A user operation method based on a survival analysis model
By using a survival analysis model, we have solved the problems of static and one-sidedness in existing user operation analysis, and achieved dynamic prediction and accurate segmentation of user behavior, thereby improving the scientific nature of operation strategies and the effective utilization of resources.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 李萌
- Filing Date
- 2026-03-06
- Publication Date
- 2026-06-05
AI Technical Summary
Existing user operation analysis methods lack a time dimension, resulting in static, single-dimensional indicators that cannot reveal the dynamic patterns of user behavior and lack predictive and causal analysis capabilities, making it difficult to provide an effective basis for differentiated operations.
Using a survival analysis model, this study describes user survival characteristics through data definition and mathematical modeling, compares the differences among different user groups, and constructs a Cox proportional hazards regression model to quantitatively analyze the factors influencing user behavior, thereby achieving dynamic prediction and precise stratification.
It enables dynamic analysis of user behavior, provides objective stratification standards, improves the accuracy of operational strategies and the optimization of resources, and enhances the effectiveness of user churn warnings and the success rate of user recall.
Smart Images
Figure CN122155774A_ABST
Abstract
Description
Technical Field
[0002] This invention belongs to the field of user operation technology, specifically a user operation method based on a survival analysis model. Background Technology
[0003] In the context of the rapid development of the internet and mobile applications, user operations have become a crucial element for all types of enterprises to enhance their core competitiveness. Users are not only the ultimate carriers of product value but also the fundamental source of sustainable growth for enterprises. Effective user operations can help enterprises extend user lifecycles, reduce churn rates, increase activity levels, and achieve differentiated competition through refined strategies. Therefore, how to scientifically, dynamically, and accurately understand and predict user behavior has become a critical issue that enterprises urgently need to address in their operational processes.
[0004] Current user operation analysis primarily relies on traditional metrics. For example, the most common metric is Daily Active Users (DAU), which is simple to calculate; any user who visits the application within a given day is counted as an active user. While this metric can provide a macro-level picture of user activity within a specific time window, it has significant limitations: Static and single-dimensional: Metrics such as DAU only reflect whether users are active at a certain point in time, lacking a continuous depiction over time and failing to reveal the dynamic patterns of user behavior changes over time.
[0005] Behavioral discrepancies masking: Under the same DAU value, user behavior patterns at different times may be completely different. For example, users who visit multiple times a day and users who visit only once cannot be distinguished under the DAU metric, leading to misjudgments of users' true activity levels; Lack of predictive and causal analysis capabilities: Existing methods are mostly based on statistical correlation analysis or aggregate indicators, which have limited predictive power for users' future behavior and are often susceptible to spurious correlations or problems such as "Simpson's paradox", lacking causal explanatory power. Insufficient support for operational strategies: Traditional indicators are difficult to provide effective basis for differentiated operations and refined interventions. Enterprises often have to formulate strategies based on experience or static rules (such as "not logging in for 30 consecutive days is judged as churn"), which are highly subjective and have limited effects.
[0006] In summary, while existing user operation techniques and methods can provide some reference information on user activity and retention rates, their static and one-sided nature, as well as the lack of a time-based dimension, result in significant shortcomings in dynamic prediction, precise segmentation, and scientific decision-making. These issues urgently require solutions through new modeling methods to achieve a more predictive and interpretable user operation analysis framework. Summary of the Invention
[0007] This invention proposes a user operation method based on a survival analysis model, comprising the following steps: S1. Data definition and mathematical modeling; S2. Descriptive analysis of user survival characteristics; S3. Comparative analysis of differences among different user groups; S4. Modeling and prediction of factors influencing user behavior; S5. Application of analysis results in operational decision-making.
[0008] Furthermore, in S1, the data definition and mathematical modeling aim to transform user operation problems into a mathematical framework for survival analysis and to mathematically model user behavior within the APP, thereby analogizing different concepts in survival analysis to the user operation field. These different concepts in survival analysis include the quantity of events, survival time, survival rate, censoring, and risks.
