Live broadcast competition compliance policy pushing method and system based on behavior promotion

Abstract

The application relates to the technical field of network live broadcast compliance supervision, and discloses a live broadcast competition compliance policy pushing method based on behavior main promotion, which comprises the following steps: S1, acquiring multi-dimensional behavior data flow of a user in a live broadcast process; S2, generating a budding period risk score; S3, judging when the budding period risk score exceeds a preset risk threshold; S4, processing a strategy scheme; and S5, generating a compliance competition strategy for the user based on a compliance policy guidance scheme. By acquiring the multi-dimensional behavior data flow of the user and constructing a time sequence feature vector, the Mahalanobis distance of an individual behavior baseline model and the Euclidean distance of a competition ecological compliance behavior paradigm are innovatively combined to calculate double deviation degrees, and then a quantitative budding period risk score is generated. The method is free from the dependence on illegal keywords or post-event, and realizes early, active and highly personalized and accurate identification effects of potential compliance risks.

Description

Technical Field

[0001] This invention relates to the field of online live streaming compliance supervision technology, specifically to a method and system for pushing live streaming competition compliance policies based on behavior. Background Technology

[0002] To maintain their ecosystem order and compliance, online information platforms typically deploy a series of technical measures to monitor information content and user behavior within the platform. These technical measures generally include content filtering systems based on preset rule bases, artificial intelligence review models for identifying illegal information, and response mechanisms for handling user complaints.

[0003] In online live streaming, especially in the specific scenario of live commerce, the aforementioned technologies are widely used. Platforms generally employ automatic speech recognition technology to convert the broadcaster's audio into text in real time, and combine this with natural language processing technology to perform keyword matching on text data such as live broadcast scripts, product titles and descriptions, and public chat comments to identify and block information containing clearly prohibited words or phrases. Simultaneously, a large-scale manual inspection team and a post-event complaint handling mechanism are used to trace and address any violations that have occurred during the live broadcast.

[0004] However, the aforementioned existing technologies have significant shortcomings in terms of the timeliness and foresight of risk identification. Current mechanisms based on keyword matching or post-event complaints are essentially a form of passive supervision, heavily reliant on clear, known characteristics of violations or facts of violations that have already occurred. This approach cannot effectively identify potential risks in their nascent stages, especially compliance risks manifested through complex, non-obvious behavioral patterns rather than specific prohibited words, making it difficult for platforms to proactively and forward-lookingly intervene before violations have a substantial impact. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for pushing live streaming competition compliance policies based on behavior-driven approaches, solving the problem of passive and delayed risk identification.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for pushing live streaming competition compliance policies based on behavior-driven approaches, comprising the following steps:

[0007] S1, S1, acquire multi-dimensional behavioral data streams and competitive ecosystem compliance behaviors of users during the live broadcast, and build a time-series feature vector of the user's current behavior and establish a baseline model of individual user behavior in real time;

[0008] S2. Based on the time-series feature vector, calculate the Mahalanobis distance between the user's current behavior and the baseline model of the user's individual behavior, as well as the Euclidean distance between the user's current behavior and the competitive ecosystem compliance behavior paradigm. Then, sum the Mahalanobis distance and the Euclidean distance with weights to generate the nascent stage risk score and set a preset risk threshold range for the nascent stage risk score.

[0009] S3. When the risk score in the nascent stage exceeds the preset risk threshold, a multi-agent game sandbox running on a computer server is triggered.

[0010] S4. Process the strategy proposals submitted by users in the multi-agent game sandbox and generate processing results. Then, based on the processing results, push a compliance policy guidance plan to the user, which includes compliance correction suggestions and effectiveness prediction indicators.

[0011] S5. Generate a compliant competitive strategy for users based on the compliance policy guidance scheme. Based on the same multi-dimensional behavioral data stream processing method as in step S1, construct the temporal feature vector of subsequent behaviors, calculate the similarity between the temporal feature vector of subsequent behaviors and the feature space of the compliant competitive strategy, and determine that the verification is successful and update the individual behavior baseline model when the similarity is greater than a preset threshold. The preset threshold is determined by the experience of the implementers.

[0012] Preferably, the individual behavior baseline model includes the mean vector and covariance matrix of the user's historical behavior; the competitive ecosystem compliance behavior paradigm includes the aggregated behavior mean vector of the peer group in the competitive ecosystem to which the user belongs.

