Urban rail transit user individual-level retention intervention planning method

Abstract

The present application relates to a kind of individual level retention intervention planning method of urban rail transit user, the method includes: based on the obtained multi-source user data, for each user i constructs the training sample containing user survival time, censored state and feature vector;RFM multi-task is trained by using training sample fusion depth survival model, for each user i generates user loss risk trajectory and predicted survival time, and calculates the user value and life cycle value of user i at current evaluation time point t, to construct the structured portrait feature of input large language model for each user i;Based on user declarative preference survey data set constructs user preference priori;Structured portrait feature, user preference priori and operation allowed incentive set are input into large language model as hard constraint, for each user i generates optimal retention plan.Compared with prior art, the present application has the advantages of improving the stability, economy and operation availability of retention decision, etc.

Description

Technical Field

[0001] This invention relates to the field of urban rail transit user operation technology, and in particular to an individual-level retention intervention planning method for urban rail transit users. Background Technology

[0002] With the expansion of urban rail transit networks, the diversification of travel scenarios, and intensified competition from ride-hailing, shared mobility, and commuting alternatives, operation and management departments face continuous pressure to improve user stickiness, reduce churn rates, and promote repeat rides and renewals. In actual operation, whether users continue to choose urban rail transit is often influenced by multiple factors, including fare and discount policies, fluctuations in service quality, changes in the travel environment, adjustments in commuting structures, and individual preferences. This manifests as behavioral signals such as decreased usage frequency, changes in travel times, alterations in transfer routes, or shifts in consumption patterns. Although operators possess large-scale transaction and travel records, they still repeatedly face several challenges in retention decisions: difficulty in timely identifying which users are at increased risk of churn, difficulty in grasping the most appropriate timing for retention efforts, difficulty in determining the upper limit of investment to avoid resource waste or excessive subsidies, and difficulty in generating retention strategies and outreach messages that can be truly implemented and align with user preferences.

[0003] Existing practices fall into two categories. One focuses on user segmentation or churn prediction, providing assessments of "risk level" or "potential churn," but often struggles to answer questions like "when is the most appropriate time to intervene?" and "how much investment is economically reasonable," making it even more difficult to translate predictions directly into actionable individual strategies. Another approach relies on operational experience and rules to directly distribute discounts or push notifications. While simple to implement, this often lacks interpretable budget boundaries, leading to issues like intervening too late with high-value users, intervening too early with low-value users, or over-investment. In recent years, large language models have demonstrated strong text generation capabilities, producing relatively natural operational language and suggestions. However, without strict time and cost constraints, and verifiable quantitative evidence, the generated content can be unenforceable, exceed budget, be inconsistent with platform rules, or lack relevance, making it difficult to use directly in high-frequency operational scenarios.

[0004] Therefore, there is an urgent need for a retention planning method that can integrate the dynamic changes in user churn risk, the potential value of users and the economically feasible investment ceiling, as well as the differences in users' preferences for incentive methods. This method should not only form clear and explainable intervention timing and investment boundaries, but also generate directly executable retention strategies and communication content under clear constraints, and be able to verify and select the feasibility of the plan, thereby improving the stability, economy and operational availability of retention decisions.

[0005] A search revealed Chinese invention patent application publication number CN116757475A, which discloses a method and system for assessing the risk of user churn throughout the entire lifecycle of urban public transport users. The method includes: obtaining a user key churn path graph model and the user's travel chain in a given behavioral state from long-term card-swiping data; obtaining a user travel chain tensor based on the parameter information of the travel chain, combined with point-of-interest data and station geospatial location information; designing user travel chain tensor feature engineering based on the user travel chain tensor and combining it with an attention mechanism model to extract the latent semantic features of the user travel chain; and designing a cyclic deep survival analysis model considering competition risk based on the latent semantic features, and combining this with user value assessment to evaluate the risk of user churn throughout the entire lifecycle. This invention can help public transport managers comprehensively understand and address the risk of churn among regular urban public transport users, promoting the sustainable development of the public transport system and improving user satisfaction. The existing patent application only indicates whether the risk exists, without providing retention solutions in terms of time and investment.

[0006] How to implement individual-level retention intervention planning for urban rail transit users and improve the stability, economy, and operational availability of retention decisions has become a technical problem that needs to be solved. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the existing technology and provide an individual-level retention intervention planning method for urban rail transit users.

[0008] The objective of this invention can be achieved through the following technical solutions: According to one aspect of the present invention, an individual-level retention intervention planning method for urban rail transit users is provided, the method comprising: Based on the acquired multi-source user data, a training sample containing user survival time, censoring status and feature vector is constructed for each user i; the multi-source user data includes historical travel behavior data of urban rail transit users, travel environment context feature data and user declarative preference survey dataset; The training samples are used to train a deep survival model that integrates RFM multi-tasks. For each user i, a user churn risk trajectory and predicted survival time are generated. The user value and lifetime value of user i at the current evaluation time point t are calculated. Thus, the individual intervention window, economic budget limit and user churn risk trajectory are calculated for each user. The structured profile features of the input large language model are constructed. Construct user preference priors based on user declarative preference survey datasets; The structured profile features, prior user preferences, and operational incentive sets are used as hard constraints to input the large language model. After feasibility verification and preference consistency scoring, an optimal retention plan is generated for each user i.