[0009] Furthermore, in S2, the descriptive analysis of user survival characteristics aims to describe the overall activity or retention characteristics of the user group from a macroscopic perspective. It uses non-parametric methods to estimate and visualize the survival function of the user group. Survival function In this scenario, the probability that a user survives longer than time t can be represented as: (1); Where k represents the number of observation times, n represents the number of samples, d represents the number of people whose events occurred at time t, and mj represents the number of people whose events were censored during the time interval (t, t+1). The survival function estimation formula is as follows: (2); Where tj represents the time point when the j-th event occurs, nj represents the number of users who were at risk before time tj, and dj represents the number of users whose events occurred at time tj. Survival curves were calculated and generated using the Kaplan-Meier method or life table method. The faster the curve drops, the shorter the time required for user revisits and the higher the activity level of the group; conversely, the flatter the curve, the longer the interval between user revisits and the lower the activity level.
[0010] Furthermore, in S3, the comparative analysis of differences among different user groups aims to compare whether there are essential differences in the activity levels of user groups with different strategies or attributes. The Log-rank test or Breslow test is used to perform statistical hypothesis testing on the survival curves of different user groups, and the chi-square value χ² is calculated. 2The p-value is used to determine the differences between multiple survival curves and to determine whether the corresponding user groups have similar or different behavioral patterns.
[0011] Furthermore, in S4, the modeling and prediction of user behavior influencing factors aims to identify key factors affecting user return visit time and construct a standard Cox proportional hazards regression model to quantitatively analyze the impact of covariates on user survival time. The risk function formula is as follows: (3); Where h(t, x) is the risk function at time t given the covariate x, h(t, 0) is the baseline risk function, and exp quantifies the multiplicative effect of different factors x on risk.
[0012] Furthermore, in S5, the application of the analysis results in operational decision-making is to transform the model analysis results into specific and executable business strategies in user operations, which can be used for user segmentation and clustering, churn warning and intervention, and personalized strategy optimization.
[0013] The beneficial effects of this invention are: 1. Achieving Dynamic and Refined User Insights: This invention integrates the "time" dimension into user behavior analysis, upgrading from a static result of "whether users are retained" to a dynamic process analysis of "when users will return." This allows operators to gain a more refined understanding of the rhythm and patterns of the user lifecycle.
[0014] 2. Provides objective and quantifiable user segmentation standards: Traditional user segmentation often relies on fixed parameters, which is subjective. This invention uses clustering and segmentation based on the overall survival curve pattern of user groups, resulting in more objective standards and stronger data-driven approaches.
[0015] 3. Empowering Precise Operations and Resource Optimization: Through methods such as the Cox Model, this invention can accurately identify key drivers influencing user activity. This enables businesses to precisely allocate limited operational budgets and resources to user groups and functional modules with the highest return on investment, achieving precise marketing and intervention.
[0016] 4. Improve the effectiveness of user churn warning: This invention can not only predict which users may churn, but more importantly, it can predict when is the best time to intervene, thereby retaining users when their churn tendency is strongest and significantly improving the recall success rate. Attached Figure Description
[0017] Figure 1 This is a flowchart of a user operation method based on a survival analysis model according to the present invention; Figure 2 This is a diagram illustrating the calculation process of the Kaplan-Meier method of this invention; Figure 3 This is a schematic diagram of the survival curve of the present invention; Figure 4 This is a diagram illustrating the calculation process of the lifetime method of this invention; Figure 5 The Log-rank test chart for the present invention is called Survival Functions. Figure 6 This invention uses the Breslow test chart for Product-Limit survival estimation. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] The following is in conjunction with the appendix Figure 1-2 The present invention will be further described as follows: A user operation method based on a survival analysis model includes the following steps: S1, Data Definition and Mathematical Modeling, aims to transform user operation problems into a mathematical framework for survival analysis and to mathematically model user behavior within the app. This allows different concepts in survival analysis (such as events, survival time, survival rate, censorship, and the quantity of risks) to be analogized to the user operation field, laying the foundation for the use of mathematical tools in subsequent steps.
[0020] Definition of an event: A key user action on a product (such as an app) is defined as an "event." The most typical event is "the user's next visit." Similarly, "the user creating an activity on the app" can also be considered an event.
[0021] Survival Time is defined as the length of time elapsed from a user's last point in time (such as the last time they used the app) to the occurrence of an "event." In the example above, the time from the start of the observation period to the user opening the app is defined as the survival time t.
