Service approval processing method and device, equipment and storage medium

Abstract

The invention relates to the technical field of service examination and approval, and discloses a service examination and approval processing method, device and equipment and a storage medium, and the method comprises the steps: receiving a user authorization request, verifying the user authorization request, and obtaining a sub-account certificate with a reading only permission; performing real-time data acquisition on the plurality of e-commerce platforms based on the sub-account credentials to obtain multi-dimensional user data including transaction records, operation data and evaluation information; inputting the multi-dimensional user data into a preset space-time sequence attention network and graph convolution hybrid model to carry out dynamic evaluation on a user behavior pattern to obtain risk quantitative indexes; inputting the risk quantitative indexes into a rule engine integrating an adaptive decision tree and a genetic algorithm for automatic decision making to obtain a graded approval decision; and performing multi-signature encryption on the execution instruction according to the graded approval decision to obtain a safety operation instruction and execution state information, thereby realizing full-process automation from data acquisition, risk assessment to decision execution.

Description

Technical Field

[0001] The present application relates to the field of business approval technology, and in particular to a business approval processing method, device, equipment and storage medium. Background Art

[0002] Traditional approval processes usually rely on manual operations, from data collection, risk assessment to final decision execution. The entire process is time-consuming, costly, and prone to human bias and omissions. With the diversification of e-commerce platforms, enterprises need to connect to data sources from multiple platforms at the same time, but these data are often in heterogeneous formats and inconsistent update frequencies, making data integration difficult and the basis for approval insufficient.

[0003] As the volume of business grows, the risk assessment pressure faced by enterprises also increases. Traditional rule-based assessment models lack flexibility and adaptability, and are difficult to cope with complex and changing business scenarios and evolving risk models. When the transaction model or market environment changes, static preset rules often cannot be adjusted in time, resulting in untimely or overly conservative risk control, affecting business efficiency and user experience. In addition, the execution of approval results often faces problems such as delayed information transmission and operational errors. The lack of an effective tracking and monitoring mechanism makes it difficult to ensure the accurate execution of approval decisions. Summary of the invention

[0004] This application provides a business approval processing method, device, equipment and storage medium, thereby realizing full process automation from data collection, risk assessment to decision execution.

[0005] The first aspect of the present application provides a business approval processing method, the business approval processing method comprising: Receiving a user authorization request, verifying the user authorization request, and obtaining a sub-account credential with read-only permission; Based on the sub-account credentials, real-time data collection is performed on multiple e-commerce platforms to obtain multi-dimensional user data including transaction records, operation data and evaluation information; Inputting the multi-dimensional user data into a preset spatiotemporal sequence attention network and graph convolution hybrid model to dynamically evaluate the user behavior pattern and obtain a risk quantification indicator; Input the risk quantification index into a rule engine that integrates an adaptive decision tree and a genetic algorithm to make an automatic decision and obtain a hierarchical approval resolution; The execution instructions are encrypted with multiple signatures according to the hierarchical approval resolution to obtain security operation instructions and execution status information.

[0006] A second aspect of the present application provides a business approval processing device, the business approval processing device comprising: A verification module, used to receive a user authorization request, verify the user authorization request, and obtain a sub-account credential with read-only permission; A collection module, used to collect real-time data from multiple e-commerce platforms based on the sub-account credentials to obtain multi-dimensional user data including transaction records, operation data and evaluation information; A dynamic evaluation module, used to input the multi-dimensional user data into a preset spatiotemporal sequence attention network and graph convolution hybrid model to dynamically evaluate the user behavior pattern and obtain a risk quantification indicator; An automatic decision-making module, used to input the risk quantification index into a rule engine that integrates an adaptive decision tree and a genetic algorithm to make an automatic decision and obtain a hierarchical approval resolution; The encryption module is used to perform multi-signature encryption on the execution instruction according to the hierarchical approval resolution to obtain the security operation instruction and execution status information.

[0007] The third aspect of the present application provides an electronic device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the electronic device executes the above-mentioned business approval processing method.

[0008] A fourth aspect of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned business approval processing method.

[0009] Compared with the prior art, the present application has the following beneficial effects: the sub-account credential mechanism with only read permissions is used to realize the non-intrusive and secure collection of e-commerce platform data, and the data collection efficiency is greatly improved by combining distributed task scheduling and multi-threaded parallel processing technology. The combination of spatiotemporal sequence attention network and graph convolution hybrid model enables risk assessment to capture the temporal dynamic changes of user behavior, analyze the spatial topological structure of transaction relationships, and identify complex risk patterns that are difficult to find in traditional models. The rule engine that integrates adaptive decision tree and genetic algorithm can automatically adjust decision rules and threshold parameters according to changes in business environment, avoiding the lag of traditional fixed rule model. In terms of security, multi-signature encryption and threshold signature algorithm ensure that execution instructions cannot be forged or tampered with, and the application of blockchain technology enables the approval process to form a complete and traceable operation chain. The present invention realizes the automation of the whole process from data collection, risk assessment to decision execution, supports all-weather uninterrupted operation, eliminates the time delay of manual review, and reasonably allocates resources through a multi-level hierarchical approval mechanism, so that low-risk businesses are automatically passed, and high-risk or borderline cases are guided to manual review, optimizing resource allocation and overall system performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0011] The structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with this technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantive technical significance. Any structural modification, change in proportion or adjustment of size, without affecting the effects and purposes that can be achieved by the present invention, should still fall within the scope of the technical contents disclosed by the present invention.

