A Smart Early Warning Method and System for User Churn on Trading Platforms

Abstract

This application relates to the technical field of e-commerce platform monitoring, and discloses a method and system for intelligent early warning of user churn on a trading platform. The method includes: collecting user operation behavior through an event processing engine, constructing a user behavior timeline, and performing sequence matching based on preset composite behavior patterns to generate a preliminary early warning signal; calculating a behavior deviation index by combining user group attribute tags, time-sensitive period identification, and a sliding window algorithm, and dynamically generating personalized early warning thresholds using a machine learning model to improve early warning accuracy; after confirming the triggering of a formal early warning, based on a hierarchical recall strategy library, integrating user tag layers, scenario layers, and strategy pool layers, matching applicable tags with dominant risk scenarios, screening a set of candidate strategies, evaluating historical conversion effects through real-time A / B testing, dynamically adjusting strategy weights, generating the optimal recall strategy combination, and executing it. This application can achieve accurate and effective user recovery.

Description

Technical Field

[0001] This application relates to the technical field of e-commerce platform monitoring, and in particular to a method and system for intelligent early warning of user churn on a trading platform. Background Technology

[0002] Currently, e-commerce platforms generally rely on static rules based on fixed thresholds for user churn warnings, such as using "not logging in for 7 consecutive days" as the criterion. These methods fail to consider differences in individual user behavior habits, making it difficult to adapt to the behavioral patterns of different user groups. This results in a high false alarm rate, delayed response, and a lack of refined identification capabilities for churn risk.

[0003] Meanwhile, most user recall strategies employ a uniform push notification approach, lacking a mechanism to adapt the strategy selection to user scenarios and profiles. This results in limited intervention effectiveness and makes it difficult to achieve precise and effective user recovery. Therefore, how to achieve precise and effective user recovery remains a challenge. Summary of the Invention

[0004] To achieve accurate and effective user recovery, this application provides a method and system for intelligent early warning of user churn on a trading platform.

[0005] Firstly, this application provides a method for intelligent early warning of user churn on a trading platform, employing the following technical solution:

[0006] A smart early warning method for user churn on a trading platform includes:

[0007] The event processing engine collects user behavior data within the platform in real time, constructing a user behavior timeline with timestamps as the axis. This behavior data includes page browsing, adding items to cart, order creation, payment attempts, and login status changes. At least one composite behavior pattern is defined as a pre-defined churn warning template, where each composite behavior pattern is a preset combination of behavior sequences. A rules engine performs pattern matching on the current user's behavior timeline, calculating the sequence matching score between the current user and each composite behavior pattern. When the sequence matching score of any composite behavior pattern exceeds a preset initial threshold, a preliminary warning signal is generated, and the current user's sequence matching score and the corresponding composite behavior pattern are identified to obtain the corresponding user identifier.

[0008] The system acquires the current user's historical behavior data and extracts the user's group attribute tags, including user type tags, purchase frequency levels, and average order value ranges. Combining this with time context information, it identifies the current time-sensitive period, including promotional periods, holidays, weekdays, and platform system downtime. A sliding window algorithm is used to calculate the current user's historical activity benchmark value within the same time period, and a behavior deviation index is calculated relative to the historical benchmark. This deviation index is weighted by the login frequency decline rate, transaction interruption frequency, and page dwell time reduction ratio. Based on the sequence matching score and the behavior deviation index, a linear regression or decision tree model is used to dynamically adjust the warning threshold, generating a personalized warning threshold for each user. If the current sequence matching score is higher than this personalized warning threshold, a formal warning is triggered.

[0009] A hierarchical recall strategy library is constructed, comprising a user tag layer storing static and dynamic tags generated based on user profiles, a scenario layer corresponding to the user's current behavioral scenario, and a strategy pool layer containing multiple executable recall strategies and their weight configurations. The user tag layer includes price-sensitive, logistics-conscious, and service-dependent users. The scenario layer includes incomplete payment, order cancellation without repurchase, and silence after negative feedback. The strategy pool layer includes issuing targeted coupons, initiating dedicated customer service outbound calls, providing free returns and exchanges, and pushing personalized recommendations. Applicable tags in the user tag layer are matched based on the composite behavioral pattern type and the official warning result, and the current dominant churn risk scenario is determined in conjunction with the scenario layer. A set of candidate strategies matching the applicable tags and the current dominant churn risk scenario is selected from the strategy pool layer. The historical conversion effect data of the candidate strategies in the candidate strategy set is evaluated through a real-time A / B testing module, the execution weight of each candidate strategy is dynamically adjusted, and the optimal recall strategy combination is generated and executed.

[0010] Optionally, the method further includes:

[0011] Obtain the recall strategy execution results corresponding to the recall strategy execution. Based on the composite behavior pattern, sequence matching score, behavior deviation index and recall strategy execution results, perform churn attribution analysis on users who triggered formal warnings. Identify key influencing factors that lead to user churn and generate attribution analysis results including the ranking of main causes. The key influencing factors include one or more of the following: number of payment failures, customer service response delay, logistics timeliness deviation, and competitor price fluctuations.

[0012] Based on the attribution analysis results, add or strengthen frequently associated behavioral sequences in the composite behavioral pattern library; adjust the weight coefficients of each dimension in the behavioral deviation index based on the attribution analysis results.

[0013] Optionally, the method further includes:

[0014] When evaluating the historical conversion performance data of each candidate strategy in the candidate strategy set through the real-time A / B testing module, a dynamic weight allocation mechanism is used to adjust the execution weight of each candidate strategy.

[0015] The dynamic weight allocation mechanism calculates the priority of a strategy based on its historical conversion rate and its matching degree with the current user behavior scenario; whereby the historical conversion rate includes user click-through rate, order conversion rate, and user retention rate.

[0016] The scoring parameters of each candidate strategy are updated in real time through an online learning mechanism. Based on the scoring parameters, the corresponding high-performance strategy is determined. The high-performance strategy is executed first when a warning is triggered.

[0017] Optionally, a user behavior correlation analysis module can be introduced into the user tag layer. Other methods include:

[0018] The user behavior association analysis module generates dynamic user tag vectors by analyzing social associations, transaction link associations, and behavioral sequence similarities among users; social associations include jointly purchasing goods, browsing similar pages, and rating interactions;

[0019] The static tags are supplemented based on the dynamic user tag vector. The static tags include price-sensitive and logistics-focused tags, and the weights of each user tag are dynamically adjusted through a context-aware weight adjustment mechanism.

[0020] Optionally, in the process of identifying time-sensitive periods, the method further includes:

[0021] By combining external environmental data with historical user behavior data, a time-series feature extraction model is used to predict the trend of user behavior activity in future time periods. External environmental data includes e-commerce platform promotion calendars, holiday databases, and regional weather data.

[0022] The time window length of the sliding window algorithm is dynamically adjusted based on the user behavior activity trend, and the abnormal misjudgment caused by time offset is eliminated by the behavior sequence alignment algorithm; in non-sensitive periods, occasional behavior deviations are smoothed by noise filtering mechanism.

[0023] Optionally, a user behavior semantic analysis module is introduced into the attribution analysis results, and the method further includes:

[0024] The user behavior semantic analysis module extracts sentiment tags by analyzing users' text comments and behaviors on the platform, and quantifies the contribution of sentiment tags to user churn. Text comments include order reviews and customer service conversation records, behavioral sequences include frequent order cancellations and repeated unpaid add-to-carts, and sentiment tags include anxiety, dissatisfaction, and hesitation.

[0025] By combining four key influencing factors—the number of payment failures, customer service response delays, logistics timeliness deviations, and competitor price fluctuations—a causal inference model is used to calculate the attribution weight of each factor and output a dynamic explanation report that includes the ranking of the main causes. The dynamic explanation report is used to optimize the generation model parameters of personalized warning thresholds and guide the priority ranking of recall strategies in the strategy pool layer.

[0026] Secondly, this application provides an intelligent early warning system for user churn on a trading platform, which adopts the following technical solution:

[0027] A smart early warning system for user churn on a trading platform, comprising:

[0028] The behavior modeling and preliminary early warning module collects user operation behavior data within the platform in real time through an event processing engine, constructing a user behavior timeline with timestamps as the axis. The operation behavior data includes page browsing, adding items to cart, order creation, payment attempts, and changes in login status. At least one composite behavior pattern is defined as a churn precursor template, and the composite behavior pattern is a preset combination of behavior sequences. Based on the rule engine, the module performs pattern matching on the current user's user behavior timeline, calculating the sequence matching score between the current user and each composite behavior pattern. When the sequence matching score of any composite behavior pattern exceeds a preset initial threshold, a preliminary early warning signal is generated, and the current user's sequence matching score and the corresponding composite behavior pattern are identified to obtain the corresponding user identifier.

