How to provide early warning for user churn?

Recalling lost users is ineffective? The attrition rate remains high? Why not take preventive measures before they happen and conduct early warning intervention?

1. Why do we need to issue a warning for lost users?

There is a life cycle for users when they use a product. Users come into contact with the product, understand the product, and use the product after experiencing its core value. However, due to various factors, they may eventually leave the product and switch to competing products or other solutions. Loss is inevitable.

Although churn is inevitable, it does not mean that we do not need to pay attention to churn. It is not easy to attract new users, but churn is easy. The cost of attracting a new user is more than 5 times the cost of maintaining an old user. We hope that the user pool can continue to expand.

Therefore, many companies focus on attracting new users, but ignore the loss of old users. The rate of user loss has even exceeded the growth rate of new users. The business has become extremely unhealthy, and all the money and efforts have been wasted.

Therefore, we cannot just focus on attracting new users, but also need to keep an eye on the status of user churn. We cannot force zero churn, but we must control the user churn rate below the risk point based on the characteristics of the product. How to control the churn rate?

The loss of a user may be due to lack of guidance for new users, and they may lose the user before they can experience the core value of the product. It may also be that the old user has experienced a major revision and feels that the product is no longer as easy to use as before. But no matter what the reason for the loss, once the user leaves the product, the ways and channels to reach the user are very limited, and it becomes very difficult to recall the user.

Therefore, we cannot be wise after the event. Instead, we must issue warnings and intervene in a timely manner when there are signs of user loss before users leave, so as to increase the possibility of winning back users.

2. How to define churn?

Before issuing a user churn warning, we must first clarify how user churn should be defined.

Generally speaking, if a user has not used a product for a long time, there is a high probability that the user has been lost.

But how long is this “long time”? This “unused product” means that the user has not taken any action? For different types of products, these two points are very different.

Take social products such as WeChat as an example. Since social scenarios naturally have the characteristics of high frequency and high stickiness, users use WeChat to communicate every day, so generally speaking, if a user does not use it for a week, he or she may be considered lost.

For tool products, such as Hellobike, users may only ride it once a month or even longer, so this time period needs to be longer.

In addition, there are differences for different types of products with "unused functions". For example, for e-commerce products, it may be the behavior of "not placing an order", while for short video products, it may be "not watching the video".

Generally speaking, this function should be the core function of the product. If the core function of the product is not used for a long time, it can be considered that the user has been lost.

3. Why do users churn?

Everything has a reason. The reason why users choose to leave is that they feel that the product fails to solve their problems and cannot meet their needs, so they turn to other products.

There may be several reasons for this:

1. Product Value

When users download and register a product, they do so with a need in mind and want to solve their own problems. However, this demand may be different. It may be just an ordinary demand of the user, or it may be a rigid demand, or it may be a pain point demand.

For example, you may feel very tired recently and want to go to Maldives to relax. This is a common demand. But when you arrive in Maldives and play for half a day, you may feel a little hungry and want to eat. This is a rigid demand.

So I looked up some recommendations online. There happened to be a restaurant nearby, but it had bad reviews and was very expensive. There was another restaurant that tasted very good but was very far away. Finding a restaurant that was close, tasted good and was affordable was a pain point.

Therefore, ordinary needs → rigid needs → pain points is a progressive process, which reflects the urgency of users' desire to solve the problem layer by layer. Therefore, if the product can solve rigid needs, do not be satisfied with only solving ordinary needs. If it can solve user pain points, do not just stop at solving rigid needs.

The more urgent the users are, the easier it is to reflect the value of the product, the user stickiness will naturally be stronger, and the probability of churn will be much smaller.

2. User Experience

Solving users’ pain points is the foundation of a product’s existence, but in the current era of serious product homogeneity, good user experience is also the core competitiveness of a product.

When all products are similar, products with poor user experience will definitely have more serious user churn. User experience is mainly reflected in the following aspects:

• Poor visual experience: The product's UI is too low-level, the color scheme is very cheap, and it looks very unrefined at first glance. The first impression is very bad, and there is no desire to continue using it.
• Poor interactive experience: The product’s interaction goes against the user’s usage habits. Although users can swipe up and down to scroll through the screen, they have to keep clicking on the next page, which makes the user experience not smooth.
• Complex registration process: A complex and lengthy registration process is a major obstacle that scares away users. Some information does not need to be collected during the registration process. The more information users fill in during registration, the more private it is, and the more likely they are to lose users during the registration process.
• User expectations are not met: The functions provided by the product fail to meet the needs of users well, or the content quality is poor, resulting in users not getting what they need and leaving.

Only by understanding the different reasons for the loss of different users can we more accurately describe the precursor characteristics before user loss, and then predict the probability of loss and recover them in advance.