[0022] Survival rate, also known as survival probability or survival function, represents the probability that, at any given moment within the observation period, the subject has not yet experienced the event. Using the example above, a higher survival rate indicates that the user has not opened the app by time t. Therefore, when characterizing active users, a lower survival rate generally indicates higher activity. This contradicts the medical field, where a higher survival rate is generally considered better.
[0023] The danger function (h(t)) represents the probability of the event occurring instantaneously after time t. In this case, it represents the probability that the user will instantly open the app after several hours.
[0024] Censoring is defined as the situation where a pre-defined "event" has not occurred by the end of a set observation period. For example, if a user does not visit the app again during the observation period, their lifespan is incomplete, and this situation is considered censored data.
[0025] Number of entities in risk: The number of entities whose status can be tracked during the observation period and for which no events have occurred. In the example above, users who did not open the app within a certain period are collectively referred to as the number of entities in risk.
[0026] Based on the above definitions, user online behavior within an app (especially in operational aspects) can be transformed into different events in survival analysis, with a specific operational time period defined as the survival time. Similarly, if an event does not occur, it is considered censored. For example, in an e-commerce promotion, a user receives a discount coupon. During the promotion, they place an order and use the coupon at checkout. In this scenario, using the coupon is a typical event, and the time from receiving the coupon to using it is the survival time. If the promotion ends and the coupon expires, and the user never uses the coupon, it is considered censored—more precisely, right-censored data.
[0027] Precautions: The data source is user behavior logs recorded by the server.
[0028] The definitions of "event" and "lifetime" must be clear and unique to ensure model consistency. For example, it should be clear whether "access" refers to opening the App or generating a valid interaction.
[0029] The purposes of survival analysis include: estimation, comparison, influencing factors, and prediction.
[0030] Estimation involves using sample survival data to estimate the overall survival rate and other relevant indicators. For example, in operations, it involves estimating when users will open the app within a certain timeframe. Starting from zero, it involves observing what percentage of users open / repeatedly open the app at each time point, and what type of users are opening the app. This type of data helps the app make decisions regarding important activities or key timing.
[0031] Comparison involves comparing the survival rates of different processing groups. For example, comparing the usage habits of different user groups of the same app can help to tailor operations to individual user preferences.
[0032] Influencing factor analysis involves exploring and understanding the factors that affect the length of time a product can be sold. For example, identifying the biggest privacy concerns influencing user habits, such as check-ins, points systems, or popular TV programs. This helps cultivate user habits and achieve targeted marketing goals.
[0033] Prediction refers to the prediction of individual survival at different levels of factors. For example, predicting the usage habits of new users for an existing app, or predicting the user base of a new app based on competitive analysis.
[0034] S2. Descriptive analysis of user survival characteristics aims to describe the overall activity or retention characteristics of the user group from a macro perspective.
[0035] Main content: This paper estimates and visualizes the survival function of a user group using a nonparametric method. In this scenario, the survival function S(t) represents the probability that a user survives longer than t (i.e., visits again after time t).
[0036] Specific calculation methods: Kaplan-Meier method or lifetime table method.
[0037] Suppose we have k observation times and N samples, where d represents the number of people whose events occur at time t, and mj represents the number of people censored during the time interval (t, t+1). Then the number of people still at risk at time t can be represented as: (1); Its survival function estimation formula is: (2); Where tj represents the time point when the j-th event occurs, nj represents the number of users at risk before time tj, and dj represents the number of users when the event occurs at time tj.
[0038] Output and Interpretation: Kaplan-Meier method or life table method: a method that explicitly describes a user's life status in the form of tables or curves.
[0039] Figure 2 Taking the Kaplan-Meier method as an example, the first three columns have relatively clear data definitions. The data in the fourth column represents the probability of surviving from the previous time period to the current time period. For example, if the survival time in the third row is 19, the fourth column represents the overall user survival probability from survival time 6 to survival time 19. The fifth column here is the overall survival rate. A simple calculation method is to divide the number of survivors by the total sample size of 19. For example, the overall survival rate in the third row = (19-2) / 19 = 0.8947.
[0040] from Figure 2 The following conclusions can be quickly drawn: While the survival rate gradually decreased at the beginning of the observation period, it increased after 42 days, indicating the effectiveness of the treatment. After treatment, the overall survival rate was approximately 0.525 by day 253.