[0029] The dynamic baseline assessment and formal warning confirmation module acquires the current user's historical behavior data and extracts the current user's group attribute tags, including user type tags, purchase frequency levels, and average order value ranges. Combined with time context information, it identifies the current time-sensitive period, including promotional periods, holidays, weekdays, and platform system outage periods. A sliding window algorithm is used to calculate the current user's historical behavioral activity benchmark value within the same time period, and a behavioral deviation index is calculated relative to the historical benchmark. This behavioral deviation index is weighted by the login frequency decline rate, transaction interruption frequency, and page dwell time reduction ratio. Based on the sequence matching score and the behavioral deviation index, a linear regression or decision tree model is used to dynamically adjust the warning threshold, generating a personalized warning threshold for each user. If the current sequence matching score is higher than this personalized warning threshold, it is used to confirm and trigger a formal warning.

[0030] The layered recall strategy generation and execution module constructs a layered recall strategy library. This library includes a user tag layer storing static and dynamic tags generated based on user profiles, a scenario layer corresponding to the user's current behavioral scenario, and a strategy pool layer containing various executable recall strategies and their weight configurations. The user tag layer includes price-sensitive, logistics-conscious, and service-dependent users. The scenario layer includes incomplete payment, order cancellation without repurchase, and silence after negative feedback. The strategy pool layer includes issuing targeted coupons, initiating dedicated customer service outbound calls, providing free returns and exchanges, and pushing personalized recommendations. Based on the composite behavioral pattern type and the official warning result, the module matches applicable tags from the user tag layer and combines this with the scenario layer to determine the current dominant churn risk scenario. It then selects a set of candidate strategies from the strategy pool layer that match the applicable tags and the current dominant churn risk scenario. Finally, a real-time A / B testing module evaluates the historical conversion effect data of the candidate strategies in the set, dynamically adjusting the execution weights of each candidate strategy to generate and execute the optimal recall strategy combination.

[0031] Thirdly, this application provides an intelligent early warning system for user churn on a trading platform, which adopts the following technical solution:

[0032] A method for intelligent early warning of user churn on a trading platform includes a processor, wherein the processor runs a program for the method for intelligent early warning of user churn on a trading platform as described in any one of the above-mentioned methods.

[0033] Fourthly, this application provides a storage medium, which adopts the following technical solution:

[0034] A storage medium storing a program for the intelligent early warning method for user churn on a trading platform as described in any one of the above.

[0035] In summary, this application includes at least one of the following beneficial technical effects:

[0036] By constructing a sequence matching mechanism that combines user behavior timelines with complex behavioral patterns, and integrating it with dynamic baseline assessment, the system achieves accurate identification of user churn risk. By introducing group attribute tags and time-sensitive periods, it utilizes a sliding window algorithm to calculate individualized historical behavioral activity benchmarks and quantifies behavioral deviation indices based on indicators such as decreased login frequency and transaction interruption frequency. Furthermore, it dynamically generates personalized early warning thresholds by integrating sequence matching scores, effectively avoiding false alarms and missed detections caused by traditional static rules, and significantly improving the accuracy and timeliness of early warnings.

[0037] Building upon this foundation, a tiered recall strategy library is constructed to achieve multi-dimensional matching of user tags, behavioral scenarios, and executable strategies, ensuring that intervention measures are highly aligned with the actual reasons for user churn. Real-time A / B testing is used to evaluate the historical conversion effects of candidate strategies and dynamically adjust their execution weights, generating the optimal recall strategy combination. This makes recovery actions such as offering discounts, customer service intervention, and logistics compensation more targeted. This mechanism achieves a shift from "broad notification" to "precise intervention," significantly improving the conversion efficiency and success rate of user recall. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating an intelligent early warning method for user churn on a trading platform, according to an exemplary embodiment.

[0039] Figure 2 This is a structural block diagram illustrating an intelligent early warning system for user churn on a trading platform, according to an exemplary embodiment. Detailed Implementation

[0040] The embodiments of this application are described in detail below, and examples of the embodiments are shown in the accompanying drawings.

[0041] In the description of this specification, the references to "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples" refer to specific features, structures, materials, or characteristics described in connection with the described embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0042] This application discloses an intelligent early warning method for user churn on a trading platform, referring to... Figure 1 ,include:

[0043] S100 uses an event processing engine to collect real-time user behavior data within the platform, constructing a user behavior timeline with timestamps as the axis. This behavior data includes page browsing, adding items to cart, order creation, payment attempts, and changes in login status. At least one composite behavior pattern is defined as a pre-defined churn indicator template, where the composite behavior pattern is a pre-defined combination of behavior sequences. A rules engine performs pattern matching on the current user's behavior timeline, calculating the sequence matching score between the current user and each composite behavior pattern. When the sequence matching score of any composite behavior pattern exceeds a pre-defined initial threshold, a preliminary warning signal is generated, and the current user's sequence matching score and the corresponding composite behavior pattern are identified to obtain the corresponding user identifier.

[0044] The execution process of S100 specifically includes the following steps:

[0045] Step 1: Collect user behavior in real time and build a behavior timeline:

[0046] The system continuously acquires all user actions on the platform through a real-time data collection module. For example, actions such as browsing product pages, adding products to the shopping cart, creating an order but not paying, failing to pay, logging in or out of the account are all recorded. Each action event is accompanied by a precise timestamp to mark the exact moment the action occurred. This behavioral data is then organized into a timeline indexed by user identity (such as user ID), arranging all action events in chronological order. For example, user U12345's timeline might include: "August 20, 2025, 9:00 AM: Browsed product A," "August 20, 2025, 9:05 AM: Added product A to the shopping cart," "August 20, 2025, 9:10 AM: Created an order but not paid," etc. The purpose of the timeline is to completely record all user actions from registration to the present, providing foundational data support for subsequent analysis.

[0047] Step 2, define a composite behavioral pattern as a template for early churn:

[0048] The system pre-defines multiple composite behavioral patterns, based on typical churn signals summarized from historical user churn cases and business experience. Each pattern consists of a series of actions in a specific sequence, constrained by a time frame. For example, Pattern 1 is defined as "a user adds an item to their shopping cart, then creates an order but doesn't pay, followed by payment failure, and finally doesn't log in for three consecutive days"; Pattern 2 is defined as "a user browses an item multiple times but doesn't pay, then leaves a negative review and remains silent"; Pattern 3 is defined as "a user frequently cancels orders, then leaves a negative review, and doesn't log in for seven consecutive days." These patterns are designed to capture complex behavioral chains that may lead to user churn, rather than just isolated behaviors. For example, a single instance of "not paying" might be a temporary issue, but when combined with the complete chain of "adding to cart → creating an order → payment failure → prolonged inactivity," it is more likely to indicate a risk of user churn.

[0049] Step 3: Perform pattern matching based on the rule engine:

[0050] The system compares the user's behavior timeline with preset composite behavior patterns one by one to determine whether the user's behavior matches a certain pattern. The matching process consists of three key steps:

[0051] Time window alignment: Based on the time range of the pattern (e.g., "no login for 3 consecutive days"), extract a subset of behaviors for the corresponding time period from the user's timeline to ensure that the analysis scope is consistent with the pattern requirements.

[0052] Behavior sequence comparison: This checks whether user actions occur in the order specified in the pattern. For example, Pattern 1 requires a user to first add items to their cart, then create an order, fail to complete the payment, and finally remain inactive for an extended period. If the order of user actions does not conform to the pattern requirements (e.g., payment fails before adding items to the cart), it is considered a mismatch.

[0053] Match score calculation: The system calculates the match score based on the following factors:

[0054] Sequential Consistency: Whether user actions occur exactly in the order specified in the pattern. For example, if a user adds items to their cart first, then fails to pay but doesn't create an order, points are deducted. Time Deviation: Whether the time interval between actions is within the pattern's allowed range. For example, Pattern 1 requires "no login within 3 days of payment failure," and if a user doesn't log in for 4 days after a payment failure, it's considered non-compliant. Missing Key Behaviors: Whether certain key behaviors are missing from the pattern. For example, Pattern 1 requires "payment failure," and if a user skips the payment step, points are deducted.

[0055] Ultimately, the system calculates a matching score between the user and each pattern based on the factors mentioned above. The higher the score, the closer the user's behavior is to that pattern.