[0012] Figure 1 It is a flowchart of a business approval processing method provided by an embodiment of the present invention; Figure 2 is a schematic block diagram of the structure of a business approval processing device provided by an embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0013] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0014] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.

[0015] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0016] It should be further understood that the term "and / or" used in the specification and appended claims of this application refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations. Figure 1 , an embodiment of the business approval processing method in the embodiment of the present application includes: Step 100: receiving a user authorization request, verifying the user authorization request, and obtaining a sub-account credential with read-only permission; It is understandable that the execution subject of the present application may be a business approval processing device, or a terminal or a server, which is not specifically limited here. The present application embodiment is described by taking a server as the execution subject as an example.

[0017] Specifically, the user authorization request is received and the platform store identification information is obtained, including the store name, platform account ID, store URL or other relevant data that uniquely identifies the store. The store identification information is standardized, including format conversion, character standardization and deduplication operations, such as converting all store IDs into string format, eliminating redundant characters, ensuring that the semantics of store identification information is consistent between different platforms, and generating standardized store identity identifiers. An authorization request data packet is constructed based on the standardized store identity identifier. The authorization request data packet contains information such as the user's authorization intent, store ID, requested permission type (such as read permission, write permission, etc.) and the validity period of the request. The data packet is attached with a timestamp and random number generated by the system to enhance the uniqueness of the data packet and the ability to prevent replay attacks. The constructed data packet is transmitted through a specific e-commerce platform API interface. In this process, the data packet is adapted and formatted according to the API specifications of different platforms to ensure that the request data is correctly parsed and responded to by the platform. When the authorization request data packet is successfully transmitted to the e-commerce platform, the system receives the response data returned by the platform. The platform response data contains the authorization result, authorization token, authorization permission scope, validity period information and other security-related metadata. After receiving these response data, the data packet is digitally signed and verified by using the public key provided by the platform or a specific signature verification algorithm to ensure that the source of the response data is credible and has not been tampered with. The digital signature verification process is based on an asymmetric encryption algorithm, such as RSA or ECDSA, and verifies the integrity of the data by comparing the signature and the hash value of the original data. After completing the digital signature verification, the authorization scope in the response data is matched to determine whether the permissions granted by the platform are consistent with the permissions requested by the user. Pay attention to whether the authorization permissions only include "read" related operation permissions, such as browsing data and querying records, but not including high-risk permissions such as modification, deletion, and writing. The permission matching analysis generates the authorization validity verification result and the corresponding permission level mark by comparing the permission list responded by the platform with the system preset permission whitelist. The permission level mark is an identifier used to indicate the security level of the current authorization permission. Based on the authorization validity verification result, the current authorization status is judged. When the authorization verification result shows valid, the permission is configured for the sub-account according to the permission level mark. In this process, only "read" related permissions are assigned to the sub-account to avoid permissions beyond the necessary scope being incorrectly granted to the sub-account. The permission configuration process involves setting up the access control list of the sub-account, which clearly stipulates that the sub-account can only perform specific query operations, but cannot perform modification or deletion operations. After completing the permission configuration, the sub-account configuration information is written into the security credential database. During the writing process, the data is timestamped. At the same time, in the security credential database, an access control list is set to restrict the sub-account's access rights to other data in the database through predefined access rules to enhance security.Generate a subaccount credential with read-only permissions.

