Marketing risk rule problem three library atlas intelligent recommendation management method

Abstract

This invention discloses an intelligent recommendation management method for a three-database graph of marketing risk rules, relating to the field of power marketing auditing and intelligent risk control technology. The method includes the following steps: constructing a multi-dimensional business behavior risk entropy baseline database; performing reverse knowledge extraction based on alarm-free abnormal behavior; performing self-verifying knowledge verification based on business closed loops; and updating the three-database graph based on a knowledge lifecycle evolution model. In this invention, the multi-dimensional business behavior risk entropy baseline database is used to independently detect alarm-free abnormal behavior. Combined with reverse knowledge extraction, knowledge gaps in the three databases are located and potential knowledge is generated. This knowledge is then transformed into formal knowledge through self-verifying verification within the business closed loop. A knowledge lifecycle evolution model is introduced to dynamically update the three-database graph. This solves the self-reinforcing paradox of knowledge blind spots in the three marketing databases, as well as the problems of information fragmentation across databases, low efficiency of manual updates, and inability to detect hidden risks in existing technologies. It achieves early risk warning, automated knowledge management throughout the entire process, and precise, scenario-based intelligent recommendation.

Description

Technical Field

[0001] This invention relates to the field of power marketing auditing and intelligent risk management technology, specifically to an intelligent recommendation management method for marketing risk rule-based questions using a three-database graph. Background Technology

[0002] With the deepening of the construction of new power systems and the reform of the power system, the business scenarios of electricity marketing are becoming increasingly complex. Violations such as incorrect electricity pricing, illegal electricity transfer, and electricity theft occur frequently, making marketing risk management increasingly difficult. China Southern Power Grid Company has built a three-tiered marketing defense system and a four-database marketing system. Among them, the risk database, rule database, and problem database are the core knowledge assets supporting marketing auditing and risk management.

[0003] Currently, those skilled in the art generally focus their research on improving the storage capacity, query speed, and basic correlation capabilities of the three databases. They believe that the main shortcomings of existing systems are fragmented information across databases, low accuracy of personalized recommendations, and poor efficiency of manual updates. The commonly adopted solution in the industry is to introduce knowledge graph technology to enhance the correlation between the three databases, optimize recommendation algorithms by combining user behavior data, and establish a manual review and feedback mechanism to improve content quality.

[0004] However, the marketing databases, as the sole source of knowledge for the three lines of defense, suffer from knowledge gaps and errors that directly prevent these lines from detecting new risks arising from these gaps and errors. Since the three lines of defense rely on the knowledge from these databases for anomaly identification, these undetected new risks do not generate any alerts or work orders, and therefore are never fed back to the databases for supplementation and correction. This creates a vicious cycle that cannot be broken by current technology: the more incomplete the knowledge in the databases, the more risks the three lines of defense miss; the more risks missed, the larger the knowledge gaps in the databases become, ultimately forming an ever-expanding black hole of hidden risks.

[0005] Current technologies rely on the premise that three lines of defense can detect all risks and report them to three databases. Therefore, no matter how much the content quality and correlation capabilities of these three databases are optimized, hidden risks cannot be fundamentally eliminated. Furthermore, with the growth of business volume and the emergence of new business scenarios, the scale of hidden risks will expand exponentially, causing incalculable economic losses and service risks to power companies. In view of this, a three-database intelligent recommendation management method for marketing risk rules is proposed to overcome the above problems. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent recommendation management method for marketing risk rules based on a three-database graph, in order to solve the problems mentioned in the background section.

[0007] To address the aforementioned technical problems, the present invention provides a marketing risk rule-based intelligent recommendation management method based on a three-database graph, comprising the following steps: S1. Construct a multi-dimensional business behavior risk entropy baseline library to detect abnormal behaviors without alarms, independent of the three lines of defense in marketing. S2. Based on alarm-free abnormal behavior, reverse knowledge extraction is performed to locate knowledge gaps in the three databases and generate potential knowledge; S3. Based on the business closed loop, conduct self-verifying knowledge verification and transform the verified potential knowledge into formal knowledge in the three databases. S4. Update the three-database graph based on the knowledge lifecycle evolution model for contextualized intelligent recommendation.