[0056] Step 4, generate a preliminary warning signal:

[0057] When a user's match score with a certain composite behavioral pattern exceeds a preset initial threshold (e.g., 80 points), the system generates an initial warning signal. This signal contains the following information: the user's identity identifier (e.g., user ID), the type of matching pattern (e.g., pattern 1), and the match score (e.g., 85 points). For example, if a user's behavior perfectly matches the conditions of pattern 1 and the match score is 100 points, the system will immediately mark the user as high-risk and trigger the warning signal. The purpose of the initial warning signal is to provide input for subsequent steps (such as dynamic baseline assessment) and to provide a basis for the formulation of recall strategies. For example, a user matching pattern 1 may be marked as "payment barrier-related churn risk," and subsequent strategies may prioritize issuing payment coupons or optimizing the payment process.

[0058] By collecting user behavior data in real time and constructing a behavior timeline, combined with a pre-set composite behavior pattern matching mechanism, it accurately identifies early signs of user churn. It analyzes user action sequences (e.g., adding to cart → payment failure → silence) and time-sensitive periods (e.g., promotional periods) to calculate a behavior deviation index and generate preliminary warning signals, providing data support for subsequent dynamic threshold adjustments and personalized recall strategies. This step achieves analysis from isolated behaviors to behavior chains, significantly improving the accuracy and timeliness of churn warnings, and is a crucial preliminary step in formulating user recovery strategies.

[0059] S200: Obtain the current user's historical behavior data and extract the current user's group attribute tags, including user type tags, purchase frequency level, and average order value range; combine with time context information to identify the current time-sensitive period, including promotional periods, holidays, weekdays, and platform system outage periods; use a sliding window algorithm to calculate the current user's historical behavior activity benchmark value within the same time period, and calculate the behavior deviation index of the current behavior timeline relative to the historical benchmark. The behavior deviation index is weighted by the login frequency decline rate, transaction interruption frequency, and page dwell time reduction ratio; based on the sequence matching score and the behavior deviation index, use linear regression or decision tree models to dynamically adjust the warning threshold and generate a personalized warning threshold for each user; if the current sequence matching score is higher than the personalized warning threshold, a formal warning is triggered.

[0060] The execution process of S200 specifically includes the following steps:

[0061] Step 1: Obtain user historical behavior data and extract group attribute tags:

[0062] The system first extracts data from the user's complete historical behavior records, including the user's daily operation trajectory (such as logging in, browsing products, adding items to the cart, and paying for orders), consumption records (such as order amount and purchase frequency), and active time periods (such as active periods on weekdays or weekends). Based on this data, the system generates a set of group attribute tags to characterize the user's basic features and consumption habits. For example:

[0063] User type tags: Based on the frequency of user behavior and consumption characteristics, users are divided into categories such as "high-frequency consumers" (ordering multiple times per week), "low-frequency consumers" (ordering only once per month), "new users" (registered for less than 30 days) or "churned users" (no behavior for 7 consecutive days).

[0064] Purchase frequency level: By counting the number of orders placed by users over the past 30 days, these orders are categorized into different levels such as "daily purchases", "weekly purchases" or "monthly purchases" to reflect the user's activity level.

[0065] Average order value range: By analyzing the average order amount of users over the past 90 days, we divide the data into ranges such as "low average order value (<100 yuan)," "medium average order value (100-500 yuan)," and "high average order value (>500 yuan)" to measure users' spending power.

[0066] Step 2: Identify the current time-sensitive period by combining the time context:

[0067] The system will combine the current time point to determine the user's current time-sensitive period, in order to distinguish whether changes in user behavior stem from their own churn tendency or external environmental factors (such as holidays or promotional activities). For example:

[0068] Promotional periods: During major promotional events such as "Double 11" and "618", user behavior usually fluctuates in the short term due to increased price sensitivity (such as delayed payment after adding items to the cart); Holidays: During holidays such as Spring Festival and National Day, users may reduce their platform usage frequency due to travel or family activities; Weekdays: User behavior may show regularity (such as high activity during morning and evening peak hours); Platform system failure periods: If the platform experiences service interruptions or functional abnormalities recently, it may lead to a sharp drop in user behavior.

[0069] By identifying these cycles, the system can filter out normal behavioral fluctuations caused by the external environment, thereby more accurately capturing user churn signals.

[0070] Step 3: Calculate the baseline value of historical behavior activity using the sliding window algorithm:

[0071] The system uses the user's current time-sensitive period as a benchmark and employs a sliding window algorithm to analyze the user's historical activity levels within similar periods. For example:

[0072] Window size: If the current period is a "promotional event", the user's behavior data during the same period of the past three years will be selected as the reference window; if the current period is a "holiday", the user's holiday data during the same period of the past three years will be selected.

[0073] Activity metrics: Within a selected window period, key user activity metrics are statistically analyzed, including login frequency, order quantity, and page browsing time, to calculate the historical behavioral activity baseline value for that period. For example, if a user's average login frequency during previous "Double 11" promotional periods is 5 times / day and the number of orders is 2 / day, then the baseline value is "5 logins and 2 orders per day".

[0074] The purpose of this benchmark is to provide a comparison standard for the user's current behavior and help the system determine whether the behavior deviates from the normal range (such as the current number of logins dropping to 1 time / day, which is far below the historical benchmark).

[0075] Step 4, calculate the deviation index of the current behavior relative to the historical baseline:

[0076] The system compares the user's current behavior data with historical activity benchmarks to calculate a behavior deviation index. This index is derived by weighting the following three core indicators:

[0077] Login frequency decline rate: The percentage decrease in the number of user logins during the current period compared to the historical baseline. For example, if the historical baseline was 5 logins per day, and the current login is only 1, then the decline rate is 80%.

[0078] Transaction interruption frequency: The increase in the number of times a user fails to complete a transaction within the current period (such as adding items to a cart without payment, or order cancellation) compared to the historical baseline. For example, if the historical baseline was 1 cancellation out of every 10 cart additions, and the current baseline is 1 cancellation out of every 5 cart additions, then the interruption frequency has doubled.

[0079] Page dwell time reduction percentage: This refers to the percentage reduction in the time users spend on key pages (such as product detail pages and shopping cart pages) compared to historical baselines. For example, if the historical baseline is an average dwell time of 3 minutes on a product detail page, and the current dwell time is only 1 minute, then the reduction percentage is 66%.

[0080] These three metrics are weighted according to business experience (e.g., login frequency has the highest weight) and combined to derive a deviation index. The higher the index, the more abnormal the user behavior. For example, a user with a deviation index of 90 points indicates that their behavior has significantly deviated from the normal range.

[0081] Step 5: Dynamically adjust the warning threshold and generate personalized warnings:

[0082] The system combines the sequence matching score generated by S100 (e.g., if the user matches the pattern "add to cart → payment failure → silence") with the currently calculated behavior deviation index, and dynamically adjusts the warning threshold using linear regression or a decision tree model. For example:

[0083] Linear regression model: Trained using historical data, this model establishes a correlation formula between "match score" and "deviation index" on churn risk, dynamically calculating the warning threshold for individual users. For example, if historical data shows that a user with a match score of 80 and a deviation index of 70 has a 60% probability of churn, the system will set a threshold of 80 for similar users.

[0084] Decision tree model: Based on user group attribute tags (such as "high-frequency consumer" and "high average order value") and time period characteristics (such as "promotion period"), different rule nodes are divided, and differentiated thresholds are set for each type of user. For example, high average order value users may have a lower warning threshold (such as 70 points) due to the higher cost of churn, while the threshold for low-frequency users may be higher (such as 90 points).

[0085] Ultimately, the system generates a personalized alert threshold for each user. If the current sequence matching score exceeds the personalized alert threshold, a formal alert signal is triggered. For example:

[0086] User A's sequence matching score is 85 points, and the personalization threshold is 80 points → triggering an alert.

[0087] User B's sequence matching score is 75 points, and the personalization threshold is 90 points → no warning is triggered.

[0088] This mechanism avoids a "one-size-fits-all" approach to early warnings, ensures that high-risk users are identified first, and provides accurate input for subsequent recall strategies.

[0089] By integrating users' historical behavioral characteristics with time-period context, and combining group attribute tags (such as purchase frequency and average order value) with dynamic behavioral deviation analysis, personalized and accurate identification of user churn risk can be achieved. A sliding window algorithm is used to compare with historical activity benchmarks to quantify the degree of current behavioral anomalies, and linear regression or decision tree models are used to dynamically adjust the warning threshold. Ultimately, differentiated warning signals are generated for each user, effectively distinguishing between external environmental fluctuations and actual churn risk, providing a reliable basis for subsequent precise recall strategies.

