An activity strategy generation method and device, a storage medium, and an electronic device

Abstract

The specification discloses an activity strategy generation method and device, a storage medium and an electronic device, wherein the method comprises: obtaining a service activity strategy recommendation description input by a server to a strategy processing large model for a platform service promotion activity; determining a target activity scene and an activity strategy generation intention based on the service activity strategy recommendation description and using the strategy processing large model; calling a service strategy knowledge base to generate an activity strategy logic structure based on the target activity scene and the activity strategy generation intention; generating activity strategy canvas configuration data that meets a preset strategy flow arrangement standard based on the strategy logic structure; and using the activity strategy canvas configuration data to perform strategy flow rendering in a strategy canvas management engine to obtain target activity strategy data for the platform service promotion activity, so that a high-quality activity strategy can be generated for the platform service promotion activity.

Description

Technical Field

[0001] This specification relates to the field of computer technology, and in particular to an activity strategy generation method, apparatus, storage medium, and electronic device. Background Technology

[0002] In related technologies, service platforms provide services such as shopping, finance, express delivery, and insurance. To offer a better service experience to platform users, promotional activities are typically developed. For example, in e-commerce, consumer finance, payment, and membership service scenarios, these promotions usually revolve around elements such as discounts, coupons, credit limits, points, or product bundles, aiming to guide users in their decision-making process regarding whether to participate in the service activities. To ensure consistent user experience and the effectiveness of promotional activities, platform service operations personnel typically need to develop and deploy complex activity strategies based on the service's development stage and customer base. Traditional strategy generation and management models heavily rely on expert experience, usually involving operations personnel manually analyzing historical data and configuring specific activity strategies within the platform service to achieve service promotion goals and provide a better user experience. Summary of the Invention

[0003] This specification provides an activity strategy generation method, apparatus, storage medium, and electronic device, the technical solution of which is as follows:

[0004] Firstly, this specification provides an activity strategy generation method applied to a service platform, the method comprising:

[0005] Obtain the service activity strategy recommendation description input from the server to the strategy processing model for platform service promotion activities;

[0006] Based on the service activity strategy recommendation description, the strategy processing big model is used to determine the target activity scenario and activity strategy generation intent, and based on the target activity scenario and activity strategy generation intent, the service strategy knowledge base is called to generate the activity strategy logical structure, and based on the strategy logical structure, activity strategy canvas configuration data that conforms to the preset strategy process orchestration standard is generated.

[0007] The activity strategy canvas configuration data is used in the strategy canvas management engine to render the strategy process and obtain the target activity strategy data for the promotion of the platform service.

[0008] Secondly, this specification provides an activity strategy generation apparatus, the apparatus comprising:

[0009] The description input module is used to obtain the service activity strategy recommendation description input by the server to the strategy processing big model for platform service promotion activities;

[0010] The strategy processing module is used to determine the target activity scenario and activity strategy generation intent based on the service activity strategy recommendation description and the strategy processing big model, and to call the service strategy knowledge base to generate the activity strategy logical structure based on the target activity scenario and activity strategy generation intent, and to generate activity strategy canvas configuration data that conforms to the preset strategy process orchestration standard based on the strategy logical structure.

[0011] The strategy display module is used in the strategy canvas management engine to render the strategy process using the activity strategy canvas configuration data to obtain the target activity strategy data for the promotion activity of the platform service.

[0012] Thirdly, this specification provides a computer storage medium storing at least one instruction adapted for loading by a processor and executing method steps of one or more embodiments of this specification.

[0013] Fourthly, this specification provides a computer program product storing at least one instruction adapted to be loaded by a processor and to execute the method steps of one or more embodiments of this specification.

[0014] Fifthly, this specification provides an electronic device that may include: a processor and a memory; wherein the memory stores a computer program adapted to be loaded by the processor and to execute the method steps of one or more embodiments of this specification.

[0015] The beneficial effects of the technical solutions provided in some embodiments of this specification include at least the following:

[0016] In one or more embodiments of this specification, by obtaining the strategy recommendation description input from the server, the target activity scenario and generation intent are analyzed using a large strategy processing model, and activity strategy canvas configuration data conforming to preset orchestration standards is generated in conjunction with the service strategy knowledge base. Finally, the process rendering is completed in the strategy canvas management engine to obtain the target activity strategy data. This approach achieves end-to-end transformation from natural language requirements to standardized strategy assets, generating high-quality activity strategies for platform service promotion activities. On the one hand, it simplifies the complex strategy configuration work into dialogue interaction, significantly reducing the technical threshold for operators and improving configuration efficiency, allowing operations to focus more on high-value analysis. On the other hand, through the knowledge base-based generation mechanism, implicit expert experience is digitally modeled and standardized for reuse, effectively breaking down the experience barriers caused by the high dependence on personal experience in the traditional model, and avoiding service risk fluctuations caused by personnel turnover. At the same time, it transforms abstract strategy logic into a visual canvas process, realizing white-box management of strategy generation and enhancing the interpretability and controllability of strategy logic. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this specification or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of a scenario for an activity strategy generation system provided in this manual;

[0019] Figure 2 This is a flowchart illustrating the generation of an activity strategy logic structure as described in this manual;

[0020] Figure 3 This is a flowchart illustrating another method for generating activity strategies provided in this manual;

[0021] Figure 4 This is a schematic diagram of the structure of an activity strategy generation device provided in this specification;

[0022] Figure 5 This is a schematic diagram of the structure of an electronic device provided in this specification. Detailed Implementation

[0023] The technical solutions in this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this specification.

[0024] In the description of this specification, it should be understood that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. In the description of this specification, it should be noted that, unless otherwise expressly specified and limited, "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices. Those skilled in the art can understand the specific meaning of the above terms in this specification based on the specific circumstances. Furthermore, in the description of this specification, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship.

[0025] In related technologies, the traditional "human-based" model has significant technical limitations in practical applications. Expert experience is often "black box" in nature, making it difficult to digitally model and standardize its transmission. Once personnel turnover occurs, the accumulated strategic experience is easily lost, affecting the stability and continuity of service activity strategy management and resulting in poor user service experience.

[0026] The present specification will now be described in detail with reference to specific embodiments.

[0027] Please see Figure 1 This is a schematic diagram of a scenario for an activity strategy generation system provided in this specification. Figure 1 As shown, the activity policy generation system may include at least a client cluster and a service platform 100.

[0028] The client cluster may include at least one client, such as Figure 1 As shown, it specifically includes client 1 corresponding to user 1, client 2 corresponding to user 2, ..., client n corresponding to user n, where n is an integer greater than 0.

[0029] Each client in a client cluster can be an electronic device with communication capabilities, including but not limited to: wearable devices, handheld devices, personal computers, tablets, in-vehicle devices, smartphones, computing devices, or other processing devices connected to a wireless modem. Electronic devices may have different names in different networks, such as: user equipment, access terminal, user unit, user station, mobile station, mobile station, remote station, remote terminal, mobile device, user terminal, electronic device, wireless communication device, user agent or user device, cellular phone, cordless phone, personal digital assistant (PDA), and electronic devices in 5G networks or future evolved networks.

[0030] The service platform 100 can be a standalone server device, such as a rack-mount, blade, tower, or cabinet-type server device, or a workstation, mainframe, or other hardware device with strong computing power; or it can be a server cluster composed of multiple servers. The servers in the service cluster can be composed in a symmetrical manner, wherein each server is functionally and hierarchically equivalent in the transaction chain, and each server can provide services independently. The independent provision of services can be understood as not requiring the assistance of other servers.

[0031] In one or more embodiments of this specification, the service platform 100 can establish a communication connection with at least one client in the client cluster, and complete the data interaction during the activity strategy generation process based on the communication connection, such as online transaction data interaction. For example, the service platform 100 can recommend content to the client based on the target neural network model obtained by the activity strategy generation method of this specification; or the service platform 100 can obtain training data from the client, such as the first training data.

[0032] It should be noted that the service platform 100 establishes a communication connection with at least one client in the client cluster for interactive communication via a network. This network can be a wireless network or a wired network. Wireless networks include, but are not limited to, cellular networks, wireless LANs, infrared networks, or Bluetooth networks. Wired networks include, but are not limited to, Ethernet, Universal Serial Bus (USB), or Controller Area Network (CAN). In one or more embodiments of the specification, technologies and / or formats including HyperText Markup Language (HTML), Extensible Markup Language (XML), etc., are used to represent data exchanged over the network (such as target compressed packets). Furthermore, conventional encryption technologies such as Secure Socket Layer (SSL), Transport Layer Security (TLS), Virtual Private Network (VPN), and Internet Protocol Security (IPsec) can be used to encrypt all or some links. In other embodiments, customized and / or dedicated data communication technologies can be used to replace or supplement the aforementioned data communication technologies.

[0033] The activity policy generation system embodiments provided in this specification and the activity policy generation methods described in one or more embodiments belong to the same concept. The execution entity corresponding to the activity policy generation method in one or more embodiments of this specification can be the aforementioned service platform 100; the execution entity corresponding to the activity policy generation method in one or more embodiments of this specification can also be the electronic device corresponding to the client, specifically determined based on the actual application environment. The implementation process of the activity policy generation system embodiments can be detailed in the following method embodiments, and will not be repeated here.

[0034] based on Figure 1 The following is a detailed description of the activity strategy generation method provided by one or more embodiments of this specification, as illustrated in the scene diagram.

[0035] Please see Figure 2 This document provides a flowchart illustrating an activity strategy generation method according to one or more embodiments. This method can be implemented using a computer program and can run on an activity strategy generation device based on the von Neumann architecture. The computer program can be integrated into an application or run as a standalone utility application. The activity strategy generation device can be a service platform.

[0036] Specifically, the method for generating this activity strategy includes:

[0037] S102: Obtain the service activity strategy recommendation description input by the server to the strategy processing model for the platform service promotion activities;

[0038] The strategy processing large model refers to a multimodal large language model pre-trained on a massive general corpus based on a machine learning model architecture, or a large language model fine-tuned using data from specific vertical service scenarios (such as financial marketing and e-commerce operations) based on an existing multimodal large language model. This model has the capabilities of natural language understanding, intent recognition, logical reasoning, and code / structured data generation, and can understand fuzzy instructions from the server and transform them into precise transaction logic.