[0018] Step 200: Real-time data collection is performed on multiple e-commerce platforms based on the sub-account credentials to obtain multi-dimensional user data including transaction records, operation data, and evaluation information; Specifically, an API call parameter set is constructed based on the sub-account credentials. The API interface documents of each e-commerce platform are parsed to extract the necessary request parameters, such as the interface access path, request method, request header information, query parameters (such as store ID, time range, data type), etc. At the same time, according to the permission requirements of each platform, the sub-account credentials are bound to the request parameters to ensure that the API request can pass the security verification of the platform. In this process, the API call parameter set is dynamically configured according to the specific needs of the data request. For example, transaction status filtering parameters will be added to transaction record requests, specific statistical dimension settings will be included for operational data requests, and paging and sorting rules will be set for evaluation information requests to obtain the latest user evaluation data. The API call parameter set is encrypted and encapsulated to obtain the platform interface access request package. The encryption and encapsulation process adopts asymmetric encryption, symmetric encryption or hybrid encryption. For example, the request data is encrypted using AES (Advanced Encryption Standard), and then the AES key is encrypted using RSA (asymmetric encryption algorithm) to achieve secure data transmission and tamper-proof. In the request package, the timestamp and random number of the request are added to prevent replay attacks and malicious use of the timeliness of the request. The platform interface access request package is distributed to the multi-threaded data engine through the distributed task scheduler to realize the parallel data extraction operation of multiple e-commerce platforms. The role of the distributed task scheduler is to reasonably distribute the data collection tasks to multiple computing nodes or threads to avoid the low data collection efficiency of a single node or thread due to excessive load. The multi-threaded data engine realizes simultaneous data capture of multiple platforms and multiple interfaces by asynchronously calling the API interface. In actual operation, a data collection task is instantiated for each e-commerce platform, and the task contains the API interface call logic, data processing logic and exception handling mechanism of the platform. For example, when the API interface response of a platform times out or returns an error, the multi-threaded data engine will automatically retry or switch to the backup interface to ensure the stability and reliability of the data collection process. During the data collection process, the original data returned by different platforms is usually in heterogeneous formats, for example, some platforms return JSON format data, and some platforms return XML format. Field mapping and type conversion are performed on the original platform data in heterogeneous formats to obtain standardized structured data. The field mapping process converts fields with the same semantics but different names in different platforms into standard fields through a predefined field mapping table. Type conversion processing standardizes the diversity of data formats, such as unifying the time format to ISO standards, converting the string form of the amount field to a numeric type, and filling or removing empty values ​​and abnormal values ​​in the data to ensure data consistency and computability. Through field mapping and type conversion, all collected data is converted into a standardized structured data format.Based on standardized structured data, a multidimensional association index is constructed to achieve fast data retrieval and multidimensional association analysis. The process of building a multidimensional association index includes creating indexes for key fields in the data table (such as store ID, order ID, product ID, timestamp, etc.) to improve the efficiency of data query. At the same time, association relationships are established for data of different dimensions, such as associating transaction records with evaluation information through order ID, and associating operation data with historical transaction data of the store through store ID, forming a multidimensional and highly associated data model. The multidimensional association index is cross-validated and integrity checked. Cross-validation is performed by comparing whether data from different sources are consistent in key fields, such as whether the order amount in the transaction record matches the sales in the operation data. The integrity check is performed by checking whether the required fields in the data are empty and whether the logical relationship of the data is reasonable. Through verification and validation operations, a real-time updated user data set is obtained. Dimension expansion and feature derivative calculation are performed on the real-time updated user data set to obtain multi-dimensional user data containing transaction records, operation data and evaluation information. Dimension expansion makes data analysis results richer and more comprehensive by adding new analysis dimensions, such as introducing time dimensions (such as weeks, months, and quarters), geographic dimensions (such as regions and cities), and behavioral dimensions (such as activity and repurchase rate). Feature derivative calculations generate new features based on existing data, such as calculating the average customer unit price through transaction records, analyzing the comprehensive score of stores through evaluation information, and predicting the sales trend of stores through operational data. These feature derivative operations rely on statistical methods, data mining algorithms, and machine learning models, such as predicting sales through time series analysis, extracting emotional features in user reviews through sentiment analysis, and discovering potential patterns in customer groups through clustering algorithms. After this series of data processing, analysis, and feature calculation, multi-dimensional user data is finally obtained.

[0019] Step 300: Input the multi-dimensional user data into a preset spatiotemporal sequence attention network and graph convolution hybrid model to dynamically evaluate the user behavior pattern and obtain a risk quantification index; It should be noted that the multi-dimensional user data is decomposed in time series and the topological structure is constructed. The complex user transaction data and operation behavior data are converted into a time series feature matrix and a graph relationship adjacency matrix that are easy for the model to process. In the process of time series decomposition, the user's transaction records are arranged in chronological order, and the continuous transaction behavior is converted into time series features through sliding windows and time segmentation technology. For example, the daily transaction frequency, transaction amount, order quantity, etc. are converted into feature dimensions in the time series. At the same time, according to the time interval of the transaction, the periodicity and seasonality of the order, the long-term and short-term time dependencies are extracted to ensure that the subsequent model accurately captures the changing trend of user behavior and potential risk signals. At the same time, based on the association information in the multi-dimensional user data, the graph structure relationship of user transaction behavior is constructed. When constructing the graph structure, each user, product, store or transaction behavior is regarded as a node in the graph, and the transaction relationship, evaluation interaction, and product association between the nodes constitute the edge in the graph. For example, when a user purchases the same product multiple times, a weighted edge is established between the user node and the product node, and the weight of the edge represents the total purchase frequency or amount. Similarly, by analyzing the user interaction data in the evaluation information, a social relationship graph between users is constructed. These topological structures are represented in the form of an adjacency matrix. Each element in the adjacency matrix represents the connection relationship and its strength between two nodes, thereby obtaining a graph relationship adjacency matrix. The time series feature matrix is ​​input into the multi-head self-attention network layer for time dimension feature extraction. In the self-attention network, the data of each time step is associated with other time steps to obtain its importance weight in the entire time series. The position encoding mechanism adds position information to the time step, so that the model maintains the consistency of the time order when processing unordered sequence data. The introduction of the mask mechanism can prevent the model from leaking future information when predicting future behavior. By masking the unobserved time steps, the validity and rationality of the model prediction are guaranteed. After being processed by the multi-head self-attention network, the feature vector of the user behavior is obtained. The graph relationship adjacency matrix is ​​input into the graph convolution network for spatial dimension feature extraction. In the graph convolution network, the information of the node and its neighbor nodes is aggregated through the message passing mechanism to generate the embedded representation of each node. In the graph structure of transaction behavior, the transaction relationship, evaluation interaction and commodity association information of user nodes are integrated so that the characteristics of each user node not only depend on its own transaction behavior, but also are affected by other nodes with which it has transaction or interaction relationships. For example, if a user has frequent transactions with multiple high-risk users, then after the graph convolution operation, the user's feature representation will also reflect a higher risk signal. The graph convolution operation includes three steps: neighborhood feature aggregation, feature transformation and nonlinear activation. Through the superposition of multi-layer graph convolution networks, deeper relationships and features in the graph structure are captured to obtain a representation of the transaction relationship.The behavior feature vector and the transaction relationship representation are input into the hybrid attention layer to achieve adaptive weight allocation of spatiotemporal features. The hybrid attention layer simultaneously pays attention to the behavior pattern in the time dimension and the relationship structure in the space dimension. By dynamically allocating attention weights, it automatically learns the features that should be given priority in different scenarios. The hybrid attention mechanism generates a user behavior representation vector by weighted calculation, comprehensively considering the influence of temporal features and spatial features. The user behavior representation vector is input into the anomaly detection module, and the behavior deviation is calculated using the contrastive learning framework. Contrastive learning measures the difference between the current behavior and the normal behavior pattern by comparing the user's current behavior with the historical behavior. For example, by calculating the deviation between the user's current transaction frequency and amount distribution and the historical average, it can be identified whether the user has abnormal transaction behavior. On the basis of calculating the behavior deviation, the deviation is standardized in combination with the historical statistical distribution, and the risks of different users and different scenarios are quantified into a unified risk quantification indicator.