[0008] Furthermore, the process of constructing a multi-dimensional business behavior risk entropy baseline library in step S1 is as follows: Collect native business data from the entire marketing business process that is unrelated to the three lines of defense marketing alerts; For each type of business data, a three-dimensional risk entropy baseline model of time-space-business type is constructed. First, the single-dimensional risk entropy is calculated using the single-dimensional business behavior risk entropy calculation formula, and then the comprehensive business behavior risk entropy of each spatiotemporal unit is obtained using the comprehensive risk entropy calculation formula. Using a sliding window statistical method, the mean and standard deviation of the comprehensive risk entropy of each spatiotemporal unit within a set period are calculated to establish a dynamic risk entropy baseline. Behaviors in which the comprehensive risk entropy deviates from the baseline by more than 3 times the standard deviation are marked as abnormal behaviors without alarms.

[0009] Furthermore, step S1 also includes a baseline dynamic self-calibration mechanism: A global consistency verification algorithm is introduced to automatically trigger baseline updates by comparing the risk entropy change trends of similar data in different regions and at different times. The calibrated risk entropy baseline is calculated based on the calibration formula. When the risk entropy change of a certain type of business in the global scope exceeds the first threshold, a full baseline update is triggered. When the risk entropy change of a certain type of business in a single region exceeds the second threshold, a baseline update for the corresponding business type in that region is triggered.

[0010] Furthermore, the specific process of reverse knowledge extraction based on alarm-free abnormal behavior in step S2 is as follows: All abnormal behaviors without alarms are clustered, the anomaly intensity and impact range of each cluster are calculated, and the clusters are sorted from high to low according to their anomaly intensity and the clusters that meet the set conditions are processed first. For each anomalous cluster, retrieve all relevant risks, rules, and issues from the three existing databases, calculate the similarity between the anomalous cluster feature vector and the knowledge feature vectors of the three databases, and mark the anomalous cluster with the highest similarity below a set threshold as a knowledge gap; A counterfactual reasoning algorithm is used to generate all possible risk hypotheses that can explain the anomalous cluster, and the reasonableness of the risk hypotheses is verified.

[0011] Furthermore, the specific process of generating latent knowledge in step S2 is as follows: For each validated risk hypothesis, a structured knowledge entry containing risk point description, triggering conditions, detection rules, and typical problem case fields is automatically generated using the structure templates of similar knowledge in the three existing databases. The generated potential knowledge is stored in a temporary knowledge base, which is only accessible to personnel participating in the grayscale verification.

[0012] Furthermore, the specific process of self-verifying knowledge verification based on the business closed loop in step S3 is as follows: The generated potential knowledge is distributed to grassroots inspectors in the form of recommendation rules, and the business processing data of the experimental group and the control group are recorded separately. Scores are calculated from three dimensions: business processing effectiveness, user feedback, and anomaly elimination effectiveness. The comprehensive verification score of potential knowledge is obtained through the comprehensive verification score calculation formula.

[0013] Furthermore, the specific process of automatic knowledge classification and storage in step S3 is as follows: Potential knowledge whose comprehensive verification score exceeds the first threshold is upgraded to formal knowledge, added to the corresponding risk base, rule base, or question base, and associated with it. Potential knowledge with a comprehensive verification score between the second and first thresholds is labeled as secondary knowledge and used for general recommendation. Potential knowledge with a comprehensive verification score below the second threshold is marked as knowledge to be optimized and pushed to the provincial marketing audit expert database for manual review.

[0014] Furthermore, the specific process of step S4 is as follows: Based on the application frequency, problem-solving success rate, and user evaluation data of each piece of knowledge, the weight of knowledge in the three-database graph is dynamically adjusted through a weight update formula, and knowledge whose weight is lower than a set threshold for consecutive periods is marked as invalid knowledge. The graph convolutional neural network algorithm is used to analyze the co-occurrence patterns of knowledge, automatically uncover the implicit relationships between knowledge and update the three-database graph; The system identifies the feature vector of the user's current business scenario, retrieves knowledge from the three databases that has a similarity greater than a set threshold to the feature vector, generates a recommendation list by weighting and sorting the knowledge based on its weight and association strength, and pushes the recommendation results to the user's business processing interface in real time.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. By detecting risk entropy anomalies independently of the three lines of defense in marketing, it proactively identifies knowledge gaps in the three databases, no longer relying on feedback from the three lines of defense to update the databases. This completely solves the vicious cycle of the more missed detections, the more knowledge is lacking, and the more knowledge is lacking, the more missed detections occur. It can cover hidden risks that existing technologies cannot detect.