[0013] Preferably, the step of generating the multi-agent game sandbox specifically includes:

[0014] Based on the feature dimensions in the time-series feature vector that exceed the baseline range, they are mapped to a preset risk type library to determine the risk type;

[0015] Based on the risk type, one or more virtual competitor agents are retrieved and instantiated from a preset strategy model library. The virtual competitor agents have a defined and executable strategy model.

[0016] Preferably, the corrected data for outputting dual-dimensional guidance specifically includes:

[0017] The compliance rule inference engine verifies the strategy scheme and outputs the violation points and structured correction instructions;

[0018] The game analysis unit simulates the game response of the virtual competitor agent and outputs a quantitative prediction effectiveness index of the strategy scheme.

[0019] Preferably, the quantitative forecasting effectiveness indicators include estimated market share and estimated profit margin.

[0020] The behavior-driven live streaming competition compliance policy push system includes:

[0021] The data acquisition and processing unit is configured to acquire multi-dimensional behavioral data streams of users during live streaming, and to build a temporal feature vector of the user's current behavior and establish a baseline model of the user's individual behavior in real time.

[0022] The risk perception unit is configured to calculate the Mahalanobis distance between the user's current behavior and the user's individual behavior baseline model, and the Euclidean distance between the user's current behavior and the competitive ecosystem compliance behavior paradigm, based on the time-series feature vector, and to weight and sum the results of the two distance calculations to generate an nascent risk score.

[0023] The sandbox generation unit is configured to trigger the generation of a multi-agent game sandbox running on a computer server when the risk score in the nascent stage exceeds a preset risk threshold. The simulation parameters of the multi-agent game sandbox are configured based on the individual behavior baseline model and the competitive ecosystem compliance behavior paradigm.

[0024] The strategy guidance unit is configured to guide users to submit strategy proposals in the multi-agent game sandbox, and to process data based on the built-in compliance rule inference engine and game analysis unit, and output corrected data containing guidance in both compliance and effectiveness dimensions.

[0025] The closed-loop verification unit is configured to support users in generating compliant competitive strategies based on the corrected data, and monitor the similarity between the user's subsequent behavior and the compliant competitive strategies in the feature space. When the similarity is greater than a preset threshold, the verification is determined to be successful and the risk perception unit is driven to update the individual behavior baseline model.

[0026] Preferably, the data acquisition and processing unit is built on a streaming computing framework and uses a sliding time window mechanism to process the multidimensional behavioral data stream to ensure the real-time performance of data processing.

[0027] Preferably, the sandbox generation unit is specifically configured to:

[0028] Based on the feature dimensions in the time-series feature vector that exceed the baseline range, they are mapped to a preset risk type library to determine the risk type;

[0029] Based on the risk type, one or more virtual competitor agents are retrieved and instantiated from a pre-indexed and optimized strategy model library.

[0030] Preferably, the game analysis unit built into the strategy guidance unit is configured to output a quantitative prediction effectiveness index containing estimated market share and estimated profit margin by simulating the game response of the virtual competitor agent.

[0031] Preferably, the closed-loop verification unit is configured to use a cosine similarity algorithm to calculate the similarity between the user's subsequent behavior and the compliance competition strategy in the feature space.

[0032] This invention provides a method and system for pushing live streaming competition compliance policies based on behavior-driven recommendations. It has the following beneficial effects:

[0033] 1. This invention acquires multidimensional behavioral data streams from users and constructs time-series feature vectors. It innovatively combines Mahalanobis distance from the individual behavioral baseline model with Euclidean distance from the competitive ecosystem compliance behavior paradigm to calculate a dual deviation, thereby generating a quantitative risk score in the nascent stage. This approach eliminates the reliance on illegal keywords or post-event events, achieving early, proactive, and highly personalized accurate identification of potential compliance risks.

[0034] 2. This invention dynamically generates a multi-agent game sandbox based on user individual models and competitive ecosystem configuration simulation parameters when identifying risks in the early stages. It also provides dual-dimensional guidance on compliance and effectiveness by utilizing a built-in compliance rule inference engine and game analysis unit. This allows users to deduce strategies in an isolated simulation environment, effectively bridging the gap between compliance policies and actual operations, and achieving an interactive and contextualized flexible guidance effect.