[0020] The transaction records in the multi-dimensional user data are sorted and segmented to obtain a time series segmented data set. According to the timestamp information in the transaction records, the transaction records are sorted in chronological order to ensure the continuity and logic of the data in the time dimension. After the sorting is completed, the continuous transaction records are divided into several time segments using a sliding window or fixed time interval segmentation method. For example, daily transaction records are used as a segment, or according to business needs, the transaction data of a week or a month are aggregated into a time series segment. Based on the time series segmented data set, the core time series features such as transaction frequency, transaction amount change rate and cumulative transaction volume are calculated to construct the initial time series feature set. The transaction frequency represents the number of transactions that occur in a time period, which is obtained by counting the number of orders in each segment. The transaction amount change rate reflects the volatility of user consumption behavior by calculating the change range of the transaction amount of adjacent segments. The cumulative transaction volume is to accumulate the historical transaction amounts to obtain the user's long-term transaction trend. After calculating these time series features, the feature data is standardized to eliminate the differences in the numerical magnitude of different features. By converting the data into a form of zero mean and unit variance (such as Z-Score standardization), the model's sensitivity to different features is enhanced, making the subsequent time series feature extraction process more stable and reliable. The initial time series feature set is decomposed at multiple scales to extract time series patterns at different frequencies. The multiscale decomposition method uses techniques such as wavelet transform or empirical mode decomposition to decompose time series data into multiple frequency components to capture short-term fluctuations and long-term trends in the data. For example, the wavelet transform decomposes the time series signal into sub-signals of different resolutions through high-frequency and low-frequency filters, thereby capturing the changing characteristics of the data in both the time domain and the frequency domain. On the basis of multiscale decomposition, the Fourier feature extraction technology is combined to analyze the periodic information in the time series data. The Fourier transform converts the time series signal from the time domain to the frequency domain, and identifies the main periodic components in the data through spectrum analysis, such as seasonal consumption patterns, monthly transaction fluctuations, and other features. Through multiscale decomposition and Fourier feature extraction, the time series feature matrix is ​​obtained. Based on the transaction records in the multi-dimensional user data, the transaction entity nodes and transaction relationship edges are extracted to construct the initial transaction relationship network. The transaction entity nodes include user nodes, product nodes, store nodes and other transaction-related entities. For example, when a user purchases a specific product, a user node and a product node are created and connected through a transaction relationship edge. The weight of the transaction relationship edge represents information such as transaction frequency, transaction amount or transaction success rate. Similarly, the transaction relationship between the store node and the user node is represented by an edge, so that the entire transaction relationship network not only reflects the interaction between users and products, but also shows the complex relationship between users and stores, and between products and stores.When constructing the initial transaction relationship network, the network structure information is stored in the form of an adjacency matrix. Each element in the adjacency matrix represents whether there is a connection between the corresponding nodes in the network and the weight of the connection. The transaction network is clustered, and the centrality index, structural similarity and connectivity characteristics are calculated to obtain the graph relationship adjacency matrix. The node clustering process divides the nodes in the network into several groups according to the similarity of transaction behavior. For example, the community discovery algorithm is used to cluster frequently interacting users into a group, identify potential group patterns in transaction behavior, and discover abnormal transaction behavior. For example, by analyzing isolated or abnormally active nodes in the clustering results, potential risk users are identified. Centrality indicators are used to evaluate the importance of nodes in the network. Centrality indicators include degree centrality, betweenness centrality and eigenvector centrality. These indicators help the system identify nodes that play a key role in the transaction network, such as high-frequency trading users or core stores. Structural similarity measures the relationship characteristics between a node and its neighbor nodes by calculating the similarity between node neighborhoods, such as using the Jaccard coefficient or cosine similarity, while connectivity characteristics reveal the overall structural characteristics of the transaction network by analyzing the number of paths between different nodes in the network, the shortest path length, and the clustering coefficient of the network. After integrating these features, the graph relationship adjacency matrix is ​​updated to not only represent the connection relationship between nodes, but also contain information about the importance, similarity, and network connectivity of nodes.

