A user loss early warning method and related products
By constructing semantic vector sequences and trajectory features of users, and combining them with dynamic updates of semantic danger zones, the problem of lag in traditional user churn warnings has been solved, enabling more accurate churn warnings and interventions, and improving the accuracy and timeliness of user churn warnings.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIAMEN NANXUN CO LTD
- Filing Date
- 2026-02-25
- Publication Date
- 2026-05-29
AI Technical Summary
Traditional user churn early warning methods rely on manually preset trigger rules, resulting in delays and low success rates in recovering lost users, making it difficult to achieve precise intervention in user churn.
Based on users' historical behavior data, semantic vector sequences are constructed. By analyzing trajectory features and the distance to semantic danger zones, churn risk is determined. A semantic trajectory prediction model is used to predict users' future behavior, and danger zones are dynamically updated for accurate early warning.
It improves the accuracy of user churn warning, reduces false alarms and missed alarms, captures churn risks in advance, and reserves a time window for subsequent intervention, thereby enhancing the accuracy and effectiveness of user churn warning.
Smart Images
Figure CN122116595A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of data processing technology, and in particular to a user churn early warning method and related products. Background Technology
[0002] In customer relationship management, user churn prediction is a core requirement for maintaining user lifetime value and ensuring the healthy operation of the business.
[0003] Traditional user churn warning methods mainly rely on manually preset trigger rules, such as triggering an alert after 60 consecutive days of no consumption. However, this method of user churn warning based on manually preset trigger rules has a serious lag. By the time the alert is triggered, the user has often already churned, resulting in a very low success rate of recovery and making it difficult to support precise user churn intervention.
[0004] Improving the accuracy of user churn prediction is a technical problem that urgently needs to be solved. Summary of the Invention
[0005] To address the aforementioned issues, this application provides a user churn early warning method and related products, with the aim of improving the accuracy of user churn early warning.
[0006] The embodiments of this application disclose the following technical solutions: The first aspect of this application provides a user churn early warning method, the method comprising: Construct a semantic vector sequence for the user based on the user's historical behavior data; Based on the user's semantic vector sequence, determine the corresponding trajectory features; Based on the trajectory features, determine the user's vector coordinates at the next moment; The risk of user churn is determined based on the distance between the vector coordinates and the semantic danger zone, and an early warning is issued based on the risk of user churn; the semantic danger zone is dynamically updated based on the behavioral data of churned users.
[0007] Optionally, determining the corresponding trajectory features based on the user's semantic vector sequence includes: Calculate the corresponding semantic speed based on the user's semantic vector sequence; Calculate the corresponding semantic acceleration based on the user's semantic vector sequence; Based on the semantic velocity and the semantic acceleration, the corresponding trajectory features are determined.
[0008] Optionally, determining the user's vector coordinates at the next moment based on the trajectory features includes: Based on the trajectory features, the semantic trajectory prediction model is used to determine the user's vector coordinates at the next moment; the semantic trajectory prediction model is used to predict the user's vector coordinates at the next moment based on the user's semantic drift trajectory.
[0009] Optionally, determining the user churn risk based on the distance between the vector coordinates and the semantically dangerous region includes: Calculate the distance between the vector coordinates and the boundary of the semantically dangerous region; If the distance is less than or equal to a preset distance threshold, it indicates that the user has entered a dangerous area and an alarm is triggered; if the distance is greater than the preset distance threshold, it indicates that the user is in a safe area and no alarm is triggered.
[0010] Optionally, after triggering an alarm if the distance is less than or equal to a preset distance threshold, indicating that the user has entered a dangerous area, the system further includes: Convert the vector coordinates into corresponding natural language labels; Based on the natural language tags, corresponding intervention actions are matched to intervene in the user's behavior.
[0011] Optionally, the method for obtaining the semantically dangerous region includes: Clustering algorithms are used to cluster the behavioral data of churned users, resulting in multiple cluster centers; Based on the multiple cluster centers and the preset distance threshold, semantically dangerous regions are determined.