[0090] S300 constructs a layered recall strategy library. The library includes a user tag layer storing static and dynamic tags generated based on user profiles, a scenario layer corresponding to the user's current behavioral scenario, and a strategy pool layer containing various executable recall strategies and their weight configurations. The user tag layer includes price-sensitive, logistics-conscious, and service-dependent users. The scenario layer includes incomplete payment, order cancellation without repurchase, and silence after negative feedback. The strategy pool layer includes issuing targeted coupons, initiating dedicated customer service outbound calls, providing free returns and exchanges, and pushing personalized recommendations. Applicable tags in the user tag layer are matched based on composite behavioral patterns and formal warning results, and the current dominant churn risk scenario is determined in conjunction with the scenario layer. A set of candidate strategies matching the applicable tags and the current dominant churn risk scenario is selected from the strategy pool layer. The historical conversion effect data of the candidate strategies in the candidate strategy set is evaluated through a real-time A / B testing module, and the execution weight of each candidate strategy is dynamically adjusted to generate the optimal recall strategy combination and execute it.

[0091] The execution process of S300 specifically includes the following steps:

[0092] Step 1, Build a layered recall strategy library:

[0093] The system first builds a layered recall strategy library, which consists of three layers: user tag layer, scenario layer, and strategy pool layer.

[0094] User Tagging Layer: Static and dynamic tags are generated based on user profiles to characterize the core features of users. Static tags include attributes that remain unchanged over time (such as "price-sensitive", "logistics-conscious", "service-dependent"), while dynamic tags are dynamically adjusted according to changes in recent user behavior (such as "recently viewed high-priced items").

[0095] Scenario Layer: Defines the user's current behavioral scenario to determine the main driving factors of churn risk. Examples include: Incomplete Payment (the user adds items to their cart but doesn't complete payment); Order Cancellation and No Repurchase (the user cancels multiple orders and doesn't subsequently repurchase); Silence After Negative Review (the user submits a negative review but remains inactive for an extended period).

[0096] Strategy pool layer: Contains various executable recall strategies and their weight configurations. For example:

[0097] Issue targeted coupons: Provide exclusive discounts for price-sensitive users.

[0098] Initiate dedicated customer service outbound calls: proactively contact service-dependent users to resolve their concerns.

[0099] Offer free returns and exchanges: This lowers the trust threshold for users who are concerned about logistics.

[0100] Personalized recommendations: Recommend products or services based on user interests.

[0101] The initial weights of the strategy are set based on business experience and are then dynamically adjusted through real-time A / B testing.

[0102] Step 2, match user tags with churn risk scenarios:

[0103] The system combines composite behavioral pattern types (such as "add to cart → payment failure → silence" as defined in S100) and formal warning results (such as personalized warning signals generated in S200) to match applicable tags from the user tag layer. For example:

[0104] If a user matches the pattern "add to cart → payment failure → silence" and the warning score is high risk, they may be marked as "price sensitive" (because payment failure may be related to price hesitation); if a user matches the pattern "frequent order cancellation → negative review → silence" and the warning score is medium risk, they may be marked as "service dependent" (because negative reviews may point to service quality issues).

[0105] At the same time, the system combines the scenario layer to determine the user's current dominant churn risk scenario. For example, if the user's most recent behavior is "payment failure", the scenario is "payment not completed"; if the user's most recent behavior is "submitting a negative review", the scenario is "silence after providing negative feedback".

[0106] Step 3: Filter the set of candidate strategies that match the tags and scenarios:

[0107] The system filters a set of candidate strategies from the strategy pool layer that match the applicable tags and the dominant churn risk scenarios. For example, if a user is tagged as "price-sensitive" and the scenario is "incomplete payment," the matching candidate strategies might include:

[0108] Issue targeted coupons (to lower the payment threshold for price-sensitive users) and push personalized recommendations (to recommend high-value products to attract repeat purchases); if a user is marked as "service-dependent" and the scenario is "silence after negative feedback", the matching candidate strategies may include: initiating dedicated customer service outbound calls (proactively contacting users to resolve negative review issues); and providing free return and exchange services (rebuilding trust by lowering the after-sales threshold).

[0109] Step 4: Dynamically adjust strategy weights through real-time A / B testing:

[0110] The system uses a real-time A / B testing module to evaluate the historical conversion performance of each strategy in the candidate strategy set. For example:

[0111] Historical conversion rate of targeted coupons: If 200 out of 1,000 similar users in the past resumed purchases due to coupons, the conversion rate is 20%.

[0112] Historical conversion rate of dedicated customer service outbound calls: If 150 out of the past 500 similar users resumed their purchases after communication with customer service, the conversion rate is 30%.

[0113] Based on the evaluation results, the system dynamically adjusts the execution weight of each candidate strategy. For example, if the conversion rate of "initiating dedicated customer service outbound calls" is higher than that of "issuing targeted coupons", the former's weight is increased to 60% and the latter's weight is reduced to 40%. If a strategy is ineffective in a specific scenario (such as "free return and exchange service" being ineffective for price-sensitive users), its weight is automatically reset to zero.

[0114] Step 5: Generate and execute the optimal recall strategy combination:

[0115] The system generates the optimal recall strategy combination based on the adjusted strategy weights and executes it immediately. For example, for users who are "price-sensitive + have not completed payment", the final strategy combination might be: issuing targeted coupons (weight 60%) and pushing personalized recommendations (weight 40%).

[0116] For users who are "service-dependent and remain silent after receiving negative feedback", the final strategy combination may be: initiating dedicated customer service outbound calls (70% weight); providing free return and exchange services (30% weight).

[0117] During execution, the system records the actual effect of each strategy (such as whether the user resumes purchasing) and continuously feeds back to the A / B testing module, forming a closed-loop optimization. For example, if a user resumes purchasing due to a "targeted coupon," the system will strengthen the weight of this strategy among similar users; if a user does not respond to a "dedicated customer service outbound call," the system will try other strategies (such as pushing personalized content).

[0118] By constructing a layered strategy library (user tag layer, scenario layer, and strategy pool layer), and combining user profiles with real-time behavioral scenarios, highly relevant recall strategies (such as targeted discounts and exclusive services) are dynamically matched. The strategy weights are continuously optimized through real-time A / B testing, forming a closed loop of "tag matching - scenario recognition - strategy execution - effect feedback". Ultimately, personalized and efficient user recovery is achieved, significantly improving the activation rate and repurchase conversion rate of churned users.

[0119] In this embodiment of the application, the method further includes: obtaining the recall strategy execution result corresponding to the recall strategy execution; performing attribution analysis on the churn reasons of users who triggered formal warnings based on composite behavioral patterns, sequence matching scores, behavioral deviation index, and recall strategy execution results; identifying key influencing factors that lead to user churn and generating attribution analysis results including the ranking of main causes, wherein key influencing factors include one or more of the following: number of payment failures, customer service response delays, logistics timeliness deviations, and competitor price fluctuations; adding or strengthening high-frequency associated behavioral sequences in the composite behavioral pattern library based on the attribution analysis results; and adjusting the weight coefficients of each dimension in the behavioral deviation index based on the attribution analysis results.

[0120] The above-mentioned scheme specifically includes the following steps:

[0121] Step 1: Obtain the execution results of the recall strategy and integrate multi-dimensional data:

[0122] After executing the recall strategy for S300, the system first collects the results of the strategy execution, including the user's actual response to the strategy (such as whether they resumed purchasing, clicked on coupons, or accepted customer service communication). Simultaneously, the system integrates the following four types of core data:

[0123] Composite Behavioral Pattern: The sequence of behaviors exhibited by users before churn (e.g., "add to cart → payment failure → silence"); Sequence Matching Score: The similarity of users matching high-risk behavioral patterns (e.g., a matching score of 90 points); Behavioral Deviation Index: The degree of abnormality between the user's current behavior and historical benchmarks (e.g., an 80% decrease in login frequency); Recall Strategy Execution Results: The actual impact of the strategy on users (e.g., coupon click-through rate, customer service communication conversion rate).

[0124] Step 2: Conduct a churn attribution analysis for users who triggered the formal alert:

[0125] The system conducts a churn attribution analysis on users who trigger formal alerts (such as users whose match scores exceed the threshold in S200) based on the integrated data mentioned above. The specific steps are as follows:

[0126] Related Behavioral Patterns and Strategy Responses: By comparing the user's composite behavioral patterns before churn (such as "payment failure → order cancellation") with the results of the recall strategy (such as coupons not being clicked), it can be determined whether the strategy effectively covers the core reasons for user churn.

[0127] Quantify key influencing factors: Extract metrics highly correlated with user churn from pre-defined key influencing factors (such as number of payment failures, customer service response delays, logistics timeliness deviations, and competitor price fluctuations). For example, if a user fails to complete a purchase after multiple payment failures, "number of payment failures" is marked as a high-influence factor; if customer service does not respond promptly after a user submits a negative review, "customer service response delays" is marked as a high-influence factor.

