Communication workflow efficiency prediction method and system based on user digital twinning

Abstract

The invention discloses a communication workflow efficiency prediction method based on user digital twinning. The method comprises the following steps: desensitizing historical user communication interaction data, extracting hierarchical behavior modes of a basic response layer, an interaction mode layer and a decision tendency layer, and constructing a standardized behavior mode template library; generating a virtual user group based on the template library, wherein each virtual user is configured with a probability state machine decision engine and a memory mechanism; a communication workflow node to be tested is converted into a standardized interaction scene, and a multi-channel service node is supported; virtual user and workflow interaction is executed in a simulation environment, interaction events are aggregated into a complete behavior track, the conversion rate of each node is analyzed and calculated through a conversion funnel, a loss path is identified, and a service efficiency index is predicted. According to the method, the workflow efficiency can be accurately predicted before actual deployment, key bottleneck nodes are identified, optimization suggestions are provided, the cost is effectively reduced, and the user conversion rate is improved.

Description

Technical Field

[0001] This invention relates to the fields of digital twin technology and user behavior modeling, and in particular to a method and system for predicting communication workflow performance based on user digital twins. Background Technology

[0002] With the rapid development of digital marketing and customer service, enterprises are increasingly relying on complex, multi-step communication workflows to reach and serve users. These workflows typically involve a combination of multiple communication channels, such as SMS, push notifications, email, and telephone, as well as dynamic branching decisions based on user responses. However, before the formal deployment of a workflow, enterprises lack effective means to predict its performance and often need to evaluate its effectiveness through A / B testing or trial runs after actual use. This post-implementation verification approach is costly, risky, and time-consuming.

[0003] In terms of user behavior modeling, existing technologies mainly have the following limitations: 1. Insufficient fidelity of traditional modeling methods: Agent-Based Modeling (ABM) is a commonly used method for simulating user behavior, but its pre-defined behavioral rules are too simplistic and fail to accurately reflect human decision-making processes in complex communication scenarios. User response behavior is often influenced by multiple factors such as content preferences, timing sensitivity, channel habits, and historical interaction experience. Traditional ABM methods struggle to accurately model this multi-dimensional complexity.

[0004] 2. Limitations of Generative AI Applications: While generative AI technologies have been able to create realistic synthetic users for single-interaction simulations in recent years, they lack a systematic framework for integration with structured, multi-step business workflows. Existing generative AI user modeling primarily focuses on single-point applications such as dialogue generation or content recommendation, and has not yet been effectively extended to predicting the overall performance of complete communication workflows.

[0005] 3. Limitations of Digital Twin Technology Applications: Current digital twin technology is mainly applied to physical assets and tangible production processes, such as engine performance monitoring and production line optimization. For abstract communication workflows centered on human-computer interaction, there is a lack of corresponding digital twin modeling methods and simulation tools. Existing technologies cannot effectively handle the modeling of human factors such as user subjective decision-making, emotional responses, and fatigue effects in communication workflows.

[0006] 4. Significant Privacy Compliance Risks: The vast majority of user modeling technologies rely on the direct processing and analysis of raw personal information. Against the backdrop of increasingly stringent global data privacy regulations such as the Personal Information Protection Act and GDPR, this practice faces significant legal risks and compliance obstacles, limiting the practical application of user modeling technologies.

[0007] In the field of workflow performance prediction, existing related technologies are fundamentally different from this invention: Existing internal process optimization technologies mainly focus on predicting internal execution efficiency indicators such as employee task processing time, which falls under the category of internal operation optimization rather than predicting communication efficiency for external users, and lack the ability to model user subjective decisions and behavioral preferences. Statistical analysis-based user churn prediction methods use a static result classification approach, which can only predict the final probability of user churn, but cannot simulate the dynamic interaction process and path dependence effect of users in a new workflow, making it difficult to provide process-oriented optimization guidance. Existing business process monitoring technologies are mainly post-event monitoring platforms, which have no pre-event prediction capabilities and cannot provide performance evaluation before workflow deployment. They can only conduct passive analysis after problems occur. While machine learning-based communication workflow analysis technology can process historical data, it mainly relies on statistical analysis methods and lacks the ability to proactively simulate and predict based on user digital twins, thus failing to simulate users' real behavioral responses in unknown scenarios.

[0008] The core problem with existing technologies lies in the lack of a systematic solution that integrates high-fidelity user modeling, abstract workflow simulation, and privacy-friendly processing. When deploying new communication marketing or service processes, enterprises still face the dilemma of "black-box testing," unable to accurately predict effects and risks before investing real resources. Therefore, there is an urgent need for a new technological solution that, while protecting user privacy, can accurately predict communication workflow performance by building a digital twin environment for users. This would help enterprises shift from the traditional "trial and error-optimization" model to a "prediction-optimization" model, significantly reducing business trial and error costs and decision-making risks. Summary of the Invention

[0009] This invention addresses the problems of existing digital twin technology, which is limited to physical asset modeling and cannot effectively simulate the performance of abstract communication workflows, as well as the contradiction between fidelity and applicability and the common privacy compliance risks of traditional user behavior simulation methods. It provides a communication workflow performance prediction method and system based on user digital twins to solve these problems.

[0010] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: In a first aspect, the present invention provides a method for predicting the performance of communication workflows based on user digital twins. The method includes: before deploying the communication workflow, de-identifying historical user communication interaction data, extracting layered behavioral patterns including a basic response layer, an interaction pattern layer, and a decision tendency layer, and constructing a behavioral pattern template library for pre-deployment prediction; generating a virtual user group based on the behavioral pattern template library, wherein each virtual user is instantiated as a synthetic user agent configured with a decision engine based on a probabilistic state machine and a memory mechanism; converting the communication nodes, decision nodes, and waiting nodes in the communication workflow to be tested into standardized interaction scenarios recognizable by the synthetic user agent, and establishing a mapping relationship between the workflow execution logic and the user decision system; executing the interaction between the virtual user group and the communication workflow in a simulation environment, aggregating the interaction events into a complete user behavior trajectory, calculating the conversion rate of each node and identifying high-frequency churn paths through conversion funnel analysis, and predicting the business performance indicators of the communication workflow.

[0011] Secondly, this invention provides a communication workflow performance prediction system based on user digital twins. The system includes: a behavior pattern extraction module, used to desensitize historical communication interaction data, construct hierarchical behavior patterns of a basic response layer, an interaction pattern layer, and a decision tendency layer, and use a composite distance metric to cluster user behavior to generate a reusable standardized behavior pattern template library; a virtual user generation module, including a template selection unit, a parameterized instantiation unit, and an agent construction unit, wherein the parameterized instantiation unit is used to select a combination of behavior templates based on the characteristic distribution of the target user group and introduce controlled random perturbations to generate individual differences; the agent construction unit configures a probabilistic state machine decision engine containing a perception module, an evaluation module, and an execution module for each virtual user, as well as a memory mechanism for maintaining short-term interaction history and long-term cumulative states; and a workflow mapping module, including a node parsing unit, a semantic analysis unit, a scenario construction unit, and an interface protocol unit, wherein the node parsing unit is used to identify communication nodes, decision nodes, waiting nodes, and multi-channel business node types. The semantic analysis unit is used to extract communication intent, value proposition elements, and sentiment features. The scenario construction unit is used to construct a standardized interactive scenario data structure containing scenario identifiers, channel types, content feature vectors, time context, and expected response options. The interface protocol unit is used to design a unified decision interface protocol and establish dynamic context management. The simulation execution module includes an interaction engine, a state manager, and a trajectory recorder, used to perform real-time interactive simulation of virtual user groups and communication workflows, handle cross-channel information consistency checks and channel switching scenario triggers, and record complete user response behaviors and decision paths. The performance prediction module includes a trajectory aggregation unit, a conversion analysis unit, a churn diagnosis unit, and a group comparison unit. The conversion analysis unit is used to calculate the conversion rate of each communication node and identify high-frequency churn paths. The churn diagnosis unit is used to analyze the churn reasons of specific nodes. The group comparison unit is used to predict the performance differences of the workflow for different user groups and output a performance evaluation result containing overall conversion rate predictions, key churn node identification, and optimization suggestions.