[0039] Platform service promotion activities refer to platform service activities conducted on a service platform, aimed at improving aspects such as user activity, conversion rate, retention rate, and user experience. Examples include credit limit increases in financial services, coupon distribution on e-commerce platforms, and re-engagement push notifications on content platforms.

[0040] The service activity strategy recommendation description refers to the raw information carrier directly input by the server into the large model to trigger strategy generation. Its form can be natural language text (Prompt), a structured parameter list, or speech-transcribed text. Its content, depending on the specific circumstances, can include the server-side operations personnel's expectations, constraints, or descriptions of the target object for the objective activity.

[0041] The service platform provides an intelligent dialogue interface to the server (such as a PC-based management backend or a mobile app). This interface is configured with text input boxes or voice capture controls, serving as the entry point for strategy recommendations.

[0042] The service platform receives natural language text (e.g., "Generate a coupon distribution strategy for users whose activity has decreased this week") input by the user through the server in the interactive interface.

[0043] The service platform preprocesses the received natural language text, including removing invalid characters, filtering sensitive words, and performing security checks. Then, the service platform concatenates the cleaned text with the current transaction context (such as current time and operator ID) to encapsulate it into a prompt that conforms to the input specifications of the large-scale model for policy processing, forming a service activity policy recommendation description.

[0044] S104: Based on the service activity strategy recommendation description, the strategy processing big model is used to determine the target activity scenario and activity strategy generation intention, and based on the target activity scenario and activity strategy generation intention, the service strategy knowledge base is called to generate the activity strategy logical structure, and the activity strategy canvas configuration data that conforms to the preset strategy process orchestration standard is generated based on the strategy logical structure.

[0045] The target activity scenario refers to the specific transaction context environment served by the strategy, including but not limited to transaction lines (such as reserve funds, consumer loans), target customer characteristics (such as dormant users, highly active users), and marketing nodes (such as Double Eleven, end-of-month sales push).

[0046] Activity strategy generation intent refers to the type of operation that users expect to achieve by describing it, such as "create a new strategy", "modify parameters on an existing strategy", "query historical preferred strategies", or "conduct an A / B experiment configuration".

[0047] The activity strategy logic structure refers to an intermediate representation between natural language and machine code. In this specification, it is represented as a directed acyclic graph (DAG) containing nodes (strategy atoms) and edges (flow conditions), such as for abstractly describing the execution flow of the strategy, without being bound to specific rendering coordinates.

[0048] Preset strategy process orchestration standards: These refer to the standardized protocol formats that downstream strategy engines can recognize and execute, such as preset standard XML files, XPDL files, or specific JSON graph definition specifications.

[0049] In one feasible implementation, the strategy can be instructed to handle the generation pattern of large models based on atomized graph construction and vector retrieval. This approach leverages the reasoning capabilities of large models and the retrieval capabilities of vector databases to achieve dynamic strategy assembly. Specifically:

[0050] The model receives service activity strategy recommendation descriptions through a large-scale strategy processing system. Using Named Entity Recognition (NER) and semantic analysis techniques, it deconstructs the "target activity scenario" (e.g., reserve fund increase) and the "activity strategy generation intent" (e.g., creating a new strategy). Simultaneously, the model extracts key constraints (e.g., risk level). <Lv3)。

[0051] By using the deconstructed scenarios and constraints as query vectors in the large model through strategy processing, a search is performed in the service strategy knowledge base (such as a knowledge base built on Lindorm / ODPS) to recall a number of strategy atoms with a matching degree greater than the matching threshold, such as "risk control access verification atom", "rights distribution atom", and "SMS delivery atom".

[0052] By leveraging a large-scale strategy processing model based on a pre-defined thought chain, the causal connections between the strategy atoms of the recall prediction are inferred, thereby constructing the logical structure of the activity strategy. For example, the model infers that "risk control verification" must precede "rights distribution".

[0053] The generated activity strategy logical structure is then mapped to activity strategy canvas configuration data that conforms to a preset strategy flow orchestration standard (such as BPMN 2.0). This process includes assigning a unique ID to each logical node, defining input and output parameters, and generating layout coordinate information for front-end display.

[0054] In one feasible implementation, a strategy can be instructed to handle large models based on template slot filling and rule engine generation patterns. This approach is suitable for scenarios where transaction rules are relatively fixed and stability requirements are extremely high. Specifically:

[0055] The input descriptions are categorized based on the policy processing model to identify specific transaction scenario IDs. Based on this scenario ID, a pre-defined policy skeleton template is retrieved from the service policy knowledge base. This template already contains a fixed process structure (e.g., start – audience screening – action – end), but the key parameters are empty (i.e., slots).

[0056] Based on the strategy processing big model, specific strategy configuration parameters (such as "disbursement amount = 50 yuan" and "target audience = active in the last 30 days") are extracted from the service activity strategy recommendation description and the historical activity strategy data of the service strategy knowledge base, and these parameters are injected into the corresponding slots of the template to form an instantiated activity strategy logical structure.

[0057] Then, the pre-built rule engine is invoked to validate the instantiated logical structure (e.g., checking if the amount disbursed exceeds the system limit). If the validation fails, automatic correction or error reporting is performed. The validated logical structure is serialized into activity strategy canvas configuration data in a pre-defined format such as JSON or XML, ensuring that its data fields are consistent with the interface definition of the downstream strategy canvas management engine.

[0058] Through the above implementation method, step S104 realizes the intelligent transformation from "unstructured requirements" to "standardized system configuration", which solves the technical problems of low efficiency and error-proneness of traditional manual configuration strategies.

[0059] S106: In the strategy canvas management engine, the activity strategy canvas configuration data is used to render the strategy process to obtain the target activity strategy data for the promotion activity of the platform service.

[0060] A strategy canvas management engine refers to a software system that integrates visual rendering components and a logic parsing kernel. It typically includes a front-end graphical editor and a back-end process definition management service. Its function is to transform abstract configuration code into a user-understandable graphical interface and to handle the storage, verification, and release of strategy versions.

[0061] Strategy flow rendering refers to the process of parsing strategy configuration data and mapping its nodes, connections, and attributes to graphical elements on the screen (such as rectangles, rhombuses, and arrows). This process includes not only visual rendering but also logical instantiation, that is, building an editable strategy object model in memory.

[0062] Target activity strategy data refers to the strategy entity data that has been rendered, manually verified, or automatically corrected and is finally generated, possessing the ability to be executed online. This data typically includes a unique strategy ID, version number, a complete XML description file, and related metadata (such as creator and effective time), and can be directly called by downstream service execution systems.

[0063] One feasible implementation method is to use a human-computer collaborative visualization rendering and interaction mode (standard mode). This approach is suitable for scenarios requiring operational personnel intervention for secondary confirmation or fine-tuning, embodying the combination of "expert experience" and "AI generation." Specifically:

[0064] The control policy canvas management engine receives active policy canvas configuration data (such as JSON or XML format) from S104. The engine's internal parser traverses this data, constructs the policy object model in memory, and establishes the topological relationship between node objects and connection objects.

[0065] The front-end rendering engine (e.g., based on X6, LogicFlow, or BPMN.js) reads the aforementioned object model. Based on a pre-defined stylesheet, it maps different types of strategy atoms (such as "decision nodes" and "action nodes") to specific icons, draws arrowed connections according to the topological relationships, and automatically calculates layout coordinates, transforming the active strategy canvas configuration data into a visual flowchart on the user interface. At this point, the strategy state changes from "canvas not rendered" to "canvas rendered."

[0066] In addition, users can interact with the system. Specifically, in response to user commands on the strategy canvas (such as dragging node positions, modifying node parameters, and adding branch links), the engine updates the object model in memory in real time, which allows server-side operators to use their own experience to correct the AI-generated strategies.

[0067] Furthermore, when the user clicks the "Save" or "Publish" command, the engine performs an integrity check on the current policy object model. If the check passes, it is serialized into a final binary or XML file, generating a unique policy ID and version number, which is then stored in the database to form the target activity policy data.

[0068] In one feasible implementation, an automated rendering and deployment mode based on simulation pre-visualization (fully automatic mode) can be used. This approach is suitable for high-frequency, real-time policy generation scenarios, focusing on logic verification and effect prediction. Specifically:

[0069] The control strategy canvas management engine loads the activity strategy canvas configuration data in the background. The engine does not draw the graphical interface, but performs "rendering" at the logical level, that is, checking the connectivity of the graph to ensure that there are no isolated nodes, infinite loops or unreachable paths.

[0070] Then, a simulation is performed. The engine calls the simulation component to "virtually run" the strategy process based on historical user data. At this time, tens of thousands of users are simulated to enter the system, and the traffic distribution of each branch and the expected budget consumption are statistically analyzed. If the simulation results show abnormalities (such as the budget being exhausted instantly), rendering is marked as failed and an alarm is triggered.

[0071] After the A / B test configuration is bound and the specific verification and simulation are passed, the engine automatically allocates a traffic bucket for the strategy and configures the A / B test parameters (such as the proportion of the control group). The system packages the strategy logic and the test configuration to obtain the strategy flow data.

[0072] Then, executable code is generated. The engine compiles the above-mentioned strategy process data into a set of rules or script code that the service platform can directly recognize, which serves as the target activity strategy data and is pushed to the execution queue in the production environment.

[0073] Through the above implementation method, step S106 realizes the transformation from abstract configuration data to concrete transaction strategy, ensuring that the AI-generated strategy can be safely, accurately, and visually implemented in the production environment.

