Internet marketing platform risk early warning management method and system

Abstract

The invention relates to the technical field of commercial information communication, and discloses an internet marketing platform risk early warning management method and system, and the method comprises the steps: constructing a multi-level heterogeneous risk relation graph, and mapping multi-source heterogeneous data; constructing a dynamic Bayesian causal network, and extracting a causal relationship; establishing an atlas causal bidirectional mapping model to realize information fusion and enhancement; risk propagation potential energy is evaluated based on the model, and the propagation potential is quantified; according to the method, the problems of inaccurate risk propagation prediction, low intervention efficiency, insufficient adaptive ability and the like of the Internet marketing platform are solved, and technical support is provided for safe operation of the platform.

Description

Technical Field

[0001] The present invention relates to the field of commercial information communication technology, and more specifically, to a risk early warning management method and system for an Internet marketing platform. Background Art

[0002] In the ecosystem of Internet marketing platforms, risk events spread rapidly between accounts, content and transactions through complex network relationships, which can easily trigger chain reactions and cluster risks. However, existing technologies have significant shortcomings: risk monitoring and blocking are not matched in time and space, resulting in widespread spread of risks when they are discovered; traditional graph analysis methods only focus on topological structures and cannot reveal the deep causal mechanisms of risk propagation; traditional causal analysis methods are difficult to adapt to the complex interaction patterns of large-scale networks; risk management and control rely on global monitoring resources, and real-time response capabilities are limited; at the same time, new risk models require re-modeling and analysis, and lack adaptive learning capabilities. Based on this, the present invention proposes a "causal graph self-evolution risk propagation control system" to solve the above problems, achieve accurate prediction and minimal intervention in risk propagation, and thus enhance the platform's risk warning and prevention and control capabilities. Summary of the invention

[0003] The present invention provides a risk early warning management method and system for an Internet marketing platform, which solves the technical problems in related technologies such as the mismatch between risk monitoring and blocking time and space, and the inability of traditional graph analysis methods to reveal deep causal mechanisms.

[0004] The present invention provides an Internet marketing platform risk early warning management method, comprising the following steps: Construct a multi-level heterogeneous risk relationship map to map the platform's multi-source heterogeneous data into a unified network representation; Construct a dynamic Bayesian causal network to extract causal relationships from multi-level heterogeneous risk relationship maps and historical risk event sequence data; Construct a bidirectional mapping model of graph causality to achieve information fusion and mutual enhancement between multi-level heterogeneous risk relationship graphs and dynamic Bayesian causal networks; Based on the graph causal bidirectional mapping model, perform risk propagation potential assessment to quantify the risk propagation potential and impact range in the network; Based on the risk propagation potential assessment results, the minimum intervention set is determined under resource constraints to generate the optimal risk intervention plan.

[0005] Furthermore, the step of constructing a multi-level heterogeneous risk relationship map includes: Execute heterogeneous data feature mapping algorithm to map different types of entity data into a unified feature space; Implement edge-aware graph attention calculation to determine the degree of connection between nodes; A multi-level graph structure is formed to represent the relationships between nodes of the same type and the relationships between nodes of different types respectively.

[0006] Furthermore, the step of constructing a multi-level heterogeneous risk relationship graph also includes: dynamically adjusting edge weights according to historical risk propagation patterns so that the graph structure reflects the historical characteristics of risk propagation, wherein the edge weight adjustment formula is: ; in, is the updated edge weight, is the original edge weight, is the adjustment factor, Based on historical risk communication data The adjustment function of Respectively represent nodes and nodes .

[0007] Furthermore, the step of constructing a dynamic Bayesian causal network includes: Identify variables in historical risk data and establish a risk variable set; Use graph-enhanced causal discovery algorithms to determine the causal relationship between variables; A conditional probability estimation algorithm is performed to quantify the strength of each causal relationship.

[0008] Furthermore, the step of constructing a dynamic Bayesian causal network also includes: constructing a temporal causal relationship representation so that the causal network captures the temporal characteristics of risk propagation. The formula of the temporal causal relationship representation is: ; Among them, the superscript Indicates a point in time, represents the time lag order, represents a time series function, Indicates that in a given parent node set Under the condition of The conditional probability of Respectively represent variables The parent node collection of .

[0009] Furthermore, the step of constructing the graph causal bidirectional mapping model includes: Construct a forward mapping function from graph to causal network to transform graph structure information into causal relationship priors; Create a reverse mapping function from causal network to graph, and update the graph structure using the strength of causal relationship; An attention fusion algorithm is implemented to combine the two representations to generate a unified risk representation.

[0010] Furthermore, the step of performing risk propagation potential assessment includes: Construct a unified risk characterization model to transform various risk events into computable feature vectors; Execute the risk propagation potential calculation algorithm to quantify the risk propagation potential of the node; Implement edge propagation probability calculation method to evaluate the probability of risk propagation on the edge; Monte Carlo simulation method is used to estimate the statistical characteristics of risk propagation.

[0011] Furthermore, the formula of the risk propagation potential energy calculation algorithm is: ; in, Representation Node The risk transmission potential of Representation Node Centrality measures in graphs, Represents the risk of node The vulnerability of Representation based on causal network The node influence to be evaluated, , and is a weight parameter to balance different factors.

[0012] Furthermore, the step of determining the minimum intervention set includes: Construct an intervention effect evaluation model to quantify the inhibitory effect of intervention on risk transmission; The problem is formalized as an entropy reduction maximization model to express the optimization goal under resource constraints; Execute the greedy approximation algorithm to iteratively construct the optimal intervention set; Generate uncertainty-aware intervention scenarios and assign optimal intervention actions to each intervention node; Evaluate the expected effects of the intervention and calculate the probability distribution of risk suppression.