[0041] generate Figure 3 The survival curve shown indicates that the faster the curve drops, the shorter the time required for users to return to their posts, and the higher the activity level of the group. Conversely, the flatter the curve, the longer the interval between user returns, and the lower the activity level.
[0042] The difference between the lifetime method and the Kaplan-Meier method is that the lifetime in this method is changed from a specific time to a time interval (second column). The calculation method for the other columns is the same as that of the Kaplan-Meier method.
[0043] like Figure 4 As shown, taking the third row as an example, when the number of years since diagnosis is greater than or equal to 4, there were 66 deaths, 1 missing person, and an initial number of 1170. The probability of death is calculated as 66 / 1070 = 0.0617, and the probability of survival is 1 - the probability of death = 0.9383. The overall survival rate is 0.9563 * 0.9596 * 0.9383 = 0.8610. The last column is the standard error of the survival rate.
[0044] S3. Comparative analysis of differences among different user groups: This step is performed when it is necessary to compare whether there are essential differences in the activity levels of user groups with different strategies or attributes.
[0045] Main content: Statistical hypothesis testing of survival curves for different user groups.
[0046] Specific calculation method: The Log-rank test or Breslow test is used. These two methods calculate the chi-square value χ². 2 The p-value is used to determine whether the differences between multiple survival curves are significant.
[0047] If the p-value is greater than 0.05 (or more strictly 0.01), it is generally considered that there is no significant difference between the curves, that is, the corresponding user groups have similar behavioral patterns.
[0048] If the p-value is <0.05 (or more strictly 0.01), the difference is considered statistically significant, and the corresponding user groups have fundamentally different behavioral patterns.
[0049] Note: The Breslow test assigns different weights to each time point (weight is the number of survivors) during the calculation, and is more sensitive to differences in earlier time points.
[0050] Log-rank test or Breslow test: When two survival curves are quite similar, making it difficult to intuitively determine which curve is better, we introduce two statistical hypothesis tests to determine which survival curve is superior. The Breslow test adds weight to the Log-rank test, setting the weight to the number of survivors at the beginning of each time point.
[0051] by Figure 5 For example, in Survival Functions, perform hypothesis testing: H0: The two groups of users exhibit the same usage behavior; H1: The two groups of users exhibit different usage behaviors.
[0052] The calculation yielded χ²=0.057, P=0.811>0.05, and H0 was not rejected at the significance level of 0.05. Therefore, it cannot be concluded that there is a difference in behavior between the two types of users. They can be considered as the same type of users.
[0053] by Figure 6 For example, in Product-Limit survival estimation, hypothesis testing is performed: H0: The two groups of users exhibit the same usage behavior; H1: The two groups of users exhibit different usage behaviors.
[0054] The calculated χ² = 7.3693, P = 0.0066 < 0.01. At the significance level of 0.01, the difference was statistically significant, and the H0 hypothesis was rejected, indicating a difference in behavior between the two groups of users. As for which group performed better, it depends on the actual needs and the specific definition of the event. For example, opening the app in a shorter time or earlier in the day.
[0055] S4. Modeling and predicting factors influencing user behavior is the core of the method of this invention, which aims to identify key factors affecting user return visit time and establish a predictive model.
[0056] Main content: Construct a standard Cox proportional hazards regression model to quantitatively analyze the impact of various covariates (such as demographic characteristics, consumption behavior, recommendation strategies, etc.) on user survival time (i.e., return visit time).
[0057] Specific calculation method: The standard Cox proportional hazards regression model is the benchmark model for analyzing influencing factors, and its hazard function formula is: (3); In a multi-factor scenario, the risk probability needs to consider both survival time T and the independent variable X. Here, h(t,x) is the risk function at time t (the probability of a user instantly returning) given the covariate x. If the independent variable is set to 0, h(t,0) is the baseline risk function, and exp quantifies the multiplicative effect of different factors x on the risk.
[0058] In the field of app data operations, the dependent variable h(t,x) represents the hazard function of the user at time t when the independent variable is x, that is, the probability that the user has not opened the app from time 0 to time t, but opens the app at time t.