[0074] In one feasible implementation, the activity strategy generation method further includes:

[0075] Step A2: After the strategy configuration data is deployed and executed in the platform service promotion activity, collect the post-implementation effect data of the strategy operation;

[0076] Post-implementation performance data refers to objective service feedback metrics collected by the system after the strategy is deployed in the production environment and actually reaches users. These include, but are not limited to, process metrics (such as message delivery rate, click-through rate (CTR), and page dwell time) and outcome metrics (such as conversion rate (CVR), credit approval rate, redemption amount, and delinquency rate).

[0077] As an illustration, an instant feedback acquisition method based on real-time stream computing can be adopted, suitable for short-term, high-concurrency strategies such as flash sales or event-triggered strategies, enabling rapid anomaly detection. Specifically, data points are pre-installed at the policy execution touchpoints (such as the app and SMS gateway). When a user performs an action (browse, click, apply), a log report is triggered. A stream computing engine (such as Flink) is used to subscribe to the log message queue. The system uses the unique strategy ID (Strategy_ID) in the target activity strategy data as the key to perform window aggregation on the real-time incoming logs to obtain a comprehensive log. Then, based on the comprehensive log, the cumulative exposure, clicks, and real-time conversion rate corresponding to the strategy ID are calculated in real time, generating post-effect data such as millisecond-level or second-level data.

[0078] Step A4: Perform strategy feedback effect analysis on the posterior effect data and the target activity strategy data to obtain strategy execution analysis data, and then feed the strategy execution analysis data back to the service strategy knowledge base for strategy knowledge update processing.

[0079] Strategy execution analysis data refers to in-depth insight data generated by comparing posterior effect data with preset strategy objectives (KPIs) and performing attribution analysis. Optionally, this data not only includes the conclusion of "good or bad results", but also the attribution logic of "why good / bad" (e.g., "the conversion rate for the 25-30 age group is significantly higher than the predicted value").

[0080] Strategy knowledge update processing refers to the process of accumulating the parsed experience into a knowledge base. Optional steps include updating feature weights in the vector database, adding new optimal strategy cases, and labeling inefficient strategies as "negative samples" to influence the subsequent inference and retrieval results of the large model.

[0081] Indicatively, a dynamic incremental update based on a vectorized service strategy knowledge base can be used to expand the knowledge base. Specifically, the difference between the "expected goal" generated in stage S104 and the "actual posterior data" obtained in step A2 is calculated to obtain the actual effect. If the actual effect is better than the expected indicator (e.g., actual conversion rate > expected conversion rate by 20%), it is judged as a "positive high-quality case"; otherwise, it is a "negative case". The complete logical structure of the strategy, the characteristics of the applicable audience, and the posterior effect data of this run are encapsulated into a new strategy knowledge object. The embedded model is called to vectorize the strategy knowledge object to generate a new feature vector. This feature vector is inserted into the vector index of the service strategy knowledge base. When the large model encounters a similar scenario next time, the large model will prioritize recalling cases through retrieval, thereby improving the accuracy of the generated strategy.

[0082] In one or more embodiments of this specification, by obtaining the strategy recommendation description input from the server, the target activity scenario and generation intent are analyzed using a large strategy processing model, and activity strategy canvas configuration data conforming to preset orchestration standards is generated in conjunction with the service strategy knowledge base. Finally, the process rendering is completed in the strategy canvas management engine to obtain the target activity strategy data. This approach achieves end-to-end transformation from natural language requirements to standardized strategy assets, generating high-quality activity strategies for platform service promotion activities. On the one hand, it simplifies the complex strategy configuration work into dialogue interaction, significantly reducing the technical threshold for operators and improving configuration efficiency, allowing operations to focus more on high-value analysis. On the other hand, through the knowledge base-based generation mechanism, implicit expert experience is digitally modeled and standardized for reuse, effectively breaking down the experience barriers caused by the high dependence on personal experience in the traditional model and avoiding transaction fluctuations caused by personnel turnover. At the same time, it transforms abstract strategy logic into a visual canvas process, realizing white-box management of strategy generation and enhancing the interpretability and controllability of strategy logic.

[0083] In one feasible implementation, the activity strategy generation method includes:

[0084] S202: Obtain the service activity strategy recommendation description input by the server to the strategy processing model for the platform service promotion activities;

[0085] S204: Input the service activity strategy recommendation description into the intent recognition agent to call the strategy processing big model to determine the activity strategy and generate intent;

[0086] An intent-recognition agent can be a software entity implemented based on a policy-processing large model, with pre-set intent-recognition prompt word templates and classification logic. Its core function is to focus on analyzing the user's natural language input, mapping it to categories of executable operation instructions (such as "create", "modify", "query", "pause"), and extracting key operation objects.

[0087] S206: Input the activity strategy generation intent and the service activity strategy recommendation description into the scene processing agent to call the strategy processing big model to determine the target activity scene;

[0088] A scenario processing agent is a software entity responsible for constructing a scenario context based on a large model of policy processing. It can combine user input descriptions and identified intents to supplement missing background information (such as product line type, target customer characteristics, and current marketing stage), thereby defining a specific and clearly defined target activity scenario.

[0089] S208: Input the service activity strategy recommendation description, the activity strategy generation intent, and the target activity scenario into the strategy processing agent to call the strategy processing big model to generate the activity strategy logical structure, and generate activity strategy canvas configuration data that conforms to the preset strategy process orchestration standard based on the strategy logical structure.

[0090] The policy processing agent refers to the software entity responsible for generating the core logic. It can be viewed as a "chief engineer," receiving structured information from upstream agents, invoking the policy processing model for reasoning, planning the steps, conditions, and actions of the policy, and transforming it into standardized configuration data.

[0091] In this specification, through the above embodiments, those skilled in the art will understand that by introducing multiple dedicated intelligent agents to call strategies to process large models and to cooperate through division of labor, the reasoning process of large models can be controlled more precisely, avoiding the problems of forgotten instructions or logical confusion when processing complex tasks by a single agent.

[0092] In one feasible implementation, the activity policy generation method further includes the following steps before performing the step of inputting the service activity policy recommendation description into the intent recognition agent to invoke the policy processing big model to determine the activity policy generation intent:

[0093] S210: Input the service activity strategy recommendation description into the security assessment agent to call the policy processing big model to perform security compliance detection on the service activity strategy recommendation description; if the detection passes, then execute the step of inputting the service activity strategy recommendation description into the intent recognition agent to call the policy processing big model to determine the intention of the activity strategy generation; if the detection fails, then intercept the service activity strategy recommendation description and output a preset rejection response.

[0094] A security assessment agent is a software entity specifically responsible for risk control of input / output content. It typically integrates a composite filtering engine consisting of a sensitive word database, a regular expression rule set, and large-scale security detection capabilities. Its function is to identify malicious instructions (such as prompt injection attacks), illegal transaction requests (such as deceptive inducements and false advertising), and sensitive information (such as user privacy data).

[0095] Security compliance testing refers to a multi-dimensional review process of input text. Review dimensions include, but are not limited to: content security, service compliance, data privacy, and instruction security.

[0096] A pre-set rejection response is a standardized fallback response returned to the user when a check fails. This response typically politely refuses to execute the instruction and prompts the user to modify their input (e.g., "Sorry, your input contains sensitive information or does not conform to transaction specifications, so we cannot generate a strategy for you.").

[0097] Indicatively, a discriminative detection method based on cue word engineering can be used, leveraging the semantic understanding capabilities of the large model itself to act as a "judge".

[0098] Prompt Construction: The security assessment agent loads a pre-built security detection prompt template (SystemPrompt). This template contains specific compliance standards (e.g., "You are a financial risk control expert. Please determine whether the user input contains fraudulent, loan inducement, or offensive language. If it does, output 'Block'; otherwise, output 'Pass'").

[0099] The model invocation process of the security assessment agent is as follows: the service activity strategy recommendation description input by the user is embedded into the above prompt word template, and the overall input is used to invoke the strategy processing model.

[0100] Result determination: The large model outputs a discrimination result (such as Block or Pass). The agent parses this result. If it is Pass, the detection is considered successful, and the original description is passed to the downstream intent recognition agent; if it is Block, the detection is considered unsuccessful, the agent interrupts the subsequent link, and directly reads the preset rejection message from the configuration center and returns it to the front end.

[0101] Through the description of the above embodiments, those skilled in the art can understand that step S210, by introducing intelligent security detection at the source of the process, effectively prevents insecure data from entering the large model inference chain, avoids the generation of illegal policies, and ensures the security and compliance of the service system.

[0102] Optionally, after executing the invocation of the strategy to process the large model generating activity strategy logic structure, the method further includes:

[0103] S212: The security assessment agent calls the policy processing big model to perform output security risk verification on the policy logic structure; if the detection passes, the step of generating activity policy canvas configuration data that conforms to the preset policy process orchestration standard based on the policy logic structure is executed; if the detection fails, the policy logic structure is intercepted and a preset rejection response is output.

[0104] As an example, a verification model based on static logic auditing can be adopted. This approach treats the large strategy processing model as a "code reviewer," performing static analysis on the generated logical structure, as follows:

[0105] The security assessment agent serializes the policy logic structure generated by S208 into a highly readable text description or pseudocode. The agent constructs system security review prompts containing a risk control checklist, such as: "Please check if the following policy logic has the following risks: 1. Infinite loop; 2. No inventory verification for rights distribution; 3. Lack of risk control interception nodes. If they exist, please output the risk points."

[0106] The system calls the policy processing model to perform a security assessment based on the above system security review prompts. If the model outputs "no risk" or the risk score is lower than the preset threshold, the detection is deemed to have passed, and the following steps are allowed to be executed: generating activity policy canvas configuration data that conforms to the preset policy process orchestration standard based on the policy logic structure. That is, the system continues to perform format conversion to generate canvas configuration data.

[0107] If the model identifies a specific risk (such as "Error: Missing inventory check before distribution node"), it determines that the detection has failed, the agent discards the logic structure, and provides feedback to the user on the specific reasons for the risk.

[0108] Through the above embodiments, those skilled in the art can understand that step S212 effectively constructs the last line of defense for strategy production, avoiding platform financial losses or compliance incidents caused by occasional logical illusions in large models.

