An internet marketing method based on big data

Abstract

The application discloses an internet marketing method based on big data, and relates to the technical field of internet marketing.The method comprises the following steps: deploying an Apache Flink distributed stream computing engine, and real-time accessing and cleaning user behavior stream, transaction stream and public opinion stream; dynamically constructing a Bayesian network graph structure based on conditional independence test and scoring search algorithm, and monitoring causal evolution mutation through time sequence substructure comparison; extracting user network topology features to calculate comprehensive influence weight, determining propagation potential value in combination with risk content propagation characteristics, and predicting propagation trend by using a discrete time SIR model and Monte Carlo simulation; constructing a combination optimization objective function with intervention cost and effect, and solving an optimal intervention node set through heuristic search; and adaptively adjusting strategy parameters through incremental effect evaluation and strategy gradient optimization after intervention.The application realizes real-time causal tracing and precise intervention of marketing risks, and improves the timeliness and effectiveness of risk control.

Description

Technical Field

[0001] This invention belongs to the technical field of internet marketing, specifically relating to an internet marketing method based on big data. Background Technology

[0002] In internet marketing scenarios, user behavior data, transaction data, and public opinion data are continuously generated at high frequency and high throughput. Marketing risks also exhibit dynamic evolution and rapid spread. How to identify risk factors from massive real-time data streams and implement precise interventions is the core challenge of marketing risk management.

[0003] Common marketing risk early warning methods typically rely on static risk maps, historical data statistics, or batch data analysis frameworks to conduct correlation analysis on risk factors and predict risks based on historical patterns. However, correlation analysis can only reveal the superficial relationships between risk factors and cannot characterize the intrinsic causal mechanisms of risk occurrence and spread. This leads to biases in the root causes of risks derived from correlation analysis, making it difficult for subsequent intervention strategies to directly target the key points of risk spread. The intervention effects are not sustainable and resources are wasted. How to achieve causal inference and risk tracing in a real-time data stream environment is a key challenge that urgently needs to be solved in the field of marketing risk early warning. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the present invention provides an Internet marketing method based on big data, which can effectively solve the problems mentioned in the background technology.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] This invention provides an internet marketing method based on big data, which includes the following steps:

[0007] Deploy a distributed stream computing engine based on Apache Flink, configure an event time processing mechanism and a periodic watermark generation strategy, and access user behavior streams, transaction streams and public opinion streams in real time for cleaning and format standardization;

[0008] Based on conditional independence test and scoring search algorithm, a Bayesian network graph structure is dynamically constructed from standardized data stream, and the node probability distribution is updated with incremental variational Bayesian method.

[0009] The historical data stream is divided into continuous time periods according to a fixed time length. A Bayesian network substructure for each time period is constructed. The confidence weights and direction changes of the current time period are compared with those of the historical baseline structure. When the mutation condition is met, a causal evolution alarm is triggered and the baseline is updated.

[0010] Perform network topology analysis on user social networks to extract degree centrality, betweenness centrality, clustering coefficient and community structure, calculate the user's comprehensive influence weight, determine the risk propagation potential value by combining risk content propagation characteristic parameters, and predict the propagation trend through discrete-time SIR model and Monte Carlo simulation.

[0011] A combinatorial optimization problem is constructed based on the total intervention cost and the total intervention effect, and a heuristic search with priority scheduling is used to solve for the optimal set of intervention nodes.

[0012] After the intervention is implemented, the intervention strategy parameters are adaptively adjusted through incremental effect evaluation and strategy gradient optimization.

[0013] Furthermore, the deployment of the Apache Flink-based distributed stream computing engine also includes:

[0014] The Apache Flink-based distributed stream computing engine is further configured with a stream-batch integrated computing architecture. Relying on a unified operator library, it abstracts the processing tasks of bounded and unbounded data into data stream operations. At the same time, it supports mixed processing logic of bounded and unbounded data. While processing real-time data, it switches to batch processing mode to perform historical data backtracking analysis and compares the currently monitored risk indicators with the historical level of the same period.

[0015] The Apache Flink-based distributed stream computing engine establishes a state backend management mechanism, which uses a distributed file system or embedded key-value storage to persistently save intermediate states during the computing process. The intermediate states include the aggregated cumulative values ​​of each time window, user session interaction sequences, and probability distribution parameters generated by subsequent causal inference steps. With the help of checkpoint and savepoint strategies, the computing task can be restored from the most recent consistent snapshot when a node fails.

[0016] Furthermore, the real-time access to user behavior streams, transaction streams, and public opinion streams, followed by cleaning and format standardization, specifically includes:

[0017] Through different topics of the unified message middleware, the stream computing engine subscribes to user behavior streams, transaction streams and public opinion streams in the form of consumer groups. For user behavior streams, it filters out expired records with timestamps earlier than the current system time exceeding the threshold and invalid records with empty user identifiers. For transaction streams, it filters out test order records with amounts of zero or negative numbers and converts the transaction amount into a base currency unit. For public opinion streams, it filters out meaningless records with too short text and performs a unified time zone conversion on the publication timestamp.

[0018] In the periodic watermark generation strategy, the water level value is the maximum value among all observed event timestamps minus the preset allowed delay time. When the water level value exceeds the end timestamp of a certain time window, the window calculation is triggered and the aggregation result is output. Late data whose actual timestamp is less than the current water level is judged as extremely late data and directly discarded.

[0019] Furthermore, the method of dynamically constructing a Bayesian network graph structure from the standardized data stream based on conditional independence testing and a scoring search algorithm specifically includes:

[0020] Based on the knowledge of experts in the field of internet marketing, a set of risk factor nodes for the Bayesian network is set in advance. The set of risk factor nodes includes marketing activity parameter nodes, user profile feature nodes, market environment variable nodes, and competitor behavior nodes. No directed edges are set between the nodes for the time being.

[0021] Within each data processing cycle, for any two nodes and a given set of conditional nodes, calculate the conditional mutual information of the two nodes under the given set of conditional nodes. When the conditional mutual information is lower than a preset independence threshold, determine that the two nodes are independent under the given set of conditional nodes.

[0022] For node pairs that are confirmed to have dependencies, a scoring search method based on the Bayesian information criterion is adopted to calculate the network score when adding a directed edge in the first direction and the network score when adding a directed edge in the second direction, respectively. The direction with the better network score is selected as the final direction of the directed edge, and directed edges are added step by step to form a directed acyclic graph structure.

[0023] Newly added directed edges are synchronously updated in the graph model. As data continues to accumulate, previously identified directed edges are either overturned by new evidence and removed or new, previously undiscovered edges are added.

[0024] Furthermore, the step of updating the node probability distribution using the incremental variational Bayesian method specifically includes:

[0025] The variational distribution approximates the true posterior distribution. Each time new data arrives, only the variational parameters of the node to which the data belongs, its parent node, and its child nodes are updated. The update is completed when the KL divergence value is lower than the preset convergence threshold by minimizing the variational distribution and the true posterior.

[0026] After a single node probability update is completed, the node probability change is propagated sequentially to all child nodes along the directed edge direction, and simultaneously propagated in the reverse direction to all parent nodes. Each node that receives the message recalculates its own probability distribution based on the conditional probability table and continues to propagate the change outward until a complete round of message propagation ends, and the probability value change of all nodes in the two adjacent iterations is lower than the preset convergence threshold.

[0027] Furthermore, the process of dividing the historical data stream into continuous time periods according to a fixed time length, constructing a Bayesian network substructure for each time period, comparing the confidence weights and direction changes of the current time period with the historical baseline structure, and triggering a causal evolution alarm and updating the baseline when a mutation condition is met, specifically includes:

[0028] The historical data stream is divided into continuous time periods with a fixed time length, and the structure learning algorithm is run independently for each time period to construct the Bayesian network substructure of that time period.

[0029] The historical baseline structure is given by the statistical aggregation results of substructures from several past time periods;

[0030] For each pair of directed edges, calculate the confidence weight of the directed edge in the current time period substructure and the confidence weight in the historical benchmark structure, and calculate the weight change as the absolute value of the difference between the two. At the same time, check whether the direction of the directed edge is consistent in the current time period and the historical benchmark.

[0031] When the weight change of at least one directed edge exceeds the preset weight change tolerance threshold, or when the direction of at least one directed edge reverses between two consecutive periods, a sudden change in causal relationship is determined, a causal evolution alarm is triggered, and the substructure of the current period is marked as the new baseline. Subsequent inference tasks are all executed based on the new graph structure.

[0032] When each data event arrives, incremental updates and inference calculations are performed on the cause-effect graph structure. The incremental updates and inference calculations include four stages: data preprocessing, node probability updates, edge weight adjustments, and inference queries.