[0128] Generate attribution analysis results for main cause ranking: Rank key influencing factors by statistical methods or machine learning models and output structured results (e.g., "Number of payment failures (weight 40%) > Logistics timeliness deviation (weight 30%) > Customer service response delay (weight 20%)").

[0129] Step 3: Update the high-frequency associated behavior sequences in the composite behavior pattern library:

[0130] Based on the attribution analysis results, the system adds or strengthens frequently associated behavioral sequences in the composite behavioral pattern library. For example, if the analysis finds that "payment failure → order cancellation → silence" is a typical path leading to churn, then this sequence is added as an independent behavioral pattern; if "logistics timeliness deviation → negative review → silence" is identified as a frequently associated pattern, then logistics timeliness-related behaviors are added to the original pattern, and their weights are increased.

[0131] Step 4: Adjust the weighting coefficients of each dimension in the behavioral deviation index:

[0132] Based on the attribution analysis results, the system dynamically adjusts the weight coefficients of each dimension in the behavioral deviation index. For example, if the analysis shows that "number of payment failures" is the main cause of churn, the weight of "frequency of transaction interruption" will be increased (e.g., from 20% to 40%) when calculating the deviation index; if "logistics timeliness deviation" is identified as a secondary influencing factor, the weight of "percentage reduction in page dwell time" will be decreased (e.g., from 30% to 15%).

[0133] The purpose of this adjustment is to make the deviation index more closely reflect the actual churn drivers and avoid misjudgments caused by unreasonable historical weighting (such as overemphasizing the login frequency of low-frequency users while ignoring the direct impact of payment failures).

[0134] Attribution analysis accurately identifies the primary causes of user churn (such as payment failures and service delays), dynamically updates the behavioral pattern library to capture emerging churn paths, and adaptively adjusts the deviation index weights to make the model more focused on key influencing factors. This ultimately forms a complete closed loop of "strategy execution → effect evaluation → model iteration," significantly improving the system's ability to dynamically optimize churn warning and recall strategies, achieving continuous improvement in long-term effectiveness and intelligence.

[0135] In this embodiment of the application, the method further includes: when evaluating the historical conversion effect data of each candidate strategy in the candidate strategy set through a real-time A / B testing module, adjusting the execution weight of each candidate strategy using a dynamic weight allocation mechanism; the dynamic weight allocation mechanism calculates the strategy priority based on the historical conversion rate of the candidate strategy and the matching degree with the current user behavior scenario; wherein, the historical conversion rate includes user click-through rate, order conversion rate, and user retention rate; updating the scoring parameters of each candidate strategy in real time through an online learning mechanism, and determining the corresponding high-performance strategy based on the scoring parameters, wherein the high-performance strategy is executed first when a warning is triggered subsequently.

[0136] The above-mentioned scheme specifically includes the following steps:

[0137] Step 1, Dynamic weight allocation mechanism: Adjust strategy weights based on historical conversion rates and scenario matching.

[0138] When evaluating the historical conversion performance of the candidate strategy set through a real-time A / B testing module, the system employs a dynamic weight allocation mechanism to adjust the execution weight of each candidate strategy. The core logic of the dynamic weight allocation mechanism is:

[0139] Historical conversion rate metrics integration: Comprehensive analysis of historical conversion rate data for candidate strategies, including user click-through rate (e.g., coupon click-through rate), order conversion rate (e.g., the percentage of orders placed after coupon use), and user retention rate (e.g., whether users repurchase within 7 days after strategy implementation). For example, if a strategy has a historical click-through rate of 40% and an order conversion rate of 20%, its overall conversion capability is quantified as 60 points (assuming a weighting of 50% click-through rate + 50% order conversion rate).

[0140] Scenario matching calculation: Based on the current user's behavioral scenario (e.g., "payment not completed" or "silence after negative feedback"), the matching degree between candidate strategies and the scenario is determined. For example: in the scenario of incomplete payment, "issuing targeted coupons" and "pushing personalized recommendations" have a high matching degree (weight 80%), while "providing free returns and exchanges" has a low matching degree (weight 30%); in the scenario of silence after negative feedback, "initiating dedicated customer service outbound calls" and "providing free returns and exchanges" have a high matching degree (weight 90%), while "targeted coupons" have a low matching degree (weight 20%).

[0141] Dynamic strategy priority calculation: The strategy priority score is calculated by weighting historical conversion rate and scenario matching degree. For example, if a strategy has a historical conversion score of 60 and a scenario matching degree of 80%, then the priority score is (60×70%+80×30%)=66 points (weight allocation can be adjusted). The higher the priority score, the greater the execution weight of the strategy.

[0142] Step 2, Online learning mechanism: Real-time updates of scoring parameters and optimization of high-performance strategies.

[0143] The system continuously optimizes the scoring parameters of candidate strategies through an online learning mechanism, ensuring that the strategy weights dynamically adapt to changes in user behavior. The specific process is as follows:

[0144] Real-time data feedback: Whenever a candidate strategy is executed (such as when a user clicks on a coupon or receives a customer service call), the system immediately records its real-time performance data (such as click-through rate, conversion rate, and user retention rate). For example, if a strategy increases the click-through rate from 40% to 50% in the "payment not completed" scenario, the system will capture this change.

[0145] Scoring Parameter Updates: Based on real-time performance data, the system dynamically adjusts the scoring parameters of strategies. For example, if the click-through rate of a strategy increases, its historical conversion rate weight will be automatically increased (e.g., from 50% to 60%); if the relevance of a strategy in a specific scenario decreases (e.g., "free return and exchange service" is ineffective for price-sensitive users), its scenario relevance weight will be automatically decreased.

[0146] High-performing strategies are prioritized: Based on updated scoring parameters, the system identifies high-performing strategies (such as the top 3 strategies with the highest conversion rates in the current scenario) and assigns them higher execution weights. For example, in the "silence after negative feedback" scenario, if "dedicated customer service outbound calls" has the highest real-time conversion rate, its execution weight is increased to 70%, while the weights of other strategies are reduced accordingly.

[0147] Step 3, Dynamic optimization closed loop: a complete loop from strategy execution to weight update.

[0148] The system continuously optimizes strategy combinations through a dynamic optimization loop of "strategy execution → effect recording → parameter update → weight adjustment". For example: In the initial stage, the execution weight of a certain strategy in the "payment incomplete" scenario is 50%; after execution: if the click-through rate and conversion rate of the strategy significantly improve (e.g., the click-through rate increases from 40% to 60%), the online learning mechanism will automatically increase its weight to 70%; subsequent triggering: when similar users trigger the alert again, the system will prioritize the execution of this strategy, forming a positive cycle of "high-performance strategy → high weight → high execution frequency".

[0149] Through a dynamic weight allocation mechanism, the system can accurately combine historical conversion rates with scenario matching to prioritize the execution of highly relevant strategies. Through an online learning mechanism, the system can update scoring parameters in real time, ensuring that strategy weights are dynamically adjusted according to changes in user behavior. This ultimately forms an intelligent closed loop of "performance-driven weight updates → high-performance strategies prioritized," significantly improving the timeliness and accuracy of strategy optimization and achieving long-term, efficient churn user reactivation.

[0150] In this embodiment, a user behavior association analysis module is introduced into the user tag layer. The method further includes: the user behavior association analysis module generates dynamic user tag vectors by analyzing social associations, transaction link associations, and behavioral sequence similarities between users; social associations include jointly purchasing goods, browsing similar pages, and rating interactions; static tags are supplemented based on the dynamic user tag vectors, wherein the static tags include price-sensitive and logistics-focused tags, and the weights of each user tag are dynamically adjusted through a context-aware weight adjustment mechanism.

[0151] The above-mentioned scheme specifically includes the following steps:

[0152] Step 1, Three types of association analysis in the user behavior association analysis module:

[0153] The system uses a user behavior correlation analysis module to uncover potential connections between users from the following three dimensions and generate dynamic user tag vectors:

[0154] Social Connections: Analyze social interaction behaviors between users, such as: jointly purchasing goods, if user A and user B purchase the same goods multiple times, they are marked as "socially connected users"; browsing similar pages, if user A and user B frequently browse the same type of product pages, they are marked as "interest-connected users"; rating interactions, if user A and user B rate products for each other or follow each other, they are marked as "socially interactive users".

[0155] Transaction link association: Analyze the association between users in the transaction process, such as: linked purchase, if user A recommends user B to buy a certain product, or user B places an order because of user A's order, it is marked as "recommendation relationship user"; supply chain association, if user A and user B's orders involve the same logistics channel or customer service process, it is marked as "service chain related user".

[0156] Behavioral sequence similarity: Compare users' behavioral path patterns. For example, if both user A and user B exhibit the behavioral sequence of "browsing → adding to cart → payment failure → silence", they are marked as "highly similar behavioral users". If both user A and user B frequently browse high-priced items but do not place an order, they are marked as "wait-and-see behavioral users".