[0021] Step 400: Input the risk quantification index into the rule engine integrating the adaptive decision tree and the genetic algorithm to make an automatic decision and obtain a hierarchical approval resolution; Specifically, the risk quantification indicators are divided into multiple layers of thresholds, and the initial rule set is constructed in combination with historical approval data to obtain the initial decision rules containing conditional judgment statements and execution actions. In this stage, the indicators are divided into different risk levels according to the distribution characteristics of the risk quantification indicators. In combination with the decision records in the historical approval data, corresponding rule actions are defined for each risk level. A multi-layer decision tree structure is constructed based on the initial rule set, and the multi-layer decision tree structure is path optimized and pruned to obtain the basic decision tree model. In the process of building the decision tree, each node represents a conditional judgment, each branch represents a different result of the condition, and the leaf node corresponds to the final decision result (such as pass, review, and reject). By recursively splitting the feature space of the risk quantification indicators, the complex rule conditions are decomposed layer by layer into simple binary decision paths, thereby converting the initial rule set into a complete decision tree model. In the construction process, features with higher information gain or Gini coefficient are preferentially selected for node splitting to maximize the decision efficiency improvement brought by each step of splitting. After the decision tree is generated, the decision tree is path optimized and pruned to reduce the risk of overfitting of the model and improve the generalization ability of the decision tree by pruning redundant branches and invalid nodes. For example, when the conditional judgments in certain branch paths have no significant impact on the final decision, the system will automatically merge or delete these paths, making the decision tree structure more concise and efficient. The node conditions and threshold parameters of the basic decision tree model are encoded into gene sequences, and the genetic algorithm is used to initialize the population and evaluate the fitness of the gene sequence to construct the initial population set. The gene encoding process mainly maps each node condition threshold parameter in the decision tree to a gene fragment, and multiple node conditions are combined to form a complete gene sequence. Each gene sequence represents a specific rule combination. When the population is initialized, multiple different gene sequences are generated by random generation, rule sampling, or heuristic methods based on historical data to form the diversity of the population. The fitness of each gene sequence (i.e., rule combination) is evaluated by its performance in historical approval cases, such as calculating key indicators such as the accuracy, false alarm rate, and approval efficiency of each rule in historical data. Gene sequences with higher fitness will have a higher probability of survival in the subsequent evolution process. The initial population set is subjected to crossover, mutation, and selection operations, the accuracy of the calculation rules is verified by historical cases, and the threshold parameters are iteratively optimized to obtain the optimal decision rule. In the genetic algorithm, the crossover operation generates a new rule combination by exchanging some fragments of two high-fitness gene sequences, thereby exploring new decision paths. The mutation operation increases the diversity of the population by randomly modifying certain conditions or thresholds in the gene sequence, preventing the model from falling into a local optimal solution. After each round of iteration, the gene sequence with higher fitness is selected based on the fitness evaluation results to be retained in the next generation, while the gene sequence with lower fitness will be eliminated.Through the iterative process of selection-crossover-mutation-evaluation, the node conditions and threshold parameters in the decision tree model are gradually optimized, so that the rule engine automatically learns and adjusts the optimal decision-making strategy under different risk scenarios. After multiple iterations of genetic algorithm optimization, the optimal decision rules are finally obtained, and the risk quantification indicators are multi-level classified based on these optimal rules to achieve automatic decision-making output of graded approval resolutions. The input risk quantification indicators are judged by each node condition of the optimal decision tree model in turn, and finally classified into decision results such as "automatic approval", "manual review" or "automatic rejection" according to different path branches.

[0022] Step 500: Perform multi-signature encryption on the execution instruction according to the hierarchical approval resolution to obtain the security operation instruction and execution status information.

