. In addition, we also need to define the prediction time window, that is, how long should we analyze the sample data? This requires combining the experience of business personnel and historical user behavior data, and then comprehensively considering the availability of data to ultimately establish a reasonable time prediction window.

During the observation period, we need to mine a group of sample users from historical data, and improve the evaluation indicators at all levels based on the three main dimensions of user portrait data, user behavior data, and user consumption data, and try to cover all aspects of field data to facilitate the evaluation of the correlation between each indicator and churn in subsequent modeling.

By obtaining the result data within the performance window, we can build the final prediction model, obtain the user churn rules and the importance ranking of each feature indicator. Commonly used early warning algorithms include decision trees, random forests, logistic regression, etc. During the forecast period, we continuously optimize the trained model and remove some features with low relevance. This improves the accuracy, hit rate, and coverage of the model. Next, we can predict the probability of user churn in the next month and output the scores and list of churned users.

3. Tiered operations and early warning user recall

1. User stratification

Building a churn warning model only circles out users who may be prone to churn, but without taking targeted recall guidance is a waste of effort. At this point, we already have data with different dimension labels, namely user life cycle * churn risk probability level * churn reason, etc. By grouping and cross-arranging multiple dimensions, we can obtain users with different marketing scenario significance, and based on this, we can establish a good early warning and recall user stratification mechanism.

2. Ways to activate and recall users at risk of churn

1. Send coupons and adjust discount amounts
2. Add user guidance within the app, scenario-based reminder copy, etc.
3. Optimize related recommendations
4. Personalized push copy, SMS, etc.
5. Other optimization solutions for specific churn reasons

There are many articles discussing various refined operations to promote user activation and recall methods, so I will not go into detail here. In addition, in actual operations, we need to focus on analyzing the user activation and recall effects, and analyzing the user recovery costs. Then analyze the overall ROI based on the recall effect and use AB experiments and other means to continuously optimize the ROI.

Above we have explained a general method for building user churn warning. Welcome to follow my WeChat public account and communicate about data analysis at any time.

Author: Zhao Xiaoluo

Source: Zhao Xiaoluoluo

Related reading:

4 ways to prevent user churn!

How to reduce product user churn rate?

User Operation | The underlying logic and operation guide of the user churn model!