[0016] 2. To achieve the creative utilization of defect resources, existing technologies are regarded as the three lines of defense against system defects. The missed detection is transformed into a precise indicator of knowledge gaps in the three databases, making the missed detection behavior the most valuable source of knowledge, significantly shortening the update cycle of the three databases and improving the efficiency of knowledge update.

[0017] 3. Possesses the capability to detect progressive hidden risks. Through a business behavior risk entropy quantification model, it accurately identifies slow changes and progressive hidden risks that cannot be detected by traditional threshold methods. For risks such as upgraded electricity theft methods and aging metering equipment, the detection time is earlier than existing technologies.

[0018] 4. Enables early warning of new risks. The system can detect the existence of new risks in advance by detecting abnormal changes in the risk entropy of business behavior before they are widely used or detected by any existing rules. It has significant risk prediction advantages in new business scenarios such as distributed photovoltaic grid connection and virtual power plant trading.

[0019] 5. Achieve fully automated knowledge management, completing the entire process from knowledge discovery and generation to verification and storage, significantly reducing the workload of manual review, while ensuring the accuracy and timeliness of knowledge. Attached Figure Description

[0020] Figure 1 This is a schematic diagram illustrating the principle of the intelligent recommendation management method for marketing risk rules based on a three-database graph, as described in this invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figure 1 The present invention provides a technical solution: See Figure 1 As shown, an embodiment of the intelligent recommendation management method for the three-database graph of marketing risk rules is presented: The core of this approach lies in reverse engineering the self-reinforcing paradox of knowledge blind spots as a data source for problem discovery. By constructing a novel, alarm-free anomaly detection and reverse knowledge generation self-verifying closed-loop verification method, it fundamentally solves the vicious cycle problem that existing technologies cannot address, while also generating a novel risk early warning effect. This method is built upon the data analysis domain of the Metrology Automation System 3.0, is fully compatible with the existing system architecture, and requires no additional hardware. The specific steps are as follows: S1. Construct a multi-dimensional business behavior risk entropy baseline library: Instead of relying on the inherent logic of defining anomalies using existing technologies and three databases of knowledge, this approach constructs a pure business behavior risk entropy baseline that does not depend on any of these three databases. This establishes anomaly detection capabilities independently of the three lines of marketing defense, fundamentally eliminating reliance on these databases. This step pre-defines a business behavior risk entropy quantification model, using dynamic changes in entropy values ​​to accurately identify progressive, latent risks that traditional threshold methods cannot detect.

[0023] S1.1, Full-link business behavior data collection: This involves collecting native business data unrelated to the three lines of defense alarms in the entire marketing process. Data sources cover the marketing management system, metering automation system, customer service system, and electricity consumption information collection system. The collected data includes time intervals at each stage of the business expansion process, the temporal characteristics of user electricity consumption curves, communication frequency of metering equipment, spatial distribution of meter reading success rate, weekly fluctuations in transformer area line loss, semantic characteristics of user complaints, and cyclical patterns of electricity bill collection. All of this data is neutral data generated by existing systems but never used for risk detection. The time granularity is uniformly set at 15 minutes, and the spatial granularity at the transformer area level.

[0024] S1.2, Spatiotemporal Related Business Behavior Risk Entropy Modeling For each type of business data, a three-dimensional risk entropy baseline model for time and space business types is constructed. First, the formula for calculating the risk entropy of a single-dimensional business behavior is defined: ; in: Indicates time ,space Business Type Business behavior risk entropy; Indicates time ,space Business Type Next, the The probability of a certain business behavior state occurring is calculated by dividing the number of times the state occurs within the statistical sliding window by the total number of states. This indicates the total number of behavioral states under this business type; Based on the single-dimensional risk entropy, the formula for calculating the comprehensive risk entropy is further defined: ; in: Indicates time ,space The entropy of comprehensive business behavior risk; Indicates business type The risk weight is determined based on the frequency and impact of risks that have occurred in the past for this business type, and the value ranges from 0 to 1. The sum of the weights of all business types is 1. Indicates the total number of business types; A sliding window statistical method was used to calculate the mean and standard deviation of the comprehensive risk entropy for each spatiotemporal unit over the past 90 days, establishing a dynamic risk entropy baseline. The sliding window step size was 1 day, and the baseline data was updated daily. The confidence interval for the baseline was set at 99.7%, meaning that when the comprehensive risk entropy of a spatiotemporal unit deviated from the baseline by more than 3 times the standard deviation, it was marked as an abnormal behavior without alarm.

