A user tagging rule management method and system supporting dynamic policies

Abstract

The application provides a user tagging rule management method and system supporting dynamic strategy, relates to the technical field of user portrait analysis, performs multi-modal feature fusion on user behavior data flow and environmental context data to obtain a dynamic user portrait vector, analyzes a strategy intention of a user's business target to obtain a strategy parameter set, constructs a strategy evolution graph based on a semantic association network and historical strategies, performs explicit rule reasoning and implicit demand mining on the dynamic user portrait vector and the strategy parameter set through the strategy evolution graph to obtain a candidate strategy package, performs simulation prediction on the candidate to obtain a deduction performance matrix, calibrates the confidence of the deduction performance matrix to obtain a dynamic tagging rule, and further updates the weight of nodes and associated edges in the strategy evolution graph, so that the application can realize closed-loop feedback of the user tagging strategy based on iterative optimization and confidence guarantee of the strategy evolution graph, thereby improving the dynamic adaptability of the user portrait.

Description

Technical Field

[0001] This application relates to the field of user profiling analysis technology, and more specifically, to a method and system for managing user tagging rules that supports dynamic strategies. Background Technology

[0002] User profiling analysis technology serves as a key support for precision marketing and personalized services. By integrating multi-source information such as user behavior data and environmental context, and by building models that describe user characteristics and preferences, it can provide data support for business strategy formulation. Among these, user tagging rule management undertakes the core task of converting abstract user characteristics into structured tags, which directly determines the accuracy of user profiles.

[0003] User tagging rule management can be implemented based on predefined rule templates or static machine learning models, which can meet basic needs in specific scenarios. However, when facing dynamically changing user behavior and business environments, the following shortcomings become apparent: First, rule updates rely on human experience intervention, resulting in response lag and subjective bias. Second, the lack of systematic mining of the correlations and evolutionary patterns between historical strategies makes it difficult to achieve continuous accumulation and reuse of strategy knowledge. Third, traditional technologies cannot effectively identify implicit user needs and lack reliable performance prediction and risk assessment mechanisms before strategy deployment. Finally, existing technical solutions generally lack a collaborative optimization loop that integrates machine intelligence and expert experience, leading to insufficient transparency and limited adaptability in the strategy decision-making process. Therefore, how to achieve closed-loop feedback of user tagging strategies based on iterative optimization and confidence guarantees of strategy evolution graphs to improve the dynamic adaptability of user profiles is a challenge facing the industry. Summary of the Invention

[0004] This application provides a user tagging rule management method and system that supports dynamic strategies. It can realize closed-loop feedback of user tagging strategies based on iterative optimization of strategy evolution graph and confidence guarantee, so as to improve the dynamic adaptability of user profiles.

[0005] Firstly, this application provides a user tagging rule management method supporting dynamic strategies, the rule management method comprising the following steps:

[0006] Collect user behavior data streams and environmental context data;

[0007] Multimodal feature fusion is performed on the user behavior data stream and the environmental context data to obtain a dynamic user profile vector, and the user's business objectives are analyzed to obtain a set of policy parameters.

[0008] The historical strategies for user tagging are obtained, and a strategy evolution graph is constructed based on the semantic association network and the historical strategies. The strategy evolution graph is used to perform explicit rule reasoning and implicit demand mining on the dynamic user profile vector and the strategy parameter set to obtain a candidate strategy package.

[0009] The candidate strategy package is placed in an isolated strategy sandbox environment for simulation and prediction to obtain the inference performance matrix. The confidence of the inference performance matrix is ​​calibrated through a human-machine collaborative optimization loop to obtain dynamic labeling rules.

[0010] The weights of nodes and associated edges in the strategy evolution graph are updated based on the execution feedback of the dynamic labeling rules.

[0011] In this embodiment, user behavior data streams and environmental context data are collected through a distributed log collection framework.

[0012] In this embodiment, the multimodal feature fusion of the user behavior data stream and the environmental context data to obtain the dynamic user profile vector specifically includes:

[0013] The user behavior data stream is serialized and encoded to obtain a behavior embedding vector;

[0014] Spatiotemporal features are extracted from the environmental context data to obtain a context feature vector;

[0015] The behavior embedding vector and the context feature vector are aligned and fused using a cross-modal attention mechanism to obtain a dynamic user profile vector.

[0016] In this embodiment, the strategy intent parsing of the user's business objectives yields a set of strategy parameters, specifically including:

[0017] Semantic parsing of business target text is performed using natural language processing technology to obtain the structured intent of the business target;

[0018] Based on a multi-objective optimization algorithm, constraint reasoning is performed on the business rule base and the structured intent to obtain a set of strategy parameters.

[0019] In this embodiment, the historical tagging strategy of users is obtained through a distributed repository.

[0020] In this embodiment, constructing a policy evolution graph based on a semantic association network and the historical policies specifically includes:

[0021] The historical strategies are subjected to feature extraction and vectorization representation to obtain a set of strategy feature vectors;

[0022] A semantic association network between labeling strategies is constructed by using a graph neural network, and then a time series analysis algorithm is used to perform strategy evolution on the semantic association network to obtain a set of strategy evolution relationships.