[0012] A second aspect of this application provides a user churn early warning device, the device comprising: The semantic vector sequence construction module is used to construct a user's semantic vector sequence based on the user's historical behavior data; The trajectory feature determination module is used to determine the corresponding trajectory features based on the user's semantic vector sequence; The vector coordinate determination module is used to determine the user's vector coordinates at the next moment based on the trajectory features. The user churn warning module is used to determine the risk of user churn based on the distance between the vector coordinates and the semantic danger zone, and to issue a warning based on the risk of user churn; the semantic danger zone is dynamically updated based on the behavioral data of churned users.
[0013] Optionally, the trajectory feature determination module is used for: Calculate the corresponding semantic speed based on the user's semantic vector sequence; Calculate the corresponding semantic acceleration based on the user's semantic vector sequence; Based on the semantic velocity and the semantic acceleration, the corresponding trajectory features are determined.
[0014] A third aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a user churn warning method as provided in any implementation of the first aspect.
[0015] The fourth aspect of this application provides a processor for running a computer program that, when running, executes a user churn warning method as provided in any implementation of the first aspect.
[0016] Compared with the prior art, this application has the following advantages: The user churn early warning method provided in this application constructs a semantic vector sequence of users based on their historical behavior data; determines the corresponding trajectory features based on the semantic vector sequence of users; determines the vector coordinates of users at the next moment based on the trajectory features; determines the user's churn risk based on the distance between the vector coordinates and the semantic danger zone, and issues an early warning based on the user's churn risk; the semantic danger zone is dynamically updated based on the behavioral data of churned users.
[0017] The above methods transform users' historical behavior data into semantic vector sequences, improving the accuracy of user churn warnings from a data representation perspective and reducing warning bias caused by misinterpretation of shallow features. Based on the semantic vector sequences, corresponding trajectory features are determined. These trajectory features depict the dynamic changes in user behavior, helping to capture churn risk as soon as negative changes in user behavior occur, significantly improving the accuracy of user churn warnings. Predicting the user's vector coordinates at the next moment based on trajectory features provides a proactive prediction of future user behavior tendencies, allowing warnings to precede actual churn behavior and reserving sufficient time for subsequent intervention, further improving the accuracy of user churn warnings. Determining user churn risk based on the distance between vector coordinates and semantic danger zones transforms churn risk from an ambiguous yes / no to a quantifiable value, more accurately reflecting the degree of alignment between user behavior and churn characteristics, reducing misjudgments and omissions caused by rigid rules, and contributing to improved accuracy of user churn warnings. Attached Figure Description
[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0019] Figure 1 A flowchart illustrating a user churn early warning method provided in this application embodiment; Figure 2 This application provides an illustration of the effect of early warning and intervention based on user behavior trajectories in an embodiment of the present application. Figure 3 This is a schematic diagram of a user churn early warning device provided in an embodiment of this application. Detailed Implementation
[0020] As described earlier, current user churn warning methods mainly rely on manually preset trigger rules, such as triggering a warning after 60 consecutive days of no consumption. However, this method of user churn warning based on manually preset trigger rules has a serious lag. By the time the warning is triggered, the user has often already churned, resulting in a very low success rate of recovery and making it difficult to support precise user churn intervention.
[0021] In view of the above problems, this application proposes a user churn early warning method and related products. The method includes: constructing a semantic vector sequence of the user based on the user's historical behavior data; determining the corresponding trajectory features based on the user's semantic vector sequence; determining the user's vector coordinates at the next moment based on the trajectory features; determining the user's churn risk based on the distance between the vector coordinates and the semantic danger zone, and issuing an early warning based on the user's churn risk; the semantic danger zone is dynamically updated based on the behavioral data of churned users.
[0022] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.
[0023] See Figure 1 This figure is a flowchart of a user churn early warning method provided in an embodiment of this application. Figure 1 As shown, the method includes the following steps: S101. Construct a semantic vector sequence for the user based on the user's historical behavior data.