[0109] Optional, please see Figure 3 , Figure 3 This is a flowchart illustrating the generation of an activity strategy logical structure according to one or more embodiments of this specification. Specifically, the process of generating the activity strategy logical structure by invoking the service strategy knowledge base based on the target activity scenario and the activity strategy generation intention can be referred to as follows:

[0110] S302: The large model is processed by the strategy to construct a generation task context based on the target activity scenario and the activity strategy generation intent. The generation task context is deconstructed into multi-dimensional constraint objectives. The multi-dimensional constraint objectives include at least target user group characteristic constraints, platform service key evaluation index constraints, and activity promotion cost constraints.

[0111] The generated task context refers to a structured data object or long text prompt that gathers all the necessary prior information during the execution of the strategy generation task.

[0112] A multidimensional constraint objective refers to transforming a transaction requirement described in natural language into a set of constraints that can be processed by a computer. This set must contain at least a three-dimensional vector or parameters:

[0113] Target user group characteristic constraints: Define the logical boundaries of the target group for the strategy (e.g., age range, user level).

[0114] Platform service key evaluation indicators constraints: the core business performance indicators or optimization directions pursued by strategy execution (such as conversion rate CVR, daily active users DAU).

[0115] Promotion cost constraints: The upper limit of resources allowed to be consumed in the execution of the strategy (e.g., subsidy amount per user, total budget pool).

[0116] Optionally, an explicit semantic slot extraction mode based on cue word engineering can be adopted, specifically:

[0117] The control strategy processing model first concatenates the target activity scenario, activity strategy generation intent, and original user description text determined by the upstream steps. Simultaneously, the system retrieves current transaction environment data (such as the current quarterly budget balance). The server uses pre-defined prompt word templates to assemble the above information (concatenated information and transaction environment information) into a generation task context containing complete instructions. Based on the generation task context, the strategy processing model utilizes Named Entity Recognition (NER) and semantic dependency analysis capabilities to identify key entities in the context and populate them into the constraint slots of predefined multidimensional constraint targets. The populated information is then converted into a standard JSON format object, which is output as the multidimensional constraint target.

[0118] Optionally, an implicit constraint derivation model based on historical data regression can be adopted, specifically:

[0119] The control strategy processing model maps the target activity scenario and the activity strategy generation intent into a high-dimensional vector, and retrieves the top-N strategy cases with the highest historical similarity from the service strategy knowledge base. The parameter distribution of these historical strategy cases is used as reference information and injected into the generation task context to achieve context enhancement. Then, the strategy processing model parses the constraint setting patterns in the historical cases to obtain historical implicit distribution features. For example, the model finds that for the "dormant user" scenario, historical strategies typically set the "cost per person" between 5 and 10 yuan, and the KPI is usually "recall rate".

[0120] Then, constraint completion is achieved through joint inference based on the user's explicit input description and historical implicit distribution. For example, for constraints explicitly specified by the user (such as "cost < 20 yuan"), the user's value is directly adopted. For constraints not mentioned by the user (such as no KPI mentioned), the large model automatically derives a recommended value (such as KPI = maximize recall) based on historical distribution, thereby completing the missing dimensions.

[0121] Finally, the explicitly extracted descriptive constraints and the implicitly derived implicit constraints are merged and deconstructed into a complete multidimensional constraint objective.

[0122] Through the above implementation method, this step transforms fuzzy natural language requirements into precise quantitative control indicators, providing a foundation for the subsequent generation of strategy logic that conforms to the boundaries of service activities.

[0123] S304: Based on the multi-dimensional constraint target, the service strategy knowledge base is invoked to perform strategy topology reasoning to obtain the activity strategy logical topology graph, and the activity strategy logical structure is generated based on the activity strategy logical topology graph.

[0124] An activity strategy logical topology graph refers to an abstract mathematical model used to describe the execution flow within a strategy, typically represented as a directed acyclic graph (DAG). In this graph, nodes represent specific atomic components of the strategy (such as "decision nodes" and "action nodes"), and edges represent causal relationships or temporal dependencies in the flow of transactions. This graph focuses on expressing the logical connectivity between nodes and has not yet been bound to specific data formats or rendered coordinates.

[0125] Policy topology reasoning refers to the process by which a policy processing model, based on given constraint boundaries, searches for the optimal path in the potential policy solution space and predicts the connectivity relationships between functional nodes. This process is similar to "building blocks," that is, deciding which building blocks (atoms) to use and how to assemble them (connectivity relationships).

[0126] The activity strategy logical structure refers to the machine-readable data object (such as a JSON object or XML tree) generated after serializing and instantiating the above topology map. It not only contains the structural information of the map, but also fills in specific parameter values, and serves as a preliminary intermediate representation (IR) for generating the final canvas configuration data.

[0127] Optionally, a generation model based on "atomic retrieval + causal chain prediction" can be adopted, specifically:

[0128] First, atomic component recall is performed, and the control strategy processes the multi-dimensional constraint objectives of the upstream output of the large model parsing. Based on the activity promotion cost constraint (e.g., "cost per person < 5 yuan") and the platform service key evaluation indicator constraint (e.g., "high response rate"), vector retrieval is performed in the strategy atomic library of the service strategy knowledge base to recall a set of candidate atoms that meet the constraints. For example, the high-cost "cash distribution atom" is excluded, while the "phone bill voucher atom" and "benefit card atom" are retained.

[0129] Then, the thought chain topology reasoning is performed. The control strategy processes the large model based on the above-recalled candidate atom list and target user group feature constraints, and uses the thought chain capability to perform reasoning to obtain the reasoning result:

[0130] For example, the thought process is as follows: "For high-risk users (feature constraint), risk control verification must be performed first (atom A); after verification, in order to improve the response rate (KPI constraint), benefits should be issued immediately (atom B); after the benefits are issued, the user should be notified (atom C)."

[0131] Next, a graph is constructed. The control strategy processing large model determines the adjacency relationship between each atom based on the reasoning results, and then constructs an activity strategy logic topology graph based on the atoms and the relationships between them.

[0132] Finally, the control strategy processes the large model to perform integrity checks on the generated graph (such as checking for the existence of isolated nodes). After passing the checks, it is serialized into a standard activity strategy logic structure.

[0133] Optionally, an assembly mode based on "logical subgraph matching + fusion" can be adopted, specifically:

[0134] First, subgraph matching is performed. The control strategy processing model decomposes the multidimensional constraint target into multiple sub-dimensions and retrieves the corresponding standard logical subgraphs in the service strategy knowledge base.

[0135] For example, based on the characteristic constraints of the target user group (such as new users), a new user access verification subgraph (including real-name authentication - new user judgment) is retrieved.

[0136] For example, based on the constraints of key evaluation indicators for platform services (such as maximizing conversion), a multi-wavelength reach sub-map (including SMS - 24-hour wait - Push) was retrieved.

[0137] Next, graph fusion and conflict resolution are performed. The control strategy processing model stitches together the multiple subgraphs retrieved above. During the stitching process, logical conflicts are resolved according to transaction rules (e.g., the risk control verification subgraph should be placed before the reach subgraph), and redundant nodes are merged to form a connected activity strategy logical topology graph.

[0138] Finally, parameter injection and instantiation are performed. The control strategy processing model assigns values ​​to the attributes of action nodes in the graph based on the specific values ​​in the constraints (such as "cost limit of 20 yuan"), and finally generates an activity strategy logic structure containing specific parameters.

[0139] In this specification, a large-scale model is constructed using strategy processing to generate a task context, which is then deconstructed into multi-dimensional constraint objectives that include at least the characteristics of the target user group, key evaluation indicators of platform services, and activity promotion costs. Based on these multi-dimensional constraint objectives, a service strategy knowledge base is invoked to perform strategy topology reasoning, resulting in an activity strategy logical topology graph and generating the activity strategy logical structure. This approach first transforms broad, non-standardized natural language transaction intentions into precise, computable quantifiable constraint boundaries, effectively converging the generation space of the large model. This ensures that the generated strategies are not only semantically fluent but also strictly adhere to rigid transaction constraints (such as budget caps) and clear optimization directions (such as maximizing conversion rates), thereby significantly improving the controllability and transaction compliance of the strategy generation results. Simultaneously, through this constraint-driven logical assembly mechanism, the large model can dynamically deduce the optimal atomic component connection paths in the knowledge base based on current specific constraints, replacing rigid static template filling. This ensures that the generated strategy logic possesses logical connectivity and feasibility while meeting specific transaction indicators, achieving personalized customization and optimal effect in strategy generation.

[0140] Optionally, the process of performing policy topology reasoning based on the multi-dimensional constraint target call service policy knowledge base to obtain the activity policy logical topology graph can be performed in the following manner:

[0141] S402: The multidimensional constraint target is mapped to a constraint target feature vector through the large model of the strategy processing. The constraint target feature vector is used to perform vector matching processing in the pre-set strategy atom library to obtain at least one candidate strategy atom. The strategy atom in the pre-set strategy atom library is a pre-encapsulated minimum strategy implementation functional component.

[0142] The constraint target feature vector refers to the numerical vector generated after encoding the structured multidimensional constraint targets (including user features, KPI indicators, cost limits, etc.) using a strategy to process large models.

[0143] The pre-built policy atom library refers to a collection of functional components that are pre-built and stored in the service policy knowledge base. Each entry in this library is a policy atom, and its corresponding semantic index vector is pre-computed and stored.

[0144] A strategy atom refers to a pre-encapsulated, minimal functional component for implementing a strategy, possessing independent execution logic. It forms the cornerstone of a complete strategy process and cannot be further divided. For example, in a financial services scenario, sending an SMS is an action atom, determining whether a payment is overdue is a condition atom, and calculating the discount amount is a calculation atom.

[0145] Vector matching refers to the process of calculating the similarity (such as cosine similarity or Euclidean distance) between two vectors in a vector space, with the aim of finding the inventory vector that is semantically closest to the target vector.