[0023] The policy evolution relationship set and the policy feature vector set are fused to obtain the policy evolution map.

[0024] In this embodiment, the candidate strategy package is obtained by performing explicit rule reasoning and implicit requirement mining on the dynamic user profile vector and the strategy parameter set through the strategy evolution graph, specifically including:

[0025] Based on the strategy evolution graph, subgraph matching is performed on the dynamic user profile vector to obtain a set of explicit rules;

[0026] The policy parameter set is expanded by using a graph neural network embedding propagation algorithm and the policy evolution graph to obtain a set of implicit requirements.

[0027] The explicit rule set and the implicit requirement set are combined and sorted to obtain a candidate strategy package.

[0028] In this embodiment, the candidate policy package is placed in an isolated policy sandbox environment for simulation and prediction, thereby obtaining the inference performance matrix, specifically including:

[0029] Construct a sandbox environment that includes virtual user groups and simulated environments;

[0030] The candidate strategy package is deployed to the sandbox environment instance, and the strategy execution process data is obtained by Monte Carlo simulation.

[0031] The process data is quantitatively analyzed to obtain an efficiency score, which in turn yields a deduced efficiency matrix.

[0032] In this embodiment, the human-machine collaborative optimization loop is used to realize the parameter correction interaction between machine intelligence and human intervention under a visual decision-making interface.

[0033] Secondly, this application provides a user tagging rule management system supporting dynamic strategies, used to execute a user tagging rule management method supporting dynamic strategies, the rule management system comprising:

[0034] The data acquisition module is used to collect user behavior data streams and environmental context data;

[0035] The strategy parsing module is used to perform multimodal feature fusion on the user behavior data stream and the environmental context data to obtain a dynamic user profile vector, and to perform strategy intent parsing on the user's business objectives to obtain a set of strategy parameters.

[0036] The strategy reasoning and mining module is used to obtain the historical strategies of user tagging, construct a strategy evolution graph based on the semantic association network and the historical strategies, and perform explicit rule reasoning and implicit requirement mining on the dynamic user profile vector and the strategy parameter set through the strategy evolution graph to obtain a candidate strategy package.

[0037] The sandbox inference module is used to place the candidate strategy package in an isolated strategy sandbox environment for simulation and prediction, thereby obtaining the inference performance matrix. The confidence of the inference performance matrix is ​​calibrated through a human-machine collaborative optimization loop to obtain dynamic labeling rules.

[0038] The feedback update module is used to update the weights of nodes and associated edges in the strategy evolution graph through the execution feedback of the dynamic labeling rules.

[0039] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects:

[0040] The process involves: collecting user behavior data streams and environmental context data; performing multimodal feature fusion on the user behavior data streams and environmental context data to obtain dynamic user profile vectors; parsing the user's business objectives to obtain a set of policy parameters; acquiring historical policies for user tagging; constructing a policy evolution graph based on a semantic association network and the historical policies; using the policy evolution graph to perform explicit rule reasoning and implicit requirement mining on the dynamic user profile vectors and the set of policy parameters to obtain candidate policy packages; placing the candidate policy packages in an isolated policy sandbox environment for simulation prediction to obtain an inference performance matrix; calibrating the confidence of the inference performance matrix through a human-machine collaborative optimization loop to obtain dynamic tagging rules; and updating the weights of nodes and associated edges in the policy evolution graph through the execution feedback of the dynamic tagging rules.

[0041] Therefore, this application demonstrates that closed-loop feedback of user tagging strategies can be achieved based on iterative optimization and confidence guarantees of the strategy evolution graph. Firstly, through multi-source data collection and multi-modal feature fusion, a comprehensive perception of user behavior patterns and environmental context can be realized, generating dynamic user profile vectors that reflect real-time changes in user state. Secondly, through strategy intent parsing, fuzzy business objectives are transformed into a quantifiable set of strategy parameters, laying a precise data foundation for subsequent intelligent decision-making. Thirdly, by constructing a strategy evolution graph that integrates semantic associations and evolutionary relationships, deep reasoning based on historical strategy knowledge can be performed. Validated explicit rules can be quickly identified through subgraph matching, and graph neural networks can be used to further enhance the algorithm. By uncovering potential implicit needs, the strategy generation process combines empirical reliability with exploratory innovation, which helps improve the breadth and depth of strategy coverage. Then, by conducting Monte Carlo simulations and multi-dimensional performance predictions on candidate strategies in a strategy sandbox environment, their short-term effects and long-term impacts can be quantitatively evaluated. Combined with confidence calibration of the human-machine collaborative optimization loop, the computational power of machine intelligence and the business knowledge of expert experience are effectively integrated, achieving a balance between risk control and innovative breakthroughs in the strategy decision-making process. Finally, by dynamically updating the weights and reconstructing the topology of the strategy evolution graph through execution feedback, a continuous learning path from strategy practice to knowledge accumulation is established, which is conducive to the continuous optimization of the tagging strategy.