[0157] Step 2, Generation of dynamic user tag vectors and supplementation of static tags:

[0158] Based on the above three types of association analysis results, the system generates dynamic user tag vectors and combines them with static tags (such as price-sensitive and logistics-focused). The specific process is as follows:

[0159] Dynamic tag generation: Transforms the results of social association, transaction link association, and behavioral sequence similarity analysis into quantifiable tags. For example, social association generates "social activity" tags (such as "high social association" and "low social association"); behavioral sequence similarity generates "behavioral imitation" tags (such as "high similarity" and "low similarity").

[0160] Static tag supplement: Static tags (such as price-sensitive and logistics-focused) are long-term, stable user characteristics that, together with dynamic tags, constitute a complete user tagging system. For example, user A's static tag is "price-sensitive," and the dynamic tag is "high social connection + high similarity behavior"; user B's static tag is "logistics-focused," and the dynamic tag is "low social connection + low similarity behavior."

[0161] Step 3, Dynamic label optimization using a context-aware weight adjustment mechanism:

[0162] The system dynamically adjusts the weights of user tags through a context-aware weight adjustment mechanism to better suit the needs of the current business scenario. The specific logic is as follows:

[0163] Context awareness: By combining the user's current behavioral context (such as "payment not completed" or "silence after negative feedback"), determine which tags have a greater impact on the current context. For example, in the "payment not completed" scenario, the weight of the "price-sensitive" tag increases, while the weight of the "social activity" tag decreases; in the "silence after negative feedback" scenario, the weight of the "service-dependent" tag increases, while the weight of the "behavioral imitation" tag decreases.

[0164] Dynamic weight allocation: The weight ratio of each tag is adjusted in real time through machine learning models or rule engines. For example, if user A shows high price sensitivity in the "payment not completed" scenario, the weight of his "price sensitive" tag will increase from 30% to 60%; if user B shows low social connection in the "silence after negative feedback" scenario, the weight of his "social activity" tag will decrease from 50% to 20%.

[0165] Step 4: Integrate label vectors for policy matching and execution:

[0166] The system applies the optimized user tag vector (including static and dynamic tags) to the recall strategy matching process:

[0167] Strategy matching optimization: During the strategy selection phase of S300, the system combines the user's complete tag vector (such as "price-sensitive + high social connection") to match more suitable strategies. For example, for "price-sensitive + high social connection" users, priority is given to "issuing targeted coupons + pushing content recommended by social friends"; for "logistics-conscious + low similarity behavior" users, priority is given to "providing free return and exchange services + pushing logistics timeliness guarantee information".

[0168] Enhanced strategy execution: Real-time updates to dynamic tags ensure that strategies can quickly respond to changes in user behavior. For example, if user A's social activity has decreased recently, the system will reduce the weight of strategies related to "social interaction" and instead strengthen the execution of "price-sensitive" strategies.

[0169] By introducing a user behavior correlation analysis module, the system can generate dynamic tags from dimensions such as social interaction, transaction links, and behavior sequences, thus supplementing the limitations of static tags. Combined with a context-aware weight adjustment mechanism, it can dynamically adapt the tag weights in different scenarios, making the strategy matching more in line with the user's current behavior and needs.

[0170] In this embodiment of the application, the method further includes the following steps in identifying time-sensitive periods: combining external environmental data and user historical behavior data, predicting the trend of user behavior activity in future time periods through a time-series feature extraction model; the external environmental data includes e-commerce platform promotion calendars, holiday databases, and regional weather data; dynamically adjusting the time window length of the sliding window algorithm based on the user behavior activity trend, and eliminating abnormal misjudgments caused by time offsets through a behavior sequence alignment algorithm; and smoothing out occasional behavior deviations through a noise filtering mechanism in non-sensitive periods.

[0171] The above-mentioned scheme specifically includes the following steps:

[0172] Step 1, Temporal Feature Extraction Model: Combining external environment and user historical behavior to predict activity trends:

[0173] The system uses a time-series feature extraction model to comprehensively analyze external environmental data and historical user behavior data to predict user activity trends in future time periods. The specific process is as follows:

[0174] External environment data integration: E-commerce platform promotional calendars, such as the timing of major promotional events like "Double Eleven," "618," and "Black Friday," indicate a potential significant increase in user activity; holiday databases, such as those for the Spring Festival and Labor Day, may trigger changes in user consumption habits (e.g., purchasing New Year's goods, booking travel); regional weather data, such as high-temperature warnings and rainstorm warnings, may affect users' offline activities or their demand for specific products (e.g., air conditioners, umbrellas).

[0175] User historical behavior analysis: Extract user behavior patterns within a historical period (such as daily login frequency, order amount, and page dwell time), and combine them with external environmental data (such as a 50% increase in user add-to-cart rate during promotional periods).

[0176] Activity Trend Prediction: Based on historical data and its correlation with the external environment, the model predicts user activity trends over the next 7 days, 14 days, or longer. For example, it predicts a sharp increase in user activity in the 3 days leading up to "Double Eleven," potentially leading to a higher payment failure rate; it also predicts a decrease in user browsing of outdoor products during the rainy season, but an increase in demand for indoor entertainment products.

[0177] Step 2, dynamically adjust the length of the sliding window to adapt to changes in activity trends.

[0178] The system dynamically adjusts the time window length of the sliding window algorithm based on predicted user behavior activity trends to more accurately capture behavioral changes. The specific logic is as follows:

[0179] High-activity periods: If a significant increase in user activity is predicted over the next 3 days (such as during promotional periods), the system will shorten the sliding window length from the default 7 days to 3 days. For example, shortening the window allows for faster identification of abnormal user behavior such as "frequent logins → frequent purchases → payment failures," triggering timely alerts.

[0180] Low-activity periods: If user activity is predicted to decline in the next 14 days (e.g., during the rainy season), the system will extend the sliding window length from the default 7 days to 14 days. For example, extending the window can smooth out occasional user login behavior (e.g., occasional page browsing) and avoid misjudging it as churn risk.

[0181] Step 3, Behavior sequence alignment algorithm: Eliminate abnormal misjudgments caused by time offset.

[0182] The system uses a behavior sequence alignment algorithm to resolve misjudgments caused by time offsets in user behavior. For example:

[0183] Time-skewed scenario: User A frequently browses products at 1 AM, while User B frequently browses products at 9 PM. If analyzed using a fixed time window (e.g., 8 AM to 6 PM), User A's behavior might be mistakenly identified as abnormal.

[0184] Alignment algorithm implementation: Standardize and align user behavior sequences according to their individual activity timelines (e.g., map user A's behavior to the "early morning active period" and user B's behavior to the "evening active period"), and then perform pattern matching; for example, if a user suddenly exhibits the behavior sequence "login → browsing → payment failure" during the "early morning active period", the algorithm will consider it normal; if the same sequence appears during the "evening active period", an alert will be triggered.

[0185] Step 4, noise filtering mechanism: smoothing out occasional behavioral deviations in insensitive cycles.

[0186] During non-sensitive periods (such as periods of stable user activity), the system uses a noise filtering mechanism to eliminate the interference of occasional behavioral deviations on judgment. For example:

[0187] Occasional behavioral scenarios: temporary user testing behavior (such as adding items to a shopping cart and then canceling the order); abnormal user behavior caused by system errors (such as repeatedly clicking on the page).

[0188] Noise filtering is implemented by using moving average or exponential smoothing techniques to smooth user behavior data. For example, if a user's daily login frequency suddenly increases from an average of 3 times to 10 times, but then drops back to 3 times in the following two days, the system determines this as an occasional occurrence and ignores it; if a user's login frequency is lower than the historical average for three consecutive days, it is determined to be a real risk of churn.

[0189] Through a time-series feature extraction model, the system can combine the external environment with users' historical behavior to predict future activity trends; dynamically adjust the sliding window length to make behavior detection more in line with cyclical fluctuations; the behavior sequence alignment algorithm eliminates misjudgments due to time offsets to ensure fairness; and the noise filtering mechanism smooths out occasional deviations in non-sensitive cycles to improve judgment accuracy.

[0190] In this embodiment of the application, a user behavior semantic analysis module is introduced into the attribution analysis results, and the method further includes the following steps:

[0191] Step 1: The system uses a user behavior semantic analysis module to extract sentiment tags from users' text comments and behavioral sequences within the platform, and quantifies their contribution to user churn. The specific process is as follows:

[0192] Text review analysis: Order reviews, analyze user feedback on goods or services (e.g., "Logistics is too slow, it hasn't arrived after three days" "Customer service attitude is terrible"), and extract sentiment tags (e.g., "dissatisfaction" "anxiety"); Customer service dialogue records, analyze user interaction text with customer service (e.g., "I've been waiting for 20 minutes, no one has replied" "This problem must be solved immediately"), and identify sentiment tags (e.g., "anxiety" "anger").