[0033] Furthermore, the calculation of the user's overall influence weight, combined with the risk content dissemination characteristic parameters to determine the risk dissemination potential value, specifically includes:

[0034] Static attribute features and dynamic behavioral features of each user node are extracted. Static attribute features include historical credit score, account level, and authentication status. Dynamic behavioral features include activity frequency, content posting volume, and interaction response rate within the recent time window. All indicators are pre-normalized to the range of 0 to 1.

[0035] For the There are n user nodes, and let the static attribute feature vector be . The dynamic behavior feature vector is , No. individual users at any time Overall influence weight ,in and For preset weighting coefficients and ;

[0036] The sensitivity level of risky content, the adaptability of the dissemination medium, and the reproducibility of the dissemination content are quantified into values ​​between 0 and 1. The arithmetic mean of the three is taken as the comprehensive adjustment coefficient. The comprehensive influence weight of each user node is multiplied by the comprehensive adjustment coefficient to obtain the dissemination potential value of each user node.

[0037] User nodes are sorted according to their propagation potential value, with the top-ranked nodes receiving the highest propagation potential value. User nodes at a certain number of positions are identified as a set of high-risk propagation nodes.

[0038] Furthermore, the prediction of the propagation trend using the discrete-time SIR model and Monte Carlo simulation specifically includes:

[0039] Using a discrete-time SIR model framework, each user node is divided into a susceptible state, a propagating state, or an immune state at any given time. Within each discrete time step, a susceptible node transforms into a propagating node with a transition probability determined by the propagation potential value, while a propagating node transforms into an immune node with a fixed recovery probability.

[0040] Based on the Monte Carlo simulation method, starting from the set of high-risk propagation nodes as the initial propagation node set, the state update is iteratively performed until the number of propagation state nodes drops to 0 or the preset maximum number of simulation steps is reached. After a single simulation, all user nodes that are ultimately infected in the user's social network are recorded. After repeating the simulation multiple times, the probability distribution of each user node being infected is statistically analyzed, and the confidence interval estimate of the propagation coverage is output.

[0041] Furthermore, the construction of a combinatorial optimization problem based on the total intervention cost and the total intervention effect, and the use of a priority-based heuristic search to solve for the optimal set of intervention nodes, specifically includes:

[0042] Define decision variables Indicates whether to intervene in user nodes 1 indicates that the intervention was selected, and 0 indicates that it was not selected.

[0043] The constraints include that the total number of user nodes selected into the intervention set shall not exceed the preset maximum number of interventionable nodes, and that the selected user nodes must be online or have a valid contact method to receive intervention information.

[0044] The objective function is ,in For the decision variable vector, For the overall effect of the intervention, To normalize the total cost of intervention, This is the cost penalty coefficient;

[0045] Overall Intervention Effect Elimination rate of transmission source Transmission path blocking rate and the rate of contraction of the spread We get the weighted sum. ,in , , For preset weighting coefficients and ;

[0046] Source elimination rate The transmission path blocking rate is the ratio of the number of high-risk transmission nodes selected for the intervention set to the total number of high-risk transmission nodes. The propagation range contraction rate is the percentage reduction in the number of remaining propagation paths relative to the original total number of propagation paths after simulated path removal with the selected intervention node as the starting or ending point. =1 minus the ratio of the transmission coverage predicted by the Monte Carlo simulation after intervention to the transmission coverage predicted without intervention;

[0047] Normalized total intervention cost It is obtained by dividing the sum of the fixed cost allocation and the marginal cost of each intervention node by the preset maximum possible total intervention cost, i.e. ,in The set of user nodes selected for the intervention set. , , For fixed costs, The maximum number of nodes that can be intervened. To pre-determine the basic marginal cost constant, To pre-determine the maximum possible total cost of intervention, For the user node The overall influence weight;

[0048] A greedy search algorithm is used, starting from an empty set and iterating. In each iteration, the search priority is determined from the candidate nodes that have not yet been selected, based on the ratio of the comprehensive influence weight of each node to its marginal cost. Nodes with higher ratios are evaluated first in the search process, and only those with the highest priority are considered. For each candidate node, the objective function increment is calculated, and the node that maximizes the objective function increment is added to the intervention set. The iteration continues until the size of the intervention set reaches the maximum number of interventionable nodes or no candidate node that can increase the objective function value can be found.

[0049] Furthermore, the adaptive adjustment of intervention strategy parameters through incremental effect evaluation and strategy gradient optimization specifically includes:

[0050] An incremental effect evaluation index system was established. Direct effect indicators include the proportion of the number of nodes covered by the intervention that changed from the propagation state to the immune state, the intervention response rate, and the intervention resistance rate. Indirect effect indicators include changes in the overall network propagation speed, changes in the strength of the inter-community propagation barriers, and the risk event decay cycle. Each indicator is configured with a baseline value and a target value. The baseline value is taken from the indicator prediction value predicted by Monte Carlo simulation under the condition of no intervention.

[0051] Using an A / B testing framework, the intervention subjects were divided into an experimental group and a control group. The experimental group performed the intervention operation at the optimal intervention node set, while the control group maintained the original propagation state without additional intervention. A two-sample t-test was used to determine whether the difference between the two groups in key effect indicators was statistically significant. When the p-value was lower than the preset significance level threshold, the intervention strategy was deemed effective.

[0052] A feedback adjustment mechanism is constructed, comprising four stages: effect monitoring, anomaly detection, strategy adjustment, and strategy distribution. The effect monitoring stage continuously collects data on various effect indicators. The anomaly detection stage determines an anomaly when the actual observed value of a certain indicator is lower than 80% of the target value. The strategy adjustment stage selects parameter fine-tuning, strategy switching, or intervention restart based on the anomaly type. Parameter fine-tuning adopts the strategy gradient optimization method to update the intervention strategy parameters along the gradient direction of the comprehensive effect evaluation value. The intervention strategy parameters include the cost penalty coefficient and the weight coefficients of the source elimination rate, the transmission path blocking rate, and the transmission range contraction rate. The strategy distribution stage pushes the adjusted strategy to the intervention implementation module for execution.

[0053] In summary, this application includes at least one of the following beneficial technical effects:

[0054] 1. This invention uses dynamic Bayesian networks to learn the causal structure and perform probabilistic reasoning on real-time data streams. It can accurately locate the key root cause nodes of risk propagation and their evolution path in the early stages of risk events, effectively reduce the misjudgment rate caused by causal inversion or false correlations, and significantly improve the accuracy and interpretability of risk root cause identification.

[0055] 2. This invention deeply integrates user social network topology analysis, comprehensive influence assessment, and discrete-time propagation dynamics model to construct a comprehensive effect quantification function that takes into account the elimination of propagation sources, the blocking of propagation paths, and the contraction of propagation scope. Combined with an intervention cost quantification model and a heuristic search algorithm with priority scheduling, it solves the optimal set of intervention nodes that maximizes comprehensive benefits under the constraint of limited marketing intervention resources. This avoids the blind deployment and excessive consumption of intervention resources and significantly improves the resource utilization efficiency and input-output ratio of marketing risk intervention.

[0056] 3. This invention constructs an incremental effect evaluation index system and introduces an A / B testing framework and a strategy gradient optimization mechanism, enabling intervention strategies to adaptively fine-tune parameters and dynamically iterate and update strategies based on real-time feedback data. At the same time, in conjunction with the state backend management and checkpoint recovery mechanism of the stream computing engine, it ensures the continuous effectiveness, robustness and rapid response capability of marketing risk management across the entire chain in complex and ever-changing environments, realizing a fundamental shift in marketing risk management from static decision-making to dynamic evolution. Attached Figure Description

[0057] Figure 1 A schematic diagram of the overall technical solution architecture for the big data-based internet marketing method provided by this invention;

[0058] Figure 2 This is a schematic diagram illustrating the core principle framework of the fusion architecture based on the Apache Flink stream computing engine and the dynamic Bayesian network causal inference model in this invention.

[0059] Figure 3 This is a complete closed-loop process framework diagram of risk perception, causal tracing, precise intervention, and effect evaluation in this invention;

[0060] Figure 4 This is a schematic diagram of the data flow framework for the interaction between the risk propagation potential assessment model and the minimum intervention set optimization algorithm in this invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with the appendix. Figure 1 To be continued Figure 4 The present invention will be further described in detail below with reference to specific embodiments.