[0042] In summary, the technical solution adopted in this application can realize closed-loop feedback of user tagging strategy based on iterative optimization of strategy evolution graph and confidence guarantee, so as to improve the dynamic adaptability of user profile. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of this application 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 for this embodiment of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart of a user tagging rule management method supporting dynamic strategies provided in this application;

[0045] Figure 2 This is an exemplary flowchart for determining dynamic user profile vectors according to the present application;

[0046] Figure 3 This is an exemplary flowchart based on the strategy evolution map provided in this application;

[0047] Figure 4 This is a module structure diagram of a user labeling rule management system that supports dynamic strategies, provided in this application. Detailed Implementation

[0048] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0049] This application provides a user tagging rule management method and system supporting dynamic strategies. Its core is to collect user behavior data streams and environmental context data; perform multimodal feature fusion on the user behavior data streams and environmental context data to obtain dynamic user profile vectors, and perform strategy intent parsing on user business objectives to obtain a set of strategy parameters; acquire historical user tagging strategies, construct a strategy evolution graph based on a semantic association network and the historical strategies, and perform explicit rule reasoning and implicit requirement mining on the dynamic user profile vectors and the strategy parameter set through the strategy evolution graph to obtain candidate strategy packages; place the candidate strategy packages in an isolated strategy sandbox environment for simulation prediction to obtain an inference performance matrix; calibrate the confidence of the inference performance matrix through a human-machine collaborative optimization loop to obtain dynamic tagging rules; and update the weights of nodes and associated edges in the strategy evolution graph through the execution feedback of the dynamic tagging rules.

[0050] Example 1: To better understand the above technical solution, the following will provide a detailed description of the technical solution in conjunction with the accompanying drawings and specific implementation methods. (Refer to...) Figure 1 As shown in the figure, this is a flowchart of a user tagging rule management method supporting dynamic strategies according to this embodiment of the present application. The rule management method includes the following steps:

[0051] In step S1, user behavior data streams and environmental context data are collected.

[0052] In practical implementation, a distributed log collection framework is used to collect user behavior data streams and environmental context data. This framework, comprised of multiple collaborative software components, is used to collect and transmit user behavior and environmental data from different data sources in real time. The distributed log collection framework includes: a lightweight data collection library, a log agent, an environmental sensor access component, and a data buffer pipeline. In actual implementation, the lightweight data collection library can be deployed on the web using a JavaScript-based event tracking SDK to capture user interaction behavior; the log agent can monitor changes in server log files to collect API call records in real time; the environmental sensor access component is used to obtain status information about the user's environment; and the data buffer pipeline is used to receive and temporarily store data streams from various collection points.

[0053] It should be noted that the user behavior data stream in this application refers to a sequence of user operation records arranged in chronological order. User operations include attributes such as event type, occurrence time, page position, and element identifier, which are used to reconstruct the user's complete operation path and preference characteristics. Environmental context data is auxiliary information describing the environmental state in which the user operates, including variables such as spatial dimension, device dimension, network dimension, and time dimension. Environmental context data can provide important reference for understanding the background conditions in which user behavior occurs.

[0054] In step S2, multimodal feature fusion is performed on the user behavior data stream and the environmental context data to obtain a dynamic user profile vector, and the user's business objectives are analyzed to obtain a set of policy parameters.

[0055] Preferably, in this embodiment, reference Figure 2 As shown, this diagram is an exemplary flowchart for determining a dynamic user profile vector according to the present application. In this embodiment, the dynamic user profile vector is obtained by multimodal feature fusion of the user behavior data stream and the environmental context data through the following steps:

[0056] First, in step S21, the user behavior data stream is serialized and encoded to obtain a behavior embedding vector;

[0057] Then, in step S22, spatiotemporal features are extracted from the environmental context data to obtain a context feature vector;

[0058] Finally, in step S23, the behavior embedding vector and the context feature vector are aligned and fused using a cross-modal attention mechanism to obtain a dynamic user profile vector.

[0059] In practical implementation, firstly, a sequence encoder in a deep learning framework can be used to model the user behavior data stream. Each behavior event in the user behavior data stream is converted into a fixed-dimensional initial vector through an embedding layer, and a positional encoding parameter is added to preserve the temporal order information of the behavior. Then, the vector output after processing by multiple layers of encoders is used as the behavior embedding vector. Next, a convolutional neural network is used to process the spatial dimension variables in the environmental context data. Spatial features are extracted through multiple layers of convolution, and fully connected layers are used to process variable features such as device dimension, network dimension, and time dimension. Dimensionality reduction processing of spatial features and processed variable features is performed through multiple fully connected layers to generate a fixed-dimensional context feature vector. Finally, feature alignment and fusion can be performed through a cross-modal attention mechanism. That is, the behavior embedding vector and the context feature vector are used as the query vector and key vector of the attention mechanism, respectively. The attention weight matrix between the query vector and the key vector is calculated, and the behavior embedding vector is weighted and fused according to the weight matrix. Then, the fused vector and the context feature vector are concatenated and normalized through a fully connected layer. Finally, the numerical vector output by the fully connected layer is used as the dynamic user profile vector.