[0024] Construct a sequence of semantic vectors for each user within a fixed time window, such as weekly.
[0025] User behavior data, such as browsing, searching, or ratings, is transformed into a continuous and time-ordered sequence of semantic vectors, enabling machines to understand the changing trends of user behavior in the same way they understand motion trajectories.
[0026] There are no restrictions on how users construct their semantic vectors, such as using natural language models like BERT or Sentence-BERT.
[0027] For example, in a scenario where a subscription-based coffee brand aims to prevent existing users from canceling their subscriptions, a semantic vector sequence for user Alice is constructed based on her actual consumption, interactions, complaints, and competitor browsing behavior data over the past four weeks: Week 1: [Likes new products, high interaction]; Week 2: [Normal consumption, no feedback]; Week 3: [Delivery delay complaint, slight anger]; Week 4: [Browsing competitor pages, increased frequency]. This semantic vector sequence is a highly condensed and tagged summary of user Alice's behavioral characteristics for each week. Arranging the four-week behavioral vectors in chronological order forms a continuous behavioral trajectory, clearly demonstrating the user's behavioral changes and helping to accurately identify user churn risk.
[0028] S102. Determine the corresponding trajectory features based on the user's semantic vector sequence.
[0029] In one feasible implementation: The corresponding semantic speed is calculated based on the user's semantic vector sequence.
[0030] Based on the user's semantic vector sequence, the corresponding semantic acceleration is calculated.
[0031] Based on the semantic velocity and the semantic acceleration, the corresponding trajectory features are determined.
[0032] Semantic velocity represents the distance between the semantic vectors of two adjacent time windows divided by the time interval, reflecting how quickly user needs change. Semantic acceleration represents the rate of change of semantic velocity, reflecting the drastic degree of change in user needs.
[0033] For example, based on the semantic vector sequence of user Alice over the past four weeks in S101, and based on the calculated semantic velocity and semantic acceleration, the corresponding trajectory feature is obtained as user Alice moving away from the loyalty cluster with high acceleration and towards the competitor swing cluster. The loyalty cluster represents a cluster formed by the aggregation of behavioral vectors of users who have long-term stable subscriptions and high brand stickiness in the semantic vector space. The competitor swing cluster represents a cluster formed by the aggregation of behavioral vectors of users whose satisfaction with the current brand has decreased and who have begun to actively pay attention to competitors, but have not yet made a final decision to cancel their subscription or switch brands.
[0034] S103. Based on trajectory features, determine the user's vector coordinates at the next moment.
[0035] In one feasible implementation: Based on the trajectory features, the semantic trajectory prediction model is used to determine the user's vector coordinates at the next moment; the semantic trajectory prediction model is used to predict the user's vector coordinates at the next moment based on the user's semantic drift trajectory.
[0036] Semantic trajectory prediction models can employ a time-series-based Transformer architecture to learn the inertia and periodicity of user semantic drift. Inertia represents the continuity of user behavior, while periodicity represents the repetitive patterns of user behavior. The time-series-based Transformer can better handle the long-term temporal dependencies and periodic changes in user behavior, accurately capturing the inertia and periodicity of user behavior, thereby more accurately predicting the vector coordinates of the next time window.
[0037] For example, the semantic trajectory prediction model predicts that the vector coordinates of user Alice in week five are extremely close to the churn cluster center, and contain high-dimensional features such as price comparison and service disappointment. The churn cluster center represents a cluster formed by the aggregation of behavioral vectors of users who have canceled their subscriptions or switched coffee brands in the semantic vector space.
[0038] S104. Based on the distance between the vector coordinates and the semantic danger area, determine the risk of user churn and issue an early warning based on the risk of user churn.
[0039] The semantic danger zone is dynamically updated based on the behavioral data of churned users.
[0040] In one feasible implementation: Calculate the distance between the vector coordinates and the boundary of the semantically dangerous region.