[0062] This invention provides a big data-based internet marketing method, which includes five core steps: deploying a distributed stream computing engine and accessing multi-source heterogeneous data streams in real time; dynamically constructing and updating the causal graph structure of risk factors; constructing a risk propagation potential assessment model; calculating the optimal set of intervention nodes; and evaluating the effectiveness of intervention measures in real time. The technical implementation of each step is described in detail below.

[0063] The first step, S1, involves deploying a distributed stream computing engine and connecting to multi-source heterogeneous data streams in real time. This establishes the data access and preprocessing infrastructure for the entire marketing risk management method. The following steps will be followed.

[0064] Step S101: Deploy an Apache Flink-based distributed stream computing engine as the core data access layer. The marketing business system front-end deploys the Apache Flink distributed stream computing engine. The engine adopts a cluster deployment mode, and achieves elastic scheduling and load balancing of computing resources through a master-slave architecture of task manager and job manager to cope with traffic surges during peak marketing campaigns.

[0065] The engine is configured with an event time processing mechanism that calculates windows based on the timestamps carried by the data itself, rather than relying on the time the data arrives at the system. Even if user network latency causes behavior logs to arrive later than the event's occurrence time, the engine can still categorize them into the correct time window based on the original timestamps in the logs, ensuring the accuracy of window aggregation metrics such as add-to-cart counts per minute or pageviews per minute.

[0066] Step S102: Configure the integrated stream-batch computing architecture and status backend management mechanism. The aforementioned stream computing engine is further configured with an integrated stream-batch computing architecture. Relying on a unified operator library, data processing tasks are abstracted into data stream operations, while supporting mixed processing logic for bounded and unbounded data. Bounded data refers to batch data, such as yesterday's full historical orders; unbounded data refers to real-time stream data, such as the current transaction payment event stream. With this architecture, this method can switch to batch processing mode for historical data backtracking analysis while processing real-time data, facilitating comparison of currently monitored risk indicators with historical levels for the same period.

[0067] The engine also establishes a state backend management mechanism, which persistently saves intermediate states during the computation process using a distributed file system or embedded key-value storage. This includes aggregated cumulative values ​​for each time window, user session interaction sequences, and probability distribution parameters generated by subsequent causal inference steps. This state backend management mechanism, combined with checkpoint and savepoint strategies, restores the computation task from the most recent consistent snapshot when a node fails, ensuring the continuity of marketing risk monitoring tasks.

[0068] Step S103: Configure a watermarking strategy to handle out-of-order data events. For data out-of-order issues caused by network latency or mobile reporting mechanisms in internet marketing environments, the stream computing engine configures a periodic watermark generation strategy. The watermark value is taken as the maximum value among all currently observed event timestamps minus the preset allowed delay duration.

[0069] The watermarking strategy works in conjunction with the event time window calculation, with the following judgment logic: The water level is continuously monitored. Once the water level value exceeds the end timestamp of a certain time window, it is considered that the data for that window has been basically collected. The window calculation is then triggered, and the aggregation result is output. Late data whose actual timestamp is less than the current water level is judged as extremely late data and discarded directly, no longer participating in any window calculations, preventing indefinite waiting from causing system state bloat.

[0070] Step S104: Real-time access to multi-source heterogeneous data streams through unified message middleware. After the engine configuration is completed, the three types of core data sources are accessed to the stream computing engine through unified message middleware.

[0071] User behavior streams include user interaction data such as browsing, clicking, saving, and sharing on marketing platforms. Each record carries at least the user identifier, operation type, operation object, timestamp, and session identifier fields.

[0072] The transaction flow includes transaction status change data such as order creation, payment completion, and refund application. Each record must include at least the transaction identifier, user identifier, product identifier, transaction amount, timestamp, and transaction status fields.

[0073] The sentiment stream includes user comments, topic discussions, and sentiment data collected from external social media, forums, and news channels. Each record includes at least the information source, the user who posted it, the content text, the posting timestamp, and the sentiment polarity score.

[0074] The three types of data streams are logically isolated through different topics in the message middleware, and the stream computing engine subscribes to each topic in the form of consumer groups. Simultaneously with data access, the engine performs preliminary cleaning and format standardization on the three types of data streams: for user behavior streams, expired records with timestamps earlier than the current system time exceeding a threshold and invalid records with empty user identifiers are filtered out; for transaction streams, test order records with zero or negative amounts are filtered out, and all transaction amounts are uniformly converted to the base currency unit; for public opinion streams, meaningless records with excessively short text are filtered out, and the publication timestamps are uniformly converted to the time zone. The cleaned and standardized data is stored in an in-memory data buffer, awaiting the invocation of the Bayesian network structure learning algorithm in subsequent step S2.

[0075] Step S105: Establish subscription relationships between data streams and computation tasks, and dynamically allocate parallelism based on data throughput. The stream computing engine routes data streams from different topics to corresponding data processing tasks according to the pre-configured subscription relationships. Each type of data stream is allocated an independent computation slot, and each computation slot corresponds to a parallel subtask instance of a processing task. The parallelism allocation is calculated according to the following formula:

[0076]

[0077] In the above formula, the subscript Indicates the data stream type index. Represents the user behavior flow. Represents transaction flow, Represents public opinion flow. For the first The degree of parallelism assigned to a data stream processing task is a dimensionless integer. This is the preset minimum parallelism constant, with a value of 2, which is a dimensionless integer. For the first The peak throughput of this type of data stream within the most recent statistical period, expressed in messages per second. Recommended processing capacity threshold for a single parallel subtask under standard resource configuration, in units of messages per second. (Symbol) This indicates the rounding up operation. This indicates that the larger of the two values ​​is taken.

[0078] This formula originates from load balancing resource allocation methods in the field of parallel computing. Its core principle is to divide the data load per unit time by the service capacity of a single processing unit, and round the quotient up to obtain the minimum number of processing units required to satisfy the current load. This method further introduces a minimum parallelism constant. The maximum value is taken to ensure that this method retains a minimum level of parallel fault tolerance under all circumstances. Dimensional review: and Both units are bars per second. The division between the two yields a dimensionless numerical value, which can be rounded up or divided by a dimensionless constant. Compare and take the maximum value, then output the final value. It is a dimensionless integer. This formula operates independently on three types of data streams, without coupling between them.

[0079] With the above allocation strategy, when user behavior streams experience explosive growth during marketing campaigns, the processing tasks for sentiment analysis and transaction streams will not be affected by resource contention. The engine distributes the cleaned data to each subtask instance according to the allocated parallelism, achieving parallel processing of multi-source data.

[0080] Step S1 completes the deployment of the distributed stream computing engine and the real-time access of multi-source heterogeneous data. Event-time processing mechanisms and watermarking strategies address data out-of-order issues in high-concurrency scenarios, while the integrated stream and batch architecture establishes a channel for real-time monitoring and historical backtracking. The parallelism allocation formula transforms dynamically changing data throughput into a deterministic allocation of computing resources.

[0081] The next step is S2, which involves dynamically constructing and updating the causal graph structure of risk factors. Based on the standardized event flow formed in step S1, step S2 uses a dynamic Bayesian network to automatically discover, update, and infer the causal relationships between risk factors in real time, providing a causal basis for subsequent dissemination assessment and intervention optimization. This is carried out according to the following steps.

[0082] Step S201: Based on domain knowledge, a set of risk factor nodes is pre-set to complete the node initialization of the Bayesian network graph model. In the startup phase, based on expert knowledge in the field of internet marketing, a set of risk factor nodes for the Bayesian network is pre-set.

[0083] The node set covers marketing campaign parameter nodes, including discount level, campaign duration, and distribution channel combination; user profile feature nodes, including user spending power level, historical purchase frequency, and activity score; market environment variable nodes, including competitor promotion intensity and industry public opinion index; and competitor behavior nodes, including competitor price change range and competitor advertising volume.

[0084] The aforementioned nodes serve as the basic building blocks of the Bayesian network graph model, providing a variable space for subsequent structure learning.

[0085] After the nodes are initialized, no directed edges are set between the nodes. The graph model is represented as a set of nodes without edges, waiting for subsequent steps to learn the structure based on real-time data.

[0086] Step S202 involves using conditional independence testing and a scoring search algorithm to dynamically discover directed edge relationships between nodes from the real-time data stream, constructing a Bayesian network graph structure. As the real-time data stream from step S1 continues to flow in, the structure learning algorithm dynamically discovers the directed edge relationships between nodes. Within each data processing cycle, the algorithm performs a conditional independence test on the current node set.

[0087] For any two nodes and and the given set of condition nodes. Algorithm calculation and In the given Conditional mutual information under conditions When the conditional mutual information is lower than a preset independence threshold, a decision is made. and In the given They are independent of each other under certain conditions, meaning there is no direct dependency between them.