[0146] As an illustration, a single-search mode using a global weighted fusion approach can be employed. This method is computationally efficient and can quickly recall components that best match the overall strategy configuration intent from a macro-level perspective. Specifically:

[0147] First, feature extraction and encoding are performed. Then, a strategy is used to process the large model by extracting information from each sub-dimension of the multi-dimensional constraint objective (such as "target user = dormant user", "cost = low", "KPI = click-through rate"). The model calls a pre-built embedding algorithm to generate embedding vectors for each sub-dimension.

[0148] Then, a weighted fusion is performed. The large model is processed according to the preset business weight configuration (e.g., cost constraint weight = 0.5, user constraint weight = 0.3, KPI constraint weight = 0.2). The above sub-vectors are weighted and averaged or concatenated to generate a comprehensive constraint target feature vector.

[0149] Then, similarity calculation and recall are performed. The comprehensive vector is used as the query vector, and a full database scan is performed in the pre-set strategy atomic library to calculate its cosine similarity with the atomic index vectors of each strategy.

[0150] Finally, a truncation and filtering process is performed to select the top N atoms with the highest similarity scores (Top-N), or to select all atoms with similarity scores greater than a preset threshold (such as 0.8), forming a candidate strategy atom set.

[0151] As an illustration, a multi-path parallel retrieval mode based on hierarchical deconstruction can be adopted, which has high recall accuracy, avoids the feature sparsity problem caused by the curse of vector dimensionality, and ensures that all types of atoms (judgments, actions, and reach) are covered. Specifically:

[0152] First, dimensional deconstruction and independent mapping are performed. The control strategy processes the large model to deconstruct the multidimensional constraint target into three independent dimensions: admission constraint, action constraint, and reach constraint. Independent constraint target feature vectors are generated for each of these three dimensions (i.e., three different query vectors are generated).

[0153] Then, parallel matching is performed on partitions. The admission constraint vector is used to match within the "conditional judgment partitions" of the policy atom library, recalling candidate atoms such as "new / old user judgment" and "blacklist verification," for example:

[0154] By using action constraint vectors (such as low cost) to match in the "rights action category partition", low cost candidate atoms (such as "interest-free coupons") are recalled, and high cost atoms (such as "cash red envelopes") are filtered out.

[0155] By using reach constraint vectors to match within the channel reach category partitions, candidate atoms such as "SMS sending" or "AppPush" are recalled.

[0156] Finally, the results obtained from the parallel retrieval of the three partitions are merged and deduplicated to obtain the final list of candidate strategy atoms.

[0157] Through the above implementation method, step S402 uses vectorization technology to achieve soft semantic matching of hard business constraints (such as cost and risk), ensuring that the selected building blocks (atoms) meet the building requirements.

[0158] S404: The strategy processes the large model to predict the causal connections between the candidate strategy atoms, constructs a directed acyclic graph for each candidate strategy atom based on the causal connections, and injects execution parameters into the candidate strategy atoms in the directed acyclic graph based on historical strategy data in the service strategy knowledge base, forming an activity strategy logical topology graph.

[0159] Causal connections refer to the logical dependencies or temporal sequences between different policy atoms in the policy execution process. These relationships define the direction of data flow and control flow. For example, the output state (pass / reject) of the risk verification atom determines whether the rights issuance atom is triggered; the former is the causal precursor of the latter.

[0160] A directed acyclic graph (DAG) is a graph-like data structure consisting of policy atoms as nodes and causal connections as directed edges. In this structure, there is no path that starts from any node, traverses several edges, and returns to that node, thus mathematically guaranteeing that the policy execution process will not enter an infinite loop.

[0161] Execution parameters refer to the specific numerical configurations required by a strategy atom at runtime. For example, issuing coupons requires parameters such as face value and validity period; sending SMS requires parameters such as template ID and sending time.

[0162] The activity strategy logic topology map refers to the complete strategy logic model after structural assembly and parameter instantiation. It includes not only the topological structure of the process but also the parameter configuration of the process, and serves as the direct basis for generating the final engine configuration data.

[0163] Optionally, S404 can adopt a structured prediction and statistical parameter injection mode based on thought chain reasoning. This method utilizes statistical laws for parameter configuration, resulting in a highly robust strategy that conforms to the general patterns of historical business. Specifically:

[0164] First, perform causal chain prediction: Based on the candidate strategy atom list, the large model is processed by the strategy described above. Using the chain-of-thought capability, the functional definition and input / output interface constraints of each atom are analyzed to deduce a reasonable business flow sequence.

[0165] Further reasoning examples yield predicted connection relationships, such as: "Atom A is a condition judgment class, and atom B is an action execution class. According to business logic, action B must be executed when condition A is satisfied. Therefore, the predicted connection relationship is: A->(True)->B."

[0166] Then, a Directed Acyclic Graph (DAG) is constructed, a node adjacency matrix is ​​built based on the predicted connectivity, and loop detection is performed. If no loops are detected, an initial DAG structure is generated.

[0167] Finally, historical statistics are injected. For each atomic node in the directed acyclic graph structure, the parameter slot to be filled is identified. Then, using "atomic type + current scenario" as the index, the historical data table in the service strategy knowledge base is queried to calculate the historical average or mode of the parameter in the scenario (for example, the query shows that the historical average coupon amount for "dormant users" is 5 yuan). Then, the statistical value is injected into the corresponding slot to form the final activity strategy logic topology graph.

[0168] Optionally, a sequence-based full-graph construction and optimal parameter mapping mode can be adopted. This approach reuses historically optimized parameters, enabling newly generated strategies to possess high service conversion potential from the outset. Specifically:

[0169] First, a serialized graph is generated. The graph construction task is treated as a sequence generation task by a strategy processing large model, and the domain-specific language (DSL) code describing the graph structure is directly determined (e.g., "Start-->Risk_Check-->|Pass|Coupon_Issue-->End").

[0170] Then, structural parsing and verification are performed. The above DSL code is parsed through a strategy to process the large model, instantiated into a graph object in memory, and the logical connectivity is verified to ensure that all paths are reachable and have an end point.

[0171] Finally, best practice mapping is performed. When injecting parameters into graph objects, instead of using average values, a selection strategy is adopted. This involves retrieving the optimal strategy case from the service strategy knowledge base that has the highest historical performance metrics and is similar to the current graph structure corresponding to the graph object. The specific parameter values ​​of this optimal strategy case (e.g., face value = 8.8 yuan, sending time = 19:00) are directly mapped and injected into the corresponding atoms of the current graph, forming the final activity strategy logical topology graph.

[0172] Through the above implementation method, step S404 realizes the intelligent assembly from scattered components to complete logic, and uses historical data to inject verified execution parameters into the strategy, ensuring the logical rigor and business effectiveness of the generated strategy.

[0173] In one feasible implementation, each strategy atom in the strategy atom library includes metadata, input slots, a logic body, and an output interface. The strategy atom structure is as follows: In this embodiment, each strategy atom is encapsulated as a standardized object containing four core elements: Metadata: A unique identifier (ID), name, functional description, and version number describing the atom, used for retrieval and management. Input Slots: Variable interfaces required for the atom to execute its logic, such as the amount disbursed, the judgment threshold, and the reach time. Slots are empty during initialization and require external injection of specific values. Logic Body: Program code or rule scripts that implement the atom's function, such as calculation formulas or API call instructions. Output Interface: The definition of the return result after the atom's execution, used to connect to the input of downstream atoms.

[0174] The execution parameters for atomically injecting into candidate strategies in the directed acyclic graph based on historical strategy data in the service strategy knowledge base can be performed in the following manner:

[0175] Step B2: Identify the input slots to be filled for each candidate strategy atom in the directed acyclic graph;

[0176] Optionally, a static parsing mode based on metadata can be used to traverse the activity strategy logic topology graph (DAG) generated by S404, accessing each candidate strategy atomic node in the graph. For each node, the interface specification defined in its metadata is read. This specification explicitly lists all input slots of the atom and their data types (such as Integer, String, Enum) and required / optional attributes. The identifiers of all slots with "required" attributes and currently empty values ​​are extracted to generate a "list of parameters to be filled".

[0177] Step B4: Retrieve from the service strategy knowledge base the strategy parameter values ​​for the associated service activity scenarios corresponding to the candidate strategy atom, as well as the strategy implementation effect parameters corresponding to the strategy parameter values.

[0178] The strategy parameter value refers to the specific value actually configured for the input slot of a certain strategy atom in historical business activities (such as 10 yuan, 20% off, 19:00).

[0179] The strategy implementation effect parameter refers to the business KPI data (such as conversion rate CVR, click-through rate CTR, and return on investment) generated when the strategy atom is configured with specific strategy parameter values ​​and runs in the actual production environment.

[0180] As an illustration, a vector similarity-based retrieval model can be used to map the current target activity scene into a scene feature vector. Then, similar scene matching is performed by retrieving a set of historical scenes from the knowledge base whose similarity to the scene feature vector exceeds a threshold. Next, cross-scene aggregation is performed, extracting the operational data of the atom in similar scenes from the historical scene set. Based on the operational data of the atom, the policy parameter values ​​and policy implementation effect parameters of these scenes are aggregated, expanding the sample space of parameter sources.

[0181] Step B6: Select historical parameter values ​​from the strategy parameter values ​​that satisfy the preset threshold for strategy implementation effect parameters, and inject the historical parameter values ​​as execution parameters into the input slots corresponding to the candidate strategy atoms.

[0182] Preset threshold: refers to a predefined quantitative standard used to measure whether the strategy's effectiveness has met the target.

[0183] Historical parameter values ​​refer to specific configuration values ​​that have actually taken effect in past business activities and are stored in the service policy knowledge base.

[0184] Injection refers to the process of assigning a defined numerical value to a property of a strategy atom object. After injection, the strategy atom transforms from an abstract logical node into an executable instance node.

[0185] Optionally, the retrieved historical record list is sorted in descending order based on a preset preferred metric (such as ROI). Then, the list is traversed, and records with strategy implementation effect parameters below a preset threshold are removed. The historical parameter value corresponding to the first record (Top-1) in the sorted list is selected and marked as the "preferred parameter," which is then assigned to the input slot corresponding to the candidate strategy atom. If the list is empty after filtering (i.e., no historical data meets the threshold), a fallback logic is triggered, injecting the system's default conservative parameter value.