[0059] By modeling this hazard function and analyzing its distribution, expected value, or other relevant data, we can obtain the time-related behavioral patterns of this type of user. For example, if the hazard function shows that it reaches its maximum value at noon, then marketing strategies targeting this group of users can be deployed in advance. Alternatively, if simultaneously opening the app would cause the system to crash, push notifications can be initiated some time before noon to distribute users.
[0060] S5. The application of analysis results in operational decision-making is to transform the model analysis results into specific and actionable business strategies in user operations, which can be used for user segmentation and clustering, churn warning and intervention, and personalized strategy optimization.
[0061] User segmentation and clustering: Based on the shape of the survival curve, users are divided into different groups such as highly active, moderately active, inactive, and churn-prone.
[0062] Churn warning and intervention: By analyzing the inflection point of the survival curve or the changes in the risk curve, determine the best time to intervene with potential churned users.
[0063] Personalized strategy optimization: Combining the influencing factors analyzed by the Cox model, personalized recall or recommendation strategies are implemented for different users. For example, if the model shows that "live stream exposure" is a key activity factor for a user, then live stream-related content can be added to the recall push notification. In summary, this invention is not limited to the specific embodiments described above. Those skilled in the art can make several modifications and alterations without departing from the spirit and scope of this invention. The scope of protection of this invention should be determined by the claims of this invention.
[0064] In summary, this invention is not limited to the specific embodiments described above. Those skilled in the art can make various modifications and alterations without departing from the spirit and scope of this invention. The scope of protection of this invention should be determined by the claims of this invention.
Claims
1. A user operation method based on a survival analysis model, characterized in that, Includes the following steps: S1. Data definition and mathematical modeling; S2. Descriptive analysis of user survival characteristics; S3. Comparative analysis of differences among different user groups; S4. Modeling and prediction of factors influencing user behavior; S5. Application of analysis results in operational decision-making.
2. The user operation method based on a survival analysis model according to claim 1, characterized in that, In S1, data definition and mathematical modeling aim to transform user operation problems into a mathematical framework for survival analysis and to mathematically model user behavior within the APP, thereby analogizing different concepts in survival analysis to the user operation field. These different concepts in survival analysis include the quantity of events, survival time, survival rate, censoring, and risks.
3. The user operation method based on a survival analysis model according to claim 2, characterized in that, In S2, the descriptive analysis of user survival characteristics aims to describe the overall activity or retention characteristics of the user group from a macroscopic perspective. It uses nonparametric methods to estimate and visualize the survival function of the user group. Survival function In this scenario, the probability that a user survives longer than time t can be represented as: （1）； Where k represents the number of observation times, n represents the number of samples, d represents the number of people whose events occurred at time t, and mj represents the number of people whose events were censored during the time interval (t, t+1). The survival function estimation formula is as follows: （2）； Where tj represents the time point when the j-th event occurs, nj represents the number of users who were at risk before time tj, and dj represents the number of users whose events occurred at time tj. The Kaplan-Meier method or life table is used to calculate and generate survival curves. The faster the curve drops, the shorter the time required for user revisits and the higher the activity level of the group. Conversely, the flatter the curve, the longer the interval between user revisits and the lower the activity level.
4. The user operation method based on a survival analysis model according to claim 3, characterized in that, In section S3, the comparative analysis of differences among different user groups aims to compare whether there are essential differences in the activity levels of user groups with different strategies or attributes. The Log-rank test or Breslow test is used to perform statistical hypothesis testing on the survival curves of different user groups, and the chi-square value χ² is calculated. 2 The p-value is used to determine the differences between multiple survival curves and to determine whether the corresponding user groups have similar or different behavioral patterns.
5. A user operation method based on a survival analysis model according to claim 4, characterized in that, In section S4, the modeling and prediction of user behavior influencing factors aims to identify key factors affecting user return visit time and construct a standard Cox proportional hazards regression model to quantitatively analyze the impact of covariates on user survival time. The risk function formula is as follows: （3）； Where h(t, x) is the risk function at time t given the covariate x, h(t, 0) is the baseline risk function, and exp quantifies the multiplicative effect of different factors x on risk.
6. A user operation method based on a survival analysis model according to claim 5, characterized in that, In S5, the application of the analysis results in operational decision-making is to transform the model analysis results into specific and executable business strategies in user operations, which can be used for user segmentation and clustering, churn warning and intervention, and personalized strategy optimization.