How to build a data-based operation system from 0 to 1

In recent years, with the rapid development of mobile Internet , the concept of big data has become increasingly popular, and many companies are promoting data-based management.

So, let’s talk about the topic of data management today.

1. Three Misconceptions of Data Management

Let’s first look at the misunderstandings of data management:

1. Does having more data necessarily drive business development?

Having a lot of data does not mean that it can drive business development, because there are many factors that prevent the data from being well applied to the business.

01 Factor 1: Data quality issues

On the one hand, many companies will collect a lot of dirty data when collecting data, such as simulators and wool parties brushing the volume. If anti-cheating identification is not performed, it is difficult to filter out this part of data;

On the other hand, if data is not collected and reported in a standardized and regular manner, it will be difficult to obtain effective information when conducting data mining and analysis. The accuracy and stability of data are the basis of data science management. If these two points cannot be met, there will be many problems in business decision-making.

02 Factor 2: Data is not closely related to business

Data is only valuable if it is strongly relevant to the business.

We have many indicators ourselves, probably more than 70.

However, in daily analysis, we only use a few of them frequently. The others are either rarely visited or will be gradually abandoned over time.

03 Factor 3: It’s easy to collect data, but difficult to use

Tencent has been engaged in data analysis for many years and has accumulated a lot of experience. It has a complete set of analysis systems of its own, so naturally there is no problem in data application. However, many traditional companies, such as operators such as China Mobile, China Unicom, and China Telecom , as well as insurance companies and banks, are particularly confused about this area: they have a lot of data, but do not know how to use it. Blindly collecting data like this actually doesn't make much sense at all.

2. With a data analysis team, can we definitely find the problem?

Not necessarily, because analysts often do not understand the business and can only provide mechanical analysis results, which cannot effectively guide business development.

For example, we have previously introduced an analyst team into our business operations, hoping to provide more ideas and guidance to help business development.

But after analyzing for a while, we found that the effect was not very good. The reason is that analysts generally give analysis results based on the data of the report dimensions.

For example, if there is a sudden fluctuation in data, the analyst may think it is caused by holidays, but in fact he may not know what kind of activities we have done in the entire business operation. Therefore, data analysis and business are separated.

To solve this problem, we made a new attempt to bring the data analysis team and the business team together and set up a joint project team to conduct a pilot.

Later, the results proved to be quite effective, because the analysts would communicate with the front-line operations staff when designing each indicator:

• What does real user behavior look like?
• During what periods and in what scenarios will these data be triggered?

In this way, you can be more targeted.

3. Can the analysis report provide an optimal solution?

I believe that everyone is producing such analysis reports every week or every day, but are our analysis results really the optimal solution for a targeted business?

The result is often negative.

There are also several problems here:

01 Problem 1: The analysis dimension is single and insufficient to support the conclusion

There are many analysis dimensions, such as version, channel , region, device attributes and specific behavioral characteristics.

At present, all APP analysis is basically based on the two dimensions of version and channel. For custom events, if they are only analyzed based on event ID, there are not enough parameters to limit them, and it is difficult to support and make decisions on the analysis results.

02 Problem 2: Data quality interferes with analysis results

One of the key factors in determining whether the analysis results can effectively reflect business development and provide effective optimization strategies is data quality. If the data quality is not up to standard, the analysis results based on the data will be greatly reduced.

Many of the apps we come into contact with have experienced inflated traffic to varying degrees. This kind of data not only causes financial waste to the company, but also affects normal data and makes data analysis difficult.

2. Ideas of Data Management

After recognizing these misunderstandings, how can we conduct data management scientifically?

Scientific data management requires not only comprehensive data analysis, but more importantly, actions based on the results of data analysis.

Therefore, below we will talk about the data management ideas from four aspects: data collection , data analysis, action strategy and rapid execution.

1. Data Collection

The dimensions we currently collect include basic data statistics, user attributes, user sources, user behavior and model data.

01 Basic Data

Basic data are the data we look at on a daily basis, including new additions, activity, number of launches, and retention .

02 User attributes

Based on the numerous apps covered by Tencent Mobile Analysis and the advantages of Tencent's big data, different labels for different groups of people are extracted.

User attributes include basic natural person attributes, such as gender, age, occupation, education, interests and hobbies; in addition, device attribute analysis is also performed, such as the model, brand, and networking method used by the user.

03 User Source

This is also a very important indicator that can help analyze where users come from. There can be many dimensions here, such as channel is one dimension, version is also a dimension, and region and model can also be counted as different dimensions.

04 User Source

User sources can help analyze product stickiness to users and help optimize product paths. This involves a lot of data, such as user usage time, usage frequency and page access path.