[0012] The beneficial effects of this invention are: (1) The invention realizes the expansion of digital twin technology from the physical world to the field of abstract communication: the invention breaks through the limitation of existing digital twin technology that is only applicable to physical assets and tangible processes, and realizes the systematic modeling, simulation and performance prediction of abstract communication workflow with human-computer interaction as the core, filling the technical gap of digital twin application in the field of communication.

[0013] (2) It solves the contradiction between the fidelity and applicability of user behavior simulation: Compared with the insufficient fidelity caused by the oversimplification of rules in the traditional agent-based modeling (ABM) method, this invention significantly improves the realism of user behavior simulation through hierarchical behavior pattern extraction and synthesis of user agents; Compared with the problem of the lack of structured workflow adaptation capability of simple generative AI, this invention achieves seamless integration of high-fidelity user model and complex workflow through standardized mapping mechanism.

[0014] (3) Achieving a technological breakthrough in replacing post-analysis with pre-prediction: This invention changes the passive mode of traditional communication marketing and service processes that rely on "trial and error-optimization" to an active mode of "prediction-optimization". It can accurately predict key performance indicators before the workflow is officially launched, including overall conversion rate, churn rate of each link, potential complaint risk points, etc., significantly reducing business trial and error costs and decision-making risks.

[0015] (4) Provides a privacy-friendly user modeling solution: This invention uses synthetic data generation technology to simulate virtual users that are fundamentally anonymous, avoiding the dependence of traditional user modeling technology on original personal information. It complies with the legal requirement of the Personal Information Protection Law that information after anonymization is not considered personal information, thus turning compliance barriers into a technological competitive advantage.

[0016] (5) Achieved large-scale and efficient simulation capabilities: Through distributed event-driven architecture and parallel computing technology, this invention supports concurrent interactive simulation of millions of virtual users, and can complete the performance data that originally required months of real testing in a few minutes, providing enterprises with real-time or near-real-time decision support capabilities.

[0017] (6) Provides multi-dimensional intelligent performance evaluation: This invention can not only predict the overall conversion index, but also automatically identify key nodes of user churn, discover potential user experience problems, generate actionable optimization suggestions, and support comparative analysis of different user groups, different channels and different time periods, providing comprehensive data support for strategy optimization.

[0018] In summary, this invention achieves digital twin modeling and performance prediction of abstract communication workflows through innovative steps such as hierarchical behavior pattern extraction, synthetic user agent generation, workflow standardization mapping, large-scale distributed simulation, and intelligent performance prediction. This improves the fidelity and applicability of user behavior simulation, enhances the design efficiency and return on investment of communication workflows, and demonstrates significant technical advantages and commercial value, especially in the performance evaluation and optimization decision-making of complex multi-step communication processes. Attached Figure Description

[0019] Figure 1This is a flowchart illustrating a communication workflow performance prediction method based on user digital twins provided by the present invention.

[0020] Figure 2 This is a schematic diagram of the structure of a communication workflow performance prediction system based on user digital twins provided by the present invention. Detailed Implementation

[0021] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description" and should not be construed as being more preferred or advantageous than other embodiments. To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention are clearly and completely described below with reference to the accompanying drawings.

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the core terms in the embodiments of this invention will be defined below. These definitions are intended to be understood in conjunction with the specific implementation of this invention and should not be considered as limitations on this invention: Communication workflow refers to a structured and sequential user interaction process designed by an enterprise to achieve specific business goals (such as marketing promotion, customer service, user activation, etc.), which includes multiple steps and may involve multiple communication channels (such as SMS, email, App push, etc.). A user digital twin refers to a virtualized model constructed by deeply analyzing and extracting patterns from the historical interaction data of real users. It can highly reproduce the behavioral patterns and decision-making logic of users in communication interaction scenarios. It is not a copy of the user's physical entity, but a computational expression of the behavioral patterns of users in a specific domain. Its core is the following hierarchical behavioral patterns. Synthetic user agents refer to virtual user entities that are instantiated based on the "user digital twin" model and can independently execute decisions. Each agent is equipped with a decision engine and memory mechanism based on a probabilistic state machine, enabling it to receive information, make judgments and execute response actions in the simulation environment. It is the basic unit of simulation prediction in this invention. The hierarchical behavior pattern refers to the structured data paradigm proposed in this invention for building user digital twins, which deconstructs complex interactive behaviors into three interrelated layers: the basic response layer that defines basic response habits, the interaction pattern layer that describes the timing and dependencies of behaviors, and the decision tendency layer that characterizes intrinsic motivation and preferences.

[0024] Example 1: As Figure 1 As shown, this embodiment of the invention provides a method for predicting communication workflow performance based on user digital twins, the method comprising: Step S100: Construct a behavioral pattern template library. Before deploying the communication workflow, historical user communication interaction data is anonymized to extract layered behavioral patterns comprising a basic response layer, an interaction pattern layer, and a decision tendency layer, thus constructing a behavioral pattern template library for pre-deployment prediction.

[0025] This step designs a hierarchical behavioral pattern extraction architecture specifically for predicting communication workflow performance. Compared to traditional user profiling techniques that only extract static user tags, this invention delves into the workflow interaction process, extracting the dynamic response characteristics and decision-making logic of users at each communication node. This hierarchical architecture can systematically and comprehensively capture the key factors influencing user behavior in multi-step communication processes, covering everything from basic response habits to complex decision-making psychology. Furthermore, because the entire process is based on completely desensitized and anonymized data for pattern extraction, the resulting template library does not contain any personally identifiable information, fundamentally solving the privacy compliance risks in user behavior modeling and transforming compliance requirements into technological advantages.

[0026] Furthermore, step S100, which involves extracting hierarchical behavioral patterns, includes: Step S110: Data anonymization. This step includes anonymizing historical communication interaction data while retaining the behavior sequence, timestamps, and response results. For example, sensitive personal information such as user identification, mobile phone number, and email address in the original data is hashed or randomized, retaining only the user behavior time sequence (e.g., "receive SMS → open → click → purchase"), the timestamp difference between each step (e.g., open within 30 minutes of receiving SMS), and the final response result (converted / not converted).

[0027] Preferably, to achieve unified analysis of cross-channel data, this step further maps anonymized interaction data from different channels such as SMS, push notifications, email, and telephone to a standardized communication event model. This model abstracts each interaction into an event containing five basic elements: node identifier (uniquely identifying the specific step in the workflow), channel type, sending time, user response type (e.g., click, reply, ignore, unsubscribe), and response latency. Through this standardized model, heterogeneous data from different sources can be time-aligned to form a structured node-response sequence dataset that can be directly used for subsequent analysis.

[0028] To ensure the quality of input data and the robustness of the model, a data quality preprocessing step is preferably performed before feature extraction. First, a minimum sample size requirement for historical data is established; for example, a single user must have at least 10 interaction records to be included in the analysis, ensuring the stability of behavioral patterns. Simultaneously, a data time span requirement is set, requiring data to cover at least a complete 3-month business cycle to capture seasonal or cyclical changes. Furthermore, an outlier detection and handling strategy must be established, such as marking or removing records with response delays exceeding 7 days, and performing data integrity checks, filling in or filtering missing values ​​in key fields according to preset rules. This provides a clean, complete, and representative dataset for subsequent modeling.

[0029] Step S120: Construction of the basic response layer. This basic response layer is constructed by extracting the user's response rate, response latency distribution characteristics, and channel preference strength at each communication touchpoint. For example, for the SMS channel, the average user response rate is calculated to be 35%, the response latency distribution follows an exponential distribution (λ=0.5), and the channel preference strength is 0.7; for the email channel, the response rate is 15%, the response latency distribution follows a normal distribution (μ=4 hours, σ=2 hours), and the channel preference strength is 0.3. The response latency distribution characteristics can refer to the statistical characteristics of the time interval between when the user receives the communication and when they first respond (e.g., open). For example, the mean, median, standard deviation, or 90th percentile of this distribution can be calculated to characterize whether a user is a "fast-response" or "delayed-response" user.