[0088] For node pairs confirmed to have a dependency relationship through the conditional independence test, the algorithm further employs a scoring search method based on the Bayesian information criterion to evaluate the contribution of directed edges in different directions to the overall network fit. For a pair of nodes confirmed to have a dependency relationship... and The algorithm calculates the added directed edges separately. Network rating and the addition of directed edges The network score at that time is used to select the direction with the higher network score as the final direction of the edge.

[0089] The network scoring function is constructed based on the Bayesian information criterion; a lower score indicates a better fit of the network to the data. The algorithm progressively adds directed edges between nodes, forming a directed acyclic graph structure.

[0090] Newly added directed edges are synchronously updated in the graph model for subsequent probability updates and inference. As data continues to accumulate, previously identified directed edges may be overturned by new evidence and removed, or previously undiscovered edges may be added, thus enabling the continuous evolution of the graph structure.

[0091] Step S203: When new observation data arrives, identify the node to which the data belongs, update the marginal probability distribution of the node using Bayes' theorem, and propagate the probability change to all parent and child nodes through a message passing algorithm to achieve a consistent update of the probability distribution of the entire graph.

[0092] Each standardized data record output in step S1 corresponds to one or more risk factor nodes. When a new record arrives, the algorithm first determines the node to which it belongs based on the field mapping relationships in the record. For the node, the algorithm uses Bayes' theorem to calculate the marginal posterior probability given the new observational evidence, replacing the original marginal probability value.

[0093] Let the prior probability distribution of the node be... New observational evidence is Then the updated posterior probability According to Bayes' theorem formula The calculation yielded the following. Represents a node The prior probability, Indicates at node Evidence observed when taking specific values The likelihood probability, As evidence The global marginal probability.

[0094] After a single node's probability is updated, the algorithm initiates a message passing mechanism. Along the directed edges, the probability change of the node is propagated sequentially to all its child nodes, and simultaneously in the reverse direction to all its parent nodes. Each node that receives the message uses its conditional probability table to... Recalculate its own probability distribution and continue to propagate the changes outward until the probability distributions of all nodes in the entire graph model converge.

[0095] in Indicates the first Each risk factor node Represents a node The set of all parent nodes This represents a node given the value of its parent node. The conditional probability distribution is determined by the following condition: after a complete round of message propagation, the change in probability values ​​of all nodes in two adjacent iterations is lower than a preset convergence threshold.

[0096] To meet the timeliness requirements of real-time processing scenarios, the above update process is implemented using an incremental variational Bayesian method. The variational Bayesian method uses a variational distribution... Approximate true posterior distribution ,in For the set of model parameters, For observed data, Indicates that in the given observation data Model parameters under conditions The posterior probability distribution.

[0097] Each time new data arrives, the algorithm only updates the variational parameters of local nodes and their neighbors, without performing a full Bayesian inference on the entire graph. The goal of the variational parameter update is to minimize the KL divergence between the variational distribution and the true posterior. The update is considered complete when the KL divergence value is lower than the preset convergence threshold.

[0098] Step S204: Establish a time-series causal evolution model, introduce a time window mechanism to characterize and monitor the evolution of causal relationships over time, and divide the historical data stream into continuous time periods according to a fixed time length.

[0099] Within each time period, this method independently runs the structure learning algorithm in step S202 to construct the Bayesian network substructure for that time period. After the time period ends, the substructure of the current time period is compared and analyzed with the historical baseline structure. The historical baseline structure is given by the statistical aggregation results of the substructures from several past time periods.

[0100] The comparative analysis is performed according to the following decision-making process. For each pair of directed edges... The algorithm calculates the confidence weight of each component in the current time period substructure. confidence weights in the historical benchmark structure Then calculate the magnitude of the weight change. ,in The confidence weight of the directed edge in the current time period. The confidence weights of directed edges in the historical baseline are used. This is the absolute value of the difference between the two. It also checks whether the direction of the directed edge is consistent with the historical baseline in the current time period.

[0101] A sudden change in causality is determined when any of the following conditions is triggered: Condition 1: There exists at least one directed edge whose weight changes by a certain magnitude. The preset weight change tolerance threshold is exceeded. Condition 2: There exists at least one directed edge whose direction is reversed between the two periods. When either condition is met, a causal evolution alarm is triggered, and the substructure of the current period is marked as the new baseline. Subsequent inference tasks are all performed based on the new graph structure.

[0102] Upon each incoming data event, regardless of whether a structural mutation alarm is triggered, incremental updates and inference calculations are performed on the cause-effect graph structure. The incremental update process includes four stages: data preprocessing, node probability updates, edge weight adjustments, and inference queries.

[0103] The data preprocessing stage receives the original records and completes field mapping and standardization; the node probability update stage executes the Bayesian update and message passing described in step S203; the edge weight adjustment stage dynamically scales the confidence weight of each directed edge according to the statistical characteristics of recent data; and the inference query stage responds to the currently received evidence variable and calculates the posterior probability of the target variable.

[0104] In step S205, the inference engine receives the current causal graph structure and probability distribution, calculates the posterior probability of the target variable based on the given set of evidence variables, and generates a risk alarm event when the probability exceeds a preset threshold. The inference engine uses the latest causal graph structure output in step S204 and the current probability distribution maintained in step S203 as input. When a new data event accessed in step S1 is identified as an evidence variable, the variable is set as an evidence node, and an approximate inference algorithm is called to calculate the posterior probability of all target variables.

[0105] The approximate inference algorithm employs a variational message passing method, using the current causal graph structure as the computation graph. Starting from the evidence node, messages are propagated layer by layer along the directed edges. After receiving messages from the parent and child nodes, each node calculates its own marginal posterior probability based on the conditional probability table, until the posterior probability values ​​of all nodes in the entire graph converge.

[0106] This method configures an independent alarm threshold for each risk factor node, which is pre-calibrated based on the frequency and impact of historical risk events. After each round of calculation, the inference engine iterates through all risk factor nodes, comparing the posterior probability of each node with the corresponding threshold. If the posterior probability of a node exceeds the threshold, a risk alarm event is immediately generated. The alarm event includes the risk factor node identifier, the current posterior probability value, the trigger timestamp, and a list of key evidence variables that led to the alarm. After the alarm event is generated, it is pushed to the propagation potential assessment module in step S3 and the intervention optimization module in step S4.

[0107] In summary, through steps S201 to S205, the entire chain of processing—including initialization of the dynamic Bayesian network causal graph structure, real-time structure learning, conditional probability updating, temporal evolution monitoring, and causal inference and alerting—is completed. This process adaptively discovers and corrects causal relationships between risk factors based on real-time data streams, enabling the identification of risk root causes to dynamically adjust in response to changes in the marketing environment. Alert events, as the output of causal tracing, are integrated into the risk propagation potential assessment model in subsequent step S3.

[0108] For step S3 in the method of this embodiment, a risk propagation potential assessment model is constructed. Based on the causal graph structure and risk alarm events generated in step S2, the risk propagation potential assessment model is constructed to quantitatively assess the propagation ability of risk in users' social networks, and to provide a quantitative basis for the selection of subsequent intervention nodes, including the following specific implementation steps.

[0109] Step S301: Design the propagation potential energy function. The risk propagation process in the user network is analogized to the propagation of a risk signal in a potential energy field. This step draws on the concept of potential energy in physics to design the propagation potential energy function to quantify the propagation capability of risk in the user's social network. The propagation potential energy function takes the user node's own attribute data and its position data in the network as input, and outputs the user node's propagation potential energy value.

[0110] A higher propagation potential energy value indicates that the user node has a stronger risk signal amplification capability and a wider propagation coverage. The specific calculation method of the propagation potential energy function is implemented in subsequent steps S303 and S304 through the step-by-step calculation of influence weight value and propagation characteristic parameters.

[0111] Step S302 involves extracting the structural features of the user's social network based on network topology analysis methods, providing topological dimension parameters for propagation potential calculation. This step performs network topology analysis on the user's social network involved in the marketing platform. Network topology analysis includes calculations in four dimensions: degree centrality analysis, betweenness centrality analysis, clustering coefficient analysis, and community structure identification.

[0112] Degree centrality reflects the direct influence of a user node in the network and is determined by the number of direct connections to that user node. Betweenness centrality reflects the criticality of a user node as a network information relay station and is determined by the frequency with which the user node appears on the shortest path between other user node pairs.