[0060] It should be noted that in this embodiment, the behavior embedding vector is a numerical representation obtained by transforming the user behavior sequence, which can capture the user's operating habits and interests; the context feature vector refers to the feature representation of the user's environment and device state, used to describe the background conditions in which the user's behavior occurs; the cross-modal attention mechanism refers to a computing mechanism that can automatically learn the degree of correlation between different data modalities, used to achieve intelligent fusion of multi-source features; the dynamic user profile vector refers to a real-time representation that integrates behavioral features and environmental features, which can dynamically reflect the user's latest state and provide accurate feature input for subsequent strategy decisions.

[0061] In this embodiment, the strategy intent parsing of the user's business objectives to obtain the set of strategy parameters can be performed in the following manner:

[0062] Semantic parsing of business target text is performed using natural language processing technology to obtain the structured intent of the business target;

[0063] Based on a multi-objective optimization algorithm, constraint reasoning is performed on the business rule base and the structured intent to obtain a set of strategy parameters.

[0064] It should be noted that the business objective text refers to business needs or operational goals described by business personnel in natural language, such as specific business demands like "increasing the repurchase rate of high-value users" or "reducing the risk of new user churn"; the business rule base refers to a data set that stores enterprise operational rules and constraints, including business rules such as resource allocation limits, risk control thresholds, and time period restrictions. These rules are mainly derived from the accumulation of historical operational data, industry standards and norms, and management decision-making experience, and are unified into constraints such as resource limits, time windows, and risk thresholds through a rule engine.

[0065] In practical implementation, firstly, semantic parsing of the business target text can be performed based on natural language processing technology. That is, an open-source Chinese pre-trained language model is used to segment and semantically understand the target text input by business personnel (such as "increase the shopping frequency of high-value users"), and the attention mechanism in the Chinese pre-trained language model is used to identify key business entities and action words. The identified keywords are then structured and organized, and a visualization algorithm is used to generate a structured intent containing fields such as core indicators, target objects, and expected effects. Then, a particle swarm optimization algorithm can be used to convert the constraints such as resource limits, time windows, and risk thresholds in the business rule base into boundary conditions of the optimization problem. Multiple business indicators in the structured intent are converted into optimization objective functions. An iterative search algorithm is used to select parameter combinations that can simultaneously optimize multiple objective functions under all constraints, and these parameter combinations are used as the strategy parameter set.

[0066] It should be noted that, in this embodiment, the structured intent is a machine-readable standardized data format converted from the business objectives described in natural language. It contains clear business indicators and expected results and can serve as a bridge connecting business language and technical parameters. The strategy parameter set refers to the data structure of executable parameters, including weight configuration, threshold setting, priority sorting, etc., which can be used to guide strategy generation and help to accurately convert business objectives into implementable technical solutions.

[0067] In step S3, the historical strategies for user tagging are obtained, a strategy evolution graph is constructed based on the semantic association network and the historical strategies, and explicit rule reasoning and implicit requirement mining are performed on the dynamic user profile vector and the strategy parameter set through the strategy evolution graph to obtain a candidate strategy package.

[0068] It should be noted that in this embodiment, the historical strategies for user tagging are obtained through a distributed repository. The distributed repository refers to a strategy data storage system built on a distributed database cluster, comprising three main parts: a strategy metadata database, a strategy execution record repository, and a strategy snapshot storage area. It is used to centrally manage and store various types of strategy data generated during the user tagging process. The historical strategies for user tagging refer to a set of tagging rules executed in the past, including three parts: strategy logic rules, strategy execution effect data, and strategy creation and modification information. The strategy logic rules record the judgment logic of the tagging conditions; the strategy execution effect data includes indicators such as accuracy and coverage; and the strategy creation and modification information records the version evolution process of the strategy. Preferably, historical strategies from the past year to the present can be selected, which helps ensure the timeliness of the strategy data while retaining a sufficiently long historical period to capture the complete pattern of strategy evolution, providing a sufficient data foundation for constructing a strategy evolution map.

[0069] Preferably, in this embodiment, reference Figure 3 As shown, this diagram is an exemplary flowchart of the strategy evolution graph provided in this application. In this embodiment, the construction of the strategy evolution graph based on the semantic association network and the historical strategies can be achieved through the following steps:

[0070] First, in step S31, the historical strategy is subjected to feature extraction and vectorization representation to obtain a set of strategy feature vectors;

[0071] Then, in step S32, a semantic association network between labeling strategies is constructed through a graph neural network, and then a time series analysis algorithm is used to perform strategy evolution on the semantic association network to obtain a set of strategy evolution relationships;

[0072] Finally, in step S33, the set of policy evolution relationships and the set of policy feature vectors are fused to obtain a policy evolution map.