[0041] If the distance is less than or equal to a preset distance threshold, it indicates that the user has entered a dangerous area and an alarm is triggered; if the distance is greater than the preset distance threshold, it indicates that the user is in a safe area and no alarm is triggered.
[0042] In one feasible implementation, the method for obtaining semantically dangerous regions includes: Clustering algorithms are used to cluster the behavioral data of churned users, resulting in multiple cluster centers.
[0043] Based on the multiple cluster centers and the preset distance threshold, semantically dangerous regions are determined.
[0044] By setting a preset distance threshold, if the distance between the predicted vector and the cluster center is less than or equal to the preset distance threshold, it indicates that the user has entered a dangerous area and an alarm is triggered; if the distance between the predicted vector and the cluster center is greater than the distance threshold, it indicates that the user is safe and no alarm is triggered.
[0045] The above method transforms the vague judgment of potential user churn into a precise numerical value of vector distance, providing an objective quantitative standard for risk warning. Different distance thresholds can be set based on the distance, subdividing user churn risk into multiple levels such as extremely high risk, high risk, medium risk, and low risk. This refined risk classification allows the operations team to match differentiated warning and intervention strategies for users with different risk levels.
[0046] The semantic danger zone is dynamically updated based on the behavioral data of the latest churned users. Its cluster center and boundary are adjusted synchronously with changes in business scenarios. This dynamic update mechanism ensures that the benchmark for risk judgment always fits the current churn characteristics, avoids warning deviations caused by benchmark lag, and allows the warning model to maintain high accuracy at different business stages.
[0047] In one feasible implementation, after triggering an alarm by indicating that the user has entered a dangerous area if the distance is less than or equal to a preset distance threshold, the system further includes: The vector coordinates are converted into corresponding natural language labels.
[0048] Based on the natural language tags, corresponding intervention actions are matched to intervene in the user's behavior.
[0049] For example, in week five, the distance between user Alice's vector and the semantic danger zone was less than a preset distance threshold, indicating a high probability that user Alice would churn within the next 7 days. The predicted main reasons were accumulated dissatisfaction due to previous logistics issues and the temptation of lower prices from competitors. At this point, user Alice had not yet canceled her subscription. The corresponding intervention was to send a special SMS message for loyal customers, including a no-minimum-spend upgrade coupon and a message thanking her for her recent understanding. This message proactively addressed the predicted logistics dissatisfaction. Ultimately, user Alice felt valued upon receiving the message and abandoned her original plan to cancel her subscription in week five.
[0050] like Figure 2 As shown, Figure 2This diagram illustrates the effectiveness of early warning and intervention based on user behavior trajectories. It clearly demonstrates the complete process from user behavior tracking and churn risk prediction to preemptive intervention using a two-dimensional semantic vector space. Semantic dimensions X and Y represent two feature dimensions of user behavior in the semantic space; for example, the X-axis represents brand loyalty and the Y-axis represents service satisfaction. Each user behavior is encoded as a point in this two-dimensional space. The blue cloud-shaped area is the safety zone, representing the cluster of brand loyalty and the aggregation of behavioral vectors from healthy users with long-term stable subscriptions and high brand loyalty. The red cloud-shaped area is the danger zone, representing the cluster of churn warning clusters and the aggregation of behavioral vectors from users who have canceled their subscriptions or show a clear tendency to churn. The semi-circular boundary formed by the red dashed lines is the semantic electronic fence, the electronic warning line for the danger zone. Figure 2 The blue dots and lines represent the changes in user behavior vectors at different points in time, forming a continuous temporal semantic trajectory. T-3, T-2, T-1, and T0 represent the user's past behavior trajectory to the present. At the current time T0, the user's behavior is still within the safe zone, but the predicted vector coordinates for the user's next time T+1 have already entered the semantic electronic fence. At this point, an early warning is issued, and preemptive intervention is implemented to redirect the user's behavior trajectory from moving towards the churn warning cluster back to the brand loyalty cluster, thereby preventing user churn.