[0030] Preferably, the response rate extraction goes beyond a general statistical analysis, refining it to multiple specific metrics such as click-through rate, response rate, and ignore rate. Response latency distribution characteristics are further quantified, divided into multiple intervals such as second-level, minute-level, hour-level, and day-level response times to more accurately depict users' response speed habits. Channel preference strength is determined by comprehensively evaluating users' response rates and speeds across different channels, classifying users as SMS-prioritized, push-prioritized, or hybrid, providing a basis for subsequent multi-channel strategy simulation.

[0031] Step S130: Construction of the interaction mode layer. Specifically, the interaction mode layer is constructed by analyzing the sequence dependency of the preceding node's response on subsequent nodes, time window sensitivity, and communication fatigue decay function. The communication fatigue decay function employs a time-dependent decay model, increasing with communication frequency and shortening time intervals. For example, the communication fatigue decay function can be expressed as:

[0032] Where t is the time interval (in hours) since the last communication, n is the cumulative number of communications, and α is the attenuation parameter (value 0.1-0.5).

[0033] Preferably, sequence dependencies are modeled by constructing a state transition probability matrix to quantify the probability that a user will execute the next node response (such as "clicking a link") after executing the response of the previous node (such as "opening a text message"). Time window sensitivity specifically analyzes the differences in user response behavior on weekdays and weekends, and at different times of the day, forming a personal daily routine profile of the user.

[0034] This step introduces modeling of the dynamics and contextual dependencies of user behavior. In particular, the introduction of a communication fatigue decay function addresses the real-world scenario where users develop resistance due to frequent interruptions, a scenario often overlooked in existing models. This is crucial for predicting the effectiveness of high-frequency marketing campaigns. The interaction pattern layer transforms virtual user behavior from isolated and static to one exhibiting temporal correlation and path dependence. The model understands that user behavior evolves over time and interaction history, significantly improving prediction accuracy in multi-step, long-cycle workflow simulations and avoiding overly optimistic estimates of user response rates.

[0035] Step S140: Construction of the decision-making tendency layer. In this process, the decision-making tendency layer is constructed by building a content preference vector, a decision-making speed distribution, and a churn risk factor. Among them, the content preference vector can be constructed by analyzing the texts of communication contents that the user has had positive interactions with (such as clicks, purchases) in history; the decision-making speed distribution refers to the statistical distribution of the time interval between when the user opens the communication content and performs a key action (such as clicking a link, submitting a form). Its characteristics can help determine whether the user makes a decision impulsively or after careful consideration; the churn risk factor is a quantitative index, such as a probability value between 0 and 1, used to represent the possibility that the user churns at the current node.

[0036] Exemplarily, the content preference vector is a 512-dimensional feature vector, and the preference weights of the user for different content themes can be extracted through TF-IDF and Word2Vec technologies; the decision-making speed distribution is modeled based on the user's historical decision-making time and is divided into three types: fast decision-making type (<1 hour, accounting for 30%), cautious decision-making type (1 - 24 hours, accounting for 50%), and delayed decision-making type (>24 hours, accounting for 20%); the churn risk factors include the number of consecutive rejections, the trend of average response time delay, etc.

[0037] Preferably, the construction of the content preference vector does not stop at the keyword level, but through a deep learning model (such as BERT) to deeply understand the semantics of the communication content, and associate the user's preferences with the value proposition elements (such as discount strength, scarcity) and emotional colors (such as motivational, warning) of the communication content, so as to more accurately classify users into promotion-sensitive type, service-oriented type or information-acquisition type. The churn risk factor is quantified as the specific abandonment probability of the user at a specific type of node.

[0038] To further improve the scientificity and accuracy of the model, preferably, before performing the clustering analysis, add steps for optimizing and calibrating key parameters. In this step, the key parameter α of the fatigue attenuation function in the interaction mode layer is not a preset value, but obtained through non-linear fitting of historical A / B test data to make it closer to the real user behavior attenuation law. When selecting the number of clusters, combine quantitative indexes such as the silhouette coefficient and the business team's requirement for the interpretability of the template to jointly determine the optimal K value. For the weights of each component in the composite distance metric, methods such as grid search based on cross-validation can be used for optimization to ensure that the final distance calculation method can maximize the clustering effect.

[0039] Step S150: Cluster Analysis to Generate Templates. In this process, cluster analysis is performed on the three layers of behavioral features based on a composite distance metric to generate multiple reusable standardized behavioral pattern templates. For example, a suitable cluster analysis algorithm can be used. Since the data types of the three layers of behavioral features are different (containing numerical, vector, and time series data respectively), the composite distance metric is calculated by weighted combination of distances applicable to different features. For example, its calculation method can be expressed as:

[0040] in, Euclidean distance is used to apply numerical features (such as response rate) in the basic response layer; To be applicable to the cosine distance of content preference vectors in the decision tendency layer; This is the Dynamic Time Warping (DTW) distance applicable to behavioral sequence features in the interaction mode layer. , , The weights are the three distances, which sum to 1 and can be determined according to the optimization method (such as grid search) described in step S140 to achieve the best clustering effect.

[0041] Preferably, the clustering analysis algorithm can be a hierarchical clustering algorithm, or an algorithm that optimizes the traditional K-means algorithm. The latter can be optimized, for example, by using the K-means++ algorithm to improve the selection of initial cluster centers, thereby enhancing clustering performance and stability. When determining the number of clusters (i.e., the template number K), internal evaluation metrics such as the silhouette coefficient can be used to help determine the optimal number of clusters (e.g., selecting from 20 to 50 categories) to ensure that the behavior within each category is highly similar while the differences between categories are significant.

[0042] Preferably, after generating standardized behavior pattern templates, to ensure their quality and usability, a rigorous template library verification and evaluation process is implemented. First, template coverage verification is performed to ensure that the generated template library can match more than 90% of historical users, avoiding model bias. Second, template discrimination evaluation is performed, using statistical significance tests to verify that the behavioral differences between different templates are clear and meaningful. Finally, and most importantly, backtesting verification is conducted, using a portion of retained recent historical data as a test set to verify the accuracy of behavioral predictions of the virtual user group generated by the templates in historical scenarios, thereby improving the reliability of the entire template library.

[0043] Step S200: Generate a virtual user group. Based on the behavior pattern template library, a virtual user group is generated, wherein each virtual user is instantiated as a synthetic user agent configured with a decision engine based on a probabilistic state machine and a memory mechanism.

[0044] The innovation of this step lies in proposing a template-driven targeted generation method. Instead of creating users randomly or out of thin air, it uses behavioral pattern templates extracted from real data in step S100 to selectively and on a large scale replicate virtual user groups with specific behavioral tendencies. This ensures that the generated virtual user groups are highly consistent with the target real user group in terms of statistical characteristics and behavioral distribution. This makes subsequent simulation experiments more than just aimless extrapolation, thus giving the prediction results real-world reference value and commercial credibility.

[0045] The technology described in this invention, which generates high-fidelity virtual user groups in a targeted manner based on a behavioral pattern template library while protecting privacy, constitutes a specific implementation of the "synthetic data generation technology" mentioned in the beneficial effects of this invention.

[0046] Furthermore, step S200, generating the virtual user group, includes: Step S210: Template Selection and Proportion Determination. This step includes selecting appropriate behavioral template combinations from the template library based on the target user group's characteristic distribution and determining the instantiation ratio. For example, for a marketing workflow targeting young users, template 1 (40%) with a high response rate and template 3 (35%) with a fast decision-making speed are selected, while a small number of conservative templates 7 (25%) are included to ensure the comprehensiveness of the simulation.

[0047] Preferably, this step first receives a definition of the target user group for the workflow to be tested. This definition includes not only demographic characteristics but also distribution requirements for user size, business scenarios, and key behavioral characteristics. Based on this definition, the system will establish constraints for template selection to ensure that the selected template combination can comprehensively cover various typical user behavior patterns that the workflow may encounter, avoiding distortion of simulation results due to sample bias.