[0023] Specifically, an original instruction data structure including operation type, execution parameters and business identifier is generated based on the hierarchical approval resolution, and the integrity check of the original instruction data structure is performed to obtain standardized execution instructions. When generating the original instruction data structure, the decision results in the hierarchical approval resolution are converted into specific operation instructions, and necessary execution parameters (such as amount, target account, operation object ID) and business identifiers (such as order number, task ID) are added in combination with the business scenario to ensure the accuracy and uniqueness of the instructions in the subsequent execution process. In order to prevent the instruction data from being tampered with during the generation and transmission process, the integrity check of the original instruction data structure is performed. By calculating the hash value of the instruction data and storing the hash value together with the instruction data, when the instruction is subsequently operated, the hash value is recalculated and compared with the original value to ensure the integrity of the data. Only instruction data that passes the integrity check will enter the next step of the standardization process. In the standardization process, the format, field type, value range, etc. of the instruction data are uniformly processed so that the execution instructions maintain consistency and compatibility when transmitted between different systems and modules, and standardized execution instructions are obtained. The standardized execution instructions are input into the permission control module, and the required number of signatures and the level of signature authority are allocated according to the operation type and risk level to obtain the multi-signature requirement configuration. In the permission control module, the basic signature requirements are determined according to the operation type of the instruction. At the same time, the number of signatures and the level of authority are dynamically adjusted in combination with the risk level of the instruction (such as low, medium, and high risk). For example, high-risk operations require more signatories to participate, and the signatories need to have higher authority levels. Based on the multi-signature requirement configuration, a signature request task is generated, and the signature request is distributed to the relevant authority holders through a distributed message queue to obtain a multi-party signature response set. When generating a signature request task, the system creates an independent signature request for each authority holder who needs to sign. The signature request contains a standardized execution instruction summary, the signer's identity authentication information, and the time limit of the signature (for example, the signature is valid within 24 hours). In order to achieve efficient distribution of signature requests, the distributed message queue is used to push the signature request in the form of an asynchronous message to the signer's approval system or signature device, such as a mobile approval APP, email, or hardware signature key. When the signer receives the signature request, he or she verifies his or her identity (such as two-factor authentication, digital certificate verification) and signs the request in the approval system. After the signing is completed, the signer's digital signature (usually a digital signature fragment encrypted using a private key to encrypt the instruction summary) is fed back to the system through a secure channel and compiled into a multi-party signature response set. The multi-party signature response set is verified for validity and signature quantity, and the threshold signature algorithm is used to combine the multi-party private key fragments to obtain digital signature authentication information. During the validity verification process, the legitimacy of each signature response is verified one by one, including signer identity verification, signature validity period check, and signature data integrity check.For example, the public key is used to verify whether the digital signature matches the original instruction summary to ensure that the signer is indeed the authorized authority holder. When verifying the number of signatures, it is determined whether the number of valid signatures in the signature response set meets the signature threshold requirements in the multi-signature requirement configuration. Only when the number of signatures reaches or exceeds the threshold, the system will continue with the subsequent operation. Through the threshold signature algorithm, the private key fragments of the multi-party signatures are combined and calculated to generate digital signature authentication information. The advantage of the threshold signature algorithm is that even if some signers do not provide signatures, as long as the threshold number is reached, the system can still successfully restore the complete digital signature, thereby ensuring the high availability and flexibility of the system. The standardized execution instructions and digital signature authentication information are encrypted and encapsulated, and the instruction content is protected by an asymmetric encryption algorithm to obtain a secure operation instruction. During the encryption and encapsulation process, the instruction content is encrypted using the public key of the instruction recipient. Only the recipient holding the corresponding private key can decrypt and view the instruction content, thereby ensuring the data security of the instruction during transmission. The encrypted secure operation instruction will be distributed to the execution system, and the encrypted instruction will be pushed to the specific execution node or execution device (such as business system, automation robot, execution server, etc.) through the instruction distribution module. During the execution of instructions, the status tracking module monitors the execution process in real time, collects execution status information (such as "executing", "completed", "failed"), and feeds back the execution results to the approval system to achieve closed-loop management and full traceability of the business approval process, and finally obtain the execution status information of the safe operation instructions.

[0024] In the embodiment of the present application, a sub-account credential mechanism with only read permissions is used to realize non-intrusive and secure collection of e-commerce platform data, and the data collection efficiency is greatly improved by combining distributed task scheduling and multi-threaded parallel processing technology. The combination of spatiotemporal sequence attention network and graph convolution hybrid model enables risk assessment to capture the temporal dynamic changes of user behavior, analyze the spatial topological structure of transaction relationships, and identify complex risk patterns that are difficult to find in traditional models. The rule engine that integrates adaptive decision trees and genetic algorithms can automatically adjust decision rules and threshold parameters according to changes in the business environment, avoiding the lag of traditional fixed rule models. In terms of security, multi-signature encryption and threshold signature algorithms ensure that execution instructions cannot be forged or tampered with, and the application of blockchain technology enables the approval process to form a complete and traceable operation chain. The present invention realizes the automation of the entire process from data collection, risk assessment to decision execution, supports uninterrupted operation around the clock, eliminates the time delay of manual review, and reasonably allocates resources through a multi-level hierarchical approval mechanism, so that low-risk businesses are automatically passed, and high-risk or borderline cases are guided to manual review, optimizing resource allocation and overall system performance.