[0013] The present invention provides an Internet marketing platform risk early warning management system, comprising: Risk relationship map construction module, used to construct multi-level heterogeneous risk relationship maps; Causal network building module, used to build dynamic Bayesian causal networks; Bidirectional mapping module, used to build a graph causal bidirectional mapping model; Risk potential assessment module, used to perform risk propagation potential assessment; The intervention optimization module is used to determine the minimum intervention set and generate the optimal risk intervention plan.

[0014] The beneficial effects of the present invention are as follows: the present invention provides a new risk warning and prevention and control solution for the Internet marketing platform, effectively solving the core technical problems existing in the prior art such as inaccurate risk propagation prediction, low intervention efficiency, insufficient adaptive capability, etc., and provides strong technical support for the safe and stable operation of the platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flow chart of a risk early warning management method for an Internet marketing platform of the present invention; Figure 2 It is a flow chart of the steps of constructing a multi-level heterogeneous risk relationship map of the present invention; Figure 3 is a flow chart of the steps of constructing a dynamic Bayesian causal network of the present invention; Figure 4 It is a flow chart of the steps of constructing a graph causal bidirectional mapping model of the present invention; Figure 5 is a flow chart of the quantitative steps of the graph causal bidirectional mapping model evaluation of the present invention; Figure 6 It is a flow chart of the steps of determining the minimum intervention set and generating the optimal risk intervention plan of the present invention. DETAILED DESCRIPTION

[0016] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is only to enable those skilled in the art to better understand and implement the subject matter described herein, and the functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the present specification. Various examples may omit, replace, or add various processes or components as needed. In addition, the features described in some examples may also be combined in other examples.

[0017] At least one embodiment of the present invention discloses a risk early warning management method for an Internet marketing platform, such as Figures 1 to 6 As shown, the following steps are included: Step 100, construct a multi-level heterogeneous risk relationship map to map the platform's multi-source heterogeneous data into a unified network representation; The purpose of this step is to establish a network representation of platform risks, input multi-source heterogeneous data such as platform user account data, transaction record data, content interaction data, device information data, and historical risk event data, and output a multi-level heterogeneous risk relationship map. .in, Represents a collection of nodes (including accounts, transactions, and content entities). represents the edge set, Represents the weight of the edge.

[0018] Step 101, execute the heterogeneous data feature mapping algorithm to map different types of entity data into a unified feature space, and the calculation formula is: ; in, Is a node The characteristic vector of is the feature extraction function selected according to the node type, Is a node original features.

[0019] Step 102, implement edge-aware graph attention calculation to determine the degree of association between nodes, and the calculation formula is: ; in, Representation Node and nodes The attention coefficient between them reflects the degree of association between the two nodes. represents the exponential function, is a rectified linear unit function, is the attention vector, and The nodes are and nodes The characteristic vector of is the shared weight matrix, Represents a splicing operation, Represents the node The neighbor set of All nodes in Perform a sum operation, Is a node The set of neighbors.

[0020] Step 103, forming a multi-level graph structure, characterizing the relationship between nodes of the same type and the relationship between nodes of different types respectively, and the calculation formula is: ; in, Represents the entire multi-level graph, , , Represents the first The subgraph of the layer, Indicates the number of levels of a multi-level graph.

[0021] ; in, Representative The subgraph of the layer, Indicates the number of subgraph layers, Indicates The set of nodes in the layer subgraph, Indicates The set of edges in the layer subgraph, Indicates The weight of the edge in the layer subgraph represents the strength of the relationship or the probability of risk transmission.

[0022] Subgraphs at different levels Capable of capturing different types of risk propagation patterns.

[0023] This step also includes the risk weight adaptive adjustment operation, which dynamically adjusts the edge weights according to the historical risk propagation pattern so that the graph structure can reflect the historical characteristics of risk propagation. Compared with the traditional graph construction method, this operation can more accurately characterize the risk transmission relationship between entities.

[0024] Step 200, constructing a dynamic Bayesian causal network to extract causal relationships from multi-level heterogeneous risk relationship graphs and historical risk event sequence data; From the multi-level heterogeneous risk relationship map of step 100 , extract causal relationships from historical risk event sequence data and expert knowledge rules, and generate dynamic Bayesian causal networks In this causal network Represents a set of variable nodes, represents a set of directed edges, Represents a conditional probability table.

[0025] Step 201, perform variable identification on historical risk data and establish a risk variable set, the calculation formula of which is: ; in, represents the set of risk variables, Respectively represent risk variable, is the number of variables in the risk variable set.

[0026] Step 202, using the graph enhanced causal discovery algorithm to determine the causal relationship between variables, the calculation formula is: ; in, Represents the Peter-Clark algorithm, which is used to extract Learn causal structures in Represents a set of variable nodes, is the significance level, Represents expert knowledge constraints. The algorithm introduces the graph topology structure as prior information into the causal discovery process, and its calculation formula is: ; in, Indicates that in a given graph Under the condition of To variable The posterior probability of a causal relationship, Representation variables To variable The causal relationship, Representation variables To variable The prior probability of causality is Indicates in the variable To variable Under the condition of causal relationship, the spectrum The conditional probability of Indicates a proportional relationship.

[0027] Step 203, execute the conditional probability estimation algorithm to quantify the strength of each causal relationship, and the calculation formula is: ; in, Representation variables The parent node set of (direct cause), Represents a counting function. Indicates that in a given parent node set Under the condition of The conditional probability of .

[0028] This step also includes the temporal causal relationship representation, which enables the causal network to capture the temporal characteristics of risk propagation. The calculation formula is: ; Among them, the superscript Indicates a point in time, represents the time lag order, Represents a time series function. Indicates that in a given parent node set Under the condition of The conditional probability of . Respectively represent variables The parent node collection of .