[0009] As a preferred technical solution, the deep survival model adopts an encoder-decoder structure that integrates RFM multi-tasks to support joint learning of RFM index prediction and survival prediction by sharing latent representations.

[0010] As a preferred technical solution, the training of the deep survival model adopts a multi-task learning structure, with the prediction of user value-related RFM indicators as an auxiliary task, and optimized by a joint loss function. The joint loss function includes RFM metric prediction loss and survival prediction loss. The RFM metric prediction loss is used to quantify the model's prediction bias on the user's RFM three-dimensional value index. The survival prediction loss is a combination of accelerated failure time loss and ranking loss, used to measure the model's prediction error on user churn time and status.

[0011] As a preferred technical solution, the deep survival model calculates an RFM three-dimensional index score for each user i, including a score based on the time since the most recent use, a score based on the frequency of use, and a score based on the amount spent, and calculates the user value based on the RFM three-dimensional index score.

[0012] As a preferred technical solution, the calculation of users is based on the predicted lifespan and user value. i At time Lifetime value : , in, For user i in time User value, The predicted survival time output by the deep survival model.

[0013] As a preferred technical solution, the individual intervention time window is used to limit the feasible time range for the intervention to occur. ,satisfy: Intervention start time: This represents the user's risk value. First time not lower than the lower threshold The earliest time t; Intervention end time: This represents the user's risk value. The first time is not lower than the upper threshold. The earliest time t; Based on clustered risk distribution Empirical cumulative distribution function The quantiles are used to determine the lower threshold of the cluster s. With upper threshold ,satisfy: in, for inverse function, , The preset probability quantiles, ; Clustered Risk Distribution Specifically: Select the current cluster And in the future The set of users who will churn within a period of time Collect the risk value sequence of these users in the period before churn. To construct clustered risk distribution .

[0014] As a preferred technical solution, the process of predicting user churn risk trajectory and survival time includes: The deep survival model outputs a streaming risk probability sequence that changes over time for each user i, determines the earliest time to trigger intervention according to set rules, i.e., the predicted survival time, and organizes the streaming risk probability sequence into a churn risk trajectory for user i to intuitively show the changes in churn risk for users at different time points.

[0015] As a preferred technical solution, the calculation of the economic budget ceiling is specifically as follows: For users Match the non-churn reference group belonging to the same subgroup s. Calculate the reference group at time point Average remaining life value and calculate user i Lifetime value in the event of impending churn The upper limit of the economic budget is determined by comparing user i with the average residual value of the non-churned reference group in the same subgroup s. ,satisfy: .

[0016] As a preferred technical solution, the similarity between user i and the respondent is retrieved from the user declarative preference survey dataset by the retrieval enhancement generation method to obtain the top-k nearest neighbor set. The user preference prior is obtained by pooling the ranking preference, attribute weight and attribute level of the nearest neighbor set.

[0017] As a preferred technical solution, the retention plan generated by the large language model uses a fixed set of output fields, which can be directly executed and audited by the operator. After the retention plan is evaluated for feasibility based on time, cost, and incentive type, and a preference consistency score is applied, the optimal retention plan is selected. The optimal retention plan includes individual intervention window, final intervention time, incentive type, incentive cost, user communication information, and evidence interpretation.

[0018] Compared with the prior art, the present invention has the following beneficial effects: 1) This invention constructs user survival samples through multi-source data fusion (travel behavior, environmental context, and declarative preference survey data), and generates dynamic churn risk trajectories and predicted survival times based on a deep survival model that integrates multiple RFM tasks. This effectively solves the core pain point of existing technologies, which can only determine "whether churn has occurred" but cannot determine "when to intervene." It calculates user value and lifetime value, individual intervention window, and economic budget ceiling, avoiding premature intervention that wastes resources or late intervention that reduces success rate. It breaks through the limitations of traditional experience-based interventions, which lack interpretable investment boundaries and are prone to over-subsidization, ensuring that retention investment does not exceed the reasonably recoverable user surplus. The value lies in constructing structured profile features for each user, summarizing similar group preferences from declarative preference data to build user preference priors, and using these, along with structured profile features and operational incentive sets, as hard constraints input into a large language model to generate an optimal retention plan for each user i. This ensures that the strategy aligns with individual user preferences to improve reach effectiveness, and feasibility verification ensures that intervention timing, cost, and incentive types comply with operational rules. This achieves individual-level, executable, and cost-effective retention intervention, suitable for batch and rolling user retention management in urban rail transit operations, significantly improving the accuracy of intervention timing, resource allocation efficiency, and strategy implementation quality.

[0019] 2) The encoder-decoder structure of this invention, which integrates multiple tasks of RFM, relies on shared latent representation to achieve joint learning of RFM user value prediction and survival churn risk prediction. This effectively solves the problem that "risk judgment" and "value assessment" are independent and their information is not interconnected in the existing technology. User value features in the RFM dimension can help the model more accurately identify the risk evolution pattern of high-value users (avoiding misjudging the ordinary behavioral fluctuations of low-value users as churn risk), while survival risk-related features (such as churn time trends) can also feed back into the dynamics of value assessment (such as predicting the rate of decay of the remaining value of high-risk users). During the training process, RFM index prediction is used as an auxiliary task and optimized with a joint loss function. This solves the defect of existing models that cannot accurately characterize "when the risk will occur and what the degree of risk is". Ultimately, it provides accurate and reliable dual support of risk and value for the subsequent determination of individual intervention windows and generation of constrained strategies, significantly improving the scientificity, economy and feasibility of retention intervention decisions.