Regarding behavior, I would like to emphasize that the data required to be reported must be standard and standardized, which will help with later analysis. Otherwise, the data collected at the end is just dirty data and has no meaning for subsequent analysis.

05 Model Data

Model data is an advanced model based on basic data analysis, which can perform more advanced analysis for users. We will perform model analysis for each stage of the user life cycle , such as the new user acquisition model, activation model and churn model.

2. Data Analysis

Data itself has no value, only analysis has value.

When doing analysis, the following aspects are mainly included:

• Multidimensional cross-analysis
• User group analysis
• Product quality analysis

When doing data analysis, multidimensional cross-analysis is a good analytical method.

When we look at the overall trend of the data, we can find out whether there are problems or opportunities in business development, but we cannot find out where the problems lie, or where the opportunities are.

We would advocate multi-dimensional drill-down analysis, such as channel, version, country, device attributes and natural attributes. These can be used as dimensions for cross-analysis to ultimately identify where the problem lies.

User population analysis can help analyze the attribute characteristics and behavioral characteristics of different groups of people. Develop differentiated operation strategies based on different population characteristics to maximize the operation effect.

Product quality analysis is naturally a data indicator used to measure product quality, including app crash conditions, power consumption, network speed and other data.

3. Action Strategy

Based on the results of data analysis, after discovering problems or opportunities, we must first formulate some action strategies. This step is the key to bridging the gap between data analysis and action.

Take the commonly used strategies of life cycle management as an example, including new customer acquisition strategies, activation strategies, and lost customer recall strategies.

4. Fast Execution

Once the strategy is formulated, the next step is to execute it quickly. Action is the final step to verify the effectiveness of data analysis results and strategies. The faster you run, the higher the chance of success, so good execution is very important.

Tencent also adheres to the principle of taking small steps and iterating quickly. It’s okay to make mistakes, as long as we ensure speed.

2. How to build a data operation system from 0 to 1

We have talked about the misunderstandings and ideas of data management before, so next let’s talk about how to build a data operation system from 0 to 1?

Phase 1: Indicator Planning

I have come into contact with many developers who all have the need for data analysis, but it is not particularly clear at the beginning what data to collect and count, so planning in this area is very necessary.

Especially in the early stages of data construction, the indicators must be clearly defined, otherwise it is easy to cause problems in the later data analysis stage.

Another thing is the dimension setting. In addition to version and channel, you can also add custom attributes, such as region, model, network type, etc. The more dimensions you design, the more helpful it will be for the subsequent refined analysis of data.

Take new additions for example. You may notice a rapid decline in new additions. If you only design the two dimensions of version and channel, you can only analyze from these two dimensions. If you want to do a deeper drill-down analysis, you will not be able to continue positioning.

The last key factor: update cycle.

Whether the data is updated in real time, daily, or weekly, it needs to be planned in advance. Because data calculation consumes a lot of resources, we should use the best resources where they are most valuable.

Phase 2: Data Collection

After the indicators are planned, the next thing to do is data collection. Data collection includes three aspects of work: field classification, data burial and data reporting.

Field classification is very important. The more detailed it is, the more helpful it is for subsequent data analysis.

Data tracking, as the name suggests, is to collect statistics on business data by tracking the data you want to collect.

Phase 3: Report Presentation

After collecting the data, we need to consider presenting the data in the form of reports to help analyze business changes.

After we have completed the above steps, if we finally verify that this system is feasible, we can consider the entire data productization and function iteration.

01 Construction method

Next, let’s talk about how to build a data analysis system.

There are currently only two ways of construction: one is self-construction, and the other is to use third-party services.

What are the benefits of self-building?

The tracking point function is flexible and convenient. You can track the points according to your needs and connect them with business data. Because only in this way can data analysis be truly applied to business.

However, I believe that some apps that are currently in their start-up stages do not have the ability to do this.

The disadvantage of self-construction is that it requires huge investment. First of all, there is the labor cost. Then, the server resources are also expensive to maintain. In addition, it cannot be connected with external operation tools . Take attracting new users as an example. You need to have a very clear understanding of users and know the attributes and behavioral characteristics of the user group.

If you place ads on an advertising platform based on the tags you have collected, the matches will often not be accurate enough. Because it is very likely that the labeling systems of both parties are not consistent.

02 Iterative Optimization

During the entire construction phase, it is not necessary to complete it all at once. It can be built on demand according to the different stages of app development and iterated step by step.

Let’s first look at the four stages of the APP development life cycle: start-up, growth, maturity and decline. The data indicators that need to be paid attention to in each stage are different, so we can build in stages:

①What indicators should we pay attention to in the initial stage?