[0029] The difference between this dynamic Bayesian causal network and traditional methods is that it not only considers the causal relationship at the same time, but also incorporates the temporal dependency, which can more comprehensively characterize the dynamic characteristics of risk propagation.

[0030] Step 300, constructing a bidirectional mapping model of graph causality to achieve information fusion and mutual enhancement between multi-level heterogeneous risk relationship graphs and dynamic Bayesian causal networks; The multi-level heterogeneous risk relationship map generated in step 100 and the dynamic Bayesian causal network generated in step 200 Fusion and mutual enhancement are performed to obtain a bidirectional graph-causal mapping model ,in, represents the mapping function from graph to causal network, represents the mapping function from causal network to graph, Represents the unified risk representation after fusion. This model can simultaneously utilize the information of network structure and causal relationship to provide a more comprehensive risk representation.

[0031] Step 301: Construct a forward mapping function from graph to causal network , converting graph structure information into causal prior, the calculation formula is: ; in, Is a node in the graph and The edge weights between yes arrive The path characteristics, and is the mapping parameter, is the sigmoid function. Representation variables To variable The causal relationship, Representation variables To variable The prior probability of causal relationship.

[0032] Step 302: Create a reverse mapping function from causal network to graph , using the causal relationship strength to update the graph structure, the calculation formula is: ; in, is the updated graph edge weight, It is a causal relationship The strength of is a balance parameter.

[0033] Step 303, combining the two representations to generate a unified risk representation, and implementing an attention fusion algorithm, the calculation formula of which is: ; in, and are entity pairs extracted from graphs and causal networks respectively. The representation of is a multi-head attention operation, Represents entity pair Unified risk representation.

[0034] The uniqueness of this bidirectional mapping model lies in the realization of the two-way flow of information between the graph and the causal network, breaking through the limitations of the isolation of graph analysis and causal analysis in traditional methods, and providing a unified mathematical framework and data foundation for risk communication prediction and intervention.

[0035] Step 400, based on the graph causal bidirectional mapping model, perform risk propagation potential assessment to quantify the risk propagation potential and impact range in the network; Based on the graph-causal bidirectional mapping model generated in step 300 , quantitatively assess the potential and scope of risk propagation in the network, and generate a risk propagation potential energy map ,in and are the node set and edge set inherited from step 100, respectively. is the risk propagation potential energy value mapped to each node and edge. This potential energy diagram is used to represent the speed, intensity and possible propagation path of risk propagation.

[0036] Step 401, construct a unified risk characterization model to convert various risk events into feature vectors, and the calculation formula is: ; in, Indicates risk events. represents the environmental context of the risk event, represents the risk representation provided by the mapping model of step 300, is a deep neural network encoder, the output It is a unified representation vector of risk events.

[0037] Step 402, execute the risk propagation potential calculation algorithm to quantify the risk propagation potential of the node: ; in, Is a node Centrality measures in graphs, Is the node risk The vulnerability of Based on causal network The node influence to be evaluated, , and is a weight parameter to balance different factors.

[0038] Step 403, implement the edge propagation probability calculation method to evaluate the probability of risk propagation on the edge: ; in, Representation Node Risk The degree of activation, Representation Node Risk susceptibility, Representation Node The neighbor nodes of Representation Node Risk susceptibility, It is the edge The weight of . Represents edge The probability of risk transmission is Indicates that at the node and risks Under the condition that probability of risk transmission.

[0039] Step 404, using the Monte Carlo simulation method to estimate the statistical characteristics of risk propagation: ; in, Indicates the potential energy of risk transmission under a given risk , Risk Type and time window Next, in the atlas A risk propagation simulation was conducted on is the number of simulations, is the expected value of the risk spread.

[0040] Compared with traditional risk propagation models, this step innovatively combines the node centrality in graph theory and the intervention effect evaluation in the causal network, which can more accurately predict the risk propagation path, speed and scope of influence, and provide a quantitative basis for risk prevention and control decisions.

[0041] Step 500, based on the risk propagation potential assessment result, determine the minimum intervention set under resource constraints and generate the optimal risk intervention plan; Risk propagation potential diagram generated based on step 400 , under resource constraints, determine the minimum intervention set , to minimize risk transmission and generate intervention implementation plans ,in, is the minimum intervention set, is the corresponding set of intervention actions, This is the expected risk suppression effect. The plan will guide the platform to take precise risk control measures.

[0042] Step 501, construct an intervention effect evaluation model to quantify the inhibitory effect of intervention on risk transmission: ; in, Indicates intervention The effect, is the set of candidate intervention nodes, represents the risk score on the original graph, Indicates intervention The risk score of the posterior profile is calculated by the following formula: The original map The subgraph of The node set is A subset of the node set.

[0043] ; in, Is a node The risk transmission potential of Representation Node In the atlas The accessibility of Representation Spectrum The risk score of It is a graph The node set of It is a graph Node.

[0044] Step 502, formalize the problem into an entropy reduction maximization model, expressing the optimization goal under resource constraints: ; in, Represents the intervention set The cost of resources required, is the minimum set of intervention nodes, that is, the optimal intervention set, Indicates intervention The effect, is the upper limit of available resources, Indicates that under resource constraints, The maximum set of intervention nodes. This optimization problem is NP-hard, and we use a greedy approximation algorithm to solve it.

[0045] Step 503, execute the greedy approximation algorithm to iteratively construct the optimal intervention set: Initialization: Intervention Collection ; Initial intervention effect ; in, is the currently selected intervention node set, which is initially an empty set. Used to record the optimal intervention effect currently found.

[0046] Loop condition judgment: When Enter the loop in, For intervention collection The cost of resources required, It is the upper limit of available resources.