[0051] Decoding the predicted vector into natural language labels, for example, predicting that a user will become extremely price-sensitive and competitively inclined next week, triggers an intervention by sending caring messages before the user actually decides not to buy or cancels their subscription. This approach proactively pushes relevant caring content before the user actually takes negative actions like not buying or canceling, intervening in the user's decision-making process in advance. It allows the user to feel the platform's targeted care or benefits before making a final negative decision, thereby changing their behavioral choices and preventing negative behavior. Moreover, each step is precisely intervened based on the user's specific behavioral tendencies, avoiding the resource waste and user resentment caused by traditional, extensive operations.
[0052] The user churn early warning method provided in this application embodiment constructs a semantic vector sequence of users based on users' historical behavior data; determines corresponding trajectory features based on the user's semantic vector sequence; determines the user's vector coordinates at the next moment based on the trajectory features; determines the user's churn risk based on the distance between the vector coordinates and the semantic danger zone, and issues an early warning based on the user's churn risk; the semantic danger zone is dynamically updated based on the behavioral data of churned users.
[0053] The above methods transform users' historical behavior data into semantic vector sequences, improving the accuracy of user churn warnings from a data representation perspective and reducing warning bias caused by misinterpretation of shallow features. Based on the semantic vector sequences, corresponding trajectory features are determined. These trajectory features depict the dynamic changes in user behavior, helping to capture churn risk as soon as negative changes in user behavior occur, significantly improving the accuracy of user churn warnings. Predicting the user's vector coordinates at the next moment based on trajectory features provides a proactive prediction of future user behavior tendencies, allowing warnings to precede actual churn behavior and reserving sufficient time for subsequent intervention, further improving the accuracy of user churn warnings. Determining user churn risk based on the distance between vector coordinates and semantic danger zones transforms churn risk from an ambiguous yes / no to a quantifiable value, more accurately reflecting the degree of alignment between user behavior and churn characteristics, reducing misjudgments and omissions caused by rigid rules, and contributing to improved accuracy of user churn warnings.
[0054] Based on the user churn early warning method described in the preceding embodiments, this application also provides a user churn early warning device. Figure 3 This is a schematic diagram of the device. Figure 3 As shown, the user churn warning device includes: The semantic vector sequence construction module 301 is used to construct the semantic vector sequence of a user based on the user's historical behavior data.
[0055] The trajectory feature determination module 302 is used to determine the corresponding trajectory features based on the user's semantic vector sequence.
[0056] The vector coordinate determination module 303 is used to determine the user's vector coordinates at the next moment based on the trajectory features.
[0057] The user churn warning module 304 is used to determine the user churn risk based on the distance between the vector coordinates and the semantic danger zone, and to issue a warning based on the user churn risk; the semantic danger zone is dynamically updated based on the behavioral data of churned users.
[0058] Optionally, the trajectory feature determination module is used for: The corresponding semantic speed is calculated based on the user's semantic vector sequence.
[0059] Based on the user's semantic vector sequence, the corresponding semantic acceleration is calculated.
[0060] Based on the semantic velocity and the semantic acceleration, the corresponding trajectory features are determined.
[0061] Optionally, the vector coordinate determination module is used for: Based on the trajectory features, the semantic trajectory prediction model is used to determine the user's vector coordinates at the next moment; the semantic trajectory prediction model is used to predict the user's vector coordinates at the next moment based on the user's semantic drift trajectory.
[0062] Optionally, the user churn warning module is used for: Calculate the distance between the vector coordinates and the boundary of the semantically dangerous region.
[0063] If the distance is less than or equal to a preset distance threshold, it indicates that the user has entered a dangerous area and an alarm is triggered; if the distance is greater than the preset distance threshold, it indicates that the user is in a safe area and no alarm is triggered.
[0064] Optionally, the device further includes: an intervention module; The intervention module is used to convert the vector coordinates into corresponding natural language labels.
[0065] Based on the natural language tags, corresponding intervention actions are matched to intervene in the user's behavior.