4. User Churn Warning Model Construction

The essence of user churn warning is to analyze the reasons why users may churn, visualize these reasons in the form of data as reasons, and then label the users with churn probability results. In abstract terms, it is a machine learning classification problem from features to labels.

Since it is a classification problem, the following key links are indispensable.

1. Sample selection and data processing

Define churn during the observation period: Since machine learning requires training sets and test sets, it is necessary to define an observation period that is long enough and has a large enough sample size, and collect user data during the observation period and samples of user churn probability as training sets and test sets.

For example, we can take user data from the past six months as samples. Since the results of whether users have churned are known, we can label users with churn probabilities. These samples are used as input samples for the classification model after feature engineering, and are an important data source for the model to learn classification rules.

Collecting user behaviors during the performance period: Once the patterns of the observation period data have been learned by the model, it is necessary to collect user behavior data for the next window and predict the probability of churn of users who engage in these behaviors based on this.

2. Feature Engineering

Following the sample selection in the previous step, the next most important and decisive step is feature engineering. The upper limit of machine learning is determined by feature engineering, and any form of tuning can only get infinitely close to this upper limit.

Feature engineering must be based on a deep understanding and analysis of the business! It must be based on a deep understanding and analysis of the business! It must be based on a deep understanding and analysis of the business! Important things should be said three times!

The effectiveness of machine learning depends on feature engineering, and the key to feature engineering lies in familiarity with the business.

Only by being familiar enough with the business can you accurately digitize and visualize the reasons that may affect user churn, find the causes at their essence rather than the symptoms of the causes, and then find the key features that affect retention.

For example, the user's active time seems to be a feature that is very relevant to churn, but the time is not the cause of user churn. It may just be a manifestation of the reason that users cannot find commonly used functions after product iteration.

Because the commonly used functions have changed their location and cannot be found, people feel that the product is not easy to use and gradually start looking for other alternative products, which leads to a shorter usage time. This is the root cause, and the process of finding the root cause undoubtedly requires a deep understanding of the business.

Generally speaking, the features we need to consider may fall into the following categories:

(1) Basic attributes of users

Different types of users may have different churn rates depending on gender, age, income level, region, etc.

(2) User product behavior

We call the product life cycle, active frequency, frequency of use of key functions, etc. basic indicators. Basic indicators are generally the manifestation of the reasons for churn and are correlated with churn, but they are not causal and are not the key features that lead to churn.

(3) Other processing indicators

Basic indicators may not be able to effectively discover the key features that affect retention. It is necessary to process new indicators based on business understanding and use them together with basic indicators as features for model training. Common processing methods are:

① Depth index

An indicator that reflects the depth of user usage. Users should not only use the product, but also use it in depth. For example, the number of times a key function is used. Some users may only use some marginal functions and churn before they get to know the key functions.

This is a pity, so this depth indicator can be used to predict whether users are likely to churn.

②Frequency index

Users must not only use the product deeply, but also use it frequently. The definition of frequency varies depending on the product type. Some products may need to be used every day, or even several times a day, while others may require use several times a week, and so on.

However, a frequency indicator can be processed based on the characteristics of the product, such as the average number of uses per day/week or the average number of days used per day/week, so that the user's usage frequency can be characterized.

③Trend indicators

The trend of user product usage changes. The trend of user usage is directly related to user churn. If a user uses it less and less, there is a high probability that the user will churn.

Therefore, some common trend indicators, such as the rate of change of the average number of active days per week in the past three months, can be understood as a slope. If the average number of active days per week is decreasing, the slope should be negative, otherwise the slope should be positive, thereby characterizing the changing trend of user usage.

3. Model selection

After the feature construction is completed, it is necessary to select a model. For classification models, commonly used ones are logistic regression, decision tree, SVM, XGboost, etc. Each model has its own advantages and disadvantages, and also has certain requirements for features. We do not need to spend too much energy on model selection.

You can pre-select some models, bring in samples for training, observe the classification effects of different models, and select the one with the best effect as the training model.

The effect here is mainly evaluated by the evaluation criteria of the classification model, such as confusion matrix, f1 value, and the generalization ability of the model.

The focus of the churn warning model construction is on feature engineering rather than model selection, so this part is not the focus and will not be expanded in detail. You can study the relevant materials if you need it.

4. Model training and prediction

After feature processing is completed and the training model is determined, the samples need to be trained and the model effect needs to be continuously optimized by adjusting parameters.

When all indicators meet the requirements and model training is completed, you can go online for prediction, make predictions for users in the performance period, evaluate their possibility of churn, and then take targeted operational actions. At this point, the construction of the user churn warning model is completed.

Author: Big Data Analysis and Operation Knowledge Planet

Source: Big Data Analysis and Operation Knowledge Planet