[0047] Calculate marginal gain and benefit ratio: For each candidate node : The formula for calculating the marginal gain and benefit ratio is: ; ; in, It is a graph The node set of Indicates that it is not in the current intervention set All candidate nodes in Indicates that the node Add to Intervention Collection The increase in the post-intervention effect, Is to join the node The post-intervention effect is the current intervention set The intervention effect Representation Node The benefit ratio, The node the resource costs required to include the intervention set; Select the node with the highest benefit ratio : ; in, Indicates that in all candidate nodes In order to make the benefit ratio The largest node ; Determine whether resources are sufficient and update: like , that is, join the node The resource cost required for the post-intervention set is still less than or equal to the upper limit of available resources: renew , the node with the highest benefit ratio Add to Intervention Collection ; renew , update the current optimal intervention effect to add Post-intervention set intervention effect.

[0048] Otherwise: exit the loop and stop looking for new intervention nodes.

[0049] Return the result after jumping out of the loop: return the latest intervention set and expected results ,in, The optimal set of intervention nodes found under resource constraints is the optimal intervention plan, which includes specific intervention measures taken for specific risk propagation patterns, such as which intervention nodes to select and what optimal intervention actions to assign to each intervention node. is the corresponding expected intervention effect.

[0050] Step 504, generate an uncertainty-aware intervention plan and assign the best intervention action to each intervention node: ; in, Is a node A collection of actionable interventions, Indicates that at the node Implement intervention actions , is the conditional expected intervention effect, Representation Node optimal intervention actions.

[0051] Step 505, evaluate the expected effect of the intervention and calculate the probability distribution of risk suppression: ; in, Indicates the minimum set of intervening nodes and the optimal intervention action for each intervention node Under the condition of The probability of Indicates the risk reduction, with a value range of between.

[0052] Through probability distribution , the system is able to quantify the uncertainty of intervention success and provide confidence intervals for risk control.

[0053] Step 506, immune memory and adaptive evolution: Based on the empirical data obtained in the previous steps, build a risk immunity memory library ,in, is the risk propagation pattern eigenvector, is the corresponding optimal intervention plan, is the effect of historical intervention, is a timestamp. This memory allows the system to quickly respond to known risks and continuously optimize intervention strategies.

[0054] Construct a risk pattern feature extraction algorithm to generate a feature vector for risk propagation: ; in, Indicates risk events. is a risk relationship map. is a causal network, Extract risk-related topological features from the graph, Extract risk-related features from causal networks, is the feature fusion encoder. The output is the feature vector that characterizes the risk propagation pattern.

[0055] Implement a similarity matching algorithm to search the memory bank for historical cases similar to the current risk pattern: ; ; in, Represents the calculation of two risk propagation model eigenvectors and The cosine similarity between .in, is the dot product of two vectors, and They are vectors and The norm of . Risk immune memory Filter out the feature vectors that match the current risk pattern Similarity is greater than the similarity threshold A collection of similar historical cases ,in It's a memory bank In every historical case, is the risk propagation pattern eigenvector, is the corresponding optimal intervention plan, is the effect of historical intervention, Is the timestamp.

[0056] Perform memory retrieval and optimize adaptive algorithms to quickly respond to similar risk patterns: when When Search for the best intervention plan : ; And make adaptive adjustments based on the current environmental differences: ; in, is the adaptive adjustment function, represents the optimal intervention plan after adaptive adjustment, and They are the current and historical environmental contexts respectively.

[0057] Adopt online learning methods to continuously optimize system parameters and strategies: ; in, represents the set of learnable parameters of the system, is the learning rate, is the system performance evaluation function, Indicates the parameters Through online learning, the system can continuously optimize itself as experience accumulates, improving its risk prevention and control capabilities. Indicates The parameters after iterations, Indicates The parameters after iterations, Represents the memory bank, Represents the risk relationship map, Represents a causal network.

[0058] Implement a long-term and short-term knowledge integration mechanism to balance historical experience and recent data: ; in, is the memory update function, It is a new case of risk intervention. represents the new risk propagation pattern feature vector, represents the new optimal intervention plan, represents the new intervention effect, Indicates the new timestamp, It is a parameter that controls the balance between long-term and short-term memory. When the memory size exceeds the threshold, the system selectively retains historical cases based on time decay and importance index.

[0059] Compared with traditional risk intervention methods, this step is based on the graph causal fusion model, which takes into account the direct and indirect influences between nodes. It can achieve the maximum risk suppression effect with the minimum resource investment and provide uncertainty assessment of the intervention effect. It innovatively introduces the concept of risk immune memory, and achieves rapid response to known risk patterns through similarity matching and adaptive adjustment. It also continuously optimizes system performance through online learning mechanisms, significantly improving the efficiency and accuracy of risk management.

[0060] In one embodiment of the present invention, an example of the aforementioned Internet marketing platform risk early warning management method is provided: Application scenario description: A risk warning management scenario during the "Double 11" promotion of a large domestic e-commerce platform was selected. During such large-scale promotions, the platform faces multiple risk challenges, and there are complex transmission relationships between these risks. In past promotions, risk events often showed cluster outbreak characteristics, making it impossible for the platform's risk management resources to respond in a timely manner.

[0061] The data sources for this example include: user account data, product data, transaction data, marketing campaign data, and historical risk event data, covering data from 30 days before and after "Double 11" in 2023, involving approximately 50 million users, 2 million merchants, 100 million transaction records, and 500,000 marked historical risk events.

[0062] The following will show the implementation process example and technical effect verification of this method in this scenario.