[0020] 3) This invention effectively solves the core pain point of ambiguous intervention timing in existing urban rail transit user retention interventions by combining risk distribution by groups with upper and lower thresholds to determine personalized intervention windows. It ensures that intervention occurs at a stage where the risk is moderate and there is still room for recovery, and it also conforms to the common behaviors of group users, significantly improving the accuracy of intervention timing. It provides strict time constraints for subsequent strategy generation, reduces ineffective resource allocation, and avoids the problems of existing technologies relying on subjective experience or static risk labels to judge intervention timing, which can easily lead to premature intervention that wastes resources or late intervention that reduces the success rate.

[0021] 4) This invention matches target users with a non-churned reference group in the same segment. The average remaining lifetime value of the reference group is used as a reasonable value benchmark. The lifetime value of users who are about to churn is subtracted to obtain a non-negative budget ceiling. This ensures that the budget ceiling is entirely derived from the value that can be added by the users who are recovered. The budget is no longer a subjective assumption, but an economic boundary supported by data. This avoids uneconomical intervention and ensures reasonable resource allocation to high-value users, thereby improving the economy and transparency of retention resource allocation.

[0022] 5) This invention obtains the top-k nearest neighbor set most similar to the target user from user declarative preference (SP) survey data through similarity retrieval. Then, it pools the incentive strategy ranking, attribute importance weights, and attribute preference levels of these nearest neighbors to form a preference prior that aligns with the target user. This preference prior is not generated out of thin air but is based on the actual preferences of similar users. It accurately captures users' acceptance tendencies regarding incentive types (e.g., one-time discounts vs. periodic offers) and attributes (e.g., discount magnitude vs. duration), providing personalized preference constraints for subsequent strategy generation. This significantly improves the user adaptability of intervention strategies, avoids strategy failure due to preference mismatch, and thus improves the effectiveness and success rate of retention interventions. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the framework of an individual-level retention intervention planning method for urban rail transit users in one embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a deep survival model that integrates RFM multi-tasks in one embodiment of the present invention; Figure 3 This is a schematic diagram of the input information of a large language model in one embodiment of the present invention; Figure 4 This is a schematic diagram of a retrieval enhancement generation and large language model constraint generation mechanism based on preference information in one embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the feasibility analysis results of a retention plan generated by a large language model in one embodiment of the present invention; Figure 6This is a schematic diagram illustrating the constraint violation and key constraint pattern analysis across passenger groups in one embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0025] To address the aforementioned issues, this invention proposes an individual-level retention intervention planning method for urban rail transit users. This method treats user retention as a "risk-to-action" planning problem, providing three types of operationally executable outputs for each target user: first, a user-level recommended intervention timeframe to guide operators in triggering intervention at a stage where risk is moderate and there is still room for recovery; second, an investment ceiling commensurate with the user's remaining value to ensure that retention investment does not exceed reasonably recoverable new value; and third, retention strategies and outreach messages generated within the aforementioned time and cost boundaries to directly guide specific implementations such as discount distribution, activity guidance, or service care.

[0026] In terms of methodology, this invention first utilizes the historical and contextual features of urban rail transit transactions and travel behavior to establish a predictive model that can characterize the trajectory of user churn risk over time. This allows operators to move beyond relying solely on static labels or single-point risk assessments and instead observe the trend and urgency of risk evolution over time. Subsequently, this invention introduces a value assessment mechanism based on dimensions such as users' recent travel frequency, travel intensity, and consumption contribution to further estimate the residual value that users can contribute in the future. Based on this, an economically feasible retention investment boundary is constructed, avoiding uneconomical interventions that "invest beyond recoverable value." Simultaneously, this invention obtains user preference information for different forms and attributes of incentives through a statement preference survey and uses a similar group preference aggregation method to form preference priors for target users. This ensures that retention strategies not only meet rule and cost requirements but also better align with users' acceptance tendencies.

[0027] In the strategy generation stage, this invention uses the intervention timing boundary determined by the aforementioned risk trajectory and the investment ceiling determined by value assessment as hard constraints input into the large language model. Combined with user profiles, prior preferences, and expert rules from the operations team, it generates a retention plan that includes suggestions for intervention timing, incentive forms and amounts, and user-facing communication scripts and explanations. The generated plan undergoes feasibility verification to ensure it meets requirements such as timeframe, investment ceiling, and available incentive rules. When multiple candidate plans exist, the plan that best aligns with user preferences and is more feasible is prioritized and output to the operations team. Through this design, this invention achieves closed-loop retention decision support from risk identification and value constraints to strategy generation, improving the ability to grasp intervention timing, resource allocation efficiency, and strategy implementation quality in large-scale operational scenarios.