[0048] Step S220: Parameterized Instantiation. This involves parameterizing and instantiating the selected template by introducing controlled random perturbations into the base parameters to generate individual differences. For example, for the response rate parameter, a ±10% Gaussian random perturbation is introduced based on the template base value of 0.35 to generate an individualized response rate ranging from 0.315 to 0.385; for the response delay parameter, a ±20% perturbation is introduced based on the exponential distribution λ=0.5 to generate an individualized delay distribution with λ values ​​ranging from 0.4 to 0.6.

[0049] Preferably, the random perturbation here is controlled, and its perturbation range (e.g., within ±15%) is set to ensure that while individual differences are generated, the stability of the population statistical characteristics is not compromised.

[0050] Step S230: Construction of the decision engine. Specifically, the decision engine includes a perception module, an evaluation module, and an execution module. The perception module is used to parse the communication content and identify the communication type, channel, and key information elements. For example, through natural language processing technology, it is identified that the communication content is "promotion type", the channel is "text message", and it contains key elements such as "time-limited discount" and "full reduction".

[0051] The evaluation module constructs a content relevance evaluation based on user behavior parameters and the current state, through content theme matching degree, content value proposition association degree, and user preference consistency, and calculates the response probability in combination with timing suitability and channel preference. Among them, the timing suitability can be calculated according to the historical interaction time distribution of the user. For example, the system can count the open / click behavior frequencies of the user in each time period of 24 hours a day (or 7 days a week) to form a time preference distribution. When planning to send a communication, the current time is matched with the user's preference distribution to obtain a suitability score. For example, if the user is often active from 8 to 10 pm, the timing suitability score for sending at this time period is high. Among them, the channel preference intensity can be quantified based on the historical response rates of the user on different channels (such as text messages, emails, App push notifications). For example, if a user's email open rate is 30% and the click-through rate of App push notifications is 5%, it can be calculated that his preference intensity for emails is significantly higher than that of App push notifications. This intensity can be a normalized score or a vector representing the response probabilities of each channel.

[0052] Preferably, the response probability can be calculated using a linear weighted model:

[0053] where - is the weight coefficient, and the sum is 1. The execution module generates a response action based on the evaluation result and a random factor. For example, when the response probability > 0.6, the "respond immediately" action is executed with an 80% probability; when 0.3 < response probability ≤ 0.6, the "respond later" action is executed with a 50% probability; when the response probability ≤ 0.3, the "do not respond" action is executed with a 90% probability. Optionally, a non-linear model (such as logistic regression, neural network, etc.) can also be used for probability calculation to adapt to the complexity requirements of different business scenarios. Among them, the evaluation result is a comprehensive response probability P calculated based on content relevance, timing suitability, channel preference, etc. To simulate the uncertainty in human decision-making, a random factor is introduced. For example, the system can generate a random number R between 0 and 1. When R < P, the virtual user executes the "respond" action; when R >= P, the "do not respond" action is executed. This decision-making method based on probability, rather than a fixed threshold, makes the simulation results closer to the statistical distribution of the real world.

[0054] Preferably, the entire decision engine is built on a probabilistic state machine model, which ensures that although each decision of a virtual user contains randomness, it is essentially in accordance with its inherent behavioral template and current memory state, so that its behavior is neither completely fixed nor without logic, but has a high degree of consistency and traceability.

[0055] Preferably, to ensure the scientific validity and accuracy of the weight coefficients in the decision engine, decision weight calibration is also included: before the simulation begins, the weight coefficients in the formula ( - The initial weights are not set based on experience, but rather on historical behavioral big data of template users. Machine learning models such as multivariate logistic regression are used to fit the data, thus determining an optimal set of initial weights. This ensures that the decision model has the ability to interpret historical data from the initial state. During simulation, reinforcement learning is preferably used to dynamically optimize the weights. In this mechanism, the virtual user agent can be considered an agent, and its key elements in the decision-making process can be defined as follows: State, which is composed of the internal state maintained by the virtual user's memory mechanism (including short-term interaction history and long-term cumulative state) and the external dynamic context in which the decision is made; Action, which is the specific response behavior generated by the decision engine based on the response probability P calculated by the evaluation module, such as "immediate response," "delayed response," or "no response"; Reward, which is the immediate business feedback obtained by the virtual user after performing the action. For example, successfully guiding the user to the final conversion node can be defined as a positive reward (e.g., +1), while negative behaviors such as churn or canceling can be defined as a negative reward (e.g., -1). Through continuous iteration, this reinforcement learning mechanism (such as Q-learning or policy gradient algorithms) can learn an optimal policy, so that the adjustment of weight coefficients is geared towards maximizing long-term cumulative rewards, thereby simulating the learning and preference evolution process of real users in the experience.

[0056] Step S240: Memory Mechanism Implementation. During this process, the memory mechanism maintains short-term interaction history and long-term cumulative state, enabling dynamic parameter adjustments based on history. Short-term memory stores interaction records from the last 7 days, including reception time, response behavior, and emotional tendencies; long-term memory maintains cumulative fatigue levels, overall satisfaction, brand awareness, and other state variables, and is dynamically updated based on new interaction results.

[0057] Preferably, the memory mechanism is further designed to be more refined and dynamic: Short-term memory stores the most recent N interactions (e.g., N ranges from 5 to 10) to capture immediate changes in behavioral patterns. Long-term memory maintains the user's cumulative state, such as key metrics like total interactions and historical response rates across various channels. The learning and adaptation mechanism introduces a reinforcement learning-based parameter fine-tuning mechanism, allowing virtual users to dynamically adjust their response tendencies based on interaction feedback (e.g., receiving rewards or experiencing negative experiences), thus simulating the learning and adaptation process of real users. The forgetting mechanism designs a forgetting curve function so that the influence of historical interaction events decays over time, ensuring that recent interactions have a more significant impact on current decisions, which aligns better with human memory patterns.

[0058] Preferably, to ensure the stability and robustness of the virtual user generation and simulation process, exception handling is also included. For template matching failures, when the received target user group characteristics cannot find enough matching behavioral templates in the template library, a degradation strategy will be initiated. Preferably, this strategy includes selecting a general template representing the average behavior of all users, or dynamically weighting and fusing multiple closest templates to generate a temporary hybrid template to ensure the simulation task can start smoothly. When parameters exceed limits, during the parameterization instantiation process in step S220, if the introduced random perturbation accidentally causes certain behavioral parameters (such as response probability) to exceed their logically reasonable range (e.g., greater than 1 or less than 0), a correction mechanism will be automatically triggered. This mechanism will forcibly correct the out-of-bounds parameters to the boundary values ​​of their valid range (e.g., 1 or 0) and record such events for subsequent evaluation of the stability of the random perturbation algorithm. When the decision engine malfunctions, in step S230, when the perception module cannot parse the communication content normally due to abnormal format or the presence of unknown elements, a default processing strategy will be executed. Preferably, the strategy is to set the content relevance evaluation item of the communication to the lowest score, simulating the behavior of real users choosing to ignore information because they cannot understand it, thereby avoiding the interruption of the entire simulation process due to the parsing failure of a single node.

[0059] Preferably, to ensure that the final generated virtual user group is of high quality and trustworthy, the system also includes virtual user quality verification, including statistical consistency testing, i.e., after the virtual user group is generated but before it is officially put into simulation, the system will conduct a rigorous statistical feature comparison; behavioral rationality check, which checks whether the behavior of virtual users conforms to basic business logic and expectations by running virtual users in a series of benchmark workflows; and diversity assessment, which evaluates the behavioral diversity of the virtual user group to avoid the generated group behavior being too homogeneous (i.e., convergence effect). Step S300: Establish workflow mapping. Specifically, the communication nodes, decision nodes, and waiting nodes in the communication workflow to be tested are converted into standardized interaction scenarios recognizable by the synthetic user agent, establishing a mapping relationship between the workflow execution logic and the user decision-making system. In this invention, a communication node refers to an execution unit in the workflow responsible for initiating one-way or two-way information transmission to the user; a decision node is a branch unit that judges and selects the workflow path based on one or more preset conditions; and a waiting node is a control unit that introduces time delays in the workflow. This step is the core bridge connecting the high-fidelity virtual user and the structured communication workflow, solving the key technical challenge of enabling virtual users to understand and execute real-world business logic.