[0025] S1.3, Baseline Dynamic Self-calibration Mechanism: A global consistency verification algorithm is introduced. When a certain type of business data in a specific region shifts, the system automatically triggers a baseline update. The calibration process is completed by comparing the risk entropy change trends of similar data in different regions and at different times, requiring no manual intervention. The calibration formula is as follows: ; in: Indicates the type of service after calibration. In the region Risk entropy baseline; Indicates the risk entropy baseline before calibration; This represents the calibration coefficient, with a value ranging from 0 to 1, depending on the region. The similarity with global business behavior features is determined, and the similarity is calculated using cosine similarity. The higher the similarity, the larger the calibration coefficient. Indicates business type The average change in risk entropy over a global scale is calculated as the arithmetic mean of the changes in risk entropy for this business type across all regions. When the risk entropy change of a certain type of business in the global scope exceeds 0.1, a full baseline update is triggered; when the risk entropy change of a certain type of business in a single region exceeds 0.2, a baseline update of the corresponding business type in that region is triggered.

[0026] S2. Reverse knowledge extraction based on alarm-free abnormal behavior: By transforming the existing technology into a systemic flaw—a missed detection in the three lines of defense for marketing—into a precise indicator of knowledge gaps in the three knowledge bases, and by analyzing the contradictions between alarm-free abnormal behavior and existing knowledge in the three knowledge bases, the missing risks, rules, and problems are automatically extracted, thus achieving knowledge generation from scratch.

[0027] S2.1 Cluster analysis of abnormal behavior without alarms: All detected abnormal behaviors without alarms were clustered using the DBSCAN algorithm. Clustering dimensions included service type, occurrence time, geographical location, involved device type, and user characteristics. The neighborhood radius of the DBSCAN algorithm was set to 0.5, and the minimum number of neighborhood samples was set to 5. For each abnormal cluster formed after clustering, its anomaly strength and impact range were calculated. The formula for calculating anomaly strength is: ; in: Indicates anomalous clusters The abnormal intensity; Indicates anomalous clusters The number of abnormal behaviors without alarms included; Indicates the first The comprehensive risk entropy of an abnormal behavior; This represents the baseline risk entropy of the corresponding spatiotemporal unit; This represents the baseline standard deviation of the corresponding spatiotemporal unit; The formula for calculating the scope of influence is: ; in: Indicates anomalous clusters Scope of influence; Indicates anomalous clusters Number of users involved; Indicates anomalous clusters Total number of users in the region; The abnormal clusters are sorted from high to low according to their anomaly intensity, and those with anomaly intensity greater than 3 and an impact range greater than 5% are processed first.

[0028] S2.2 Knowledge Gap Location and Reverse Reasoning: For each anomaly cluster, the system automatically retrieves all relevant risks, rules, and issues from the three existing databases to determine if there is any knowledge that can explain the anomaly. The retrieval process uses a cosine similarity algorithm to calculate the similarity between the anomaly cluster feature vector and the feature vectors of the knowledge in the three databases. If the highest similarity is less than 0.3, it is marked as a knowledge gap.

[0029] The system employs a counterfactual reasoning algorithm to generate all possible risk hypotheses that can explain the anomaly. The steps of counterfactual reasoning are as follows: First, extract the core features of the anomaly cluster, including data change trends, spatiotemporal distribution patterns, and related business characteristics; then, based on the marketing business causal logic library, generate all risk causes that conform to the core features; finally, perform a rationality check on each risk cause, eliminating hypotheses that contradict existing business rules.

[0030] S2.3, Generation of potential knowledge structure: For each validated risk hypothesis, the system automatically generates a corresponding structured knowledge entry. The generation process utilizes structural templates for similar knowledge in the three existing databases to ensure compatibility between the newly generated knowledge and the existing knowledge system. Each structured knowledge entry contains four fields: a risk point description field including the risk phenomenon, scope of impact, and severity; a triggering condition field including time conditions, spatial conditions, and data thresholds; a detection rule field containing rule expressions executable by the metrology automation system; and a typical problem case field including the event time, location, handling process, and result.