• The growth rate depends on whether the data growth is linear and whether there is potential in the future. In the early stage, it is basically enough to do some additions, activities, startup times, and retention;
• The source of users, it is necessary to know which channels have better promotion effects in order to optimize the promotion effects of attracting new users.

② During the growth stage, we should not only pay attention to the user growth rate, but also pay attention to user behavior data

During the growth stage, I need to pay attention not only to the user growth rate, but also to the user behavior data, because I need to identify the quality of users.

Take promotion as an example. We should not only look at the overall growth of data, but also see whether the core users have grown. Then we need to identify who our core users are from the dimension of user behavior.

Then it is necessary to build data on dimensions such as usage frequency, usage duration, page access path, and consumption behavior.

③ The demand for data increases in the mature stage, and it is necessary to deeply explore user value

When it reaches the mature stage, the demand for data will become more and more in-depth, and it will be necessary to deeply explore the value of users. Then at this stage we need to consider making some data models.

For example, in the active model, although active users are a whole, the quality of active users is different. For example, for those who have been active for more than 100 days, more than 300 days or even more than two years, personalized operation strategies are needed for different user groups.

Another model is the churn model. When the user base reaches a certain scale, it is difficult to avoid loss. When it reaches the mature stage, the activity of some users will gradually decline and eventually they will be lost. At this time, it is necessary to use some means to interfere with users, such as message push, effective incentives, and text messages.

There is also portrait insight. When it reaches the mature stage, I believe that all data will be considered for monetization. To monetize, we must first know what the user looks like, so the construction of portrait insight can be put on the agenda.

③ Serious user loss during the recession

When it reaches the recession stage, the APP has basically begun to experience large-scale user loss, and it is very difficult to attract users back, so it is necessary to pay attention to interest shifts and look for new business growth points.

03 MTA indicator system

This is the MTA indicator system, which is divided into basic indicators, user attributes, user sources, user behavior, and model data.

New additions, active users, and churn have been mentioned in the previous article, so I will not repeat them here. I will focus on quality monitoring. Quality monitoring mainly counts the crash situations when users use the app, as well as network speed monitoring and interface call situations.

3. In-depth data analysis

Next, let’s talk about the in-depth data analysis part.

1. Multi-dimensional drill-down analysis

The advantage of multi-dimensional drill-down analysis is that you can discover problems from the entire trend, ultimately locate the problem through more fine-grained analysis, and then formulate corresponding execution strategies.

There are actually many dimensions of analysis, such as: channels, versions, regions, pages, tags, user groups, which can all be used as dimensions to analyze our users.

2. Funnel conversion analysis

Next, let’s take a look at the funnel model that is often used in daily life.

The funnel model is a very important means. The funnel can not only help analyze the key path from the first step to the final conversion result, but also help analyze the conversion rate between each step. There will basically be loss in every step of the conversion funnel, so there is no 100% funnel.

How to use the funnel? A single funnel analysis is meaningless. The significance of funnel analysis can only be reflected through trend observation, dimension comparison and dimension segmentation.

Next we will give an example.

Using Tencent Mobile Analytics can be roughly divided into three steps: registration-testing-launching.

The first step is the conversion from registration to testing, the second step is the conversion from testing to launch, and the third step is the conversion from testing to launch.

Analyzing the data , we found that the conversion rates in May and June were OK. However, looking at the data in July, the conversion rate from registration to testing was only 21.7%, but from testing to launch it was 22.6%, which showed no change. The overall conversion rate was 4.9%. Compared with May and June, the data has declined.

What is the cause of this problem?

Let’s analyze the data related to this funnel.

The first data is the new additions, the second is our testing, and the third is the number of applications that are finally launched.

We found that there was only a large increase in new applications in July, but it did not provide us with good test data. There were only 1,300 applications tested, with no obvious growth, and we finally launched only 294 of them. This shows that there is an increase in new users, but the conversion rate of this part of users is not very high.

Overall, this should be the reason for the surge in new users. We then conducted a drill-down analysis on the channel dimension and found that a large number of new users came from the official website.

When analyzing related promotional activities, we found that we held a quiz event with prizes this month, which led to many users registering accounts and creating apps. These users came for the prizes, not our users, so the conversion rate was very poor.

Finally, the reason for the decrease in data conversion was identified.

Data analysis is the basis for formulating user management strategies. After talking about analysis, let’s talk about strategy-related content.