[0193] Behavioral sequence analysis: Frequent order cancellations: If a user cancels an order multiple times (e.g., "add to cart → cancel → add to cart → cancel"), it is marked with the "hesitant" emotion label; Repeated additions to cart without payment: If a user adds items to their cart multiple times but does not complete payment (e.g., "browse → add to cart → exit → repeat operation"), it is marked with the "hesitant" or "dissatisfied" emotion label.

[0194] Emotion tag quantification contribution: The intensity score of emotion tags is calculated using natural language processing models (such as sentiment analysis algorithms) (e.g., "anxiety" has an intensity score of 80 points, and "dissatisfaction" has an intensity score of 60 points); combined with the frequency and intensity of emotion tags, their total contribution to user churn is quantified (e.g., the "anxiety" tag has a cumulative contribution of 40% in the 7 days before churn, ranking first among the reasons for churn).

[0195] The role of the user behavior semantic analysis module is to capture the subjective motivations behind users' emotional changes and behaviors, make up for the blind spots of traditional indicators (such as the number of payment failures) in terms of users' psychological state, and provide a more comprehensive perspective for attribution analysis.

[0196] Step 2, Causal Inference Model: Calculate attribution weights by combining key influencing factors and sentiment labels.

[0197] The system uses a causal reasoning model to combine four key influencing factors (number of payment failures, customer service response delays, logistics timeliness deviations, and competitor price fluctuations) with emotional tags (anxiety, dissatisfaction, and hesitation), calculates the attribution weight of each factor, and generates a dynamic explanatory report. The specific steps are as follows:

[0198] Key influencing factors integration: Number of payment failures, counting the frequency of order interruptions due to payment failures before user churn (e.g., 3 consecutive payment failures); Customer service response delays, calculating the average customer service response time after a user submits an inquiry or complaint (e.g., average waiting time exceeding 30 minutes); Logistics timeliness deviations, comparing the difference between the actual delivery time of a user's order and the promised delivery time (e.g., a 2-day delay); Competitor price fluctuations, monitoring changes in the prices of competitors' products that users are interested in (e.g., a 20% decrease in competitor prices).

[0199] Causal reasoning calculation: The model analyzes the correlation between emotion tags and key influencing factors based on historical data. For example, if the emotion tag "anxiety" is often accompanied by "payment failures ≥ 3 times", the model determines that "payment failures" is a strong driving factor of "anxiety"; if the emotion tag "dissatisfaction" is highly correlated with "customer service response delay ≥ 20 minutes", the model determines that "customer service response delay" is the core cause of "dissatisfaction".

[0200] Using causal graphs or Bayesian networks, calculate the direct and indirect contributions of each factor to churn. For example, "number of payment failures" may directly lead to churn (contribution 50%), while indirectly exacerbating churn by triggering "anxiety" (contribution 20%).

[0201] Dynamic Explanation Report Generation: Output a structured report with the main cause ranking (e.g., "Number of payment failures (weight 40%) > Customer service response delay (weight 30%) > Competitor price fluctuations (weight 20%)"); The report also marks the impact path of emotion tags (e.g., "Anxiety is directly caused by payment failures, accounting for 60% of the total reasons for churn").

[0202] Step 3, Dynamic Interpretation of Report Application: Optimize Early Warning Thresholds and Strategy Priorities.

[0203] The system optimizes the generation model parameters for personalized early warning thresholds based on dynamic interpretation reports and guides the priority ranking of recall strategies in the strategy pool layer. The specific implementation is as follows:

[0204] Personalized alert threshold optimization: If the report indicates that "payment failure times ≥ 3 times" is the main reason for churn, the system will dynamically adjust the alert threshold for this dimension (e.g., increase the default threshold from 2 times to 3 times); if the report shows that "customer service response delay ≥ 20 minutes" has a significant impact on "dissatisfaction" sentiment, the system will reduce the alert sensitivity for this dimension (e.g., adjust the threshold from 30 minutes to 20 minutes).

[0205] Recall strategy priority ranking: Based on the primary cause, strategies corresponding to high-weight factors are matched first. For example: if "number of payment failures" is the primary cause, then "targeted coupons + payment failure compensation plan" will be implemented first; if "customer service response delay" is the secondary cause, then "outbound customer service calls + service satisfaction improvement plan" will be implemented second priority.

[0206] Dynamic adjustment of strategy priority: When the ranking of the main factors changes (such as "competitor price fluctuations" becoming the second main factor), the system automatically adjusts the execution order of the strategies (such as increasing the execution weight of the "competitor price comparison reminder" strategy).

[0207] Through the user behavior semantic analysis module, the system can capture the implicit correlation between emotion tags and behavioral patterns and quantify their contribution to churn. Combined with the causal reasoning model, the system can reveal the multiple causal relationships between key influencing factors and emotional motivations, and generate dynamic explanatory reports. Finally, by optimizing the early warning threshold and strategy priority, the system can achieve multi-dimensional driving force and dynamic response for churned users to be recalled, significantly improving the system's intelligence level and long-term effectiveness.

[0208] This application discloses an intelligent early warning system for user churn on a trading platform, referring to... Figure 2 ,include:

[0209] The Behavior Modeling and Preliminary Early Warning Module 001 collects user action data within the platform in real time through an event processing engine, constructing a user behavior timeline with timestamps as the axis. Action data includes page browsing, adding items to cart, order creation, payment attempts, and changes in login status. At least one composite behavior pattern is defined as a pre-emptive churn indicator template, where the composite behavior pattern is a preset combination of behavior sequences. Based on a rule engine, pattern matching is performed on the current user's behavior timeline, calculating the sequence matching score between the current user and each composite behavior pattern. When the sequence matching score of any composite behavior pattern exceeds a preset initial threshold, a preliminary early warning signal is generated, and the current user's sequence matching score and the corresponding composite behavior pattern are identified to obtain the corresponding user identifier.

[0210] The Dynamic Baseline Assessment and Formal Early Warning Confirmation Module 002 acquires the current user's historical behavior data and extracts the current user's group attribute tags, including user type tags, consumption frequency levels, and average order value ranges. Combined with time context information, it identifies the current time-sensitive period, including promotional periods, holidays, weekdays, and platform system outage periods. It uses a sliding window algorithm to calculate the current user's historical behavioral activity benchmark value within the same time period and calculates the behavioral deviation index of the current behavior timeline relative to the historical benchmark. The behavioral deviation index is weighted by the login frequency decline rate, transaction interruption frequency, and the proportion of reduced page dwell time. Based on the sequence matching score and the behavioral deviation index, it dynamically adjusts the early warning threshold using linear regression or a decision tree model to generate personalized early warning thresholds for individual users. If the current sequence matching score is higher than this personalized early warning threshold, it is used to confirm and trigger a formal early warning.

[0211] The tiered recall strategy generation and execution module 003 constructs a tiered recall strategy library. This library includes a user tag layer storing static and dynamic tags generated based on user profiles, a scenario layer representing the user's current behavioral context, and a strategy pool layer containing various executable recall strategies and their weight configurations. The user tag layer includes price-sensitive, logistics-conscious, and service-dependent users. The scenario layer includes incomplete payment, order cancellation without repurchase, and silence after negative feedback. The strategy pool layer includes issuing targeted coupons, initiating dedicated customer service outbound calls, providing free returns and exchanges, and pushing personalized recommendations. Based on composite behavioral patterns and official alert results, the module matches applicable tags from the user tag layer and combines this with the scenario layer to determine the current dominant churn risk scenario. It then selects a set of candidate strategies from the strategy pool layer that match the applicable tags and the current dominant churn risk scenario. Finally, a real-time A / B testing module evaluates the historical conversion performance data of the candidate strategies in the set, dynamically adjusting the execution weights of each candidate strategy to generate and execute the optimal recall strategy combination.

[0212] This application also discloses an intelligent early warning system for user churn on a trading platform, including a processor, wherein the processor runs a program for the intelligent early warning method for user churn on a trading platform as described in any one of the above embodiments.

[0213] This application also discloses a storage medium storing a program for the intelligent early warning method for user churn on a trading platform as described in any one of the above embodiments.