[0060] Furthermore, step S300, which involves converting workflow nodes into standardized interaction scenarios, includes: Step S310: Workflow structured parsing, wherein the workflow is structured and parsed to identify communication nodes, decision nodes, and waiting node types. For example, the workflow definition file in BPMN format is parsed using the workflow engine API to identify the "send SMS" node as a communication node, the "determine whether to click" node as a decision node, and the "wait for user response" node as a waiting node. Preferably, the parsing result is constructed as a directed acyclic graph (DAG) representation of the workflow to clearly reveal the dependencies and execution order between nodes.

[0061] Step S320: Semantic analysis of communication content. This step includes performing semantic analysis on the content of communication nodes to extract communication intent, value proposition elements, and sentiment features. For example, using a BERT pre-trained model to analyze the SMS content "Limited-time offer for 3 days, 20% off everything, new users only," the communication intent is extracted as "promotion," the value proposition elements are ["limited-time offer," "discount offer," "exclusive for new users"], and the sentiment features are "urgency + attractiveness." Preferably, this analysis can further identify explicit "call-to-action" (CTA) elements and evaluate content complexity metrics, providing richer input for virtual users' decision-making.

[0062] Step S330: Construction of standardized scene data structure. This step includes constructing a standardized interactive scene data structure, which includes scene identifier, channel type, content feature vector, time context, and expected response options. Preferably, the standardized interactive scene data structure supports a dynamic parameter injection mechanism, which can dynamically adjust the specific content of the scene according to the real-time status of the virtual user (such as accumulated fatigue) to achieve higher fidelity interaction.

[0063] Step S340: Design a unified decision interface protocol. This step includes designing a unified decision interface protocol, defining a request format that includes a scenario description, user context, and optional actions, and a response format that includes decision results and response latency. For example, the request format includes inputs such as the user's current state, scenario parameters, and historical interaction records; the response format includes outputs such as response actions, response time, and confidence level. Preferably, the protocol includes a decision routing mechanism to distribute different decision requests to the corresponding evaluation modules, and can improve the response efficiency of high-frequency scenarios through decision caching.

[0064] Step S350: Dynamic context management, wherein dynamic context management is established to maintain the session state and historical path information of each virtual user. For example, state variables are maintained for each virtual user, including: "current workflow position", "cumulative number of interactions", "last response time", "emotional state value", etc., and are updated in real time after each interaction. Preferably, the dynamic context management mechanism may include time context awareness capabilities, thereby being able to distinguish the different impacts of absolute time (such as weekends or weekdays) and relative time (such as the time since the last interaction) on decision-making.

[0065] Furthermore, when the communication workflow to be tested includes multiple channel business nodes, the construction of the standardized interaction scenario also includes: Step S351: Multi-channel Node Type Identification. During this process, the types of multi-channel business nodes in the workflow are identified, including parallel outreach nodes, channel cascading nodes, and channel complementary nodes. Parallel outreach nodes send the same information simultaneously through multiple channels (e.g., sending both SMS and push notifications). Channel cascading nodes select subsequent channels based on the response status of preceding channels (e.g., sending an email if there is no SMS response). Channel complementary nodes deliver different levels of information content through different channels (e.g., sending a discount code via SMS and sending detailed activity rules via email).

[0066] Step S352: Multi-channel scenario template establishment. Specifically, a standardized scenario template is established for each type of multi-channel business node, transforming the abstract channel combination logic into a decision tree structure executable by virtual users, including channel selection conditions, response priorities, and status transmission rules. For example, for channel cascading nodes, a decision tree is established: if the SMS response rate < threshold and the waiting time > the set value, the email channel is triggered; if the email still has no response and certain conditions are met, a telephone outbound call is triggered. The multi-channel scenario template is part of a broader cross-channel collaborative decision-making framework. Preferably, the multi-channel scenario module framework can also be set to be responsible for uniformly managing the response priorities of each channel and tracking the cumulative fatigue across channels.

[0067] Step S353: Construction of a cross-channel information consistency check mechanism. This mechanism ensures semantic consistency in the expression of the same business intent across different channels, preventing virtual users from making abnormal decisions due to information conflicts. This mechanism achieves precise control by establishing a multi-layered consistency verification system. At the semantic level, a pre-trained BERT model is used to calculate the semantic vectors of content from different channels, and cosine similarity is used to assess semantic consistency. For example, a threshold of 0.85 or higher can be set to pass verification. At the information element level, a key business element extractor is established to automatically identify and compare core information elements such as price, time limits, discount conditions, and target user groups, ensuring complete consistency of numerical information across channels. For example, when the SMS content is "20% off for 3 days only" and the email content is "20% off for 72 hours only," the system can identify "3 days" and "72 hours" as consistent expressions. However, if the email displays "25% off," an inconsistency alarm will be triggered.

[0068] Step S354: Design of Channel Switching Scenario Trigger Conditions. Specifically, this involves designing the scenario trigger conditions for channel switching, defining the trigger thresholds and decision logic for virtual users migrating from low-interaction-cost channels to high-interaction-cost channels. This design is based on channel cost-benefit analysis theory and establishes a quantitative channel cost hierarchy system. For example, SMS can be set as the baseline cost unit 1, push notifications as 1.2, email as 1.5, and outbound calls as 5.0, achieving cost comparability between different channels. The design of the trigger conditions comprehensively considers three core elements: time dimension, response probability, and user characteristics. In terms of the time dimension, progressive waiting thresholds can be set; for example, if there is no response after 4 hours of SMS sending, consider switching to email; if there is no response after 24 hours of email waiting, consider outbound calls.

[0069] For example, for a user group with a 30% SMS response rate, if the response rate falls below the expected 50% within 4 hours of sending an SMS, the email channel is automatically activated. The decision logic also integrates the user's channel preference strength; for users who prefer email (preference strength > 0.6), email is prioritized as the primary reach channel, while for users who prefer SMS, the waiting time for the SMS channel is extended. Preferably, this mechanism also includes an intelligent channel combination strategy, capable of dynamically selecting parallel or cascading reach modes based on business urgency and expected conversion value. When the business value exceeds a set threshold, multiple channels are activated simultaneously to improve reach success rate.

[0070] Preferably, a scenario mapping validity verification mechanism can be established to ensure that the constructed standardized interactive scenarios can be correctly parsed and executed by the virtual user agent. This verification mechanism includes three levels: syntax integrity check, logical consistency verification, and execution feasibility test. The syntax integrity check verifies the format of each standardized scenario using a predefined scenario data structure pattern, ensuring that all required fields are correctly filled and the data types meet the requirements. Logical consistency verification checks the logical relationships within the scenario, such as whether the response options cover all possible user behavior paths, whether the time constraints are reasonable, and whether the preconditions match the subsequent actions. The execution feasibility test verifies the actual executability of the scenario by running a sample virtual user agent in an isolated environment.

[0071] Step S400: Perform simulation prediction. This step includes executing the interaction between the virtual user group and the communication workflow in a simulation environment, aggregating the interaction events into a complete user behavior trajectory, calculating the conversion rate of each node through conversion funnel analysis and identifying high-frequency churn paths, and predicting the business performance indicators of the communication workflow. Preferably, the simulation environment can be built based on a distributed, event-driven simulation engine, supporting concurrent interaction of millions of virtual users, enabling large-scale testing to be completed in a short time, and supporting dynamic adjustment and A / B testing of the workflow.