[0028] Example 1 This embodiment relates to an individual-level retention intervention planning method for urban rail transit users, aiming to address the following: In urban rail transit operations, how to accurately identify individual churn risks and determine appropriate intervention timing and economically reasonable investment limits based on generative artificial intelligence and large language models, thereby generating individualized retention strategies and communication content that are directly executable and conform to operational rules and user preferences. Based on ticketing data and user declarative preference survey data, the method includes: predicting user churn risk trajectories based on deep survival models, calculating intervention time windows based on churn risk trajectories, calculating value and generating constraints based on RFM value models, constructing preference priors and injecting domain knowledge, and generating and verifying constrained strategies. Specifically, it includes the following steps: S1, Multi-source user data acquisition, sample construction, and overall framework: Multi-source user data includes at least: (1) historical travel behavior data, used to characterize user usage intensity, frequency trends, spatiotemporal patterns and consumption-related characteristics (user identification, entry and exit time, payment amount or ticket information, frequency and time distribution of rides, etc.); (2) contextual feature data of the established travel environment and services, used to characterize the user's spatial environment, accessibility of alternative travel modes and service quality, etc.; (3) user stated preference (SP) survey data, used to characterize users' preferences for different forms of incentives and incentive attributes, including users' socioeconomic attributes, usage behavior characteristics, incentive strategy preference ranking, importance weight of incentive attributes and attribute preference level.

[0029] Based on the above data, for each user Building a Survival Log ,in: Indicates the lifespan (the time from the user's first use to the end of the observation period, or the time until churn). The censorship status indicates whether the user has churned during the observation period: 1 indicates churned, 0 indicates no churn / data truncation. The censorship status is determined by judging whether the last use time is earlier than the observation end date minus the dormancy period threshold. The feature vector is composed of behavioral and contextual variables, which is formed by concatenating travel behavior indicators (such as frequency and amount), travel environment context (such as accessibility of alternative modes), and user static attributes (such as whether they have participated in SP surveys). It does not contain any future information such as preference ranking. The model input feature vector is constructed synchronously based on this information. To form a training sample set , which serves as the original feature vector input to the deep survival model.

[0030] S2, Deep Survival Model Training and Risk Trajectory Prediction: A deep survival model is trained based on the sample set I obtained from S1. The deep survival model outputs an individual-level churn probability curve. Used to dynamically reflect user behavior i At any moment t The probability of loss.

[0031] like Figure 2 The deep survival model employs an encoder-decoder architecture that integrates multiple RFM tasks to share latent representations. Supports joint learning of RFM indicator prediction and survival prediction. This refers to the collective term for RFM features, which are the input urban rail transit user feature vectors. θ represents the neural network parameters corresponding to this model. This deep survival model is not a completely existing general model, but rather an improved design based on the classic survival analysis framework, specifically for the scenario of urban rail transit user churn prediction. Its core innovation lies in introducing a multi-task learning mechanism, using the prediction of user value-related RFM indicators as an auxiliary task, sharing the underlying encoder representation with the survival prediction task, thereby enhancing the model's ability to characterize user churn risk. The encoder part uses the user feature vector z... i As input, it is mapped into a shared latent representation through a multi-layer neural network. This latent representation simultaneously connects to two task branches: the RFM metric prediction task outputs predicted values ​​of the user's three-dimensional metrics (i.e., recent usage time, usage frequency, and spending amount) through a fully connected layer, while the survival prediction task outputs the user's churn probability curve p over time based on this representation. i (t). During training, the two tasks are optimized using a joint loss function, ensuring that the latent representation retains information related to churn time while incorporating features from the user value dimension, thereby improving the accuracy and interpretability of survival prediction. This encoder-decoder combined with multi-task learning structure is the improvement of this invention for the user churn prediction problem.

[0032] like Figure 2 The training employs a multi-task learning structure, using user value-related RFM three-dimensional metrics (Recency, Frequency, Monetary metrics) as auxiliary tasks, and optimizing them with a joint loss function. Satisfaction formula: , Among them, the RFM index predicts loss. Used to measure the error of the model on user value metrics, satisfying: , in, , and These are the loss weights for the three tasks, used to balance the contributions of different RFM dimensions to the learning of the latent representation; Mean square error, and These are the actual and predicted values ​​of the Recency index, respectively. and These are the actual and predicted values ​​of the Frequency index, respectively. and and are the actual and predicted values ​​of the Monetary index, respectively, where i represents the i-th sample.

[0033] Survival prediction loss To measure the model's prediction error for user churn time and status, a combination of accelerated time to failure (AFT) loss and ranking loss is used to adapt to non-proportional risk data and enhance risk ranking capabilities, satisfying the following: , in, t i and t j They are users i and users j The actual time of loss (or the observation cutoff time). It is a model for log t i The linear prediction value, β x is the linear prediction parameter. i For feature vectors; It is a balance coefficient between the two types of loss, used to control the proportion of Rank loss in the total loss; and users respectively i and users j The non-censoring indicator variable is used to indicate the censoring status; This is a smoothing term hyperparameter used to control the steepness of the loss curve.