[0073] In specific implementation, firstly, existing text vectorization models can be used to process the textual descriptions of policy rules, converting the conditional statements and action descriptions of each policy into 128-dimensional numerical vectors, and extracting policy execution performance indicators as numerical features, including accuracy, coverage, and execution frequency. Then, the text vectors and numerical features are concatenated and dimensionality reduced using principal component analysis to obtain 64-dimensional policy feature vectors. The set of all policy feature vectors is then used as the policy feature vector set. Next, a graph neural network is used, treating each policy in the historical policies as a graph node and the policy feature vector as a node feature, to calculate the semantic similarity between policies (the similarity can be measured using Euclidean distance). When the similarity exceeds... When a preset threshold is set, a semantic association network is formed by establishing node connection edges through a graph neural network. Then, a time-series analysis algorithm is used to sort the strategies by their creation time, calculate the changing trends of strategy feature vectors in adjacent time periods, identify the derivative relationships and version evolution paths between strategies, and obtain a set of strategy evolution relationships including parent-child relationships, substitution relationships, and enhancement relationships. Finally, the nodes in the set of strategy evolution relationships are mapped to feature vectors in the set of strategy feature vectors, and the evolution relationships are used as directed edges in the graph. The weights of the edges are set according to the relationship type and evolution intensity. The node features and edge relationships are integrated using graph structure data to obtain a graph structure containing node features, edge relationships, and weights. This graph structure is then used as a strategy evolution graph.

[0074] It should be noted that, in this embodiment, the strategy feature vector set refers to the set of features that convert strategy rules and performance indicators into machine-readable features through numerical methods, used to quantify the similarity and correlation between strategies; the semantic association network refers to a network model that represents the semantic relationships between strategies through a graph structure, which can reveal the inherent connections between strategies in terms of function and application scenarios; the strategy evolution relationship set refers to a data structure that records the evolutionary relationships of strategies over time, such as derivation and substitution, reflecting the evolutionary path of the strategy system; and the strategy evolution graph refers to a comprehensive knowledge graph that integrates strategy features, semantic associations, and evolutionary history, providing a complete knowledge foundation for strategy reasoning and optimization.

[0075] In this embodiment, the candidate strategy package is obtained by performing explicit rule reasoning and implicit requirement mining on the dynamic user profile vector and the strategy parameter set through the strategy evolution graph. Specifically, this can be done in the following way:

[0076] Based on the strategy evolution graph, subgraph matching is performed on the dynamic user profile vector to obtain a set of explicit rules;

[0077] The policy parameter set is expanded by using a graph neural network embedding propagation algorithm and the policy evolution graph to obtain a set of implicit requirements.

[0078] The explicit rule set and the implicit requirement set are combined and sorted to obtain a candidate strategy package.

[0079] In practical implementation, firstly, the query function of the graph database can be used to calculate the similarity between the dynamic user profile vector and the feature vectors of all policy nodes in the policy evolution graph. That is, Euclidean distance is used to find policy nodes with a similarity greater than 0.8. Then, a graph traversal algorithm is used to find the connected subgraphs containing these nodes, and the rule definitions and parameter configurations of all policy nodes in the subgraphs are extracted. The extracted rule definitions and parameter configurations are then used as the explicit rule set. Next, a neighborhood expansion is performed using a graph neural network embedding propagation algorithm. That is, the weight configurations in the policy parameter set are used as initial features, and a three-layer neighborhood propagation is performed in the policy evolution graph. A graph convolutional network can be used to compute the similarity of each... The nodes are embedded and represented, and then features are aggregated based on the edge weights between nodes. The new node features obtained after propagation are compared with the original features to filter out nodes with significant feature changes. Then, the strategy rules and parameters corresponding to the nodes with significant feature changes are used as the implicit requirement set. Finally, a multi-objective optimization algorithm is used to combine and sort the explicit rule set and the implicit requirement set. That is, the strategies in the explicit rules and implicit requirements are scored according to three dimensions: effect prediction value, risk coefficient, and implementation cost. The comprehensive score of each strategy is calculated, and the strategies are sorted from high to low. The top 20 strategies are selected to form a candidate strategy list, and the list and its score data set are used as the candidate strategy package.

[0080] It should be noted that the explicit rule set refers to the strategy combination that is highly related to the current user characteristics and is obtained through direct matching, reflecting the clear correspondence between user needs and strategies; the implicit need set refers to the potential related strategy combination discovered through graph reasoning, reflecting the needs that users may not have explicitly expressed; and the candidate strategy package refers to the strategy recommendation set after comprehensive evaluation and ranking, which includes specific strategy content, implementation parameters and priority scores, providing a basis for subsequent strategy verification and selection.

[0081] In step S4, the candidate strategy package is placed in an isolated strategy sandbox environment for simulation and prediction to obtain the inference performance matrix. The confidence of the inference performance matrix is ​​calibrated through a human-machine collaborative optimization loop to obtain dynamic labeling rules.