[0031] The generated potential knowledge is stored in a temporary knowledge base, awaiting subsequent verification. The knowledge in the temporary knowledge base is only accessible to those participating in the gray-scale verification and is not displayed to ordinary users.

[0032] S3. Self-verifying knowledge verification based on business closed loop: Instead of requiring manual review, the correctness of knowledge is automatically verified by applying potential knowledge to business practices and observing its effects. This achieves a deep integration of knowledge verification and business processing, completely solving the problems of low efficiency and inconsistent quality of manual review.

[0033] S3.1, Gradual Release of Potential Knowledge: The generated potential knowledge will be released to grassroots inspectors in the form of recommendation rules through a phased rollout. The phased rollout uses a stratified random sampling method, selecting 10% of grassroots inspectors from each of the 21 cities in the province to form an experimental group, and the remaining personnel to form a control group. The phased rollout period is 14 days, during which the system records all relevant business processing data for both the experimental and control groups.

[0034] When auditors in the experimental group processed business related to this potential knowledge, the system automatically pushed the potential knowledge and indicated that it was knowledge to be verified. After processing, they were required to provide feedback on its validity. When auditors in the control group processed the same business, this potential knowledge was not pushed to them.

[0035] S3.2 Multi-dimensional effect verification: The system automatically collects data from three dimensions to verify the effectiveness of the potential knowledge, with each dimension corresponding to a quantitative score.

[0036] The formula for calculating the business processing performance score is as follows: ; in: This indicates the score for the effectiveness of business processing; This indicates the problem detection rate of the experimental group's inspectors; This indicates the problem detection rate of the control group's inspectors; The formula for calculating user feedback scores is: ; in: This indicates the user feedback score; This represents the average user satisfaction with the processing of work orders in the experimental group. This represents the average user satisfaction with the processing of work orders in the control group. The formula for calculating the anomaly elimination effect score is: ; in: The score represents the effectiveness of anomaly elimination. This indicates the incidence rate of similar abnormal behaviors without alarms before verification; This indicates the incidence of similar abnormal behaviors without alarms after verification; The formula for calculating the comprehensive verification score is as follows: ; S3.3 Automatic hierarchical knowledge entry into the database: When the comprehensive verification score of a piece of potential knowledge exceeds 20%, the system automatically upgrades it to formal knowledge, adds it to the corresponding risk base, rule base, or question base, and establishes a relationship with other related knowledge. The knowledge classification standard is as follows: a comprehensive verification score of 50% or higher is classified as Level 1 knowledge, which is given priority in recommendation; a comprehensive verification score of 20% or higher but less than 50% is classified as Level 2 knowledge, which is given ordinary recommendation; a comprehensive verification score of less than 20% is classified as knowledge that needs optimization and is pushed to the provincial marketing audit expert database for manual review.

[0037] Once knowledge is added to the knowledge base, the knowledge graph is automatically updated, and the corresponding entries in the temporary knowledge base are deleted. For knowledge that passes manual review and is subject to optimization, its comprehensive verification score is adjusted to above 20%, and it is upgraded to level 2 knowledge; for knowledge that fails review, it is deleted from the temporary knowledge base.

[0038] S4. Dynamic Evolution and Intelligent Recommendation of Three-Database Maps: Based on existing knowledge graph technology, a knowledge lifecycle evolution model is introduced, enabling the three-database graph to automatically adjust the weights and relationships of knowledge based on feedback from business practices, thereby achieving the self-evolution of the knowledge system.

[0039] S4.1 Dynamic Adjustment of Knowledge Weights: The weight of each piece of knowledge in the knowledge graph is dynamically adjusted based on data such as its application frequency in actual business, problem-solving success rate, and user feedback. The weight update formula is: ; in: Representing knowledge In time The weights; Representing knowledge In time The weights; This represents the learning rate, with a value of 0.1. Representing knowledge The standardized application frequency over the past 30 days is calculated by dividing the number of times the knowledge was used by the total number of times all knowledge was used. Representing knowledge The problem-solving success rate over the past 30 days is calculated by dividing the number of times the knowledge was used to solve a problem by the total number of times it was used. Representing knowledge Standardized user ratings over the past 30 days, calculated by dividing the average rating of the knowledge by 5; The weight update cycle is 1 day, and it is automatically executed every morning at midnight. For knowledge with a weight below 0.01 for 90 consecutive days, the system automatically marks it as invalid knowledge, pushes it to the administrator for review, and returns it after the review is approved.