[0066] Optionally, the method for obtaining the semantically dangerous region includes: Clustering algorithms are used to cluster the behavioral data of churned users, resulting in multiple cluster centers.
[0067] Based on the multiple cluster centers and the preset distance threshold, semantically dangerous regions are determined.
[0068] Furthermore, embodiments of this application also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the user churn warning method as described in any of the method embodiments.
[0069] Furthermore, this application embodiment also provides a processor for running a computer program, which executes the user churn warning method as described in any of the aforementioned method embodiments.
[0070] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for the device embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiments. The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separate. The components indicated as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment solution according to actual needs. Those skilled in the art can understand and implement this without creative effort.
[0071] The above description is merely one specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A user churn early warning method, characterized in that, include: Construct a semantic vector sequence for the user based on the user's historical behavior data; Based on the user's semantic vector sequence, determine the corresponding trajectory features; Based on the trajectory features, determine the user's vector coordinates at the next moment; Based on the distance between the vector coordinates and the semantic danger zone, the risk of user churn is determined, and an early warning is issued based on the risk of user churn. The semantic danger zone is dynamically updated based on the behavioral data of churned users.
2. The method according to claim 1, characterized in that, The step of determining the corresponding trajectory features based on the user's semantic vector sequence includes: Calculate the corresponding semantic speed based on the user's semantic vector sequence; Calculate the corresponding semantic acceleration based on the user's semantic vector sequence; Based on the semantic velocity and the semantic acceleration, the corresponding trajectory features are determined.
3. The method according to claim 1, characterized in that, Determining the user's vector coordinates at the next moment based on the trajectory features includes: Based on the trajectory features, the semantic trajectory prediction model is used to determine the user's vector coordinates at the next moment; the semantic trajectory prediction model is used to predict the user's vector coordinates at the next moment based on the user's semantic drift trajectory.
4. The method according to claim 1, characterized in that, The determination of user churn risk based on the distance between the vector coordinates and the semantically dangerous region includes: Calculate the distance between the vector coordinates and the boundary of the semantically dangerous region; If the distance is less than or equal to a preset distance threshold, it indicates that the user has entered a dangerous area and an alarm is triggered; if the distance is greater than the preset distance threshold, it indicates that the user is in a safe area and no alarm is triggered.
5. The method according to claim 4, characterized in that, If the distance is less than or equal to a preset distance threshold, indicating that the user has entered a dangerous area and triggering an alarm, the following steps are also included: Convert the vector coordinates into corresponding natural language labels; Based on the natural language tags, corresponding intervention actions are matched to intervene in the user's behavior.
6. The method according to claim 1, characterized in that, The methods for obtaining the semantically dangerous regions include: Clustering algorithms are used to cluster the behavioral data of churned users, resulting in multiple cluster centers; Based on the multiple cluster centers and the preset distance threshold, semantically dangerous regions are determined.
7. A user churn early warning device, characterized in that, include: The semantic vector sequence construction module is used to construct a user's semantic vector sequence based on the user's historical behavior data; The trajectory feature determination module is used to determine the corresponding trajectory features based on the user's semantic vector sequence; The vector coordinate determination module is used to determine the user's vector coordinates at the next moment based on the trajectory features. The user churn warning module is used to determine the risk of user churn based on the distance between the vector coordinates and the semantic danger area, and to issue a warning based on the risk of user churn. The semantic danger zone is dynamically updated based on the behavioral data of churned users.
8. The apparatus according to claim 7, characterized in that, The trajectory feature determination module is used for: Calculate the corresponding semantic speed based on the user's semantic vector sequence; Calculate the corresponding semantic acceleration based on the user's semantic vector sequence; Based on the semantic velocity and the semantic acceleration, the corresponding trajectory features are determined.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the user churn warning method as described in any one of claims 1-6.
10. A processor, characterized in that, Used to run a computer program, which, when running, executes the user churn warning method as described in any one of claims 1-6.