[0082] In this embodiment, the candidate policy package is placed in an isolated policy sandbox environment for simulation and prediction, thereby obtaining the inference performance matrix. Specifically, this can be achieved in the following manner:

[0083] Construct a sandbox environment that includes virtual user groups and simulated environments;

[0084] The candidate strategy package is deployed to the sandbox environment instance, and the strategy execution process data is obtained by Monte Carlo simulation.

[0085] The process data is quantitatively analyzed to obtain an efficiency score, which in turn yields a deduced efficiency matrix.

[0086] In practice, the process involves several steps. First, containerization technology is used to create an independent testing environment. Based on historical user behavior data, a test group of 10,000 virtual users is generated through random sampling. Each virtual user includes basic attributes, behavioral characteristics, and preference tags. The core functional modules of the online business system are replicated to build a simulation environment, including user interaction and data recording functions, forming a complete testing platform. Next, a configuration management tool is used to batch load the policy rules from the candidate policy package into the policy engine of the sandbox environment. A unified test time window is set, and Monte Carlo simulation is used to execute policy deductions. A random number generator simulates the behavioral responses of different user groups, conducting 1,000 independent simulation experiments. Detailed information such as the policy triggering conditions, execution results, and user feedback are recorded for each experiment to obtain policy execution process data. Finally, statistical analysis tools are used to calculate the performance of each policy on key indicators such as conversion rate and retention rate. An averaging algorithm is used to synthesize the simulation results to obtain the performance score for each policy. The scores of different policies on different indicators are organized into a two-dimensional table, where the rows represent policy numbers and the columns represent evaluation indicators. The filled two-dimensional table serves as the deduction performance matrix.

[0087] It should be noted that, in this embodiment, the sandbox environment refers to a testing platform isolated from the actual environment, which can safely verify the effectiveness of the strategy without affecting real users; Monte Carlo simulation refers to a statistical method that evaluates the effectiveness of the strategy through a large number of random sampling experiments, which can fully take into account the uncertainty of user behavior; process data refers to the information of the entire process of strategy execution recorded in the simulation experiment, which provides a basis for effect evaluation; and the inference performance matrix refers to a data structure that systematically records the predicted performance of each strategy on multi-dimensional indicators, which can provide a quantitative basis for strategy selection.

[0088] In this embodiment, the confidence level of the inference performance matrix is ​​calibrated through a human-machine collaborative optimization loop to obtain the dynamic labeling rules. Specifically, this can be achieved in the following manner:

[0089] The inference efficiency matrix is ​​displayed in multiple dimensions through a visual decision-making interface for human-machine collaborative optimization loop, and comparative analysis results are obtained.

[0090] Obtain parameter correction instructions, and then use the parameter correction instructions to set weight preferences for the comparison analysis results to obtain an expert optimization strategy set;

[0091] The confidence level of the expert optimization strategy set is evaluated and ranked based on the historical decision accuracy to obtain dynamic labeling rules.

[0092] It should be noted that, in this embodiment, the human-machine collaborative optimization loop is used to realize parameter correction interaction between machine intelligence and human experts under a visual decision-making interface. The human-machine collaborative optimization loop refers to a decision optimization mechanism in which the machine intelligence system and human experts interact bidirectionally through a visual interface. Machine intelligence is responsible for providing strategy recommendations and effect predictions based on data analysis, while human experts are responsible for strategy correction and parameter tuning based on business experience and domain knowledge. The two work together to complete strategy optimization decisions through an iterative interaction process. Specifically, during the parameter correction interaction in the visual decision-making interface, machine intelligence first displays the strategy deduction results to the expert in chart form. The expert then adjusts key parameters such as strategy thresholds and weight allocation using parameter adjustment controls on the interface, achieving human-machine collaborative optimization. Preferably, the human-machine collaborative optimization loop in this application is used for initial deployment or system updates. Human-machine collaboration can be performed in real time, and the labeling strategy can be dynamically adjusted by updating the strategy evolution graph.

[0093] In practical implementation, firstly, an interactive interface is developed using web front-end technology to display the various indicators in the performance matrix in the form of data tables and radar charts. The data tables list the specific values ​​of each strategy across different evaluation dimensions, while the radar charts display the relative performance of each strategy across multiple indicators. By comparing the differences between different strategies on key indicators, comparative analysis results between strategies are obtained. Secondly, a slider control and numerical input boxes are provided in the visualization interface, allowing experts to adjust the weight coefficients of each evaluation indicator according to business needs. The system records expert modification instructions for strategy parameters in real time, including adjustments to indicator weights, changes to strategy thresholds, and changes to activation conditions. Based on the modified parameters, the strategy scores are recalculated to obtain the expert's optimized strategy set. Finally, the decision record data is analyzed... The database extracts expert strategy adjustment records from the past three months. Confidence weights are calculated using the following steps: The actual improvement in performance after each strategy adjustment is statistically analyzed. When a key performance indicator (such as conversion rate) improves by more than 10% after the strategy is implemented compared to before the adjustment, the adjustment is considered accurate. The number of accurate adjustments is divided by the total number of adjustments to obtain the accuracy percentage, which is then used as the confidence weight. A weighted average algorithm is used to merge the machine recommendation score and the expert adjustment score according to the confidence weight: the machine recommendation score is multiplied by (1 - confidence weight) plus the expert adjustment score multiplied by the confidence weight, resulting in a confidence score for each strategy. The strategies are then sorted from highest to lowest based on their final scores. The sorted strategy list and the corresponding confidence scores are used as dynamic labeling rules.