[0035] 3. This invention uses a cosine similarity algorithm to monitor the feature space similarity between the user's subsequent real behavior and the generated compliance strategy after the user completes the sandbox simulation. After successful verification, the strategy is updated to the user's individual behavior baseline model. This constructs a complete technical closed loop from risk perception and guided simulation to effect verification, and realizes the technical effects of adaptive learning of the governance system and self-optimization of personalized models. Attached Figure Description

[0036] Figure 1 This is a flowchart of the present invention;

[0037] Figure 2 This is a system architecture diagram of the present invention. Detailed Implementation

[0038] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] Example 1:

[0040] Please see the appendix Figure 1 This invention provides a method for pushing live streaming competition compliance policies based on behavior-driven recommendations, specifically including the following steps:

[0041] Step S1: Acquisition of multidimensional behavioral data stream and construction of temporal feature vectors

[0042] This step aims to transform the unstructured, multimodal real-time data generated by users during live streaming into structured data that can be processed by computers.

[0043] First, the system's data acquisition and processing unit obtains various real-time data streams during the user's live broadcast through an interface. These data streams include at least:

[0044] Audio data stream: Contains all of the user's voice content during the live stream.

[0045] Transaction data stream: Includes user actions on product links, such as price setting, price modification, inventory adjustment, and discount strategy configuration.

[0046] Text data stream: Includes live stream comment section, bullet comments, and Q&A interaction text between users and viewers.

[0047] After acquiring the data stream, perform real-time feature engineering processing:

[0048] Audio processing: Automatic speech recognition technology is used to transcribe audio data streams into text in real time. Subsequently, the natural language processing module processes the transcribed text to extract speech features, including keyword frequency, sentiment polarity score, and syntactic complexity.

[0049] Transaction processing: Perform statistical analysis on transaction event data to extract transaction characteristics, such as the frequency of price modifications per unit time, the distribution of price adjustment magnitudes, the complexity of coupon settings, and discount structure parameters.

[0050] Text processing: Analyze the comment section and Q&A text to extract interaction features, such as the proportion of negative comments from viewers and the frequency of questions about specific products.

[0051] Finally, the extracted verbal features, transaction features, and interaction features are normalized and concatenated in a preset order to form a unified, high-dimensional user behavior temporal feature vector. ,in Represents a specific user, and t represents a point in time. It is the mathematical expression of a snapshot of a user's behavior at time t.

[0052] Step S2: Calculation of risk score in the budding stage

[0053] The core of this step is to objectively assess the risk level of a user's current behavior deviating from normal and compliant status through quantitative calculation. This calculation does not rely on known and explicit violations, but focuses on abnormal behavioral patterns.

[0054] The computation process consists of two parallel parts:

[0055] Individual deviation calculation:

[0056] The system for each user Maintain an individual behavior baseline model This model is obtained through statistical learning of a large number of time-series feature vector sequences of users under normal operating conditions, specifically derived from the mean vector of historical behavior. Covariance Matrix constitute.

[0057] Individual deviation The calculation is performed using Mahalanobis distance, and the formula is as follows:

[0058]

[0059] The advantage of using Mahalanobis distance is that it can take into account the correlation between different features, thus more accurately measuring a new observation point. The true difference between the distribution of historical behavior patterns and the actual differences effectively identifies abrupt changes in users' intrinsic behavior.

[0060] Ecological deviation calculation:

[0061] The system also maintains a competitive ecosystem compliance behavior paradigm. The system first determines a competitive ecosystem peer group for user U based on dimensions such as product category and number of users. Then, the system filters out users with good compliance records and healthy operations from this group, and aggregates the behavioral feature vectors of these users to calculate an aggregated behavioral mean vector. This serves as an ideal paradigm for compliance behavior within this competitive ecosystem.

[0062] Ecological deviation The Euclidean distance is used to measure the direct distance between the user's current behavior and the ideal paradigm.

[0063]

[0064] This metric is used to identify discrepancies between user behavior and generally healthy competitive practices in the industry.

[0065] Finally, the two deviations are weighted according to preset weighting coefficients. and Perform a weighted summation to generate a risk score for the nascent stage. This score combines the stability of individual behavior with the compliance of group behavior, achieving a comprehensive assessment of potential risks.