[0113] Clustering coefficients reflect the density of information transmission around a user node, determined by the actual tightness of connections between the user node's neighbors. Community structure identification uses a modularity optimization algorithm to divide the user's social network into several communities with tight internal connections and sparse external connections, providing a foundation for subsequent hierarchical propagation modeling.

[0114] The calculation results of the above four dimensions are combined to form the topological feature vector of each user node, which serves as one of the input parameters for subsequent influence weight calculation.

[0115] Step S303 employs a dynamic influence assessment algorithm, integrating users' static attribute characteristics and dynamic behavioral characteristics to calculate the influence weight value of each user node at the current moment. This step calculates the comprehensive influence weight for each user node. The input to the dynamic influence assessment algorithm consists of two types of data: the static attribute characteristics and the dynamic behavioral characteristics of the user nodes.

[0116] Static attribute features include three indicators: the user's historical credit score, account level, and authentication status. Each indicator has been pre-normalized to the range of 0 to 1. Dynamic behavioral features include three indicators: the user's activity frequency, content posting volume, and interaction response rate within a recent time window. Each indicator has also been normalized to the range of 0 to 1.

[0117] The algorithm integrates static and dynamic features through a time-weighted fusion mechanism. For the... Let there be *n* user nodes, and let their static attribute feature vectors be *i* and *j*. The dynamic behavior feature vector is ,in Indicates the current moment. , , These represent the normalized values ​​of historical credit score, account level, and authentication status, respectively. , , These represent the normalized values ​​of activity frequency, content publication volume, and interaction response rate within the recent time window, respectively. individual users at any time Overall influence weight Calculate using the following formula:

[0118]

[0119] In the above formula, and The preset weighting coefficients satisfy... , The value is 0.4. The value is 0.6. and All of them are dimensionless constants. This represents the arithmetic mean of each dimension of the static attribute feature vector. This represents the arithmetic mean of each dimension of the dynamic behavior feature vector; both are dimensionless scalars. The formula performs scalar operations on each dimension independently for each user node before summing.

[0120] By performing the weighted summation described above, we obtain the comprehensive influence weight value of each user at the current moment. This value falls within the range of 0 to 1. The overall influence weight value is updated periodically over time, with the update frequency consistent with the data processing cycle of the stream computing engine.

[0121] Step S304: Combining the risk content propagation characteristic parameters, the influence weights of the propagation characteristic parameters are superimposed on the influence weight values ​​to calculate the risk propagation potential energy distribution and identify high-risk propagation nodes. This step will use the influence weight values ​​of each user node output in step S303. A comprehensive calculation is performed in conjunction with the characteristics of risky content dissemination parameters.

[0122] The parameters for the characteristics of risky content dissemination include three factors: the sensitivity level of the risk information. The degree of compatibility of the transmission medium The replicability of the disseminated content The above three factors are each quantified as a value between 0 and 1, among which... A higher value indicates a greater sensitivity to risk information. A higher value indicates a better match between the communication medium and the target user group. A higher value indicates that the risky content is more easily copied and forwarded by users.

[0123] For the Each user node, its overall influence weight value Multiplied by a comprehensive adjustment coefficient based on the risk content dissemination characteristic parameters The propagation potential value of the user node is obtained. Comprehensive adjustment coefficient Based on sensitivity level Degree of compatibility with transmission media Reproducibility of disseminated content The arithmetic mean of the three values ​​is given:

[0124]

[0125]

[0126] In the above formula, , , All are dimensionless scalars and all are in the interval between 0 and 1. The arithmetic mean of the three values ​​also falls within the range of 0 to 1. and Both are dimensionless values, and their multiplication remains dimensionless. The formula performs scalar multiplication independently for each user node.

[0127] The above calculations yield the propagation potential energy distribution of the entire user network. This method sorts user nodes according to the propagation potential energy distribution, assigning the nodes with the highest propagation potential energy values. User nodes at a certain number of positions are identified as a set of high-risk propagation nodes. The value is determined based on the upper limit of the total intervention resources. The set of high-risk transmission nodes serves as the key input to the intervention optimization module.

[0128] Step S305: Based on the propagation dynamics equation and Monte Carlo simulation method, predict the risk propagation trend and impact range. This step predicts the risk propagation trend and impact range based on the propagation potential energy value distribution output in step S304.

[0129] The propagation trend prediction employs a discrete-time SIR model framework, classifying each user node into three states at any given time: susceptible, propagating, and immune. A susceptible state indicates the user has not yet been exposed to risk information but is at risk of infection; a propagating state indicates the user has been exposed to risk information and possesses the ability to spread it to other users; and an immune state indicates the user has lost the ability to spread the information or has undergone intervention. Within each discrete time step, susceptible nodes transform into propagating nodes with a certain probability, while propagating nodes transform into immune nodes with a fixed recovery probability. The transition probability is determined by the propagation potential value.

[0130] The impact assessment is based on Monte Carlo simulations to generate a large number of propagation path samples. Each Monte Carlo simulation follows the state transition rules described above, starting from the initial set of propagation nodes and iteratively updating the state until the number of propagation nodes drops to 0 or the preset maximum number of simulation steps is reached. After each simulation, all user nodes ultimately infected in the user's social network are recorded. After repeating the simulation multiple times, the probability distribution of infection for each user node is statistically analyzed, and the confidence interval estimate of the propagation coverage is output.

[0131] Step S305 takes the propagation potential energy distribution as input and uses a combination of propagation dynamics equations and Monte Carlo simulations to extend the assessment from static potential energy to dynamic propagation prediction.

[0132] In summary, step S3 completes the entire chain of processing, including defining the propagation potential function, extracting network topology features, calculating dynamic influence weights, calculating the distribution of propagation potential values, and predicting propagation trends. This process maps the risk alarm events output in step S2 into a risk propagation map with a clear order of propagation potential values ​​in the user's social network.

[0133] The next step, S4, involves using a heuristic optimization strategy to calculate the optimal set of intervention nodes based on the intervention cost-effectiveness quantification model. This formalizes the intervention resource optimization problem into a combinatorial optimization problem. The optimal set of intervention nodes is calculated using the cost-effectiveness quantification model and a heuristic search algorithm, providing a basis for decision-making in the implementation of subsequent intervention measures. This includes the following steps.

[0134] Step S401 formalizes the intervention resource optimization problem into a combinatorial optimization problem, determining the decision variables, objective function, and constraints. This step expresses the selection problem of intervention nodes as a standard combinatorial optimization problem. The decision variables are binary indicator variables indicating whether each user node is selected into the intervention set. For the _th ... Each user node defines a binary variable. ,in This indicates that the user node has been selected into the intervention set. This indicates that the candidate was not selected.

[0135] The objective function is defined as the weighted difference between the total intervention effect and the normalized total intervention cost. Let the total intervention cost be... The overall effect of the intervention was ,in Let the vector of binary decision variables be the sum of all user nodes. The total number of nodes in the user's social network. Objective function The expression is:

[0136]

[0137] in This is the cost penalty coefficient, with a value of 0.5, used to balance the trade-off between maximizing the effect and minimizing the cost. This represents the normalized total intervention cost. (Objective function value) The larger the value, the better the overall effectiveness of the intervention program.

[0138] The constraints include two items. The first is the upper limit constraint on the total number of intervention resources, that is, the total number of user nodes selected into the intervention set shall not exceed the preset maximum number of interventionable nodes. Mathematical expression is .

[0139] The second constraint is the technical feasibility requirement for selecting intervention nodes. Selected user nodes must meet basic technical intervention conditions, such as being online or having a valid contact method to receive intervention information. Nodes that do not meet the technical feasibility constraint will be excluded. Forced to be set to 0.

[0140] Step S402: Establish an intervention cost quantification function to calculate the resource consumption required to implement intervention measures for each user node. This step calculates the intervention cost required for each user node to be selected into the intervention set. The intervention cost includes two components: fixed cost and marginal cost. Fixed cost The fixed costs include the cost of developing the intervention plan, the cost of training intervention personnel, and the cost of preparing intervention tools. These costs are independent of the number of intervention points and are constant throughout the entire intervention task.

[0141] Marginal cost is related to the specific attributes of the intervention node. For the first... Each user node, marginal cost The difficulty of intervention by user nodes is determined by the overall influence weight of user nodes. The value of marginal cost is the basic marginal cost constant. Multiply .in The preset basic marginal cost constant, This refers to the comprehensive influence weight calculated in step S303. When... The higher the value, the higher the marginal cost, reflecting the fact that high-influence nodes need to invest more communication resources.