[0040] S4.2 Automatic Relationship Mining: This algorithm utilizes graph convolutional neural networks to analyze the co-occurrence patterns of knowledge in real-world business scenarios and automatically uncover implicit relationships between knowledge points. The algorithm takes the adjacency matrix of the knowledge graph and the knowledge feature vectors as input and outputs the knowledge embedding vectors. It calculates the cosine similarity between any two knowledge embedding vectors; when the similarity is greater than 0.7, a relationship is established between them, with the relationship strength equal to the similarity value.

[0041] The association update cycle is 7 days. Each time it is updated, the similarity of all knowledge pairs is recalculated, old associations with similarity below 0.6 are deleted, and new associations with similarity above 0.7 are added.

[0042] S4.3, Contextualized Intelligent Recommendation: Based on the user's current business scenario and historical behavior data, combined with the weights and relationships of the knowledge graph, accurate knowledge recommendations are provided to the user. The recommendation process is as follows: First, the feature vector of the user's current business scenario is identified, which includes four dimensions: work order type, business process, user type, and geographical location; then, knowledge with a similarity greater than 0.5 to the feature vector is retrieved from the knowledge graph; the knowledge is weighted and sorted according to its weight and association strength to generate a Top 10 recommendation list; at the same time, other user processing records related to this knowledge are recommended, including processing time, result, and remarks.

[0043] Recommendation results are pushed to the user's business processing interface in real time. Users can click on the recommended knowledge to view detailed content and directly call the detection rules in the knowledge to check for anomalies.

[0044] Summarize: By detecting risk entropy anomalies independently of the three lines of defense in marketing, gaps in the knowledge base are proactively identified. This eliminates the need to rely on feedback from the three lines of defense to update the knowledge base, thus completely resolving the vicious cycle of more missed detections leading to more knowledge gaps, and more knowledge gaps leading to more missed detections.

[0045] By transforming the failures of existing technologies in the three lines of defense for marketing—which are considered the biggest weakness—into the most valuable source of knowledge, defect resource utilization has been achieved, resulting in technological effects that existing technologies could not achieve.

[0046] The business behavior risk entropy model can detect slow changes and gradual deviations from the normal range that traditional threshold methods cannot detect. For gradual risks such as upgraded electricity theft methods and aging metering equipment, this method can detect them earlier than existing technologies.

[0047] The system can detect new risks in advance through risk entropy anomalies, before they are exploited on a large scale or detected by any existing rules, thus achieving proactive risk prevention. This method can detect risks early in emerging business scenarios such as distributed photovoltaic grid connection and virtual power plant trading.

[0048] It has automated the entire process of knowledge discovery, generation, verification, and storage, significantly reducing labor costs while ensuring the accuracy and timeliness of knowledge.

Claims

1. A three-database graph-based intelligent recommendation management method for marketing risk rules, characterized in that... Includes the following steps: S1. Construct a multi-dimensional business behavior risk entropy baseline library to detect abnormal behaviors without alarms, independent of the three lines of defense in marketing. S2. Based on alarm-free abnormal behavior, reverse knowledge extraction is performed to locate knowledge gaps in the three databases and generate potential knowledge; S3. Based on the business closed loop, conduct self-verifying knowledge verification and transform the verified potential knowledge into formal knowledge in the three databases. S4. Update the three-database graph based on the knowledge lifecycle evolution model for contextualized intelligent recommendation.

2. The marketing risk rule problem three-database graph intelligent recommendation management method as described in claim 1, characterized in that: The process of constructing the multi-dimensional business behavior risk entropy baseline library in step S1 is as follows: Collect native business data from the entire marketing business process that is unrelated to the three lines of defense marketing alerts; For each type of business data, a three-dimensional risk entropy baseline model of time-space-business type is constructed. First, the single-dimensional risk entropy is calculated using the single-dimensional business behavior risk entropy calculation formula, and then the comprehensive business behavior risk entropy of each spatiotemporal unit is obtained using the comprehensive risk entropy calculation formula. Using a sliding window statistical method, the mean and standard deviation of the comprehensive risk entropy of each spatiotemporal unit within a set period are calculated to establish a dynamic risk entropy baseline. Behaviors in which the comprehensive risk entropy deviates from the baseline by more than 3 times the standard deviation are marked as abnormal behaviors without alarms.