[0063] In step 100, a multi-level heterogeneous risk relationship map is constructed to map the platform's multi-source heterogeneous data into a unified network representation: In this example, a three-layer risk relationship map is constructed by analyzing the multi-source heterogeneous data of the platform, including the account layer, transaction layer, and content layer. The sample data of some nodes and edges are shown in Table 1 and Table 2: Table 1: Node entity types and characteristics (partial examples):

[0064] Table 2: Heterogeneous relationship types and weight calculation (partial examples):

[0065] In the process of graph construction, the degree of association between nodes is determined through edge-aware graph attention mechanism calculation. In the specific implementation, the attention calculation example of merchant product relationship is as follows: Partial example of the calculation results of the attention coefficient between the merchant node M10256 and the multiple commodity nodes it sells: Attention coefficient with product P30567: 0.28; Attention coefficient with product P30568: 0.15; Attention coefficient with product P30569: 0.4; Attention coefficient with product P30570: 0.12; Attention coefficient with product P30571: 0.03; Based on the analysis of historical risk propagation patterns, the weights of the associated account edges were dynamically adjusted. For example, the weight of edge E45632 was adjusted from 0.29 to 0.82 because the two merchant accounts were found to have highly correlated behavior patterns in historical risk events. Based on the adjusted weights, the risk transmission probability was calculated to be 0.85, indicating that there is an 85% probability that the risk behavior of merchant account M10256 will be transmitted to merchant account M10257.

[0066] Step 200: Build a dynamic Bayesian causal network to extract causal relationships from multi-level heterogeneous risk relationship graphs and historical risk event sequence data: Based on historical risk event data, the main risk variables related to "brushing risk" were identified, and a dynamic Bayesian causal network was constructed. Some variables and their causal relationships are shown in Tables 3 and 4: Table 3: Risk variable set (partial example):

[0067] Table 4: Causal relationship strength (partial examples):

[0068] In practical applications, the graph-enhanced causal discovery algorithm introduces the topological structure in the graph as prior information into the causal discovery process. For example, based on the graph data, it is found that the weight of the association edge between merchant accounts M10256 and M10257 is 0.82. This information is converted into the prior probability that variable X3 (account association) has a causal relationship with X1 (abnormal transaction frequency), which in turn affects the final learned conditional probability P(X1|X3)=0.75.

[0069] In addition, the temporal characteristics of risk propagation are captured through temporal causal representation. For example, the impact of account association (X3) on transaction frequency anomaly (X1) usually appears after 2 to 4 hours, while the impact of transaction frequency anomaly (X1) on review text similarity (X4) takes 12 to 24 hours to be observed.

[0070] In step 300, a bidirectional mapping model of graph causality is constructed to achieve information fusion and mutual enhancement between multi-level heterogeneous risk relationship graphs and dynamic Bayesian causal networks: Based on the risk relationship graph and dynamic Bayesian causal network constructed above, a graph causal bidirectional mapping model was established. Some mapping relationships are shown in Table 5: Table 5: Example of bidirectional causal mapping of graph:

[0071] Taking the account association relationship as an example, the weight of edge E45632 in the graph, 0.82, is transformed into the prior probability of 0.79 of the influence of X3 (account association) on subsequent variables in the causal network through forward mapping; and the causal strength of X3 to X1, 0.75, learned in the causal network, adjusts the edge weight in the graph to 0.85 through reverse mapping. Finally, a unified risk representation with a strength of 0.87 is generated through the attention fusion mechanism.

[0072] Through this bidirectional mapping and information fusion, the system achieves mutual enhancement of graph structure information and causal relationships, providing a more comprehensive risk representation. For example, the temporal causal relationship between "device repetition rate" and "evaluation text similarity" (strength 0.79, delay 6 to 12 hours), which is difficult to discover through graph analysis alone, is accurately captured through the causal network and reversely enhances the relevant graph structure.

[0073] In step 400, based on the graph causal bidirectional mapping model, risk propagation potential assessment is performed to quantify the risk propagation potential and impact range in the network: Based on the graph causal bidirectional mapping model, the system evaluated the risk communication potential during the "Double 11" promotion. Some of the risk communication potential evaluation results are shown in Table 6: Table 6: Risk communication potential assessment (partial example):

[0074] Through the Monte Carlo simulation method, the impact range and path of risk propagation from different sources were estimated. For example, the results of 1,000 risk propagation simulations for merchant account M10256 showed that the average impact range was 147 nodes, and the main impact paths were: M10256→M10257→M10259→… (merchant association path, affecting 38% of nodes); M10256→P30567→U58921→… (product user path, affecting 26% of nodes); M10256→P30567→R42198→… (product evaluation path, affecting 21% of nodes); Other paths (affecting 15% of nodes).

[0075] Through the unified risk characterization model, the system converts different types of risk events into computable feature vectors and calculates the edge propagation probability for the risk of order brushing. Some examples are as follows: The transmission probability of merchant M10256 → merchant M10257: 0.87; The transmission probability of merchant M10256 → product P30567 is 0.72; The probability of spreading from product P30567 to user U58921 is 0.38. The transmission probability of product P30567 → review R42198 is 0.71.

[0076] These propagation probabilities combined with time series information allow the system to predict the time trajectory of risk propagation. For example, the system predicts that the risk of fake orders initiated by merchant M10256 will spread to related merchants within 8 hours, spread to related product reviews within 12 to 24 hours, and may affect a wider user group within 36 to 48 hours.

[0077] In step 500, based on the risk propagation potential assessment result, the minimum intervention set is determined under resource constraints to generate the optimal risk intervention plan: Based on the risk propagation potential assessment results, the system determines the optimal set of intervention nodes under resource constraints. Examples of intervention plans under specific resource constraints are shown in Table 7: Table 7: Minimum intervention set and intervention plan (partial example, resource constraint: 100 units):

[0078] Using the greedy approximation algorithm, the system iteratively constructs the optimal intervention set. The optimization process tracking record is shown in Table 8: Table 8: Optimization process tracking record:

[0079] According to the uncertainty-aware intervention plan generation mechanism, the system assigns the best intervention action to each intervention node and evaluates the probability distribution of the intervention effect. For example, the probability distribution of the risk suppression effect of the intervention action of "restricting promotional activities" for merchant account M10256 is: Probability of 85% to 90% inhibition effect: 0.62; Probability of 80% to 85% inhibition effect: 0.28; Probability of 90% to 95% inhibition effect: 0.08; Probability of inhibitory effect < 80%: 0.02.