[0066] Step S3: Dynamic generation and configuration of the multi-agent game sandbox

[0067] When the budding stage risk score is calculated in S2 Exceeding a dynamically adjustable preset risk threshold When the system identifies a risk requiring intervention, it automatically triggers this step.

[0068] The purpose of this step is to create an isolated, highly contextualized simulation environment for users to conduct risk-free strategy simulations.

[0069] Risk Qualitative Assessment:

[0070] The system first analyzes the specific feature dimensions that constitute the high-risk score. For example, if the time series feature vector... If the value of the dimension related to the frequency of price modifications is abnormally high, the system will refer to this signal by consulting a preset risk type library, i.e., a key-value pair mapping table, and determine that the risk type is price strategy risk.

[0071] Simulation parameter configuration:

[0072] This is key to binding the sandbox to the user's real-world context; the system will trigger business context data when risks occur, such as the user's individual behavioral baseline model. Compliance behavior paradigm in the competitive ecosystem The product information currently being explained, real-time prices, etc., are loaded into the sandbox environment as initial parameters, which ensures that the starting point of the deduction in the sandbox is completely consistent with the user's actual business status.

[0073] Virtual agent instantiation:

[0074] Based on the determined risk type, the system retrieves and instantiates one or more virtual competitor agents from a pre-set strategy model library. This model library stores a variety of encoded, deterministic competitive strategy models. For example, for price strategy risk, the system may instantiate a conservative pricer agent and an aggressive promoter agent. These agents are program entities that can act autonomously in the sandbox according to built-in strategy logic.

[0075] Through the above process, the system dynamically generates a multi-agent game sandbox on the computer server that is exclusive to the user, targets specific risks, and has a game environment.

[0076] Step S4: Interactive processing and correction of strategy scheme data output

[0077] In the sandbox environment, users submit their own strategy proposals for the core simulation tasks given by the system. The proposal is a structured data object containing various operational parameters envisioned by the user.

[0078] After receiving the strategy proposal, the system performs parallel two-dimensional data processing:

[0079] Compliance processing: The built-in compliance rule reasoning engine verifies the data of the strategy plan. This engine contains a series of formalized logical rules, such as the reference original price of the promotional price must be the lowest transaction price within the seven days prior to this promotion. If the verification fails, the engine will output a structured violation point, that is, indicate which specific parameter is non-compliant and the correction instruction, that is, provide a compliant parameter range or suggested value.

[0080] Effectiveness processing: The built-in game analysis unit takes the user's strategy as input and drives the virtual competitor agent in the sandbox to execute its game response. After one or more rounds of simulated confrontation, the unit calculates and outputs quantitative predictive effectiveness indicators of the user's strategy under this competitive landscape, such as estimated market share and estimated profit margin.

[0081] The outputs of the two processing units are integrated into a unified set of correction data and returned to the user interface. This data not only tells the user where the error occurred, but also objectively demonstrates how to improve compliance and the possible market effects of different compliance solutions.

[0082] Step S5: Application Validation and Closed-Loop Update of Compliance Competition Strategies

[0083] This step is the closed-loop part of achieving system self-optimization.

[0084] First, based on the corrected data output by S4, the system provides operational support on the user interface, such as highlighting suggested parameters and automatically filling in compliance-related script templates, to help users easily adopt and generate the final compliance competition strategy. , It is also a feature vector recorded by the system.

[0085] Subsequently, the system enters a post-hoc observation period, continuously monitoring user behavior during the actual live stream, i.e., obtaining subsequent real-time feature vectors. .

[0086] Next, application verification was conducted: the system uses the cosine similarity algorithm to calculate the true behavior vector. Compliance competition strategy vectors generated in the sandbox The system measures the similarity between the two pairs of data. When the similarity score is greater than a preset verification threshold that indicates a high degree of correlation, the system determines that the verification for this guidance is successful.

[0087] Finally, model update: Once the verification is successful, the system will perform two update operations:

[0088] Will As a new and validated sample of compliant behavior, it is integrated into the user individual behavior baseline model. In the update calculation, its mean vector is adjusted. Covariance Matrix This measure can prevent the system from issuing false alarms about compliance behaviors that users have already mastered in the future.