[0034] After training, targeting users Output a time-varying streaming risk probability sequence And according to established rules, determine the predicted survival time (i.e., find the earliest time to trigger intervention): , in, This corresponds to the set of observation times. Further, the streaming risk probability sequence is organized into a user churn risk trajectory. It visually displays the changes in user churn risk at different points in time, providing a basis for subsequent intervention timing and strategy planning.

[0035] S3, Calculation and Grouping of Three-Dimensional Index Scores for Urban Rail Transit Users: At the current assessment point Calculate users Three-dimensional index scores Users are divided into different groups based on a preset grouping method (such as RFM grouping or K-means). The clustering is used to estimate the risk threshold distribution and reference value level under similar behavioral structures, thereby enabling differentiated intervention timing and budget constraints.

[0036] S4, User Value Calculation: According to user At time The RFM three-dimensional index score is used to calculate user value. Its satisfying formula: , in, These are weighting parameters determined by expert evaluation, used to reflect the importance of each dimension; The score is based on the Recency dimension, which is the score based on how close the last time it was used. Score the Frequency dimension, i.e., use frequency scores; The score is for the Monetary dimension, which is the score of consumption amount / usage intensity.

[0037] S5, Lifetime Value Calculation: Predicted survival time based on S2 And the user value obtained by S4 Calculate users i At time Lifetime value Its satisfying formula: , in, For user i in time User value, The predicted survival time is the survival time obtained by the deep survival model.

[0038] S6, Value at Risk Construction: To balance churn probability and residual value, a risk value is constructed for intervention triggering. Representing users i In time t The risk value represents the expected value that the platform would lose if a user churns at this point. Its equation is: , in, The point-in-time probability output provided by the survival prediction module makes the risk value change synchronously with the probability and value, which can be closer to the operational intuition of "worth retaining and becoming dangerous".

[0039] S7, Construction of Cluster Risk Distribution and Determination of Thresholds: Grouping Select the current cluster And in the future The set of users who will churn within a period of time Collect the risk value sequence of these users in the period before churn, and construct a cluster risk distribution. This allows subsequent thresholds to be obtained from historical data and sample statistical distributions rather than subjective settings, improving the robustness of window determination and satisfying the following: , based on Empirical cumulative distribution function quantiles, to determine the threshold for clustering. With upper threshold ,satisfy: in, for inverse function, , For the preset probability quantiles (e.g.) =0.25、 =0.75), used to divide low, medium, and high risk zones. Lower threshold Set minimum value constraints to avoid premature intervention and waste of resources.

[0040] S8, Determining the individual intervention window: For target users (Its subgroup is) Using the value-at-risk trajectory obtained in S6 and the threshold obtained in S7, the start and end times of the recommended intervention are determined. and ,satisfy: Intervention start time: This indicates the earliest moment when a user's risk value first does not fall below the lower threshold, used to avoid premature intervention; Intervention end time: This indicates the earliest moment when a user's risk value first does not fall below the upper threshold, used to avoid intervention that is too late, leading to a decrease in success rate or uneconomical costs.

[0041] Thus, the individual intervention window is obtained. This is used to define the feasible timeframe for an intervention to occur.

[0042] S9, Economic Budget Cap Calculation, is used to assess the economic rationale for intervention and calculates the maximum budget that can be invested for a single user.

[0043] In grouping Inside, for target users Match the non-churn reference group belonging to the same subgroup s. Calculate the reference group at time point Average remaining life value And calculate the lifetime value of target users in the event of impending churn. The upper limit of the economic budget is determined by comparing the average residual value of the target users with that of the non-churned reference group in the same segment. ,satisfy: , when At that point, it was determined that there was no economically reasonable room for investment to retain talent.

[0044] S10, Construction of Structured Profile Features (Structured Evidence Input): like Figure 3 In determining With user profiles Then, for each target user Constructing structured profiling features for large language models ,satisfy: , Among them, user profile It should include at least the descriptions of the cluster labels and three-dimensional indicators, as well as a summary of behavioral features that can be used for strategy personalization; This serves as a user churn risk trajectory, reflecting the dynamic changes and urgency of the risk over time. This serves as a budget cap, used to limit incentive investments to no more than the explainable economic return.

[0045] S11, Preferred Prior Construction: Based on user declarative preference survey dataset Generate preference priors for the target users.

[0046] Used to characterize users j The preference structure for different incentive schemes and key attributes satisfies: , in, Socioeconomic attributes (including at least gender, age, income, vehicle ownership, etc.) Use behavioral characteristics (including at least usage frequency stratification, most recent use, etc.). For the set of incentive strategies The complete ranking, i.e., the ranking of incentive strategy preferences; To assign importance weights to incentive attributes, a Likert scale is used to represent the degree of importance users attach to incentive attributes. Incentive attributes include at least the single discount amount, cumulative discount amount, incentive randomness, and incentive duration. This represents the preference level for the corresponding attribute (e.g., percentage level). N This represents the total number of respondents in the SP dataset.

[0047] like Figure 3 and Figure 4 Under the matching dimensions specified by expert knowledge, targeting the target user It employs a search-enhanced generation method, using a set of matching dimensions specified by experts. Retrieve similar user groups from the SP dataset and summarize their preference information to obtain prior preferences for the target users. ,satisfy: , in, This is used to specify which profile dimensions are most critical for preference matching, in order to avoid irrelevant dimensions interfering with the retrieval and to improve the rationality of preference prior behavior; The core function for enhanced retrieval takes as input the profile features of the target user. X i SP dataset D SP Matching dimension set specified by experts The output is the prior preferences of the target user.