[0072] Furthermore, step S400, which involves analyzing and predicting business performance indicators based on user behavior trajectories and conversion funnels, specifically includes: Step S410: Initialize the core components of the simulation engine. Specifically, establish a distributed simulation execution framework including an interaction engine, a state manager, and a trajectory recorder to support concurrent interaction processing of a large-scale virtual user group. This framework preferably adopts a distributed event-driven architecture. This architecture means that each component of the system (such as virtual user agents and workflow engine nodes) acts as an independent event producer or consumer, communicating through asynchronous messages or events rather than directly making synchronous calls. This loosely coupled architecture can effectively handle high concurrency and asynchronous interaction scenarios, providing good scalability for large-scale simulations.

[0073] In this architecture, the interaction engine coordinates real-time interactions between virtual users and workflow nodes, employing an event-driven architecture to handle asynchronous communication scenarios and supporting a high concurrency capability of processing 100,000 interaction events per second. The state manager maintains real-time state information for each virtual user, including current workflow position, accumulated interaction history, fatigue status, and decision context, ensuring data consistency and fast access through distributed caching technology. The trajectory recorder uses a streaming data processing architecture to capture and store the user's complete behavioral trajectory in real time, supporting subsequent batch analysis and real-time monitoring.

[0074] Preferably, the framework also includes a load balancing mechanism and fault recovery capabilities, which can automatically migrate tasks to other nodes when a single computing node fails, ensuring the continuity and stability of the simulation process. The simulation engine also supports dynamic scaling, automatically adjusting the allocation of computing resources according to the simulation scale. The advantage is that this framework provides a solid technical foundation for large-scale, high-fidelity user behavior simulation, ensuring the accuracy and reliability of the simulation results.

[0075] Step S420: Cross-channel simulation processing mechanism. Specifically, during simulation execution, cross-channel information consistency checks and channel switching scenario triggers are handled in real time to ensure correct simulation execution of multi-channel business nodes. When a virtual user encounters a multi-channel business node, the information consistency check mechanism established in step S353 is first invoked to verify the semantic consistency of the content across channels. If inconsistencies are found, the simulation process for the current user is paused and the abnormal event is recorded. For channel switching scenarios, the user's response status is monitored in real time according to the trigger conditions designed in step S354. When the switching conditions are met, the channel migration logic is automatically executed.

[0076] For example, when a virtual user waits for 4 hours without a response on the SMS channel, the system automatically triggers activation of the email channel, updates the user's channel status, and recalculates the response probability. This mechanism also maintains cumulative fatigue calculations across channels, ensuring that the user's interaction history across different channels correctly influences subsequent decisions. Preferably, the system establishes a dynamic channel priority adjustment mechanism, adjusting the channel preference weights for different user groups based on real-time simulation results to improve the simulation's adaptability. The advantage is that this mechanism ensures the accuracy and completeness of simulations in multi-channel business scenarios, avoiding prediction biases caused by errors in channel processing logic.

[0077] Step S430: High-Frequency Churn Path Identification. In this process, high-frequency churn path patterns are identified using statistical methods. The frequency and impact of different churn paths are quantified. Sequence pattern mining algorithms are used to analyze the behavioral trajectories of all virtual users, identifying churn path sequences whose frequency exceeds a set threshold. For example, the definition criteria for a high-frequency churn path may include a path frequency greater than 10% of the total churned users, and an average churn rate on that path higher than 1.5 times the overall average churn rate.

[0078] For example, analysis revealed that the most frequent churn path was "receive SMS → open SMS → browse webpage → leave," accounting for 25% of all churned users, with an average churn rate of 85%. Further calculations were made to determine the business impact of each high-frequency churn path, quantifying its optimization priority by multiplying its frequency by the loss of value per user. Here, "loss of value per user" is a quantifiable metric that can be defined according to specific business scenarios, used to assess the potential economic loss caused by the churn of a single user in this workflow. For example, it can be defined as: the target order value or average transaction amount for this communication workflow; the estimated lifetime value (LTV) based on historical data; or the average profit contribution per user under a specific business scenario. This definition transforms the abstract risks of different churn paths into comparable financial indicators, thereby more scientifically guiding the prioritization of optimization efforts.

[0079] Preferably, a time clustering analysis of churn paths is established to identify changes in churn patterns within a specific time period, such as the difference in churn paths between weekdays and weekends.

[0080] Step S440: Behavioral Trajectory Aggregation, whereby the response records of the virtual user at each node of the communication workflow are aggregated into a complete behavioral trajectory, including the complete link of receiving communication, opening, clicking, and executing the target action. For example, a single user trajectory is recorded as: "Receive SMS at time T0 → Open SMS at time T1 → Click link at time T2 → Browse page at time T3 → Add to cart at time T4 → Complete purchase at time T5", while recording the time interval and user state changes of each step. Preferably, the aggregated massive trajectory data can be organized into a multidimensional data cube (OLAP Cube) to support subsequent multidimensional, interactive slicing and drill-down analysis.

[0081] Step S450: Conversion Funnel Analysis. In this process, a conversion funnel model is constructed based on aggregated user behavior trajectory data. By calculating the conversion rate and churn rate between each communication node, bottleneck nodes and optimization opportunities in the workflow are systematically identified. A conversion funnel is an analytical method that visualizes the user's behavioral path in a multi-step business process. By arranging the key nodes of the workflow according to business logic, a funnel-shaped structure from wide to narrow is formed, intuitively showing user retention and churn at each stage. The construction process of the conversion funnel includes three core steps: node definition, path mapping, and metric calculation. In the node definition stage, the workflow is divided into key conversion nodes according to business objectives, such as "receiving communication," "opening content," "clicking a link," "browsing a page," and "performing the target action." In the path mapping stage, the behavioral trajectories of all virtual users are standardized according to a predefined node sequence, identifying the specific path traversed by each user and the behavioral results at each node. For example, in the metric calculation stage, the conversion rate between each node is calculated using statistical methods. The conversion rate calculation formula is:

[0082] The churn rate is:

[0083] For example, in a simulation involving 10,000 virtual users, the system identified five key conversion nodes and calculated corresponding metrics: the conversion rate from node 1 "SMS sending" (base number 10,000) to node 2 "SMS opening" was 30% (3,000 users passed, 7,000 users dropped); the conversion rate from node 2 to node 3 "Link clicking" was 40% (1,200 users passed, 1,800 users dropped); the conversion rate from node 3 to node 4 "Page browsing" was 75% (900 users passed, 300 users dropped); and the conversion rate from node 4 to node 5 "Purchase completion" was 33% (300 users passed, 600 users dropped). Funnel analysis clearly shows that the dropout rate from node 1 to node 2 is the most severe, becoming the main bottleneck in the entire workflow.

[0084] Step S460: Churn Analysis. Specifically, analyze the reasons for user churn at specific points, including mismatch between communication content and user preferences, communication sending time deviating from the user's historical active time pattern, or the cumulative fatigue effect from previous communication interactions exceeding the user's acceptance threshold. For example, it was found that 70% of churn occurs in the "open to click" stage, with the main reasons including: content relevance score <0.4 (40%), sending time during a period of low user activity (30%), and cumulative fatigue level >0.8 (30%). Step S470: Group Comparison Analysis. Specifically, compare the conversion performance of different virtual user groups to predict the differences in workflow effectiveness across different user segments. For example, the overall conversion rate for young users (18-25 years old) is 4.5%, for middle-aged users (26-40 years old) it is 2.8%, and for elderly users (40 years and older) it is 1.2%, providing data support for personalized workflow optimization.

[0085] Step S480: Output Performance Evaluation Results. Specifically, the output includes a performance evaluation result comprising the overall conversion rate prediction, key churn node identification, and optimization suggestions. For example, the generated report includes: an expected overall conversion rate of 3.2% ± 0.3%, the key churn node being the "content page browsing" stage, and optimization suggestions including actionable recommendations such as "adjusting the sending time to the user's active period," "optimizing the personalization of content," and "reducing communication frequency to avoid fatigue." Through the above technical solution, performance prediction of the communication workflow before formal deployment is achieved, significantly reducing business trial-and-error costs and improving the scientific nature and accuracy of decision-making.