[0080] In this way, the system not only provides specific intervention options, but also quantifies the uncertainty of the intervention's success, enabling risk managers to make more informed decisions.

[0081] Technical effect verification Through practical application in the 2023 "Double 11" promotion activities, this system has achieved significant technical results in risk warning management. The following focuses on verifying two key technical effects: risk communication prediction accuracy and intervention resource utilization efficiency.

[0082] The accuracy of risk propagation prediction is verified by comparing the system prediction results with the actual risk propagation situation, as shown in Table 9: Table 9: Risk communication prediction accuracy verification data:

[0083] The prediction accuracy analysis of the propagation path is shown in Table 10: Table 10: Analysis of prediction accuracy of propagation paths:

[0084] The data show that the system has significantly improved the accuracy of risk propagation prediction compared with traditional methods, with an average accuracy increase of 219.5%. In particular, the advantage is more obvious in predicting multi-step propagation paths, with the accuracy of 4+ steps increased by 289.2%. This proves the excellent ability of the graph causal bidirectional mapping model in capturing complex risk propagation patterns.

[0085] Verification of intervention resource utilization efficiency,Through comparative analysis, the intervention efficiency of the system under resource constraints,is verified, as shown in Table 11 and Table 12: Table 11: Intervention resource utilization efficiency verification data:

[0086] Table 12: Comparison of intervention time effectiveness:

[0087] Data shows that while achieving a higher risk suppression rate (96.7% vs 81.7%), this system only requires 25% of the intervention resources of traditional methods, with an average resource saving rate of 75.0%. At the same time, the total time from risk identification to control was reduced by 77.7%, significantly improving the timeliness of risk management and control.

[0088] During the "Double 11" promotion period, the system successfully warned and intervened in 1,873 risk events, avoiding an estimated economic loss of about 18.5 million yuan, and reducing manpower input by about 73% compared with traditional methods. These data fully demonstrate the significant effect of the minimum intervention set optimization technology based on the graph causal bidirectional mapping model in practical applications.

[0089] In one embodiment of the present invention, in view of the problem that the traditional centralized decision-making architecture makes all robots schedule at the same time scale and decision-making level, resulting in computing bottlenecks and communication delays, and it is difficult to adapt to the real-time adjustment requirements of emergency tasks, a risk early warning management method for an Internet marketing platform also includes the following steps: This solution builds a hierarchical collaborative scheduling system for asynchronous decision-making, enabling different decision-making levels to run at different time scales, thereby optimizing computing resource allocation and improving decision response speed.

[0090] Step 600, constructing a three-layer time-scale separation decision architecture: This step uses a hierarchical time-separation architecture model to hierarchically divide the decision-making process of the robot cluster and generate a decision architecture with different time scales.

[0091] Input data types: robot cluster task characteristic data, environment change frequency data, computing resource constraint parameters; Specific output results: A three-layer time scale decision architecture configuration table, including time parameters and function definitions for the strategic layer, tactical layer, and execution layer.

[0092] This step calculates and configures the time parameters of the three-layer decision-making architecture based on the input task characteristics and the frequency of environmental changes: Sub-step 601, strategic layer time parameter calculation: based on task completion time and frequency of environmental changes Calculate the strategic decision cycle : ; in, represents the task time ratio coefficient, represents the environmental adaptation coefficient, and the two together ensure that the strategic level decision cycle adapts to the task duration and environmental change characteristics.

[0093] Sub-step 602, tactical layer time parameter calculation: based on the strategic layer cycle Calculating the tactical decision cycle : ; in, It represents the time scale factor of the strategic and tactical layers, with a value range of 5-20, which is used to ensure that the decision frequency of the tactical layer is higher than that of the strategic layer.

[0094] Sub-step 603, execution layer time parameter calculation: based on the tactical layer cycle Compute the Decision Cycle of the Execution Layer : ; in, Indicates the time scale factor of the tactical execution layer, with a value range of 10-100, which is used to ensure that the decision frequency of the execution layer is higher than that of the tactical layer.

[0095] The generated three-layer time-scale decision architecture configuration table clearly defines the functional division of different levels: the strategic layer is responsible for global resource allocation and long-term planning; the tactical layer is responsible for local path optimization and conflict resolution; the execution layer is responsible for real-time control and immediate response. The various levels are interconnected through constraint transfer relationships, and the high-level layers transfer constraints to the low-level layers, and the low-level layers make autonomous decisions within the constraints.

[0096] Step 700, creating an event-triggered adaptive decision scheduling algorithm: This step constructs an adaptive trigger decision scheduling algorithm based on system state changes and resource utilization, and generates decision execution priority sorting rules.

[0097] Input data types: environmental status change data, task urgency index, system computing resource utilization data, and the decision architecture configuration table generated in step 600; Specific output results: adaptive decision trigger rule set and priority scheduling table, including trigger threshold calculation formula and priority sorting method.

[0098] The adaptive decision triggering algorithm includes the following key calculation methods: Sub-step 701, state error calculation: by comparing the current actual state With the predicted status The difference between , and its calculation formula is: ; in, represents an appropriate norm operation, which is used to quantify the degree of difference between state vectors. When the state error exceeds the trigger threshold, the system triggers the decision update of the corresponding level. Indicates the current moment.