[0089] The anonymized and structured data from this successful case will be fed back to the backend to optimize compliance behavior paradigms in the competitive ecosystem. The calculation and preset strategy model library content is used to improve the overall governance efficiency of the entire system.

[0090] Example 2:

[0091] Please see the appendix Figure 2 This invention provides a live streaming competition compliance policy push system based on behavior-driven recommendations. In terms of hardware, it includes general computer components such as a processor, memory, and network interface. The memory stores a computer program, and when the processor executes the program, it implements the steps of the method described in Embodiment 1 of this invention. In terms of software architecture, the system includes the following core modules:

[0092] 1. Data Acquisition and Processing Unit

[0093] This unit is the perception foundation of the entire system, responsible for extracting structured behavioral features from multi-source heterogeneous live streaming data.

[0094] Functional components:

[0095] Logically, this unit consists of multiple data processing pipelines, each targeting a specific type of data stream, such as an audio processing pipeline, a transaction processing pipeline, and a text processing pipeline.

[0096] Technical principle:

[0097] It connects to the underlying services of the live streaming platform through a network interface, subscribing to users' real-time audio streams, transaction event bus, and comment and bullet screen message queues.

[0098] To ensure real-time processing and scalability, this unit can be built on a streaming computing framework that processes data immediately upon receipt, rather than waiting for data batches.

[0099] Each pipeline is equipped with a corresponding feature extraction algorithm. For example, the audio pipeline integrates an automatic speech recognition engine and a natural language processing toolkit; the trading pipeline executes time series analysis algorithms to calculate statistics such as the frequency and magnitude of price changes.

[0100] To control computational resource consumption and focus on recent behavior, this unit employs a sliding time window mechanism. The system only processes and retains data streams within the most recent time window, such as the last 5 minutes, to ensure the real-time performance and relevance of the computation.

[0101] The final output of this unit is a standardized temporal feature vector representing a snapshot of the user's behavior at time point t. .

[0102] 2. Risk Perception Unit

[0103] This unit is the system's decision-making center, responsible for quantitatively assessing the potential risks of user behavior.

[0104] Functional components:

[0105] This unit mainly includes an individual deviation calculation submodule, an ecological deviation calculation submodule, and a risk score synthesis submodule.

[0106] Technical principle:

[0107] This unit connects to a model library that stores baseline models of each user's individual behavior. It includes the mean vector Covariance Matrix And the competitive ecosystem compliance paradigms under each category. .

[0108] When a new time series feature vector is received Then, the individual deviation calculation submodule calls the corresponding By performing Mahalanobis distance calculation, the individual deviation is obtained. .

[0109] Meanwhile, the ecological deviation calculation submodule calls the corresponding [module name] based on the user's category. Ecological deviation is obtained by performing Euclidean distance calculation. .

[0110] After obtaining the two deviation values, the risk score synthesis submodule calculates the deviation score based on the preset weights stored in the configuration center. and Perform a weighted summation operation to generate the final nascent stage risk score. .

[0111] This unit outputs a floating-point risk score and compares it with a dynamically configurable preset risk threshold. The scores are compared, and if the score exceeds the threshold, a risk trigger signal is generated and passed to the sandbox generation unit.

[0112] 3. Sandbox Generation Unit

[0113] This unit is one of the core execution units of this invention, responsible for dynamically building a personalized strategy deduction environment for users.

[0114] Functional components:

[0115] This unit includes a risk assessment submodule, a parameter configuration submodule, and a virtual agent instantiation submodule.

[0116] Technical principle:

[0117] After receiving a risk trigger signal, the risk qualitative submodule analyzes the time-series feature vector that caused the risk score to exceed the limit. The specific dimensions are determined, and a risk type library is queried. This library is a lookup table that implements the mapping relationship between the combination of feature dimensions and risk semantic tags. Through this mapping, the specific type of risk is determined.

[0118] The parameter configuration submodule is responsible for the initialization of the sandbox. It obtains the user's individual behavior baseline model and competitive ecosystem compliance behavior paradigm from the risk perception unit, and obtains the user's current business context data from other data services. This data is formatted and used as initial parameters to configure the simulation environment to be generated, ensuring that the starting point of the simulation is consistent with the user's real situation.