4. User Management Strategy

1. User lifecycle management

User lifecycle management can be divided into six stages:

1. Potential user stage
2. Beginner stage
3. Effective active stage
4. Active descent phase
5. About to lose stage
6. Loss Stage

In different user life cycles, we need to implement differentiated operation strategies to maximize user value.

for example:

• In the potential user stage, it is necessary to formulate a new user acquisition strategy based on user characteristics;
• In the novice stage, it is necessary to provide users with real-time user care and develop a suitable novice task system;
• In the effective and active stage, we provide unique product exploration guidance and positive incentives;
• Conduct anti-loss intervention in the active decline stage; conduct loss warning in the impending loss stage;
• Initiate attrition care during the attrition stage.

Next, we will explain with examples.

01 Accurately attract new customers

In the potential user stage, when attracting new users, sort out the relevant fields based on historical data and develop a corresponding new user acquisition model using historical players as sample data.

By comparing the promotion effect of the experimental group with that of the overall market, the effect of attracting new customers is evaluated.

In fact, after such analysis, the effect of attracting new customers will be significantly improved. In the end, the experimental group had an improvement of 30% to 60% compared to the random group.

In fact, the effect of this model is not very obvious in the initial stage, and it requires continuous training to achieve such an effect. Therefore, data analysis needs to be iteratively upgraded step by step, rather than achieving a good effect overnight.

02 Newbie Care

Appropriate new user care methods can help users stay longer. For example, you can set up a personalized novice task system and personalized level rewards based on the user's interests and preferences; for example, you can give different gifts based on the user's gender and design different task difficulties.

03 Active Growth

Good content recommendations and growth systems can increase user activity, and can recommend appropriate content by portraying user portraits .

For example, in games , adding social attributes can increase user activity. Then, appropriate team information can be recommended based on user attributes and behavioral characteristics. Once a user joins a team, the activity level can be effectively improved.

04 Anti-loss warning

Some behavioral data can be used to determine whether users are at risk of churn. For example, a decrease in activity and an increase in usage time intervals are typical characteristics. Then, after identifying this part of users, you can use some operational tools to recall lost users .

For example, you can reach users through message push, text messages or advertising platforms. If the activity level decreases, you can send them more props or introduce more interesting gameplay to prevent user loss.

05 Lost and Returned

Based on experience, once users are lost, it is difficult to get them back. Therefore, instead of spending energy on recovering lost users, it is better to analyze the user’s interest shifts and find new business growth points.

2. User group management

User group analysis is a very good method. In addition to helping with population analysis and problem location, it can sometimes help uncover needs that users themselves are not even aware of.

So how do we create clusters?

One is based on user attributes, and the other is based on user behavior characteristics.

For example, the age and gender listed in the PPT, or non- paying users , multiple-paying users, etc.

With a diverse group of people, what can we do?

In addition to analyzing the characteristics of different groups of people, we can also make differentiated operation plans and accurately attract new users for different users.

Next, let’s look at a case.

This is an e-commerce app that uses our services. The user growth is good, but the transaction volume has been sluggish, so they came to ask, how should they use data to drive business development?

So we did a round of analysis based on this case.

First, we established clusters to analyze the characteristics of transacted users and high-value users; then we compared them with the general market of users and found that there are more men in the entire market, but the proportion of women among transacted users and high-value people is higher, which means that women are more likely to complete transactions.

In addition, we made a comparison of population preferences for these three user groups.

Compared with the general users, transaction users and high-value users are more interested in shopping and finance. This is an analysis conclusion we have reached.

Then back to the question, the volume did not meet expectations. The first possibility is that there are quality issues with channel users , and the second possibility is that there are problems with product positioning. There are more male users in the entire market, so maybe there are not enough product recommendations for men, or there are fewer product categories for men. These are all possibilities.

Compared with the other two reasons, the verification cycle of the first one will be shorter, so let’s analyze the first one first.

As shown in the figure above, these are several channels we have obtained. We can see that D and E are the main sources of traffic , and D’s weekly retention rate is also good.

But judging from the transaction volume, A’s data is also good, so A may also be a high-quality channel.

This is one of our analysis results. Then we need to verify whether channel A meets the characteristics of transaction population and high-value population.

According to the data, the proportion of women in channel A is 62%, and their interest in shopping is also higher than the market as a whole. This is in line with the characteristics of our transaction users and high-value user groups.

Based on this analysis result, we recommended that developers adjust the strategy of distribution channels, increase the distribution ratio of A, and reduce the distribution ratio of B and C. One week after the strategy was implemented, I tracked the data.

This graph shows the overall conversion rate. The front graph is before optimization, and the back graph is after optimization. The overall conversion rate has increased by more than ten percent. This is the whole analysis process, I hope it can inspire everyone.

The above is my sharing. If you have other questions, you can communicate with me at any time.