[0025] In a specific embodiment, the process of executing step 100 may specifically include the following steps: Receive user authorization requests and obtain platform store identification information, normalize the platform store identification information, and obtain a standardized store identity; Construct an authorization request data packet based on the standardized store identity, and transmit the authorization request data packet to the e-commerce platform API interface to obtain platform response data; Perform digital signature verification and authorization scope matching analysis on the platform response data to obtain the authorization validity verification result and permission level mark; The authorization status is determined based on the authorization validity verification result. When the authorization is valid, the permission level tag is configured to obtain the sub-account configuration information. The sub-account configuration information is written into the security credential database for timestamp marking and access control list setting, and a sub-account credential with read-only permission is obtained.

[0026] Specifically, a user authorization request is received, which includes user identity information, platform store identification information, and requested permission type. Suppose the request data is represented as a vector ,in Indicates the user ID, Indicates store identification information Indicates the permission type requested. After receiving the request data, the system extracts the platform store identification information , and normalize it to ensure that identification information in different formats can be represented uniformly. The normalization process is represented as a function , used to convert the platform store identification information into a standardized format. The formula is: in It is a standardized store identity logo. It is a specific normalization operation function, including character standardization, format conversion, and uniqueness check. After normalization, a standardized store identity is obtained. , and build an authorization request packet based on it. The authorization request packet includes a standardized store identity , User ID , the type of permission requested , Timestamp and random numbers And other key information, data packets The data structure is represented as: Data Packet Encryption is performed to ensure the security of data during transmission. The encryption process is implemented through an asymmetric encryption algorithm, for example: in Indicates the encrypted data packet. is the public key of the e-commerce platform. The encryption function Encrypt uses the public key encryption algorithm to encrypt the data packet. The data is transmitted to the e-commerce platform through the API interface. After receiving the request, the platform will decrypt the data and return the response data. Response data Include authorization status , Authorization Token , Scope of Authorization And the digital signature Sig of the platform, expressed as: After receiving the platform response data, the system verifies the digital signature Sig to verify the authenticity and integrity of the response data. To verify, the verification formula is: If the signature verification passes, the system will perform a matching analysis of the authorization scope to ensure that the permissions granted by the platform are valid. Requested permissions from the user The matching degree analysis is obtained by calculating the ratio of the intersection of the requested permission and the authorized scope to the requested permission: in Indicates the matching degree, when It indicates an exact match when indicates a partial match or no match. Calculate the authorization validity verification result and permission level tags After obtaining the authorization validity verification result, determine the authorization status , when authorization is valid, mark according to the permission level Perform permission configuration and generate sub-account configuration information. Sub-account configuration information includes sub-account ID , permission level and the validity period of the authorization , expressed as: Configure the sub-account information Write to the security credential database, and perform timestamp marking and access control list setting during the writing process. Timestamp marking Used to record the creation time of configuration information. The ACL is set to read-only permission. Generate sub-account credentials with read-only permission , data security is ensured through encryption and signing. The formula for generating sub-account credentials is: in It is the private key of the system, which encrypts and encapsulates the configuration information and signature through an asymmetric encryption algorithm to generate the final security credentials.

[0027] In a specific embodiment, the process of executing step 200 may specifically include the following steps: Build an API call parameter set based on the sub-account credentials, encrypt and encapsulate the API call parameter set, and obtain a platform interface access request package; The platform interface access request package is distributed to the multi-threaded data engine through the distributed task scheduler, and parallel data extraction operations are performed on multiple e-commerce platforms to obtain the original platform data in heterogeneous formats; Perform field mapping and type conversion on the original platform data in heterogeneous formats to obtain standardized structured data; Build a multidimensional association index based on standardized structured data, cross-validate and verify the integrity of the multidimensional association index, and obtain a real-time updated user data set; Perform dimension expansion and feature derivation calculations on the real-time updated user data set to obtain multi-dimensional user data including transaction records, operation data and evaluation information.