[0119] This submodule connects to a pre-defined strategy model library that has been indexed and optimized. Based on the risk type determined by the risk qualitative submodule, it efficiently retrieves the corresponding competitive strategy model and uses these models as a blueprint to instantiate one or more virtual competitor intelligent agent program objects in the server memory. Each intelligent agent object encapsulates a defined behavioral logic.

[0120] The output of this unit is an instance handle or access address of a multi-agent game sandbox that is successfully running on a computer server and contains specific initial parameters and active virtual agents.

[0121] 4. Strategy Guidance Unit

[0122] This unit serves as the interface for deep interaction between the user and the system, responsible for providing real-time, two-dimensional guidance within the sandbox.

[0123] Functional components:

[0124] This unit mainly consists of a compliance rule inference engine and a game analysis unit.

[0125] Technical principle:

[0126] This unit receives structured strategy scheme data submitted by users on the client through an API interface.

[0127] This is an expert system built on formal logic. It has a built-in set of updatable, structured compliance rules. When a policy proposal is received, the engine matches each rule and performs logical verification. If an inconsistency is found, it generates data containing violation parameter identifiers and structured correction suggestions as part of the compliance guidance.

[0128] This unit controls the simulation process within the sandbox. It broadcasts the user's strategy to all virtual agents and triggers these agents to execute their internal strategy models to engage in game-playing responses. After the simulation ends, the unit collects the simulation results and calculates quantitative predictive effectiveness indicators, such as estimated market share and estimated profit margin, as part of the effectiveness guidance.

[0129] This unit integrates compliance guidance and effectiveness guidance data into a unified set of correction data, which is then returned to the front-end user interface for display via an interface.

[0130] 5. Closed-loop verification unit

[0131] This unit is crucial for ensuring the system has self-learning and self-optimization capabilities, and is responsible for verifying the guidance effect and driving model updates.

[0132] Functional components:

[0133] This unit includes a behavior tracking submodule, a similarity calculation submodule, and a model update driving submodule.

[0134] Technical principle:

[0135] Based on the revised data, users ultimately determine and generate a compliant competitive strategy. Then, the unit receives and stores the feature vector of this strategy.

[0136] The behavior tracking submodule continuously acquires the user's real time-series feature vector from the data acquisition and processing unit within a preset observation period after the strategy is generated. .

[0137] After receiving the actual subsequent behavior vector, the similarity calculation submodule immediately uses the cosine similarity algorithm to calculate its similarity to the stored compliance competition strategy vector. The cosine value of the angle between the vectors in the vector space.

[0138] When the calculated similarity value exceeds a preset threshold indicating high correlation, the submodule determines that the verification is successful and generates a model update instruction. This instruction drives the risk perception unit to call the corresponding algorithm to update the model. As a new training sample, the user's individual behavior baseline model is updated. At the same time, the anonymized data of this success case is also sent to the backend model training service for regular optimization of the competitive ecosystem compliance behavior paradigm and the preset strategy model library.

[0139] The main output of this unit is to issue model update instructions to other modules in the system, thus forming a complete data and logic closed loop.

[0140] Through the collaborative work of the above modules, the system of this embodiment constitutes a complete technical solution that can perceive risks, generate personalized guidance environments, provide interactive simulations, and ultimately verify effects to achieve self-optimization.

[0141] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A live competition compliance policy pushing method based on behavior main promotion, characterized in that, The method comprises the following steps: S1, acquiring multi-dimensional behavior data flow of a user in a live broadcast process and competitive ecological compliance behavior, and constructing a time sequence feature vector of current behavior of the user and establishing a user individual behavior baseline model in real time; S2, calculating Mahalanobis distance between the current behavior of the user and the user individual behavior baseline model and Euclidean distance between the current behavior of the user and the competitive ecological compliance behavior norm based on the time sequence feature vector, and weighting and summing the Mahalanobis distance and the Euclidean distance to generate a sprouting period risk score, and setting a preset risk threshold interval for the sprouting period risk score; S3, triggering generation of a multi-agent game sandbox running on a computer server when the sprouting period risk score exceeds the preset risk threshold; S4, processing a strategy scheme submitted by the user in the multi-agent game sandbox and generating a processing result, and then pushing a compliance policy guidance scheme including a compliance correction suggestion and an effectiveness prediction index to the user according to the processing result; S5, generating a compliance competition strategy for the user based on the compliance policy guidance scheme, constructing a time sequence feature vector of subsequent behavior based on the same multi-dimensional behavior data flow processing method as step S1, calculating a feature space similarity between the time sequence feature vector of the subsequent behavior and the compliance competition strategy, and determining that verification is successful and updating the individual behavior baseline model when the similarity is greater than a preset threshold, wherein the preset threshold is determined by an execution personnel experience.