[0048] The enhanced retrieval generation method utilizes a set of expert matching dimensions. Next, calculate the similarity between target user i and SP respondent j to obtain the top-k nearest neighbor set. It satisfies the following formula: , in, and These are the feature vector representations after mapping the profile features of target user i and respondent j to the feature space of the SP dataset, and the feature vector representation of the respondent. ,in s j Representative of the interviewees j The socio-economic attributes of b j Represents usage behavior characteristics; This is a similarity function, which can be the inverse of cosine similarity, weighted Euclidean distance, or other similarity functions; Indicates the highest similarity k One user.

[0049] User preference prior By analyzing the nearest neighbor set The ranking preferences, attribute weights, and attribute levels are pooled to obtain This allows it to guide both the selection of incentive types and the configuration of incentive attributes, satisfying the following formula: , , , in, Choose the probability for the incentive type. To implement the incentive strategies of respondent j sorting The converted weighted average score; the incentive attribute weights are obtained through normalization. ; For neighboring users j For the q The importance weight of each incentive attribute; The activation attribute preference level serves as a reference benchmark for the target user's preference attribute level; the scale symbol ∝ indicates that the final result will be normalized so that the sum of the probabilities of all incentive types is 1.

[0050] S12, Constrained policy generation: The structured profile features of the target user obtained in S10 Obtained preference priors The operator allows incentive aggregation Together with its discrete denomination rules, these serve as generation conditions within the fixed decoding parameters of the large language model. Next, a retention plan is generated from the large language model. ,satisfy: , The retention plan output structure uses a fixed set of output fields, allowing operators to directly execute and audit the plan. , in, The recommended intervention time (should fall within the recommended window) is specified. The incentive type should be within the set permitted by the operator. Incentive costs (should not exceed the budget limit). Communicate information to users (for direct outreach and delivery). For the purpose of evidence interpretation (used to explain to the operator why the intervention was carried out at this time and why this incentive and amount were selected, and the interpretation should at least cite one or more of the following information: risk trajectory, budget cap and preference prior).

[0051] S13, Feasibility Verification and Preference Consistency Score: Perform hard constraint validation on the generated candidate plans and define the feasible solution set for the target user. satisfy: , That is, for any candidate plan Y, the intervention time must fall within the individual's intervention window. Within this period, incentive costs shall not exceed the budget limit. Furthermore, the incentive type must belong to the set of incentives permitted by operations. This is to ensure that the strategy is executable under operational rules and economic constraints.

[0052] Based on the hard constraints, a preference consistency score is calculated to measure the degree of matching between the generated plan and the user's prior preferences. satisfy: , in, As a positive term, it is used to enhance the preference for incentive types that are higher. In the preference prior, the higher the log probability, the higher the score; This is a negative penalty item; Used to penalize candidate programs Y The value of the incentive attribute q and the user preference level Differences, weights The higher the score, the greater the difference, the heavier the penalty, and the lower the score.

[0053] S14, Multiple Candidate Optimization and Output: When multiple candidate retention plans exist, in the feasible set The final retention plan is obtained by performing optimization according to formula (10). : , in, To decode parameters Large Language Model Generation Retention Plan Y The logarithmic probability reflects the model's confidence in the plan; The weighting coefficient is used to adjust the influence of preference items on the final choice, thereby achieving the optimal solution selection of "feasibility first, preference weighting".

[0054] The final output includes: individual intervention window Final Intervention Time Incentive Types Incentive costs User communication information and explanation of evidence It is used by urban rail transit operators to carry out individual-level, executable, and auditable user retention interventions.

[0055] As described above, this embodiment first establishes the risk trajectory and economic budget boundary, and then generates and verifies retention strategies under hard constraints, thereby realizing risk, value, and preference-driven urban rail transit user retention planning.

[0056] This method is based on urban rail transit ticketing transaction and travel behavior data, integrates contextual information such as travel environment and service exposure, and combines user preference information for different incentive methods to form interpretable intervention timing and resource investment boundaries. Under clear constraints, it generates directly executable retention strategies and communication content to support urban rail transit operators in making rolling, large-scale, and auditable user retention decisions.

[0057] Compared with existing technologies, this invention has the following advantages: First, it can transform the dynamic changes in user risk into a clear intervention trigger time range, reducing resource waste and decreased success rate caused by premature or late intervention; second, it can provide an upper limit of investment that matches the user's residual value, making the allocation of retention resources more economical and interpretable; third, it generates directly executable retention strategies and communication scripts under clear time and cost constraints, reducing the risk of unexecutability caused by free generation; fourth, it introduces user preference information, making the incentive forms and expressions more tailored to user differences, improving the effectiveness and stability of retention outreach, and is suitable for batch, rolling user retention management in urban rail transit operation departments.

[0058] Example 2 This embodiment also relates to an individual-level retention intervention planning method for urban rail transit users, such as... Figures 1 to 6 As shown, this embodiment provides a data verification test process and result display for real urban rail transit scenarios, based on embodiment 1.