[0142] No. Total intervention cost when a user node is selected into the intervention set It is the sum of the fixed cost allocation and the marginal cost, i.e.:

[0143]

[0144] in This is the preset maximum number of nodes that can be intervened. This means that fixed costs are evenly distributed across each interveneable node. All selected nodes. The total cost of intervention is:

[0145]

[0146] To achieve dimensional consistency between the effect and cost terms in the objective function, the following measures are taken: Normalization is performed: ,in This is the preset maximum possible total cost constant for intervention.

[0147] Step S403: Construct an effectiveness quantification function to evaluate the inhibitory effect of intervention measures on risk transmission. This step, for a given set of intervention nodes, quantifies and evaluates the inhibitory effect of the intervention plan on risk transmission. The effectiveness quantification function is evaluated based on the changes in the state of the transmission network after intervention, including three evaluation indicators: source elimination rate, transmission path blocking rate, and transmission range contraction rate.

[0148] Source elimination rate The extent to which intervention measures cover the high-risk transmission nodes identified in step S304 is measured. The source elimination rate is calculated as: the number of high-risk transmission nodes selected for the intervention set divided by the total number of high-risk transmission nodes identified in step S304. High-risk transmission nodes are those ranked highest in transmission potential value. User nodes of bits, among which The value is the same as the maximum number of interventionable nodes in step S401.

[0149] Transmission path blocking rate The effectiveness of interventions in disrupting critical transmission pathways is measured. The transmission pathway blocking rate is calculated as follows: after removing simulated pathways using the selected intervention node as the starting or ending point, the percentage reduction in the number of remaining transmission pathways relative to the total number of original transmission pathways is calculated. The total number of original transmission pathways is statistically derived from all transmission pathway samples generated in step S305 under non-intervention conditions using Monte Carlo simulation.

[0150] Spread range contraction rate This measures the reduction in the actual spread after intervention relative to the expected spread without intervention. The spread reduction rate is calculated as: 1 minus the ratio of the spread coverage predicted by Monte Carlo simulation after intervention to the spread coverage predicted by Monte Carlo simulation without intervention.

[0151] Overall effect evaluation value The calculation is based on the weighted sum of the above three indicators:

[0152]

[0153] in , , The preset weighting coefficients satisfy... , The value is 0.3. The value is 0.4. The value is 0.3, and all weighting coefficients are dimensionless constants. The above three indicators... , , All values ​​are within the range of 0 to 1, therefore the overall effect evaluation value is... It also falls within the 0 to 1 range.

[0154] Step S404: An improved heuristic search algorithm is used to solve for the optimal set of intervention nodes under the condition of limited intervention resources. This step is based on the objective function defined in step S401, the cost quantification function established in step S402, and the effect quantification function established in step S403. An improved heuristic search algorithm is used to solve for the optimal set of intervention nodes.

[0155] The optimal criterion is: in all cases where... Among the candidate intervention sets with constraints, the objective function is to... The set of interventions that achieves the maximum value is the optimal set of intervention nodes. The preset cost penalty coefficient is 0.5.

[0156] The specific execution flow of the heuristic search algorithm is as follows. The algorithm starts iterating from an empty set. In each iteration, a node is selected from the currently unselected candidate nodes and added to the intervention set. The selection criterion is to maximize the objective function after the node is added. The increment is the largest. The algorithm continues to iterate until the size of the intervention set reaches its maximum. Or unable to find a way to make The search continues until candidate nodes are added. When the search process terminates, the current set of intervention nodes is the optimal set of intervention nodes output by the algorithm.

[0157] The improved heuristic search algorithm introduces a priority scheduling mechanism on top of the basic greedy search. This priority scheduling mechanism is based on the ratio of the comprehensive influence weight of each node to the intervention cost. The search priority is determined, with nodes having higher ratios being evaluated first during the search process. In each iteration, the algorithm only selects nodes with the highest priority. The candidate nodes of position are evaluated, among which The preset search window size has a value of [value]. Priority scheduling mechanisms reduce the computational cost of each iteration without significantly sacrificing solution quality.

[0158] Step S405: Output the optimal set of intervention nodes as input to the intervention implementation module. This step uses the target function output by the search algorithm in step S404. Maximizing the binary decision variable vector Convert to the corresponding user node set This is the optimal set of intervention nodes. The user nodes in the optimal set of intervention nodes are the intervention targets that, under limited intervention resources, can achieve the optimal balance between intervention effectiveness and intervention cost. After the optimal set of intervention nodes is output, it is pushed to the intervention implementation and effect evaluation module in step S5.

[0159] Step S4 transforms the risk propagation map output from Step S3 into an executable intervention plan. This process takes the set of high-risk propagation nodes and the distribution of propagation potential values ​​as input, and the optimal set of intervention nodes as the final output, completing the transformation from risk identification to intervention decision-making. The optimal set of intervention nodes output in Step S4 will be directly used for the implementation and dynamic optimization of intervention measures in the subsequent Step S5.

[0160] Finally, in step S5, the actual effect of the intervention measures is evaluated in real time and adaptively optimized through incremental effect comparison analysis. The effects of the implemented intervention measures are evaluated and the strategies are optimized to form an adaptive adjustment closed loop from intervention decision to effect feedback. This is carried out according to the following steps.

[0161] Step S501: Establish an incremental effect evaluation index system to quantify changes in risk indicators before and after the implementation of intervention measures. This step establishes an incremental effect evaluation index system to quantify changes in risk indicators before and after the implementation of intervention measures. The incremental effect evaluation index system includes two categories: direct effect indicators and indirect effect indicators.

[0162] Direct effect indicators include changes in the propagation status of intervention-covered nodes, intervention response rate, and intervention resistance rate. Changes in the propagation status of intervention-covered nodes refer to the proportion of user nodes selected for the intervention set whose propagation status changes from a propagation state to an immune state after the intervention is implemented.

[0163] Intervention response rate refers to the proportion of users who receive intervention information and actually take a response action, including reading the intervention information, clicking the intervention link, or performing the intervention guidance operation. Intervention resistance rate refers to the proportion of users who explicitly refuse or block intervention information.

[0164] Indirect effect indicators include changes in the overall network propagation speed, changes in the strength of inter-community propagation barriers, and the risk event decay period. Changes in the overall network propagation speed refer to the change in the average speed at which risk information spreads within the network before and after the intervention. Changes in the strength of inter-community propagation barriers refer to the change in the resistance to the spread of risk information between different communities before and after the intervention. The risk event decay period refers to the length of time from the triggering of a risk alarm event to the number of nodes in the propagation state falling below a preset threshold.

[0165] Each indicator in the indicator system is configured with a baseline value and a target value. The baseline value is taken from the indicator prediction value predicted by the Monte Carlo simulation in step S305 under the condition of no intervention. The target value is preset by the business personnel based on historical intervention experience and expected intervention effects, and is used to evaluate the degree to which the intervention effect is achieved.

[0166] Step S502: Design an A / B testing framework to compare the differences in the effectiveness of different intervention strategies. This step involves designing an A / B testing framework to compare the effectiveness of different intervention strategies. The A / B testing framework divides the intervention subjects into an experimental group and a control group. The experimental group implements the intervention strategy to be evaluated, that is, performs the intervention operation on the optimal intervention node set obtained in step S404. The control group maintains its original risk propagation state and does not receive any additional intervention.

[0167] The A / B testing framework uses statistical hypothesis testing to determine whether the differences between the two groups on key effect indicators are statistically significant. The statistical hypothesis testing method uses a two-sample t-test; when the p-value obtained from the test is lower than the preset significance level threshold of 0.05, the intervention strategy is considered effective.

[0168] The A / B testing framework supports parallel testing of multiple strategies. When multiple candidate intervention strategies exist, the A / B testing framework randomly divides the subjects into multiple groups, each implementing a different intervention strategy. By comparing the performance metrics of each group pairwise, the strategy with the best performance is selected for wider adoption. The criterion for determining the best performance is the overall performance evaluation value. Highest.

[0169] Step S503 involves adjusting the intervention strategy parameters based on the effect evaluation results using the policy gradient optimization method. This step utilizes the policy gradient optimization method to adjust the intervention strategy parameters according to the effect evaluation results output by the A / B testing framework in step S502. The intervention strategy parameters include the cost penalty coefficient of the objective function in step S401. and the weight coefficients of each indicator in the effect quantification function in step S403. , , .