2. The live competition compliance policy push method based on behavior primary push according to claim 1, characterized in that, The individual behavior baseline model comprises a mean vector and a covariance matrix of historical behavior of the user; and the competitive ecological compliance behavior norm comprises an aggregated behavior mean vector of a peer group of the competitive ecology in which the user is located. 3.The live competition compliance policy pushing method based on behavior main pushing according to claim 1, characterized in that, The step of generating the multi-agent game sandbox specifically comprises: mapping a feature dimension exceeding a baseline range in the time sequence feature vector to a preset risk type library to determine a risk type; retrieving and instantiating one or more virtual competitor agents from a preset strategy model library according to the risk type, wherein the virtual competitor agents have a determined and executable strategy model.

4. The live competition compliance policy push method based on behavioral primary push of claim 1, wherein, The step of pushing the compliance policy guidance scheme including the compliance correction suggestion and the effectiveness prediction index to the user specifically comprises: verifying the strategy scheme by using a compliance rule reasoning engine and outputting a rule violation point and a structured correction instruction; simulating a game response of the virtual competitor agent by using a game analysis unit and outputting a quantitative prediction effectiveness index of the strategy scheme.

5. The live competition compliance policy push method based on behavioral primary push of claim 4, wherein, The quantitative prediction effectiveness index comprises an estimated market share and an estimated profit rate.

6. The live competition compliance policy push method based on behavioral primary push of claim 1, wherein, In step S1, the multi-dimensional behavior data flow is processed by using a sliding time window mechanism to ensure real-time data processing.

7. The live competition compliance policy push method based on behavioral priming of claim 1, wherein, In step S2, the Mahalanobis distance and the Euclidean distance are weighted and summed by using preset individual deviation weight coefficients and ecological deviation weight coefficients for weighted calculation.

8. The live competition compliance policy push method based on behavioral primary push of claim 3, wherein, In step S5, the feature space similarity between the time sequence feature vector of the subsequent behavior and the compliance competition strategy is calculated by using a cosine similarity algorithm.

9. The live competition compliance policy push method based on behavioral priming of claim 1, wherein, The preset risk threshold in the step S3 is a dynamic threshold, and a value of the dynamic threshold is periodically adjusted according to overall compliance performance of a competitive ecological peer group to which the user belongs.

10. A live-streaming competition compliance policy push system based on behavior-driven recommendations, characterized in that: Comprise: A data acquisition and processing unit configured to acquire a multi-dimensional behavior data stream of the user in a live broadcast process, construct a time sequence feature vector of a current behavior of the user in real time, and establish an individual behavior baseline model of the user; A risk perception unit configured to calculate Mahalanobis distance between the current behavior of the user and the individual behavior baseline model of the user and Euclidean distance between the current behavior of the user and a compliance behavior paradigm of a competitive ecology based on the time sequence feature vector, and to generate a sprouting stage risk score by weighted summation of calculation results of the two distances; A sandbox generation unit configured to trigger generation of a multi-agent game sandbox running on a computer server when the sprouting stage risk score exceeds a preset risk threshold, and to configure simulation parameters of the multi-agent game sandbox based on the individual behavior baseline model and the compliance behavior paradigm of the competitive ecology; A strategy guidance unit configured to guide the user to submit a strategy scheme in the multi-agent game sandbox, and to output modified data containing two-dimensional guidance of compliance and effectiveness by data processing of an embedded compliance rule reasoning engine and a game analysis unit; A closed-loop verification unit configured to support the user to generate a compliance competitive strategy according to the modified data, and to monitor similarity of subsequent behaviors of the user and the compliance competitive strategy in a feature space, and to determine verification success and drive the risk perception unit to update the individual behavior baseline model when the similarity is greater than a preset threshold.

Images