[0186] Optionally, the retrieved historical records can be grouped according to historical parameter values. For example, all records with "coupon amount = 5 yuan" can be grouped together, and those with "coupon amount = 10 yuan" can be grouped together. For each group, the statistical distribution characteristics (such as mean and standard deviation) of the corresponding strategy implementation effect parameters can be calculated.

[0187] Set a composite preset threshold (e.g., average > X and standard deviation < Y), and filter out the parameter groups that meet this composite threshold. If there are multiple groups that meet the criteria, select the group with the highest "average / standard deviation" ratio (a similar indicator to the Sharpe ratio). Inject the historical parameter values corresponding to the selected preferred group as execution parameters into the input slot.

[0188] Through the above method, the automated and intelligent configuration of policy parameters is achieved, ensuring that each parameter setting is based on evidence and avoiding the subjective blindness of manual configuration.

[0189] In a feasible implementation manner, before generating the active policy logic structure based on the active policy logic topology map described in S304, the following steps are also referred to:

[0190] Step C2: Vectorize the active policy logic topology map based on the policy processing large model to obtain a policy intermediate representation, and start an instantiated adversarial simulation mechanism, where the adversarial simulation mechanism includes a generation role and a discrimination role;

[0191] The policy intermediate representation refers to a data format designed specifically for large model reasoning and adversarial simulation, which is between the graphical topology and machine code. It usually appears as a high-dimensional semantic vector (SemanticVector) or a serialized structured text (such as DSL code), and can completely map the node attributes, connection relationships, and logical constraints of the policy map, facilitating the operation of the large model in the latent space (LatentSpace).

[0192] The adversarial simulation mechanism refers to a virtual game environment built inside the large model inference engine. This mechanism is based on the generative adversarial network or the red-blue army drill idea, and simulates the performance of the policy in extreme or malicious environments by presetting opposing role goals.

[0193] The generation role refers to the party responsible for "construction" in the adversarial mechanism. Its function is to maintain the integrity of the policy logic, and repair and optimize the map according to the feedback of the discrimination role, with the goal of making the policy pass all tests.

[0194] The discrimination role refers to the party responsible for "attack" in the adversarial mechanism. Its function is to use means such as counterfactual reasoning and boundary value injection to find logical loopholes or security risks in the current policy, with the goal of falsifying the robustness of the current policy.

[0195] Optionally, a mode based on graph embedding and multi-way Prompt instantiation can be adopted, specifically as follows:

[0196] The algorithm uses a large-scale policy processing model to call a graph neural network model (such as GraphSAGE or GAT) or the encoder of the large-scale policy processing model to read the activity policy logic topology graph generated upstream. The algorithm aggregates and encodes the node features (such as atom type and parameter values) and edge features (such as flow conditions) in the graph, mapping them to a high-dimensional dense vector of fixed dimension, i.e., the intermediate representation of the policy.

[0197] Then, character instantiation is performed: within the inference context of the large-scale policy processing model, two independent dialogue instances (Sessions) or thought chain branches are initialized in parallel using system-level prompts.

[0198] The generated character instance loads a prompt message, such as "You are the commander, responsible for refining the strategy logic based on feedback."

[0199] The system identifies the role instance loading prompt, such as "You are an auditor, responsible for using extreme case attack strategy logic."

[0200] Then, the mechanism is activated, a message channel is established between the two instances, and the intermediate representation of the strategy is shared with the two roles as the initial state, marking the official start of the adversarial simulation mechanism.

[0201] Step C4: Control the large model of the strategy processing under the discriminator role to inject abnormal scenario vectors or boundary condition vectors into the intermediate representation of the strategy based on the counterfactual reasoning logic chain, so as to determine whether the strategy execution result of the intermediate representation of the strategy under the abnormal scenario vector satisfies the preset security robustness rules.

[0202] Counterfactual reasoning refers to a "hypothesis-verification" deep reasoning model. The large model is not based on the current reality, but rather constructs a hypothetical premise that contradicts or is extreme from the facts (e.g., "What if this high-quality user suddenly becomes a fraudster..."), and deduces subsequent causal chains based on this assumption.

[0203] An abnormal scenario vector refers to a high-dimensional feature vector representing a system environment in an abnormal state. For example, a state vector representing "the budget pool balance suddenly becomes zero", "the downstream risk control interface response times out", or "concurrent traffic surges by 100 times".

[0204] Boundary condition vectors refer to feature vectors representing business input data at critical points. For example, a data vector representing "a user's credit score is exactly equal to the admission threshold", "a user's registration time is exactly 0 seconds", or "the application amount is negative".

[0205] Predefined security robustness rules refer to the bottom-line principles that the system must adhere to under all circumstances. Examples include: no financial loss under any abnormal situation; all network anomalies must have a fallback branch; and granting privileges to blacklisted users is prohibited.

[0206] Optionally, a black-box attack model based on data boundary mutation can be adopted, specifically:

[0207] First, attack vectors are constructed. The large-scale strategy processing model scans the input parameter definitions of the strategy under the role discrimination. Based on boundary value analysis, a set of boundary condition vectors is automatically constructed. For example, if the strategy requires "age > 18", the role discrimination constructs test vectors for "age = 18" and "age = null"; if the strategy involves "identity", an adversarial vector for "identity = black market account" is constructed.

[0208] Then, counterfactual injection and deduction are performed to determine whether the role injects the above vector into the intermediate representation of the strategy through thought chain instructions: "Assuming that the user who triggers the current strategy has the above characteristics (counterfactual), please deduce the execution path of the graph."

[0209] Then, path tracing and adjudication are performed, with the strategy processing model simulating the strategy flow. If the simulation results show that the "black market account" successfully flows to the "fund disbursement" node, it is determined that the preset security robustness rules are not met (there is a vulnerability). If the simulation results show that the process is blocked at the "risk control node" or enters the "rejection branch", it is determined that the rules are met.

[0210] Optionally, a white-box attack model based on environmental fault simulation can be adopted, specifically:

[0211] First, construct scenario vectors to identify the external services (such as API atomic calls) invoked by the role in the strategy, and construct abnormal scenario vectors. For example, for "query quota atomic call", construct vectors for "Return_Code=500 (server error)" or "Latency=10s (timeout)".

[0212] Next, conduct logical stress testing: the large model for policy processing injects environmental fault information into the intermediate representation of the policy under the role discrimination, such as "assuming that when executing step 3, the external interface returns a 500 error, how should the subsequent logic handle it?"

[0213] Then, robustness verification is performed: check whether there is a capture mechanism or default fallback path for the "environmental fault information" in the policy logic.

[0214] If the simulation shows that the process is directly interrupted, throws an unhandled exception, or causes the system to freeze, it is determined that the preset security robustness rules are not met (lack of fault tolerance).

[0215] If the simulation shows that the process jumps to a fallback node such as "send a reassuring text message" or "end the process", then the rule is deemed to be satisfied.

[0216] Through the above implementation method, this step utilizes the counterfactual reasoning capability of the large model to successfully simulate a malicious attack scenario before the strategy goes live, identify hard-to-detect logical flaws, and provide accurate targets for subsequent automatic repair (step C5).

[0217] Step C5: If not satisfied, control the strategy processing model under the generation role to perform graph feature correction processing on the logical topology structure features of the intermediate representation of the strategy to obtain the corrected activity strategy logical topology graph.

[0218] Logical topological features refer to the skeletal elements that constitute the policy graph, including node types, node attributes, and connecting edges between nodes. Modifying these features means changing the execution flow of the policy.

[0219] Graph feature correction processing refers to incremental editing operations performed on the strategy graph. Common operations include "node insertion", "connection redirection", "attribute modification" and "branch pruning".

[0220] Optionally, an incremental correction mode based on defensive node insertion can be adopted. This approach is suitable for compensating for risk control deficiencies or logical loopholes, preserving as much of the original business logic as possible, and only adding checkpoints. Specifically:

[0221] First, vulnerability localization is performed, and a "successful attack path" is generated to receive and verify the output of the role. For example, the path shows "black market user -> action node A (payment)". The generated role is located on the preceding path to action node A, but the necessary verification nodes are missing.

[0222] Then, patch generation is performed, and the strategy is controlled to process the large model to search the knowledge base and find defense strategy atoms (such as Atom_Risk_Check) that can defend against the attack vector.

[0223] Then, topology reconstruction is performed based on the defense strategy atoms:

[0224] Disconnect: Cut off the connection edge between action node A and its original predecessor node affected by the attack vector.

[0225] Insertion: Insert a new "verification atom", also known as a defense strategy atom, at the breakpoint.

[0226] Connection: Establish a new connection: original predecessor - verification atom - pass - action node A; at the same time, add a fallback path: verification atom - failure - rejection / end.

[0227] Finally, the representation is updated, and the modified topology is remapped to an activity strategy logic topology graph, in preparation for the next round of verification.

[0228] Optionally, a path redirection-based logical circuit breaker mode can be used. This approach is suitable for handling systemic anomalies or serious logical errors that cannot be resolved through simple checks; specifically:

[0229] The control strategy processes the large model to generate reports on role-based anomaly identification and analysis, and analyzes whether there are irreparable deadlocks or serious compliance risks in the current path (for example, a business branch will inevitably lead to resource over-exploitation under certain conditions).

[0230] The control strategy processes the large model to identify the entire logical branch (sub-graph) that leads to risk, marks it as an "invalid path," and redirects the ingress traffic of that branch to a safe "circuit breaker node" or "manual review node."

[0231] The operation logic for redirection execution is as follows: Modify the condition attribute of the branch entry judgment node, and redirect the condition that originally flowed to the high-risk branch to Atom_Human_Review (transfer to human) or Atom_End_Process (direct termination).

[0232] Finally, a graph is generated, outputting a logical topology graph of the activity strategy after pruning and redirection, ensuring that risk paths are physically unreachable.