[0170] Policy gradient optimization methods parameterize the intervention policy as a weight vector of the policy network. The strategy gradient optimization method calculates the comprehensive performance evaluation value. Relative to policy parameters The gradient direction is used to update the policy network weights, allowing the policy parameters to gradually converge to a more effective configuration during iterative optimization. After each round of intervention, the effect data of that round of intervention is collected, and the comprehensive effect evaluation value is calculated. For strategy parameters gradient Then, the policy parameters are updated along the gradient direction. The update formula is:

[0171]

[0172] in For the updated policy parameter vector, The policy parameter vector before the update. The preset learning rate constant has a value of 0.01. This is the gradient vector of the overall performance evaluation value with respect to the policy parameter vector. Each component corresponds to a policy parameter, and each component is calculated by applying a small perturbation to each individual parameter and then obtaining it through numerical difference. There is no cross-coupling relationship between the dimensions. The update formula performs scalar operations independently on each of the four policy parameter dimensions.

[0173] The policy gradient optimization method supports an online learning mode, enabling continuous feedback on the effects and real-time adjustment of policy parameters during the intervention process, without waiting for a complete offline experimental cycle.

[0174] Step S504: Construct a feedback adjustment mechanism to achieve adaptive dynamic optimization of the intervention strategy. This step involves constructing a feedback adjustment mechanism to achieve adaptive dynamic optimization of the intervention strategy. The feedback adjustment mechanism includes four stages: effect monitoring, anomaly detection, strategy adjustment, and strategy distribution.

[0175] The effect monitoring process continuously collects data on the direct and indirect effect indicators defined in step S501, with the collection frequency consistent with the data processing cycle of the stream computing engine.

[0176] The anomaly detection step statistically examines the indicator data collected in the performance monitoring step to identify instances where the indicators deviate from the expected range. The expected range is determined jointly by the baseline value and the target value set in step S501. When the actual observed value of an indicator is lower than 80% of the target value, the indicator is deemed to have deviated abnormally.

[0177] When an anomaly is detected, the strategy adjustment process selects the appropriate adjustment action based on the type and severity of the anomaly. The adjustment actions include three types: parameter fine-tuning, strategy switching, and intervention restart. Parameter fine-tuning is suitable for situations where the indicator deviation is minor, i.e., adjusting the strategy parameters in step S503. .

[0178] Strategy switching is applicable when a particular strategy continues to underperform, i.e., switching from the current strategy to a validated alternative strategy within the A / B testing framework. Intervention restart is applicable when a fundamental change in the risk profile renders the current intervention plan completely ineffective, i.e., re-executing steps S401 to S404 to generate a new intervention plan.

[0179] The strategy distribution phase pushes the adjusted strategy to the intervention implementation module for execution. The feedback adjustment mechanism forms an adaptive adjustment closed loop, achieving dynamic matching between the intervention strategy and the risk situation.

[0180] In summary, this method completes the entire chain of processing, from quantifying effect indicators, A / B testing and verification, gradient optimization of strategy parameters to constructing a feedback adjustment closed loop. The optimal set of intervention nodes output in step S4 is transformed into a dynamic intervention strategy with adaptive optimization capabilities in step S5, which, together with the risk perception in step S2, the causal tracing in step S3, and the precise intervention in step S4, constitutes a complete closed loop.

[0181] To facilitate understanding, the implementation process of the method of the present invention will be described in detail below with reference to specific application scenarios.

[0182] In promotional activities on internet marketing platforms, when the platform launches limited-time discount activities, it collects user behavior flow data in real time to identify user browsing, clicking, adding to cart and other behavioral patterns; it collects transaction flow data in real time to monitor transaction indicators such as order creation volume and payment conversion rate; and it collects public opinion flow data in real time to analyze the discussion heat and sentiment trend of users on social media.

[0183] The stream computing engine performs real-time correlation analysis on the three types of data streams mentioned above, extracting risk characteristic indicators related to promotional activities. The dynamic Bayesian network causal inference model, based on real-time incoming characteristic indicator data, identifies the causal relationships between promotional activities and abnormal user behavior, transaction fluctuations, and changes in public opinion, outputting a causal graph structure of risk factors.

[0184] The propagation potential assessment model comprehensively considers the influence weight, social relationship strength, and historical propagation patterns of key users participating in promotional activities, calculates the potential value distribution of each user in the propagation of promotional risks, and identifies a set of high-risk propagation nodes.

[0185] The intervention optimization module is based on the intervention cost-effectiveness quantification model. It calculates the optimal set of intervention nodes that can achieve better inhibition under limited intervention resources and pushes customized intervention information to these nodes.

[0186] The effectiveness evaluation module continuously monitors changes in risk indicators after intervention implementation, compares the differences in effectiveness of different intervention strategies through A / B testing, and dynamically optimizes intervention strategy parameters based on feedback data.

[0187] The entire process forms a complete closed loop of risk perception, causal tracing, precise intervention, and effect evaluation, achieving dynamic control of internet marketing risks across the entire chain.

[0188] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0189] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A big data-based internet marketing method, characterized in that, include: Deploy a distributed stream computing engine based on Apache Flink, configure an event time processing mechanism and a periodic watermark generation strategy, and access user behavior streams, transaction streams and public opinion streams in real time for cleaning and format standardization; Based on conditional independence test and scoring search algorithm, a Bayesian network graph structure is dynamically constructed from standardized data stream, and the node probability distribution is updated with incremental variational Bayesian method. The historical data stream is divided into continuous time periods according to a fixed time length. A Bayesian network substructure for each time period is constructed. The confidence weights and direction changes of the current time period are compared with those of the historical baseline structure. When the mutation condition is met, a causal evolution alarm is triggered and the baseline is updated. Perform network topology analysis on user social networks to extract degree centrality, betweenness centrality, clustering coefficient and community structure, calculate the user's comprehensive influence weight, determine the risk propagation potential value by combining risk content propagation characteristic parameters, and predict the propagation trend through discrete-time SIR model and Monte Carlo simulation. A combinatorial optimization problem is constructed based on the total intervention cost and the total intervention effect, and a heuristic search with priority scheduling is used to solve for the optimal set of intervention nodes. After the intervention is implemented, the intervention strategy parameters are adaptively adjusted through incremental effect evaluation and strategy gradient optimization.

2. The internet marketing method based on big data according to claim 1, characterized in that, The deployment based on the Apache Flink distributed stream computing engine also includes: The Apache Flink-based distributed stream computing engine is further configured with a stream-batch integrated computing architecture. Relying on a unified operator library, it abstracts the processing tasks of bounded and unbounded data into data stream operations. At the same time, it supports mixed processing logic of bounded and unbounded data. While processing real-time data, it switches to batch processing mode to perform historical data backtracking analysis and compares the currently monitored risk indicators with the historical level of the same period. The Apache Flink-based distributed stream computing engine establishes a state backend management mechanism, which uses a distributed file system or embedded key-value storage to persistently save intermediate states during the computing process. The intermediate states include the aggregated cumulative values ​​of each time window, user session interaction sequences, and probability distribution parameters generated by subsequent causal inference steps. With the help of checkpoint and savepoint strategies, the computing task can be restored from the most recent consistent snapshot when a node fails.

3. The internet marketing method based on big data according to claim 1, characterized in that, The real-time access to user behavior streams, transaction streams, and public opinion streams, followed by cleaning and format standardization, specifically includes: Through different topics of the unified message middleware, the stream computing engine subscribes to user behavior streams, transaction streams and public opinion streams in the form of consumer groups. For user behavior streams, it filters out expired records with timestamps earlier than the current system time exceeding the threshold and invalid records with empty user identifiers. For transaction streams, it filters out test order records with amounts of zero or negative numbers and converts the transaction amount into a base currency unit. For public opinion streams, it filters out meaningless records with too short text and performs a unified time zone conversion on the publication timestamp. In the periodic watermark generation strategy, the water level value is the maximum value among all observed event timestamps minus the preset allowed delay time. When the water level value exceeds the end timestamp of a certain time window, the window calculation is triggered and the aggregation result is output. Late data whose actual timestamp is less than the current water level is judged as extremely late data and discarded directly.

4. The internet marketing method based on big data according to claim 1, characterized in that, The algorithm based on conditional independence testing and scoring search dynamically constructs a Bayesian network graph structure from a standardized data stream, specifically including: Based on the knowledge of experts in the field of internet marketing, a set of risk factor nodes for the Bayesian network is set in advance. The set of risk factor nodes includes marketing activity parameter nodes, user profile feature nodes, market environment variable nodes, and competitor behavior nodes. No directed edges are set between the nodes for the time being. Within each data processing cycle, for any two nodes and a given set of conditional nodes, calculate the conditional mutual information of the two nodes under the given set of conditional nodes. When the conditional mutual information is lower than a preset independence threshold, determine that the two nodes are independent under the given set of conditional nodes. For node pairs that are confirmed to have dependencies, a scoring search method based on the Bayesian information criterion is adopted to calculate the network score when adding a directed edge in the first direction and the network score when adding a directed edge in the second direction, respectively. The direction with the better network score is selected as the final direction of the directed edge, and directed edges are added step by step to form a directed acyclic graph structure. Newly added directed edges are synchronously updated in the graph model. As data continues to accumulate, previously identified directed edges are either overturned by new evidence and removed or new, previously undiscovered edges are added.