[0059] like Figure 1 The method includes: Step 1: Data Acquisition of User Declarative Preference Survey: Data from a user declarative preference survey of urban rail transit conducted in a certain city in January 2026 was selected as the basis for implementation. This survey targeted urban rail transit users and collected 428 valid questionnaires, covering user groups with different travel frequencies, usage habits, and commuting characteristics, which can well reflect the subjective preferences of rail transit users in the retention intervention scenario.

[0060] The survey primarily included users' preferences for various rail transit incentive methods, such as single-ride discounts, multi-ride discounts, and periodic ticket discounts. It also included users' subjective evaluations of incentive attributes (such as discount strength, validity period, and difficulty of obtaining them). This survey data was used to characterize the differences in users' incentive acceptance levels, providing a preference reference for the generation of subsequent retention intervention strategies.

[0061] Step 2, Predicting changes in user retention risks In view of the differences in travel stability among urban rail transit users, this embodiment introduces the prediction results of changes in user future retention risk to identify the appropriate time range for intervention. Figure 2 The diagram illustrates a model structure used to predict changes in user retention risk. Based on users' existing behavioral characteristics such as travel frequency and time distribution, the model characterizes the trend of risk changes over a future period. The prediction results can identify time periods when users may experience a decrease in travel frequency or an increase in the risk of travel disruption, thus providing a time reference for retention interventions and avoiding unnecessary incentive campaigns during periods when users' travel intentions are already strong.

[0062] Step 3, Introduction of User Incentive Preference Information After determining the potential intervention timeframe for users, this embodiment incorporates user preference information regarding incentive type selection into the retention intervention strategy generation process, based on the aforementioned declarative preference survey results. Statistical analysis of the survey results identifies the general tendencies of users with different travel frequencies in their incentive choices. For example, some users with lower travel frequencies prefer more immediate single-trip discounts, while users with more stable travel patterns are more likely to accept periodic tickets or cumulative discounts. This preference information guides the selection of incentive schemes, reducing situations where the generated results are significantly inconsistent with users' subjective acceptance.

[0063] Step 4: Construct the information required for generating the intervention plan. This embodiment organizes relevant input information in a unified manner before generating a retention intervention plan. This information includes a description of the user's basic travel characteristics, the predicted intervention time range, the acceptable incentive input range, and the set of incentive methods currently available in the urban rail transit system. This explicit organization of input information ensures that the subsequent generation process is always based on known conditions, avoiding incentive suggestions that are significantly inconsistent with actual operating conditions. Figure 3 This demonstrates the input format of relevant information during the generation process. This information organization helps improve the stability and consistency of the generated results.

[0064] Step 5: Generation and explanation of retention intervention strategies Based on the above information, the generation process of the specific retention intervention strategy involved in this embodiment is as follows: Figure 4 As shown. The generated results mainly include the suggested intervention time, incentive type and its basic parameters, supplemented by a brief explanation of the reasons for the applicability of the scheme. During the generation process, the system will refer to historical survey sample summaries similar to the target user's preference characteristics and existing operational experience, so that the generated content is closer to the actual application scenario of urban rail transit in terms of expression and incentive selection.

[0065] S6, Verification of the feasibility of retention intervention programs This embodiment verifies the feasibility of the generated retention intervention schemes from the perspective of actual urban rail transit operation. The verification process mainly focuses on whether the suggested intervention time is within a reasonable range, whether the selected incentive method is available from the current system, and whether the incentive cost is within an acceptable range. Verification results show that after incorporating user risk prediction and preference information, the vast majority of generated schemes can meet the above basic implementation conditions, indicating that the method has good feasibility in real rail transit scenarios. For the few schemes that do not meet the conditions, corrections can be made by adjusting incentive parameters or delaying the intervention time without affecting the overall effectiveness of the method.

[0066] Furthermore, this embodiment presents the feasibility analysis results of an urban rail transit user retention intervention plan generated based on a large language model, such as... Figure 5 As shown in the figure, this figure statistically analyzes the satisfaction of generated plans in terms of timing, incentive costs, and incentive forms at the overall sample level. It reflects that after incorporating user travel risk prediction results and preference information, the generated retention plans can meet the actual operational constraints of rail transit in most cases. This result indicates that by introducing explicit constraint information during the generation process, the proportion of unsolvable plans can be effectively reduced, making the retention plans closer to actual application needs.

[0067] Meanwhile, this embodiment statistically analyzes the constraint violations and their main manifestations under different rail transit user group conditions, such as... Figure 6 As shown in the figure, this diagram reflects the distribution characteristics of situations where the generated solutions fail to meet constraints among user groups with different travel frequencies and risk levels. For example, in some low-frequency or budget-constrained user groups, the incentive costs are more likely to exceed the acceptable range. The above analysis can provide a basis for setting differentiated intervention strategies or parameter ranges for different user groups in subsequent actual operations, thereby further improving the overall applicability of retention intervention solutions in rail transit systems.