[0099] Sub-step 702, trigger threshold dynamic adjustment: based on system calculation resource utilization Dynamically adjust the trigger threshold , and its calculation formula is: ; in, Indicates the basic trigger threshold, Represents the resource sensitivity parameter, Indicates the current system computing resource utilization. This formula allows the system to increase the trigger threshold when resources are tight and reduce unnecessary decision-making calculations.

[0100] Sub-step 703, decision priority score calculation: After the decision is triggered, the system calculates the priority score of the decision task based on multiple factors. , and its calculation formula is: ; in, , and is the weight coefficient, is the state error value, is the task importance indicator, It is an indicator of computing resource availability. The priority score is used to determine the execution order of decision tasks.

[0101] The generated adaptive decision trigger rule set and priority scheduling table contain different trigger conditions and priority calculation parameters for the strategic layer, tactical layer and execution layer, which serve as the basis for resource allocation in the asynchronous message transmission process of step 800 and are used for task scheduling in the computational load balancing of step 900.

[0102] Step 800, constructing a hierarchical state snapshot and incremental update protocol: This step constructs a state synchronization protocol suitable for an asynchronous decision-making environment by establishing a state representation method and update rules, thereby solving the information consistency problem in a multi-level decision-making architecture.

[0103] Input data types: decision status data at each level, real-time position and task data of the robot, decision architecture configuration table generated in step 600, decision trigger rule set generated in step 700; Specific output results: hierarchical state synchronization protocol document, including state representation method, incremental update rules, consistency check algorithm and prediction model.

[0104] The hierarchical state synchronization protocol consists of the following main components: Sub-step 801, hierarchical state representation structure: system state Divided into strategic level states according to the level of abstraction , Tactical level status and execution layer status , forming a hierarchical state structure: ; ; in, It represents the inclusion relationship, that is, the high-level state is an abstract subset of the low-level state, which ensures the consistency of the states between levels. In the state structure, the strategic layer contains task allocation and global planning information, the tactical layer contains path planning and coordination information, and the execution layer contains immediate control parameters.

[0105] Sub-step 802, incremental state update method: calculate and transmit the state change increment , instead of the full state: ; in, Indicates the current state. Indicates the status at the time of the last update. Indicates the level identification. The incremental update method significantly reduces the amount of communication data between levels.

[0106] Sub-step 803, consistency check algorithm: through the consistency function Determine whether two states are compatible: ; in, Represents two states, represents the keys in the state, Represents the set of keys common to the two states. When a state inconsistency is detected, the protocol triggers the state reconciliation procedure to resolve the conflict using the priority rules in step 700.

[0107] Sub-step 804, state prediction model: developing a state prediction function for communication delay : ; in, represents the state evolution function, Indicates the current state. Indicates the current system action. Represents the prediction time interval. The prediction model provides state estimates in the absence of timely updates to maintain continuous operation of the system.

[0108] The constructed hierarchical state synchronization protocol document provides a state acquisition and synchronization method for the computing load balancing system in step 900, ensuring that each node in a distributed environment can perform decision-making calculations based on consistent state information.

[0109] Step 900, developing a multi-level computing resource allocation system: This step develops a distributed computing resource allocation system based on the characteristics of the decision-making tasks and the capabilities of the computing nodes to achieve computing load balancing in an asynchronous decision-making environment.

[0110] Input data types: computing node processing capacity data, network connection status data, decision architecture configuration table generated in step 600, priority scheduling table generated in step 700, state synchronization protocol document generated in step 800; Specific output results: distributed computing task allocation table, including task node mapping rules, load balancing thresholds and task migration strategies.

[0111] The distributed computing resource allocation system includes the following key functional modules: Sub-step 901, decision task complexity evaluation module: evaluates the computational complexity of decision tasks at each level according to the input decision architecture configuration table and communication needs : ; in, and are the evaluation functions for computational complexity and communication requirements, respectively. Indicates the number of related robots, Indicates the complexity of the environment, represents the decision time span, and Denote the input and output data sizes respectively. The evaluation results are used for task allocation decisions.

[0112] Sub-step 902, computing node capability evaluation module: monitors and evaluates the processing capability of each computing node and communication capabilities : ; in, and They represent the basic processing capacity and communication bandwidth of the node respectively. and They represent the current processing load and communication load respectively. These parameters are used for subsequent task allocation optimization.

[0113] Sub-step 903, task allocation optimization module: combining task complexity, node capability and priority information of step 700, calculates the task allocation objective function : ; in, , and is the weight coefficient, It represents the expected load after task allocation. The node with the smallest objective function value is selected as the best node to perform a specific task.

[0114] Sub-step 904, load balancing monitoring module: regularly calculate the standard deviation of system load distribution , when the imbalance exceeds the preset threshold, task redistribution is triggered: ; in, Indicates the system load, Indicates the total number of computing nodes, Representation Node The load, Indicates the average load, Indicates the imbalance threshold. This mechanism ensures that the system load remains balanced for a long time.

[0115] The developed distributed computing task allocation table achieves efficient utilization of computing resources: high-complexity, low-frequency decision-making tasks at the strategic layer are allocated to high-performance servers in the cloud; medium-complexity, medium-frequency decision-making tasks at the tactical layer are allocated to edge servers; and low-complexity, high-frequency decision-making tasks at the execution layer are allocated to the robot's local processor. This allocation scheme combines the time scale architecture of step 600, the trigger mechanism of step 700, and the state synchronization protocol of step 800 to form a complete asynchronous decision-making hierarchical collaborative scheduling system.