[0086] Example 2: Figure 2 As shown, based on the same inventive concept as the communication workflow performance prediction method based on user digital twins provided in Embodiment 1, this embodiment of the invention also provides a communication workflow performance prediction system based on user digital twins, the system comprising: The behavior pattern extraction module 100 is used to desensitize historical communication interaction data. By constructing hierarchical behavior patterns of basic response layer, interaction pattern layer and decision tendency layer, user behavior is clustered using composite distance metric to generate a reusable standardized behavior pattern template library. The virtual user generation module 200 includes a template selection unit, a parameterized instantiation unit, and an agent construction unit. The parameterized instantiation unit is used to select a combination of behavioral templates based on the characteristic distribution of the target user group and introduce controlled random perturbations to generate individual differences. The agent construction unit configures a probabilistic state machine decision engine containing a perception module, an evaluation module, and an execution module for each virtual user, as well as a memory mechanism for maintaining short-term interaction history and long-term cumulative state. The workflow mapping module 300 includes a node parsing unit, a semantic analysis unit, a scenario construction unit, and an interface protocol unit. The node parsing unit is used to identify communication nodes, decision nodes, waiting nodes, and multi-channel business node types. The semantic analysis unit is used to extract communication intent, value proposition elements, and sentiment features. The scenario construction unit is used to construct a standardized interactive scenario data structure that includes scenario identifiers, channel types, content feature vectors, time context, and expected response options. The interface protocol unit is used to design a unified decision interface protocol and establish dynamic context management. Furthermore, the workflow mapping module 300 in this embodiment of the invention also includes a multi-channel processing unit, specifically designed to handle complex workflows containing multi-channel business nodes. Specifically, the multi-channel processing unit includes a channel type identification subunit, used to automatically identify three types of multi-channel business nodes: parallel reach nodes, channel cascade nodes, and channel complementary nodes; a cross-channel consistency check subunit, used to construct a semantic consistency verification mechanism, preferably using a BERT model to calculate the semantic similarity of content across different channels to ensure the consistency of business information expression across channels; and a channel switching logic construction subunit, used to establish channel switching decision rules based on time thresholds, response probabilities, and user preferences, realizing an intelligent migration strategy from low-cost channels to high-cost channels. This multi-channel processing unit collaborates with other units through standardized data interfaces to ensure accurate mapping and efficient execution of complex multi-channel business logic.

[0087] The simulation execution module 400 includes an interaction engine, a state manager, and a trajectory recorder. It is used to perform real-time interaction simulation between virtual user groups and communication workflows, handle cross-channel information consistency checks and channel switching scenario triggers, and record complete user response behaviors and decision paths. Furthermore, the simulation execution module 400 in this embodiment of the invention also includes a cross-channel coordination unit, specifically responsible for coordinating simulation execution in multi-channel business scenarios. In this process, the cross-channel coordination unit includes a consistency check executor, which calls the consistency check mechanism of the workflow mapping module 300 in real time during simulation runtime to dynamically verify multi-channel content; a channel switching trigger, which automatically executes channel switching logic and updates the user's channel preference state based on preset trigger conditions and the real-time response state of the virtual user; and a fatigue accumulation calculator, which maintains the cumulative fatigue state of the virtual user in different channels, ensuring that the impact of cross-channel interaction on user decisions is accurately modeled. This coordination unit works in conjunction with the interaction engine and state manager through an event bus mechanism to ensure the accuracy and continuity of simulation in multi-channel business scenarios.

[0088] The performance prediction module 500 includes a trajectory aggregation unit, a conversion analysis unit, a churn diagnosis unit, and a group comparison unit. The conversion analysis unit is used to calculate the conversion rate of each communication node and identify high-frequency churn paths. The churn diagnosis unit is used to analyze the churn reasons of specific nodes. The group comparison unit is used to predict the performance differences of the workflow for different user groups and outputs a performance evaluation result that includes the overall conversion rate prediction value, key churn node identification, and optimization suggestions.

[0089] Furthermore, the conversion analysis unit in this embodiment of the invention also includes a high-frequency churn path miner, specifically designed to identify and analyze high-frequency patterns of user churn. Specifically, this miner employs a sequence pattern mining algorithm to analyze the behavioral trajectories of all virtual users, defining high-frequency churn paths by setting a frequency threshold (a proportion greater than 10% of total churned users) and an impact threshold (churn rate higher than 1.5 times the average). The miner also includes a path clustering analysis function, merging and classifying similar churn paths to identify representative churn patterns; a path impact assessment function, quantifying the business impact of different churn paths by calculating the product of path frequency and single-user value loss; and a path optimization suggestion generation function, automatically generating targeted optimization suggestions for each high-frequency churn path based on churn cause analysis. The analysis results of this miner are output in a structured data format, supporting visualization and further business analysis.

[0090] Furthermore, in this embodiment of the invention, the behavior pattern extraction module 100 is also used to perform the following steps: Specifically, the data anonymization processing subunit performs privacy protection processing on the original interaction data to ensure the security of personal sensitive information; the feature extraction subunit constructs multi-dimensional features for the basic response layer (response rate, latency distribution, channel preference), the interaction pattern layer (sequence dependency, time sensitivity, fatigue decay), and the decision tendency layer (content preference, decision speed, churn risk); preferably, the clustering analysis subunit uses an improved K-means algorithm combined with a composite distance metric to generate 8-12 representative behavioral pattern templates.

[0091] Furthermore, in this embodiment of the invention, the virtual user generation module 200 is also used to perform the following steps: The template selection unit selects the most matching combination of behavioral templates from the template library based on the age structure, consumption habits, and active time periods of the target user group, and determines the instantiation ratio of each template. The parameterized instantiation unit introduces Gaussian random perturbation (perturbation amplitude 10%-20%) on the basic parameters of the selected templates to generate virtual user parameters with individual differences. The agent construction unit builds a decision engine for each virtual user that includes three stages: perception, evaluation, and execution, as well as a memory mechanism to maintain short-term (7 days) and long-term (cumulative) interaction history.

[0092] Furthermore, in this embodiment of the invention, the workflow mapping module 300 is also used to perform the following steps: This step includes: the node parsing unit identifies various node types in the workflow through a BPMN parser, supporting complex multi-channel business logic; the semantic analysis unit uses pre-trained models such as BERT to perform deep semantic understanding of the communication content and extract key information elements; the scenario construction unit converts abstract workflow nodes into structured interactive scenario descriptions to ensure that the virtual user agent can accurately understand and respond; and the interface protocol unit establishes standardized API interfaces to achieve seamless integration between the workflow engine and the user agent system.

[0093] Furthermore, in this embodiment of the invention, the simulation execution module 400 is also used to perform the following steps: Among them, the interaction engine supports concurrent interaction simulation of millions of virtual users and adopts an event-driven architecture to ensure high performance; the state manager maintains the state information of each virtual user in real time and supports complex state transitions and context passing; the trajectory recorder fully records the user's behavior trajectory, including response time, decision path, state changes and other detailed information, providing a data foundation for subsequent analysis.

[0094] Furthermore, in this embodiment of the invention, the performance prediction module 500 is also used to perform the following steps: The trajectory aggregation unit aggregates scattered interaction events according to user and time dimensions to form a complete user behavior trajectory; the conversion analysis unit calculates key indicators such as conversion rate and churn rate for each node and builds a conversion funnel model; the churn diagnosis unit analyzes the reasons for churn through machine learning algorithms and identifies key factors affecting conversion; and the group comparison unit supports multi-dimensional group comparison analysis to provide data support for personalized optimization.

[0095] Preferably, embodiments of the present invention further include a system coordination controller for overall management of data flow and execution coordination among modules. Specifically, the system coordination controller includes a data flow management unit responsible for standardizing data transmission formats and interface standards between modules; an execution scheduling unit that dynamically schedules the execution priority and resource allocation of each module according to the simulation scale and computing resources; a configuration management unit that uniformly manages the global parameter configuration of the system, including key parameters such as simulation scale, accuracy requirements, and time settings; and an exception handling unit that monitors the operating status of each module and automatically executes recovery strategies or generates alarm information when an exception occurs. This controller ensures the high availability, high performance, and high reliability of the entire system, providing a solid technical guarantee for large-scale commercial applications.