[0094] It should be noted that, in this embodiment, the comparative analysis results refer to the performance difference analysis report between strategies presented in a visual manner, which helps experts quickly identify the advantages and disadvantages of strategies; the expert optimized strategy set refers to the set of strategies adjusted by incorporating the business experience of experts, which is conducive to human-machine collaborative decision-making; the dynamic labeling rule refers to the final strategy output after confidence calibration. The dynamic labeling rule includes both specific execution rules and confidence assessment, which can provide a reliable basis for strategy deployment.

[0095] In step S5, the weights of nodes and associated edges in the strategy evolution graph are updated through the execution feedback of the dynamic labeling rules.

[0096] In this embodiment, updating the weights of nodes and associated edges in the strategy evolution graph through the execution feedback of the dynamic labeling rules can be achieved in the following way:

[0097] Collect the execution effect data of the dynamic labeling rules;

[0098] By performing collaborative conflict analysis on the execution effect data through reinforcement learning mechanism, a set of node weights and a set of edge weights are obtained.

[0099] Based on the node weight set and the edge weight set, the weights of the nodes and associated edges in the strategy evolution graph are updated to realize dynamic strategy adjustment of user labeling rules.

[0100] In practice, the process involves several steps. First, core metrics such as tagging accuracy, user acceptance, and conversion rate are extracted from the user's historical tagging strategies to obtain performance data. Then, a reinforcement learning algorithm is used, treating strategy nodes as states, relationships between strategies as actions, and metrics from the performance data as reward signals. The algorithm iteratively calculates the value of each strategy node over multiple rounds, using the reciprocal of the value as the node weight, resulting in a set of node weights. Euclidean distance is then used to calculate the strength coefficient of associated edges, which is used as the edge weight, resulting in a set of edge weights. Finally, the weights of the strategy evolution graph are updated based on these weights. Specifically, the update function of the graph database is used to update the values ​​in the node weight set to the weight attributes of the corresponding strategy nodes, and the values ​​in the edge weight set to the weight attributes of the corresponding associated edges. Topological optimization is performed on nodes and edges with significant weight changes, completing the iteration of the strategy evolution graph and enabling dynamic adjustment of user tagging rules.

[0101] It should be noted that, in this embodiment, the execution effect data refers to the set of effect indicators generated by the strategy running in the actual business environment, reflecting the true performance and value of the strategy; the node weight set refers to the set of strategy importance scores obtained through reinforcement learning, representing the relative value of the strategy in the system; the edge weight set refers to the set of strength coefficients of the relationship between strategies, used to describe the degree of effectiveness of strategy cooperation; in addition, in this application, the weight update mechanism based on execution feedback is conducive to realizing the dynamic learning and self-optimization of the tagging strategy.

[0102] In summary, the technical solution adopted in this application can realize closed-loop feedback of user tagging strategy based on iterative optimization of strategy evolution graph and confidence guarantee, so as to improve the dynamic adaptability of user profile.

[0103] Example 2: This application provides a user tagging rule management system that supports dynamic strategies, referencing... Figure 4 As shown, this figure is a module structure diagram of a user tagging rule management system supporting dynamic strategies according to the present application. The rule management system includes:

[0104] Data acquisition module 100 is used to collect user behavior data streams and environmental context data;

[0105] The strategy parsing module 200 is used to perform multimodal feature fusion on the user behavior data stream and the environmental context data to obtain a dynamic user profile vector, and to perform strategy intent parsing on the user's business objectives to obtain a set of strategy parameters.

[0106] The strategy reasoning and mining module 300 is used to obtain the historical strategies of user tagging, construct a strategy evolution graph based on the semantic association network and the historical strategies, and perform explicit rule reasoning and implicit requirement mining on the dynamic user profile vector and the strategy parameter set through the strategy evolution graph to obtain a candidate strategy package.

[0107] The sandbox inference module 400 is used to place the candidate strategy package in an isolated strategy sandbox environment for simulation and prediction, thereby obtaining an inference performance matrix. The confidence of the inference performance matrix is ​​calibrated through a human-machine collaborative optimization loop to obtain dynamic labeling rules.

[0108] The feedback update module 500 is used to update the weights of nodes and associated edges in the strategy evolution graph through the execution feedback of the dynamic labeling rules.