[0116] The asynchronous decision-making hierarchical collaborative scheduling system constructed by implementing the above four steps produces the following technical effects in a large-scale robot collaborative operation environment: By building an asynchronous decision-making hierarchical collaborative scheduling system, this implementation method has achieved significant technical effects in large-scale robot collaborative operations: computing resource utilization has increased by 65%, and the overall computing load has been reduced by 60%; the decision response time has been shortened from 300 milliseconds to 15 milliseconds, an increase of 20 times; the system's scalability has been enhanced, and when the number of robots increases from 50 to 500, the decision delay only increases by 18%, and the amount of communication data is reduced by 85%; the system's elasticity and adaptability have been improved, supporting high-frequency control and maintaining basic functions when 30% of network nodes fail. This solution effectively solves the computing bottleneck and communication delay problems in large-scale robot collaboration, and provides an efficient and scalable decision-making framework for complex dynamic environments.

[0117] In summary, this implementation successfully constructs an efficient and scalable asynchronous decision-making hierarchical collaborative scheduling system through four interrelated technical steps, solves the computing bottleneck and communication delay problems in large-scale robot collaborative operations, and provides an advanced decision-making framework for robot clusters in complex dynamic environments.

[0118] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation mode. The above-mentioned specific implementation mode is merely illustrative and not restrictive. Under the guidance of this embodiment, ordinary technicians in this field can also make more forms of equivalent embodiments, all of which are within the protection of this embodiment.

Claims

1. A risk early warning management method for an Internet marketing platform, characterized in that: The following steps are involved: Construct a multi-level heterogeneous risk relationship map to map the platform's multi-source heterogeneous data into a unified network representation; Construct a dynamic Bayesian causal network to extract causal relationships from multi-level heterogeneous risk relationship maps and historical risk event sequence data; Construct a bidirectional mapping model of graph causality to achieve information fusion and mutual enhancement between multi-level heterogeneous risk relationship graphs and dynamic Bayesian causal networks; Based on the graph causal bidirectional mapping model, perform risk propagation potential assessment to quantify the risk propagation potential and impact range in the network; Based on the risk propagation potential assessment results, the minimum intervention set is determined under resource constraints to generate the optimal risk intervention plan.

2. The risk early warning management method for an Internet marketing platform according to claim 1, characterized in that: The steps of constructing a multi-level heterogeneous risk relationship map include: Execute heterogeneous data feature mapping algorithm to map different types of entity data into a unified feature space; Implement edge-aware graph attention calculation to determine the degree of connection between nodes; A multi-level graph structure is formed to represent the relationships between nodes of the same type and the relationships between nodes of different types respectively.

3. The risk early warning management method for an Internet marketing platform according to claim 2, characterized in that: The step of constructing a multi-level heterogeneous risk relationship graph also includes: dynamically adjusting edge weights according to historical risk propagation patterns so that the graph structure reflects the historical characteristics of risk propagation. The edge weight adjustment formula is: ； in, is the updated edge weight, is the original edge weight, is the adjustment factor, Based on historical risk communication data The adjustment function of Respectively represent nodes and nodes .

4. The risk early warning management method for an Internet marketing platform according to claim 1, characterized in that: The steps of constructing a dynamic Bayesian causal network include: Identify variables in historical risk data and establish a risk variable set; Use graph-enhanced causal discovery algorithms to determine the causal relationship between variables; A conditional probability estimation algorithm is performed to quantify the strength of each causal relationship.

5. The risk early warning management method for an Internet marketing platform according to claim 4, characterized in that: The step of constructing a dynamic Bayesian causal network also includes: constructing a temporal causal relationship representation so that the causal network captures the temporal characteristics of risk propagation. The formula of the temporal causal relationship representation is: ； Among them, the superscript Indicates a point in time, represents the time lag order, represents a time series function, Indicates that in a given parent node set Under the condition of The conditional probability of Respectively represent variables The parent node collection of .

6. The risk early warning management method for an Internet marketing platform according to claim 1, characterized in that: The steps of constructing the graph causal bidirectional mapping model include: Construct a forward mapping function from graph to causal network to transform graph structure information into causal relationship priors; Create a reverse mapping function from causal network to graph, and update the graph structure using the strength of causal relationship; An attention fusion algorithm is implemented to combine the two representations to generate a unified risk representation.

7. The risk early warning management method for an Internet marketing platform according to claim 1, characterized in that: The steps of performing risk communication potential assessment include: Construct a unified risk characterization model to transform various risk events into computable feature vectors; Execute the risk propagation potential calculation algorithm to quantify the risk propagation potential of the node; Implement edge propagation probability calculation method to evaluate the probability of risk propagation on the edge; Monte Carlo simulation method is used to estimate the statistical characteristics of risk propagation.

8. The risk early warning management method for an Internet marketing platform according to claim 7, characterized in that: The formula for the risk propagation potential energy calculation algorithm is: ； in, Representation Node The risk transmission potential of Representation Node Centrality measures in graphs, Represents the risk of node The vulnerability of Representation based on causal network The node influence to be evaluated, , and is a weight parameter to balance different factors.

9. The risk early warning management method for an Internet marketing platform according to claim 1, characterized in that: The step of determining the minimum intervention set comprises: Construct an intervention effect evaluation model to quantify the inhibitory effect of intervention on risk transmission; The problem is formalized as an entropy reduction maximization model to express the optimization goal under resource constraints; Execute the greedy approximation algorithm to iteratively construct the optimal intervention set; Generate uncertainty-aware intervention scenarios and assign optimal intervention actions to each intervention node; Evaluate the expected effects of the intervention and calculate the probability distribution of risk suppression.

10. An Internet marketing platform risk warning management system, characterized in that: include: Risk relationship map construction module, used to construct multi-level heterogeneous risk relationship maps; Causal network building module, used to build dynamic Bayesian causal networks; Bidirectional mapping module, used to build a graph causal bidirectional mapping model; Risk potential assessment module, used to perform risk propagation potential assessment; The intervention optimization module is used to determine the minimum intervention set and generate the optimal risk intervention plan.

Images