3. The marketing risk rule problem three-database graph intelligent recommendation management method as described in claim 2, characterized in that: Step S1 also includes a baseline dynamic self-calibration mechanism: A global consistency verification algorithm is introduced to automatically trigger baseline updates by comparing the risk entropy change trends of similar data in different regions and at different times. The calibrated risk entropy baseline is calculated based on the calibration formula. When the risk entropy change of a certain type of business in the global scope exceeds the first threshold, a full baseline update is triggered. When the risk entropy change of a certain type of business in a single region exceeds the second threshold, a baseline update for the corresponding business type in that region is triggered.

4. The marketing risk rule problem three-database graph intelligent recommendation management method as described in claim 1, characterized in that: The specific process of reverse knowledge extraction based on alarm-free abnormal behavior in step S2 is as follows: All abnormal behaviors without alarms are clustered, the anomaly intensity and impact range of each cluster are calculated, and the clusters are sorted from high to low according to their anomaly intensity and the clusters that meet the set conditions are processed first. For each anomalous cluster, retrieve all relevant risks, rules, and issues from the three existing databases, calculate the similarity between the anomalous cluster feature vector and the knowledge feature vectors of the three databases, and mark the anomalous cluster with the highest similarity below a set threshold as a knowledge gap; A counterfactual reasoning algorithm is used to generate all possible risk hypotheses that can explain the anomalous cluster, and the reasonableness of the risk hypotheses is verified.

5. The marketing risk rule problem three-database graph intelligent recommendation management method as described in claim 4, characterized in that: The specific process of generating latent knowledge in step S2 is as follows: For each validated risk hypothesis, a structured knowledge entry containing risk point description, triggering conditions, detection rules, and typical problem case fields is automatically generated using the structure templates of similar knowledge in the three existing databases. The generated potential knowledge is stored in a temporary knowledge base, which is only accessible to personnel participating in the grayscale verification.

6. The marketing risk rule problem three-database graph intelligent recommendation management method as described in claim 1, characterized in that: The specific process of self-verifying knowledge verification based on the business closed loop in step S3 is as follows: The generated potential knowledge is distributed to grassroots inspectors in the form of recommendation rules, and the business processing data of the experimental group and the control group are recorded separately. Scores are calculated from three dimensions: business processing effectiveness, user feedback, and anomaly elimination effectiveness. The comprehensive verification score of potential knowledge is obtained through the comprehensive verification score calculation formula.

7. The marketing risk rule problem three-database graph intelligent recommendation management method as described in claim 6, characterized in that: The specific process of automatic knowledge classification and storage in step S3 is as follows: Potential knowledge whose comprehensive verification score exceeds the first threshold is upgraded to formal knowledge, added to the corresponding risk base, rule base, or question base, and associated with it. Potential knowledge with a comprehensive verification score between the second and first thresholds is labeled as secondary knowledge and used for general recommendation. Potential knowledge with a comprehensive verification score below the second threshold is marked as knowledge to be optimized and pushed to the provincial marketing audit expert database for manual review.

8. The marketing risk rule problem three-database graph intelligent recommendation management method as described in claim 1, characterized in that: The specific process of step S4 is as follows: Based on the application frequency, problem-solving success rate, and user evaluation data of each piece of knowledge, the weight of knowledge in the three-database graph is dynamically adjusted through a weight update formula, and knowledge whose weight is lower than a set threshold for consecutive periods is marked as invalid knowledge. The graph convolutional neural network algorithm is used to analyze the co-occurrence patterns of knowledge, automatically uncover the implicit relationships between knowledge and update the three-database graph; The system identifies the feature vector of the user's current business scenario, retrieves knowledge from the three databases that has a similarity greater than a set threshold to the feature vector, generates a recommendation list by weighting and sorting the knowledge based on its weight and association strength, and pushes the recommendation results to the user's business processing interface in real time.

Images