[0096] The system of this invention, through modular design, achieves high scalability and maintainability, supporting communication workflow performance prediction needs of varying scales and complexities, and providing strong technical support for enterprises' digital marketing and customer service. Through the detailed description of the above embodiments, this invention realizes communication workflow performance prediction based on user digital twin technology, solving the technical challenges of traditional methods in terms of fidelity, applicability, and privacy compliance. It provides enterprises with technical means for pre-prediction and scientific decision-making, possessing significant technological advancement and commercial application value.

[0097] This invention is not an isolated innovation of a specific technology (such as machine learning algorithms or simulation technology), but rather a creative and systematic integration of the digital twin concept, deep behavioral modeling, and communication service process prediction, thereby solving a problem that existing technologies cannot systematically address.

Claims

1. A method for predicting communication workflow performance based on user digital twins, characterized in that, include: Before deploying the communication workflow, the historical user communication interaction data is anonymized, and hierarchical behavioral patterns containing basic response layer, interaction pattern layer and decision tendency layer are extracted to build a behavioral pattern template library for pre-deployment prediction. Based on the aforementioned behavior pattern template library, a virtual user group is generated, wherein each virtual user is instantiated as a synthetic user agent configured with a decision engine based on a probabilistic state machine and a memory mechanism. The communication nodes, decision nodes, and waiting nodes in the communication workflow to be tested are converted into standardized interaction scenarios that can be recognized by the synthetic user agent, and a mapping relationship between the workflow execution logic and the user decision system is established. The interaction between the virtual user group and the communication workflow is executed in the simulation environment. The interaction events are aggregated into a complete user behavior trajectory. The conversion rate of each node is calculated through conversion funnel analysis and high-frequency churn paths are identified to predict the business performance indicators of the communication workflow.

2. The communication workflow performance prediction method based on user digital twins according to claim 1, characterized in that, The steps for extracting hierarchical behavioral patterns include: Historical communication and interaction data is anonymized while retaining the behavior sequence, timestamp, and response results; The basic response layer is constructed by extracting the user's response rate, response latency distribution characteristics, and channel preference intensity at each communication touchpoint. The interaction mode layer is constructed by analyzing the sequential dependency of the previous node's response on the subsequent nodes, the time window sensitivity, and the communication fatigue decay function. The communication fatigue decay function adopts a time-related decay model, which increases with the increase of communication frequency and the shortening of time interval. The decision preference layer is constructed by building a content preference vector, a decision speed distribution, and a churn risk factor. Cluster analysis is performed on the above three-layer behavioral features based on a composite distance metric to generate multiple reusable standardized behavioral pattern templates.

3. The communication workflow performance prediction method based on user digital twins according to claim 1, characterized in that, The generated virtual user group includes: Based on the target user group's characteristic distribution, select appropriate behavioral template combinations from the template library and determine the instantiation ratio; The selected template is parameterized and instantiated, and controlled random perturbations are introduced on the basic parameters to generate individual differences; The decision engine includes a perception module, an evaluation module, and an execution module. The perception module is used to parse communication content and identify communication type, channel, and key information elements. The evaluation module constructs a content relevance assessment based on user behavior parameters and current state through content topic matching degree, content value proposition relevance degree, and user preference consistency, and calculates the response probability by combining timing suitability and channel preference. The execution module generates response actions based on the evaluation results and random factors. The memory mechanism maintains short-term interaction history and long-term cumulative state, enabling dynamic parameter adjustment based on history.

4. The communication workflow performance prediction method based on user digital twins according to claim 1, characterized in that, The process of converting workflow nodes into standardized interaction scenarios includes: The workflow is structured and parsed to identify the types of communication nodes, decision nodes, and waiting nodes. Semantic analysis is performed on the content of communication nodes to extract communication intent, value proposition elements, and sentiment features; Construct a standardized interactive scenario data structure, including scenario identifier, channel type, content feature vector, time context, and expected response options; Design a unified decision interface protocol, defining a request format that includes scenario description, user context, and optional actions, as well as a response format that includes decision results and response latency; Establish dynamic context management to maintain the session state and historical path information of each virtual user.

5. The communication workflow performance prediction method based on user digital twins according to claim 4, characterized in that, When the communication workflow to be tested contains multi-channel business nodes, the construction of the standardized interaction scenario further includes: Identify the types of multi-channel business nodes in the workflow, including parallel reach nodes, channel cascade nodes, and channel complementary nodes. Parallel reach nodes send the same information through multiple channels simultaneously, channel cascade nodes select subsequent channels based on the response of preceding channels, and channel complementary nodes transmit different levels of information content through different channels. Establish standardized scenario templates for each type of multi-channel business node, and transform the abstract channel combination logic into a decision tree structure that can be executed by virtual users, including channel selection conditions, response priorities, and status transmission rules; Establish a cross-channel information consistency check mechanism to ensure that the expression of the same business intent remains semantically consistent across different channels, and to prevent virtual users from making abnormal decisions due to information conflicts; Design the scenario triggering conditions for channel switching, and define the triggering threshold and decision logic for virtual users migrating from low-interaction-cost channels to high-interaction-cost channels.

6. The communication workflow performance prediction method based on user digital twins according to claim 1, characterized in that, The aforementioned business performance indicators based on user behavior trajectories and conversion funnel analysis specifically include: The response records of virtual users at each node of the communication workflow are aggregated into a complete behavioral trajectory, including the complete link of receiving communication, opening, clicking and executing the target action; By analyzing the conversion funnel, the pass rate of each communication node is calculated to identify the node with the most severe user churn in the process from initial contact to final conversion; Analyze the reasons for user churn at specific points, including mismatch between communication content and user preferences, communication sending time deviating from the user's historical active time pattern, or the cumulative fatigue effect of previous communication interactions exceeding the user's acceptance threshold. By comparing the conversion performance of different virtual user groups, we can predict the differences in the effectiveness of workflows for different user segments. The output includes a performance evaluation result that includes overall conversion rate predictions, identification of key drop-off points, and optimization suggestions.

7. A communication workflow performance prediction system based on user digital twins, characterized in that, include: The behavior pattern extraction module is used to de-identify historical communication interaction data. By constructing hierarchical behavior patterns of basic response layer, interaction pattern layer and decision tendency layer, user behavior is clustered using composite distance metric to generate a reusable standardized behavior pattern template library. The virtual user generation module includes a template selection unit, a parameterized instantiation unit, and an agent construction unit. The parameterized instantiation unit is used to select a combination of behavioral templates based on the characteristic distribution of the target user group and introduce controlled random perturbations to generate individual differences. The agent construction unit configures a probabilistic state machine decision engine containing a perception module, an evaluation module, and an execution module for each virtual user, as well as a memory mechanism for maintaining short-term interaction history and long-term cumulative state. The workflow mapping module includes a node parsing unit, a semantic analysis unit, a scenario construction unit, and an interface protocol unit. The node parsing unit is used to identify communication nodes, decision nodes, waiting nodes, and multi-channel business node types. The semantic analysis unit is used to extract communication intent, value proposition elements, and sentiment features. The scenario construction unit is used to construct a standardized interactive scenario data structure that includes scenario identifiers, channel types, content feature vectors, time context, and expected response options. The interface protocol unit is used to design a unified decision interface protocol and establish dynamic context management. The simulation execution module, including the interaction engine, state manager, and trajectory recorder, is used to perform real-time interaction simulation between virtual user groups and communication workflows, handle cross-channel information consistency checks and channel switching scenario triggers, and record complete user response behaviors and decision paths. The performance prediction module includes a trajectory aggregation unit, a conversion analysis unit, a churn diagnosis unit, and a group comparison unit. The conversion analysis unit is used to calculate the conversion rate of each communication node and identify high-frequency churn paths. The churn diagnosis unit is used to analyze the churn reasons of specific nodes. The group comparison unit is used to predict the performance differences of the workflow for different user groups and outputs a performance evaluation result that includes the overall conversion rate prediction value, key churn node identification, and optimization suggestions.

Images