[0233] Through the above implementation method, this step realizes the automated iterative repair of strategy logic, dynamically fills logical loopholes by utilizing the generation capability of large models, and significantly improves the security and robustness of the final online strategy.

[0234] Step C6: If satisfied, then execute the step of generating the activity strategy logic structure based on the activity strategy logic topology graph.

[0235] This specification introduces an instantiation-based adversarial simulation mechanism based on intermediate policy representations during the policy generation phase. By controlling the discriminator role and injecting abnormal scenarios or boundary condition vectors using counterfactual reasoning, it proactively conducts high-intensity adversarial stress testing on the policy logic. Furthermore, the generation role automatically corrects and self-heals vulnerabilities by modifying graph features to address those that do not meet robustness rules. This mechanism achieves "self-censorship" and "closed-loop repair" of the policy logic, accurately identifying and eliminating potential logical vulnerabilities (such as missing risk control mechanisms or infinite loops) before policy deployment. This significantly improves the security, robustness, and logical rigor of AI-generated policies in extreme business scenarios, effectively mitigating security violations and risks caused by the illusion of large models.

[0236] The following will combine Figure 4This manual provides a detailed description of the activity strategy generation device. It should be noted that... Figure 4 The activity strategy generation device shown is used to execute this specification. Figures 1-3 The methods of the embodiments shown are illustrated only in connection with this specification for ease of explanation. For specific technical details not disclosed, please refer to this specification. Figures 1-3 The example shown.

[0237] Please see Figure 4 This diagram illustrates the structure of the activity strategy generation device 1 described herein. The activity strategy generation device 1 can be implemented as all or part of an electronic device through software, hardware, or a combination of both. According to some embodiments, the activity strategy generation device 1 includes:

[0238] Description input module 11 is used to obtain the service activity strategy recommendation description input by the server to the strategy processing big model for the platform service promotion activities;

[0239] The strategy processing module 12 is used to determine the target activity scenario and activity strategy generation intent based on the service activity strategy recommendation description and the strategy processing big model, and to call the service strategy knowledge base to generate the activity strategy logical structure based on the target activity scenario and activity strategy generation intent, and to generate activity strategy canvas configuration data that conforms to the preset strategy process orchestration standard based on the strategy logical structure.

[0240] The strategy display module 13 is used to render the strategy process in the strategy canvas management engine using the activity strategy canvas configuration data to obtain the target activity strategy data for the promotion activity of the platform service.

[0241] Optionally, the step of determining the target activity scenario and activity strategy generation intent using the strategy processing model based on the service activity strategy recommendation description, and generating an activity strategy logical structure by calling the service strategy knowledge base based on the target activity scenario and activity strategy generation intent, and generating activity strategy canvas configuration data conforming to the preset strategy flow orchestration standard based on the strategy logical structure, includes:

[0242] The service activity strategy recommendation description is input into the intent recognition agent to call the strategy processing big model to determine the activity strategy and generate intent;

[0243] The activity strategy generation intent and the service activity strategy recommendation description are input into the scene processing agent to call the strategy processing big model to determine the target activity scene;

[0244] The service activity strategy recommendation description, the activity strategy generation intent, and the target activity scenario are input into the strategy processing agent to call the strategy processing big model to generate the activity strategy logical structure, and based on the strategy logical structure, generate activity strategy canvas configuration data that conforms to the preset strategy process orchestration standard.

[0245] Optionally, before inputting the service activity policy recommendation description into the intent recognition agent to call the policy processing big model to determine the activity policy generation intent, the method further includes: inputting the service activity policy recommendation description into the security assessment agent to call the policy processing big model to perform security compliance detection on the service activity policy recommendation description; if the detection passes, then the step of inputting the service activity policy recommendation description into the intent recognition agent to call the policy processing big model to determine the activity policy generation intent is executed; if the detection fails, then the service activity policy recommendation description is intercepted and a preset rejection response is output.

[0246] After invoking the strategy to process the large model and generate the activity strategy logic structure, the process also includes:

[0247] The security assessment agent invokes the policy processing big model to perform output security risk verification on the policy logic structure; if the detection passes, the step of generating activity policy canvas configuration data that conforms to the preset policy process orchestration standard based on the policy logic structure is executed; if the detection fails, the policy logic structure is intercepted and a preset rejection response is output.

[0248] Optionally, the step of generating the logical structure of the activity strategy based on the target activity scenario and activity strategy generation intent by calling the service strategy knowledge base includes:

[0249] The large model is processed by the strategy described above to construct a generation task context based on the target activity scenario and the activity strategy generation intent. The generation task context is deconstructed into multi-dimensional constraint objectives, which include at least target user group characteristic constraints, platform service key evaluation index constraints, and activity promotion cost constraints.

[0250] Furthermore, based on the multi-dimensional constraint target, the service strategy knowledge base is invoked to perform strategy topology reasoning to obtain an activity strategy logical topology graph, and an activity strategy logical structure is generated based on the activity strategy logical topology graph.

[0251] Optionally, the step of obtaining the activity strategy logical topology graph by performing strategy topology reasoning based on the multi-dimensional constraint target call service strategy knowledge base includes:

[0252] The multidimensional constraint target is mapped to a constraint target feature vector through the large model of the strategy processing. The constraint target feature vector is used to perform vector matching processing in a pre-set strategy atom library to obtain at least one candidate strategy atom. The strategy atom in the pre-set strategy atom library is a pre-encapsulated minimum strategy implementation functional component.

[0253] The strategy processes the large model to predict the causal connections between the candidate strategy atoms, constructs a directed acyclic graph for each candidate strategy atom based on the causal connections, and injects execution parameters into the candidate strategy atoms in the directed acyclic graph based on historical strategy data in the service strategy knowledge base, thus forming an activity strategy logical topology graph.

[0254] Optionally, each policy atom in the policy atom library includes metadata, input slots, a logical body, and an output interface. The injection of execution parameters into candidate policy atoms in the directed acyclic graph based on historical policy data from the service policy knowledge base includes:

[0255] Identify the input slots to be filled for each candidate strategy atom in the directed acyclic graph;

[0256] Retrieve from the service strategy knowledge base the strategy parameter values ​​for the associated service activity scenarios corresponding to the candidate strategy atom, as well as the strategy implementation effect parameters corresponding to the strategy parameter values.

[0257] Historical parameter values ​​that satisfy a preset threshold are selected from the strategy parameter values, and these historical parameter values ​​are injected as execution parameters into the input slots corresponding to the candidate strategy atoms.

[0258] Optionally, before generating the activity strategy logic structure based on the activity strategy logic topology graph, the method further includes:

[0259] Based on the aforementioned strategy processing big model, the activity strategy logic topology graph is vectorized to obtain the intermediate representation of the strategy, and an instantiated adversarial simulation mechanism is initiated, which includes generating roles and identifying roles.

[0260] The control strategy processing model, under the judgment role, injects abnormal scenario vectors or boundary condition vectors into the intermediate representation of the strategy based on the counterfactual reasoning logic chain, so as to determine whether the strategy execution result of the intermediate representation of the strategy under the abnormal scenario vector satisfies the preset security robustness rules.

[0261] If not satisfied, the strategy processing model is controlled to perform graph feature correction processing on the logical topology structure features of the intermediate representation of the strategy under the generation role, so as to obtain the corrected activity strategy logical topology graph.

[0262] If the conditions are met, then the step of generating the activity strategy logic structure based on the activity strategy logic topology graph is executed.

[0263] Optionally, the method further includes:

[0264] After the strategy configuration data is determined and the platform service promotion activity is deployed and executed, post-implementation effect data of the strategy operation is collected.

[0265] The posterior effect data and the target activity strategy data are analyzed to obtain strategy execution analysis data, and the strategy execution analysis data is fed back to the service strategy knowledge base for strategy knowledge update processing.

[0266] It should be noted that the activity strategy generation device provided in the above embodiments is only illustrated by the division of the above functional modules when executing the activity strategy generation method. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the activity strategy generation device and the activity strategy generation method embodiments provided in the above embodiments belong to the same concept, and the implementation process is detailed in the method embodiments, which will not be repeated here.

[0267] The serial numbers in this specification are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0268] This specification also provides a computer storage medium capable of storing multiple instructions adapted to be loaded and executed by a processor as described above. Figures 1-3 The activity strategy generation method described in the illustrated embodiment can be found in the following documentation for its specific execution process. Figures 1-3 The specific details of the illustrated embodiments will not be elaborated here.

[0269] This specification also provides a computer program product that stores at least one instruction, said at least one instruction being loaded and executed by the processor as described above. Figures 1-3 The activity strategy generation method described in the illustrated embodiment can be found in the following documentation for its specific execution process. Figures 1-3 The specific details of the illustrated embodiments will not be elaborated here.

[0270] Please refer to Figure 5 This is a structural block diagram of an electronic device provided in an embodiment of this specification. The electronic device in this specification may include one or more of the following components: a processor 1010, a memory 1020, an input device 1030, an output device 1040, and a bus 1050. The processor 1010, memory 1020, input device 1030, and output device 1040 may be connected to each other via the bus 1050.

[0271] Processor 1010 may include one or more processing cores. Processor 1010 connects to various parts of the electronic device using various interfaces and lines, and performs various functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in memory 1020, and by calling data stored in memory 1020. Optionally, processor 1010 may be implemented using at least one hardware form of digital signal processing (DSP), field-programmable gate array (FPGA), or programmable logic array (PLA). Processor 1010 may integrate one or a combination of central processing unit (CPU), graphics processing unit (GPU), and modem. The CPU mainly handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the displayed content; and the modem is used for wireless communication. It is understood that the modem may also not be integrated into processor 1010 and may be implemented separately through a communication chip.

[0272] The memory 1020 may include random access memory (RAM) or read-only memory (ROM). Optionally, the memory 1020 may include non-transitory computer-readable storage medium. The memory 1020 may be used to store instructions, programs, code, code sets, or instruction sets.