[0214] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for intelligent early warning of user churn on a trading platform, characterized in that, include: The event processing engine collects user behavior data in real time within the platform, constructing a user behavior timeline with timestamps as the axis; the behavior data includes page browsing, adding products to cart, order creation, payment attempts, and login status changes; Define at least one composite behavior pattern as a churn precursor template, wherein the composite behavior pattern is a preset combination of behavior sequences; perform pattern matching on the current user's user behavior timeline based on the rule engine, and calculate the sequence matching score between the current user and each composite behavior pattern; when the sequence matching score of any composite behavior pattern exceeds a preset initial threshold, generate an initial warning signal, and identify the current user's sequence matching score and the corresponding composite behavior pattern to obtain the corresponding user identifier; The system acquires historical behavior data of the current user and extracts group attribute tags, including user type tags, purchase frequency levels, and average order value ranges. It then identifies the current time-sensitive period by combining this data with time context information. Time-sensitive periods include promotional periods, holidays, weekdays, and platform system downtime. A sliding window algorithm is used to calculate the current user's historical activity benchmark value within the same time period. The system also calculates the behavior deviation index relative to the historical benchmark, which is weighted by the login frequency decline rate, transaction interruption frequency, and page dwell time reduction ratio. Based on the sequence matching score and the behavior deviation index, a linear regression or decision tree model is used to dynamically adjust the warning threshold, generating personalized warning thresholds for individual users. If the current sequence matching score is higher than the personalized alert threshold, a formal alert will be triggered. A hierarchical recall strategy library is constructed, which includes a user tag layer storing static and dynamic tags generated based on user profiles, a scenario layer corresponding to the user's current behavioral scenario, and a strategy pool layer containing multiple executable recall strategies and executable recall strategy weight configurations; the user tag layer includes price-sensitive, logistics-focused, and service-dependent users; the scenario layer includes incomplete payment, no repurchase after order cancellation, and silence after negative feedback; The strategy pool layer includes issuing targeted coupons, initiating dedicated customer service outbound calls, providing free returns and exchanges, and pushing personalized recommendation content; Based on the composite behavior pattern type and the formal warning result, the applicable tags in the user tag layer are matched, and the current dominant churn risk scenario is determined in combination with the scenario layer; a set of candidate strategies that match the applicable tags and the current dominant churn risk scenario are selected from the strategy pool layer. The real-time A / B testing module evaluates the historical conversion performance data of candidate strategies in the candidate strategy set, dynamically adjusts the execution weight of each candidate strategy, generates the optimal recall strategy combination, and executes it.

2. The intelligent early warning method for user churn on a trading platform according to claim 1, characterized in that, The method also includes: Obtain the recall strategy execution results corresponding to the recall strategy execution. Based on the composite behavior pattern, sequence matching score, behavior deviation index and recall strategy execution results, perform churn attribution analysis on users who triggered formal warnings. Identify key influencing factors that lead to user churn and generate attribution analysis results including the ranking of main causes. The key influencing factors include one or more of the following: number of payment failures, customer service response delay, logistics timeliness deviation, and competitor price fluctuations. Based on the attribution analysis results, add or strengthen frequently associated behavioral sequences in the composite behavioral pattern library; adjust the weight coefficients of each dimension in the behavioral deviation index based on the attribution analysis results.

3. The intelligent early warning method for user churn on a trading platform according to claim 1, characterized in that, The method also includes: When evaluating the historical conversion performance data of each candidate strategy in the candidate strategy set through the real-time A / B testing module, a dynamic weight allocation mechanism is used to adjust the execution weight of each candidate strategy. The dynamic weight allocation mechanism calculates the priority of a strategy based on its historical conversion rate and its matching degree with the current user behavior scenario; whereby the historical conversion rate includes user click-through rate, order conversion rate, and user retention rate. The scoring parameters of each candidate strategy are updated in real time through an online learning mechanism. Based on the scoring parameters, the corresponding high-performance strategy is determined. The high-performance strategy is executed first when a warning is triggered.

4. The intelligent early warning method for user churn on a trading platform according to claim 1, characterized in that, Introducing a user behavior correlation analysis module into the user tag layer, the method also includes: The user behavior association analysis module generates dynamic user tag vectors by analyzing social associations, transaction link associations, and behavioral sequence similarities among users; social associations include jointly purchasing goods, browsing similar pages, and rating interactions; The static tags are supplemented based on the dynamic user tag vector. The static tags include price-sensitive and logistics-focused tags, and the weights of each user tag are dynamically adjusted through a context-aware weight adjustment mechanism.

5. The intelligent early warning method for user churn on a trading platform according to claim 2, characterized in that, In the process of identifying time-sensitive periods, the method also includes: By combining external environmental data with historical user behavior data, a time-series feature extraction model is used to predict the trend of user behavior activity in future time periods. External environmental data includes e-commerce platform promotion calendars, holiday databases, and regional weather data. The time window length of the sliding window algorithm is dynamically adjusted based on the user behavior activity trend, and the abnormal misjudgment caused by time offset is eliminated by the behavior sequence alignment algorithm; in non-sensitive periods, occasional behavior deviations are smoothed by noise filtering mechanism.

6. The intelligent early warning method for user churn on a trading platform according to claim 5, characterized in that, The attribution analysis results incorporate a user behavior semantic analysis module, and the method further includes: The user behavior semantic analysis module extracts sentiment tags by analyzing users' text comments and behaviors on the platform, and quantifies the contribution of sentiment tags to user churn. Text comments include order reviews and customer service conversation records, behavioral sequences include frequent order cancellations and repeated unpaid add-to-carts, and sentiment tags include anxiety, dissatisfaction, and hesitation. By combining four key influencing factors—the number of payment failures, customer service response delays, logistics timeliness deviations, and competitor price fluctuations—a causal inference model is used to calculate the attribution weight of each factor and output a dynamic explanation report that includes the ranking of the main causes. The dynamic explanation report is used to optimize the generation model parameters of personalized warning thresholds and guide the priority ranking of recall strategies in the strategy pool layer.

7. A smart early warning system for user churn on a trading platform, characterized in that, include: The behavior modeling and preliminary early warning module collects user operation behavior data in real time within the platform through the event processing engine, and constructs a user behavior timeline with timestamps as the axis; the operation behavior data includes page browsing, adding products to cart, order creation, payment attempts, and login status changes; Define at least one composite behavior pattern as a churn precursor template, wherein the composite behavior pattern is a preset combination of behavior sequences; perform pattern matching on the current user's user behavior timeline based on the rule engine, and calculate the sequence matching score between the current user and each composite behavior pattern; when the sequence matching score of any composite behavior pattern exceeds a preset initial threshold, it is used to generate a preliminary warning signal, and the current user's sequence matching score and the corresponding composite behavior pattern are identified to obtain the corresponding user identifier; The dynamic baseline assessment and formal early warning confirmation module acquires the current user's historical behavior data and extracts the current user's group attribute tags, including user type tags, consumption frequency levels, and average order value ranges. Combined with time context information, it identifies the current time-sensitive period, including promotional periods, holidays, weekdays, and platform system outage periods. A sliding window algorithm is used to calculate the current user's historical behavioral activity benchmark value within the same time period, and a behavioral deviation index is calculated relative to the historical benchmark. This behavioral deviation index is weighted by the login frequency decline rate, transaction interruption frequency, and page dwell time reduction ratio. Based on the sequence matching score and the behavioral deviation index, a linear regression or decision tree model is used to dynamically adjust the early warning threshold, generating personalized early warning thresholds for individual users. If the current sequence matching score is higher than the personalized warning threshold, it is used to confirm the triggering of a formal warning; The layered recall strategy generation and execution module constructs a layered recall strategy library. The layered recall strategy library includes a user tag layer storing static and dynamic tags generated based on user profiles, a scenario layer corresponding to the user's current behavioral scenario, and a strategy pool layer containing multiple executable recall strategies and executable recall strategy weight configurations. The user tag layer includes price-sensitive, logistics-focused, and service-dependent users. The scenario layer includes incomplete payment, order cancellation without repurchase, and silence after negative feedback. The strategy pool layer includes issuing targeted coupons, initiating dedicated customer service outbound calls, providing free returns and exchanges, and pushing personalized recommendation content; Based on the composite behavior pattern type and the formal warning result, the applicable tags in the user tag layer are matched, and the current dominant churn risk scenario is determined in combination with the scenario layer; a set of candidate strategies that match the applicable tags and the current dominant churn risk scenario are selected from the strategy pool layer; the historical conversion effect data of the candidate strategies in the candidate strategy set are evaluated through the real-time A / B testing module, and the execution weight of each candidate strategy is dynamically adjusted to generate the optimal recall strategy combination and execute it.

8. A smart early warning system for user churn on a trading platform, characterized in that, Includes a processor, wherein the processor runs a program of the intelligent early warning method for user churn of a trading platform as described in any one of claims 1-6.

9. A storage medium, characterized in that, The program stores the intelligent early warning method for user churn of a trading platform as described in any one of claims 1-6.

Images