[0068] Step 7: Comprehensive analysis of verification results and explanation of their application value. A comprehensive analysis of the results generated from different user samples reveals that the method employed in this embodiment can generate targeted retention intervention suggestions based on differences in user travel stability and preference characteristics. For example, for users with low travel frequency and rapidly increasing risk, the generated results tend to favor short-term, highly stimulating incentives; while for users with more regular travel patterns, more continuous preferential schemes are generated. In summary, this embodiment demonstrates that the urban rail transit user individual-level retention intervention strategy planning method based on deep survival risk trajectory and large language model generation described in this invention can generate reasonable and implementable intervention suggestions under real rail transit data conditions, and has good practical application value.

[0069] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for planning individual-level retention interventions for urban rail transit users, characterized in that, The method includes: Based on the acquired multi-source user data, a training sample containing user survival time, censoring status and feature vector is constructed for each user i; the multi-source user data includes historical travel behavior data of urban rail transit users, travel environment context feature data and user declarative preference survey dataset; The training samples are used to train a deep survival model that integrates RFM multi-tasks. For each user i, a user churn risk trajectory and predicted survival time are generated. The user value and lifetime value of user i at the current evaluation time point t are calculated. Thus, the individual intervention window, economic budget limit and user churn risk trajectory are calculated for each user. The structured profile features of the input large language model are constructed. Construct user preference priors based on user declarative preference survey datasets; The structured profile features, prior user preferences, and operational incentive sets are used as hard constraints to input the large language model. After feasibility verification and preference consistency scoring, an optimal retention plan is generated for each user i.

2. The method for individual-level retention intervention planning of urban rail transit users according to claim 1, characterized in that, The deep survival model adopts an encoder-decoder structure that integrates RFM multi-tasks to support joint learning of RFM metric prediction and survival prediction by sharing latent representations.

3. The method for individual-level retention intervention planning of urban rail transit users according to claim 2, characterized in that, The deep survival model is trained using a multi-task learning structure, with the prediction task of user value-related RFM indicators as an auxiliary task, sharing the underlying encoder representation with the survival prediction task, and using the joint loss function of the two tasks for optimization. The joint loss function includes RFM metric prediction loss and survival prediction loss. The RFM metric prediction loss is used to quantify the model's prediction bias on the user's RFM three-dimensional value index. The survival prediction loss is a combination of accelerated failure time loss and ranking loss, used to measure the model's prediction error on user churn time and status.

4. The individual-level retention intervention planning method for urban rail transit users according to claim 2, characterized in that, The deep survival model calculates an RFM three-dimensional index score for each user i, including a score based on the time since the most recent use, a score based on the frequency of use, and a score based on the amount spent, and calculates the user value based on the RFM three-dimensional index score.

5. The individual-level retention intervention planning method for urban rail transit users according to claim 4, characterized in that, Calculate the user based on the predicted lifespan and user value. i At time Lifetime value : ， in, For user i in time User value, The predicted survival time output by the deep survival model.

6. The individual-level retention intervention planning method for urban rail transit users according to claim 1, characterized in that, The individual intervention time window is used to define the feasible time range for the intervention to occur. ,satisfy: Intervention start time: This represents the user's risk value. First time not lower than the lower threshold The earliest time t; Intervention end time: This represents the user's risk value. The first time is not lower than the upper threshold. The earliest time t; Based on clustered risk distribution Empirical cumulative distribution function The quantiles are used to determine the lower threshold of the cluster s. With upper threshold ,satisfy: in, for inverse function, , The preset probability quantiles, ; Clustered Risk Distribution Specifically: Select the current cluster And in the future The set of users who will churn within a period of time Collect the risk value sequence of these users in the period before churn. To construct clustered risk distribution .

7. The method for individual-level retention intervention planning of urban rail transit users according to claim 1, characterized in that, The process of determining user churn risk trajectory and predicted survival time includes: The deep survival model outputs a streaming risk probability sequence that changes over time for each user i, determines the earliest time to trigger intervention according to set rules, i.e., the predicted survival time, and organizes the streaming risk probability sequence into a churn risk trajectory for user i to intuitively show the changes in churn risk for users at different time points.

8. The individual-level retention intervention planning method for urban rail transit users according to claim 1, characterized in that, The calculation of the aforementioned economic budget ceiling is as follows: For users Match the non-churn reference group belonging to the same subgroup s. Calculate the reference group at time point Average remaining life value and calculate user i Lifetime value in the event of impending churn The upper limit of the economic budget is determined by comparing user i with the average residual value of the non-churned reference group in the same subgroup s. ,satisfy: 。 9. A method for individual-level retention intervention planning for urban rail transit users according to claim 1, characterized in that, The similarity between user i and respondent i is retrieved from the user declarative preference survey dataset using a retrieval enhancement generation method to obtain the top-k nearest neighbor set. The user preference prior is obtained by pooling the ranking preference, attribute weight, and attribute level of the nearest neighbor set.

10. A method for individual-level retention intervention planning for urban rail transit users according to claim 1, characterized in that, The retention plan generated by the large language model uses a fixed set of output fields, which can be directly executed and audited by the operators; After the retention plan is evaluated for feasibility based on time, cost, and incentive type, and a preference consistency score is applied, the optimal retention plan is selected. The optimal retention plan includes individual intervention window, final intervention time, incentive type, incentive cost, user communication information, and evidence interpretation.

Images