[0109] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0110] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compactdisc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0111] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

Claims

1. A method for managing user tagging rules that supports dynamic strategies, characterized in that, The rule management method includes the following steps: Collect user behavior data streams and environmental context data; Multimodal feature fusion is performed on the user behavior data stream and the environmental context data to obtain a dynamic user profile vector, and the user's business objectives are analyzed to obtain a set of policy parameters. The historical strategies for user tagging are obtained, and a strategy evolution graph is constructed based on the semantic association network and the historical strategies. The strategy evolution graph is used to perform explicit rule reasoning and implicit demand mining on the dynamic user profile vector and the strategy parameter set to obtain a candidate strategy package. The candidate strategy package is placed in an isolated strategy sandbox environment for simulation and prediction to obtain the inference performance matrix. The confidence of the inference performance matrix is ​​calibrated through a human-machine collaborative optimization loop to obtain dynamic labeling rules. The weights of nodes and associated edges in the strategy evolution graph are updated through the execution feedback of the dynamic labeling rules. Specifically, the process of obtaining a candidate strategy package by performing explicit rule reasoning and implicit requirement mining on the dynamic user profile vector and the strategy parameter set through the strategy evolution graph includes: performing subgraph matching on the dynamic user profile vector based on the strategy evolution graph to obtain an explicit rule set; performing neighborhood expansion on the strategy parameter set through a graph neural network embedding propagation algorithm and the strategy evolution graph to obtain an implicit requirement set; and combining and sorting the explicit rule set and the implicit requirement set to obtain a candidate strategy package.

2. The user tagging rule management method supporting dynamic strategies as described in claim 1, characterized in that, User behavior data streams and environmental context data are collected through a distributed log collection framework.

3. The user tagging rule management method supporting dynamic strategies as described in claim 1, characterized in that, The process of fusing multimodal features between the user behavior data stream and the environmental context data to obtain a dynamic user profile vector specifically includes: The user behavior data stream is serialized and encoded to obtain a behavior embedding vector; Spatiotemporal features are extracted from the environmental context data to obtain a context feature vector; The behavior embedding vector and the context feature vector are aligned and fused using a cross-modal attention mechanism to obtain a dynamic user profile vector.

4. The user tagging rule management method supporting dynamic strategies as described in claim 1, characterized in that, The strategy intent is analyzed to determine the user's business objectives, resulting in a set of strategy parameters, which specifically includes: Semantic parsing of business target text is performed using natural language processing technology to obtain the structured intent of the business target; Based on a multi-objective optimization algorithm, constraint reasoning is performed on the business rule base and the structured intent to obtain a set of strategy parameters.

5. The user tagging rule management method supporting dynamic strategies as described in claim 1, characterized in that, The historical tagging strategies of users are obtained through a distributed repository.

6. The user tagging rule management method supporting dynamic strategies as described in claim 1, characterized in that, The construction of a policy evolution graph based on semantic association networks and historical policies specifically includes: The historical strategies are subjected to feature extraction and vectorization representation to obtain a set of strategy feature vectors; A semantic association network between labeling strategies is constructed by using a graph neural network, and then a time series analysis algorithm is used to perform strategy evolution on the semantic association network to obtain a set of strategy evolution relationships. The policy evolution relationship set and the policy feature vector set are fused to obtain the policy evolution map.

7. The user tagging rule management method supporting dynamic strategies as described in claim 1, characterized in that, The candidate policy package is placed in an isolated policy sandbox environment for simulation and prediction, and the resulting inference performance matrix specifically includes: Construct a sandbox environment that includes virtual user groups and simulated environments; The candidate strategy package is deployed to the sandbox environment instance, and the strategy execution process data is obtained by Monte Carlo simulation. The process data is quantitatively analyzed to obtain an efficiency score, which in turn yields a deduced efficiency matrix.

8. The user tagging rule management method supporting dynamic strategies as described in claim 1, characterized in that, The human-machine collaborative optimization loop is used to realize the parameter correction interaction between machine intelligence and human intervention in a visualized decision-making interface.

9. A user tagging rule management system supporting dynamic strategies, used to execute a user tagging rule management method supporting dynamic strategies as described in any one of claims 1 to 8, characterized in that, The rule management system includes: The data acquisition module is used to collect user behavior data streams and environmental context data; The strategy parsing module is used to perform multimodal feature fusion on the user behavior data stream and the environmental context data to obtain a dynamic user profile vector, and to perform strategy intent parsing on the user's business objectives to obtain a set of strategy parameters. The strategy reasoning and mining module is used to obtain the historical strategies of user tagging, construct a strategy evolution graph based on the semantic association network and the historical strategies, and perform explicit rule reasoning and implicit requirement mining on the dynamic user profile vector and the strategy parameter set through the strategy evolution graph to obtain a candidate strategy package. The sandbox inference module is used to place the candidate strategy package in an isolated strategy sandbox environment for simulation and prediction, thereby obtaining the inference performance matrix. The confidence of the inference performance matrix is ​​calibrated through a human-machine collaborative optimization loop to obtain dynamic labeling rules. The feedback update module is used to update the weights of nodes and associated edges in the strategy evolution graph through the execution feedback of the dynamic labeling rules.

Images