[0273] The input device 1030 is used to receive input instructions or data, and includes, but is not limited to, a keyboard, mouse, camera, microphone, or touch device. The output device 1040 is used to output instructions or data, and includes, but is not limited to, a display device and a speaker. In this embodiment, the input device 1030 can be a temperature sensor for acquiring the operating temperature of the electronic device. The output device 1040 can be a speaker for outputting audio signals.

[0274] In addition, those skilled in the art will understand that the structure of the electronic device shown in the above figures does not constitute a limitation on the electronic device. The electronic device may include more or fewer components than shown, or combine certain components, or have different component arrangements. For example, the electronic device may also include radio frequency circuits, input units, sensors, audio circuits, wireless fidelity (WIFI) modules, power supplies, Bluetooth modules, etc., which will not be described in detail here.

[0275] In the embodiments of this specification, the executing entity for each step can be the electronic device described above. Optionally, the executing entity for each step can be the operating system of the electronic device. The operating system can be Android, iOS, or other operating systems; this specification does not limit this.

[0276] exist Figure 5 In the electronic device, the processor 1010 can be used to call a program stored in the memory 1020 and execute it to implement the activity strategy generation method as described in the various method embodiments of this specification.

[0277] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory, or random access memory, etc.

[0278] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.), and signals involved in the embodiments of this specification are all authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, descriptions of service activity strategy recommendations are obtained with full authorization.

[0279] The above-disclosed embodiments are merely preferred embodiments of this specification and should not be construed as limiting the scope of this specification. Therefore, any equivalent variations made in accordance with the claims of this specification shall still fall within the scope of this specification.

Claims

1. A method for generating activity strategies, characterized in that, Applied to a service platform, the method includes: Obtain the service activity strategy recommendation description input from the server to the strategy processing model for platform service promotion activities; Based on the service activity strategy recommendation description, the strategy processing big model is used to determine the target activity scenario and activity strategy generation intent, and based on the target activity scenario and activity strategy generation intent, the service strategy knowledge base is called to generate the activity strategy logical structure, and based on the strategy logical structure, activity strategy canvas configuration data that conforms to the preset strategy process orchestration standard is generated. The activity strategy canvas configuration data is used in the strategy canvas management engine to render the strategy process and obtain the target activity strategy data for the promotion activity of the platform service. The step of generating the activity strategy logical structure by invoking the service strategy knowledge base based on the target activity scenario and activity strategy generation intent includes: The large model is processed by the strategy described above to construct a generation task context based on the target activity scenario and the activity strategy generation intent. The generation task context is deconstructed into multi-dimensional constraint objectives, which include at least target user group characteristic constraints, platform service key evaluation index constraints, and activity promotion cost constraints. Based on the multi-dimensional constraint target, the service strategy knowledge base is invoked to perform strategy topology reasoning to obtain the activity strategy logical topology graph. Based on the strategy processing big model, the activity strategy logical topology graph is vectorized to obtain the strategy intermediate representation. Then, the instantiation adversarial simulation mechanism is started. The adversarial simulation mechanism includes generating roles and judging roles. The control strategy processing model, under the judgment role, injects abnormal scenario vectors or boundary condition vectors into the intermediate representation of the strategy based on the counterfactual reasoning logic chain, so as to determine whether the strategy execution result of the intermediate representation of the strategy under the abnormal scenario vector satisfies the preset security robustness rules. If not satisfied, the strategy processing model under the generation role will perform graph feature correction processing on the logical topology structure features of the intermediate representation of the strategy to obtain the corrected activity strategy logical topology graph. If satisfied, the step of generating the activity strategy logical structure based on the activity strategy logical topology graph will be executed. The activity strategy logic structure is generated based on the activity strategy logic topology graph.

2. The method according to claim 1, characterized in that, The process involves using the service activity strategy recommendation description to determine the target activity scenario and activity strategy generation intent using the strategy processing model, and then, based on the target activity scenario and activity strategy generation intent, calling the service strategy knowledge base to generate the activity strategy logical structure. Finally, it generates activity strategy canvas configuration data that conforms to a preset strategy flow orchestration standard based on the strategy logical structure, including: The service activity strategy recommendation description is input into the intent recognition agent to call the strategy processing big model to determine the activity strategy and generate intent; The activity strategy generation intent and the service activity strategy recommendation description are input into the scene processing agent to call the strategy processing big model to determine the target activity scene; The service activity strategy recommendation description, the activity strategy generation intent, and the target activity scenario are input into the strategy processing agent to call the strategy processing big model to generate the activity strategy logical structure, and based on the strategy logical structure, generate activity strategy canvas configuration data that conforms to the preset strategy process orchestration standard.

3. The method according to claim 2, characterized in that, Before inputting the service activity strategy recommendation description into the intent recognition agent to call the policy processing big model to determine the activity strategy generation intent, the method further includes: inputting the service activity strategy recommendation description into the security assessment agent to call the policy processing big model to perform security compliance detection on the service activity strategy recommendation description; if the detection passes, the step of inputting the service activity strategy recommendation description into the intent recognition agent to call the policy processing big model to determine the activity strategy generation intent is executed; if the detection fails, the service activity strategy recommendation description is intercepted and a preset rejection response is output. After invoking the strategy to process the large model and generate the activity strategy logic structure, the process also includes: The security assessment agent invokes the policy processing big model to perform output security risk verification on the policy logic structure; if the detection passes, the step of generating activity policy canvas configuration data that conforms to the preset policy process orchestration standard based on the policy logic structure is executed; if the detection fails, the policy logic structure is intercepted and a preset rejection response is output.

4. The method according to claim 1, characterized in that, The process of obtaining the activity strategy logical topology graph by reasoning the strategy topology structure based on the multi-dimensional constraint target call service strategy knowledge base includes: The multidimensional constraint target is mapped to a constraint target feature vector through the large model of the strategy processing. The constraint target feature vector is used to perform vector matching processing in the pre-set strategy atom library to obtain at least one candidate strategy atom. The strategy atom in the pre-set strategy atom library is a pre-encapsulated minimum strategy implementation functional component. The strategy processes the large model to predict the causal connections between the candidate strategy atoms, constructs a directed acyclic graph for each candidate strategy atom based on the causal connections, and injects execution parameters into the candidate strategy atoms in the directed acyclic graph based on historical strategy data in the service strategy knowledge base, thus forming an activity strategy logical topology graph.

5. The method according to claim 4, characterized in that, Each policy atom in the policy atom library contains metadata, input slots, a logical body, and an output interface. The injection of execution parameters into candidate policy atoms in the directed acyclic graph based on historical policy data from the service policy knowledge base includes: Identify the input slots to be filled for each candidate strategy atom in the directed acyclic graph; Retrieve from the service strategy knowledge base the strategy parameter values ​​for the associated service activity scenarios corresponding to the candidate strategy atom, as well as the strategy implementation effect parameters corresponding to the strategy parameter values. Historical parameter values ​​that satisfy a preset threshold are selected from the strategy parameter values, and these historical parameter values ​​are injected as execution parameters into the input slots corresponding to the candidate strategy atoms.

6. The method according to claim 1, characterized in that, The method further includes: After the strategy configuration data is determined and the platform service promotion activity is deployed and executed, post-implementation effect data of the strategy operation is collected. The posterior effect data and the target activity strategy data are analyzed to obtain strategy execution analysis data, and the strategy execution analysis data is fed back to the service strategy knowledge base for strategy knowledge update processing.

7. An activity strategy generation device, characterized in that, The device, applied to a service platform, includes: The description input module is used to obtain the service activity strategy recommendation description input by the server to the strategy processing big model for platform service promotion activities; The strategy processing module is used to determine the target activity scenario and activity strategy generation intent based on the service activity strategy recommendation description and the strategy processing big model, and to call the service strategy knowledge base to generate the activity strategy logical structure based on the target activity scenario and activity strategy generation intent, and to generate activity strategy canvas configuration data that conforms to the preset strategy process orchestration standard based on the strategy logical structure. The strategy display module is used in the strategy canvas management engine to render the strategy process using the activity strategy canvas configuration data to obtain the target activity strategy data for the promotion activity of the platform service. The step of generating the activity strategy logical structure by invoking the service strategy knowledge base based on the target activity scenario and activity strategy generation intent includes: The large model is processed by the strategy described above to construct a generation task context based on the target activity scenario and the activity strategy generation intent. The generation task context is deconstructed into multi-dimensional constraint objectives, which include at least target user group characteristic constraints, platform service key evaluation index constraints, and activity promotion cost constraints. Based on the multi-dimensional constraint target, the service strategy knowledge base is invoked to perform strategy topology reasoning to obtain the activity strategy logical topology graph. Based on the strategy processing big model, the activity strategy logical topology graph is vectorized to obtain the strategy intermediate representation. Then, the instantiation adversarial simulation mechanism is started. The adversarial simulation mechanism includes generating roles and judging roles. The control strategy processing model, under the judgment role, injects abnormal scenario vectors or boundary condition vectors into the intermediate representation of the strategy based on the counterfactual reasoning logic chain, so as to determine whether the strategy execution result of the intermediate representation of the strategy under the abnormal scenario vector satisfies the preset security robustness rules. If not satisfied, the strategy processing model under the generation role will perform graph feature correction processing on the logical topology structure features of the intermediate representation of the strategy to obtain the corrected activity strategy logical topology graph. If satisfied, the step of generating the activity strategy logical structure based on the activity strategy logical topology graph will be executed. The activity strategy logic structure is generated based on the activity strategy logic topology graph.

8. A computer storage medium, characterized in that, The computer storage medium stores a plurality of instructions, which are adapted to be loaded by a processor and executed as described in any one of claims 1 to 6.

9. A computer program product, characterized in that, The computer program product stores at least one instruction, which is loaded by a processor and executed as described in any one of claims 1 to 6.

10. An electronic device, characterized in that, include: A processor and a memory; wherein the memory stores a computer program adapted to be loaded by the processor and executed as described in any one of claims 1 to 6.

Images