5. The internet marketing method based on big data according to claim 1, characterized in that, The method of updating the node probability distribution using incremental variational Bayesian method specifically includes: The variational distribution approximates the true posterior distribution. Each time new data arrives, only the variational parameters of the node to which the data belongs, its parent node, and its child nodes are updated. The update is completed when the KL divergence value is lower than the preset convergence threshold by minimizing the variational distribution and the true posterior. After a single node probability update is completed, the node probability change is propagated sequentially to all child nodes along the directed edge direction, and simultaneously propagated in the reverse direction to all parent nodes. Each node that receives the message recalculates its own probability distribution based on the conditional probability table and continues to propagate the change outward until a complete round of message propagation ends, and the probability value change of all nodes in the two adjacent iterations is lower than the preset convergence threshold.

6. The internet marketing method based on big data according to claim 1, characterized in that, The process involves dividing the historical data stream into continuous time periods according to a fixed time length, constructing a Bayesian network substructure for each time period, comparing the confidence weights and direction changes of the current time period with the historical baseline structure, triggering a causal evolution alarm and updating the baseline when a mutation condition is met, specifically including: The historical data stream is divided into continuous time periods with a fixed time length, and the structure learning algorithm is run independently for each time period to construct the Bayesian network substructure of that time period. The historical baseline structure is given by the statistical aggregation results of substructures from several past time periods; For each pair of directed edges, calculate the confidence weight of the directed edge in the current time period substructure and the confidence weight in the historical benchmark structure, and calculate the weight change as the absolute value of the difference between the two. At the same time, check whether the direction of the directed edge is consistent in the current time period and the historical benchmark. When the weight change of at least one directed edge exceeds the preset weight change tolerance threshold, or when the direction of at least one directed edge reverses between two consecutive periods, a sudden change in causal relationship is determined, a causal evolution alarm is triggered, and the substructure of the current period is marked as the new baseline. Subsequent inference tasks are all executed based on the new graph structure. When each data event arrives, incremental updates and inference calculations are performed on the cause-effect graph structure. The incremental updates and inference calculations include four stages: data preprocessing, node probability updates, edge weight adjustments, and inference queries.

7. The internet marketing method based on big data according to claim 1, characterized in that, The calculation of the user's overall influence weight, combined with the risk content dissemination characteristic parameters to determine the risk dissemination potential value, specifically includes: Static attribute features and dynamic behavioral features of each user node are extracted. Static attribute features include historical credit score, account level, and authentication status. Dynamic behavioral features include activity frequency, content posting volume, and interaction response rate within the recent time window. All indicators are pre-normalized to the range of 0 to 1. For the There are n user nodes, and let the static attribute feature vector be . The dynamic behavior feature vector is , No. individual users at any time Overall influence weight ,in and For preset weighting coefficients and ; The sensitivity level of risky content, the adaptability of the dissemination medium, and the reproducibility of the dissemination content are quantified into values ​​between 0 and 1. The arithmetic mean of the three is taken as the comprehensive adjustment coefficient. The comprehensive influence weight of each user node is multiplied by the comprehensive adjustment coefficient to obtain the dissemination potential value of each user node. User nodes are sorted according to their propagation potential value, with the top-ranked nodes receiving the highest propagation potential value. User nodes at a certain number of positions are identified as a set of high-risk propagation nodes.

8. The internet marketing method based on big data according to claim 7, characterized in that, The prediction of propagation trends using discrete-time SIR models and Monte Carlo simulations specifically includes: Using a discrete-time SIR model framework, each user node is divided into a susceptible state, a propagating state, or an immune state at any given time. Within each discrete time step, a susceptible node transforms into a propagating node with a transition probability determined by the propagation potential value, while a propagating node transforms into an immune node with a fixed recovery probability. Based on the Monte Carlo simulation method, starting from the set of high-risk propagation nodes as the initial propagation node set, the state update is iteratively performed until the number of propagation state nodes drops to 0 or the preset maximum number of simulation steps is reached. After a single simulation, all user nodes that are ultimately infected in the user's social network are recorded. After repeating the simulation multiple times, the probability distribution of each user node being infected is statistically analyzed, and the confidence interval estimate of the propagation coverage is output.

9. The internet marketing method based on big data according to claim 1, characterized in that, The aforementioned method constructs a combinatorial optimization problem based on the total intervention cost and the total intervention effect, and employs a heuristic search with priority scheduling to solve for the optimal set of intervention nodes. Specifically, this includes: Define decision variables Indicates whether to intervene in user nodes 1 indicates that the intervention was selected, and 0 indicates that it was not selected. The constraints include that the total number of user nodes selected into the intervention set shall not exceed the preset maximum number of interventionable nodes, and that the selected user nodes must be online or have a valid contact method to receive intervention information. The objective function is ,in For the decision variable vector, To determine the overall effect of the intervention, To normalize the total cost of intervention, This is the cost penalty coefficient; Overall Intervention Effect Elimination rate of transmission source Transmission path blocking rate and the rate of contraction of the spread We get the weighted sum. ， in , , For preset weighting coefficients and ; Source elimination rate The transmission path blocking rate is the ratio of the number of high-risk transmission nodes selected for the intervention set to the total number of high-risk transmission nodes. The propagation range contraction rate is the percentage reduction in the number of remaining propagation paths relative to the original total number of propagation paths after simulated path removal with the selected intervention node as the starting or ending point. =1 minus the ratio of the transmission coverage predicted by the Monte Carlo simulation after intervention to the transmission coverage predicted without intervention; Normalized total intervention cost It is obtained by dividing the sum of the fixed cost allocation and the marginal cost of each intervention node by the preset maximum possible total intervention cost, i.e. ,in The set of user nodes selected for the intervention set. , , For fixed costs, The maximum number of nodes that can be intervened. To pre-determine the basic marginal cost constant, To pre-determine the maximum possible total cost of intervention, For the user node The overall influence weight; A greedy search algorithm is used, starting from an empty set and iterating. In each iteration, the search priority is determined from the candidate nodes that have not yet been selected, based on the ratio of the comprehensive influence weight of each node to its marginal cost. Nodes with higher ratios are evaluated first in the search process, and only those with the highest priority are considered. For each candidate node, the objective function increment is calculated, and the node that maximizes the objective function increment is added to the intervention set. The iteration continues until the size of the intervention set reaches the maximum number of interventionable nodes or no candidate node that can increase the objective function value can be found.

10. The internet marketing method based on big data according to claim 1, characterized in that, The adaptive adjustment of intervention strategy parameters through incremental effect evaluation and strategy gradient optimization specifically includes: An incremental effect evaluation index system was established. Direct effect indicators include the proportion of the number of nodes covered by the intervention that changed from the propagation state to the immune state, the intervention response rate, and the intervention resistance rate. Indirect effect indicators include changes in the overall network propagation speed, changes in the strength of the inter-community propagation barriers, and the risk event decay cycle. Each indicator is configured with a baseline value and a target value. The baseline value is taken from the indicator prediction value predicted by Monte Carlo simulation under the condition of no intervention. Using an A / B testing framework, the intervention subjects were divided into an experimental group and a control group. The experimental group performed the intervention operation at the optimal intervention node set, while the control group maintained the original propagation state without additional intervention. A two-sample t-test was used to determine whether the difference between the two groups in key effect indicators was statistically significant. When the p-value was lower than the preset significance level threshold, the intervention strategy was deemed effective. A feedback adjustment mechanism is constructed, comprising four stages: effect monitoring, anomaly detection, strategy adjustment, and strategy distribution. The effect monitoring stage continuously collects data on various effect indicators. The anomaly detection stage determines an anomaly when the actual observed value of a certain indicator is lower than 80% of the target value. The strategy adjustment stage selects parameter fine-tuning, strategy switching, or intervention restart based on the anomaly type. Parameter fine-tuning adopts the strategy gradient optimization method to update the intervention strategy parameters along the gradient direction of the comprehensive effect evaluation value. The intervention strategy parameters include the cost penalty coefficient and the weight coefficients of the source elimination rate, the transmission path blocking rate, and the transmission range contraction rate. The strategy distribution stage pushes the adjusted strategy to the intervention implementation module for execution.

Images