[0028] Specifically, an API call parameter set is constructed based on the sub-account credentials. The sub-account credentials are represented as ,in Indicates the sub-account ID. Indicates the permission level. Indicates the validity period of the credential. When constructing the API call parameter set, the required request parameters are generated according to the API interface specifications of different e-commerce platforms, such as request method, request URL, request header information (such as authentication information, authorization token), query parameters (such as time range, data type), etc. Assume that the API call parameter set is ,in Indicates the request URL. Represents authentication information (such as a signature or token). Indicates specific query parameters (such as date, data type, etc.). In order to ensure the security of the API call parameter set during transmission, the parameter set Encryption encapsulation. The encryption process uses an asymmetric encryption algorithm, such as the platform's public key The data is encrypted, and the formula is expressed as: in It is an encrypted platform interface access request packet. is the encryption function, This is the public key provided by the e-commerce platform. This encryption method ensures that the request package will not be tampered with or leaked during network transmission. Only the e-commerce platform holding the corresponding private key can decrypt and view the request content. The distributed task scheduler is used to distribute tasks to multi-threaded data engines to perform parallel data extraction operations on multiple e-commerce platforms. The role of the distributed task scheduler is to reasonably distribute tasks to multiple nodes or threads to avoid overloading a single node and affecting data collection efficiency. Suppose the task scheduling model is ,in Indicates The request package of an e-commerce platform, Indicates the assigned computing node. The scheduling process is done through a load balancing algorithm Implementation, so that the task load of each computing node is evenly distributed, for example: in Representation Node The current load, The multi-threaded data engine obtains the original data returned by each platform through asynchronous API calls, and obtains the original platform data set in heterogeneous formats. , these data include different formats such as JSON, XML, CSV, etc. The data field name and data structure of each format are different, which increases the complexity of data processing. In order to unify the data format, the original platform data in heterogeneous formats are processed by field mapping and type conversion to obtain standardized structured data. In the field mapping process, predefined field mapping rules are used , the data field returned by the platform Mapped to standard fields used internally by the system The type conversion process converts the data field into a consistent data type. The standardized structured data set is represented as , the field names and data types of all data have been unified. Based on standardized structured data, a multi-dimensional correlation index is constructed to achieve fast data retrieval and multi-dimensional correlation analysis. Multi-dimensional correlation index By creating index structures for key fields (such as order ID, user ID, product ID, timestamp, etc.), such as B-tree index, inverted index, or multidimensional data index (such as R-tree, KD-tree), etc. The index construction formula is expressed as: in For fields The index constructed, It is a collection of all standardized fields. It performs cross-validation and integrity check on multi-dimensional association indexes. It determines whether the key fields of different data sources match through cross-validation. For example, it verifies the correspondence between transaction records and evaluation information through order IDs. It confirms whether there are null values ​​or abnormal values ​​in the necessary fields in the data set (such as order amount, transaction status, and evaluation score) through integrity check, and obtains real-time updated user data sets. . Perform dimension expansion and feature derivation calculations on the real-time updated user data set to build multi-dimensional user data that includes transaction records, operational data, and evaluation information. The dimension expansion process makes the data analysis results more comprehensive by adding new analysis dimensions, such as time dimensions (year, month, day), geographic dimensions (region, city), user dimensions (age, gender), etc. Feature derivation calculations generate new features from existing data, such as calculating the average consumption amount of customers through transaction records. ,in Indicates Transaction amount, calculate the store's sales growth rate through operational data ,in and Represents the sales of the two time periods before and after, and calculates the average rating of the product through the evaluation information ,in It is The ratings of the reviews. Through these data processing and feature calculation, multi-dimensional user data is obtained.

[0029] In a specific embodiment, the process of executing step 300 may specifically include the following steps: Perform time series decomposition and topological structure construction on multi-dimensional user data, convert transaction records into time series features, map related transactions into graph structure relationships, and obtain time series feature matrices and graph relationship adjacency matrices; The temporal feature matrix is ​​input into the multi-head self-attention network layer to extract the temporal dimension features, and the long-term and short-term dependencies are captured through the position encoding and masking mechanism to obtain the behavior feature vector; The graph relationship adjacency matrix is ​​input into the graph convolutional network layer to extract spatial dimension features, and the node neighborhood information is aggregated through the message passing mechanism to obtain the transaction relationship representation; The behavior feature vector and transaction relationship representation are input into the hybrid attention layer, and the spatiotemporal features are adaptively weighted to obtain the user behavior representation vector. The user behavior representation vector is input into the anomaly detection module, the behavior deviation is calculated through the comparative learning framework, and normalized in combination with the historical statistical distribution to obtain the risk quantification index.

[0030] Specifically, the multi-dimensional user data is decomposed into time series and the topological structure is constructed, and the complex user transaction behavior data is converted into a time series feature matrix and a graph relationship adjacency matrix that are easy to process by the model. In the time series decomposition process, the user's transaction records are arranged in chronological order, and the time series features are extracted through the sliding window method. For example, for the transaction amount ,in Indicates The amount of the transaction, by window size Calculating Moving Averages To smooth a time series: in represents the current time step, It is a window In this way, the trend characteristics and fluctuation characteristics in the time series are extracted to generate the time series feature matrix ,in Indicates The time step feature values, including multiple time series features such as amount, transaction frequency, and rate of change. At the same time, a graph structure relationship is constructed based on the transaction records in the multi-dimensional user data, and the related transactions are mapped into a graph structure relationship. In the process of graph structure construction, each user, product, store or transaction behavior is regarded as a node in the graph, and the transaction relationship, evaluation interaction, and product association between nodes constitute the edges in the graph. For example, suppose Represents a transaction relationship graph, where is a collection of nodes, is an edge set. For user nodes and product nodes , if the user Purchased product , then create an edge in the graph , the edge weight Calculated based on the number of purchases or transaction amount, for example: in Indicates the number of purchases. Indicates the cumulative transaction amount. It is a parameter used to adjust the weight. The graph relationship adjacency matrix is ​​generated by constructing nodes and edges. ,in Representation Node and nodes The connection strength between them. The multi-head self-attention network layer is input to extract time dimension features, and the long-term and short-term dependencies are captured through position encoding and masking mechanisms. The self-attention mechanism scores the importance of specific moments in the time series by calculating the attention weights between the input feature vectors. For example, the attention weight calculation formula is: in , and denote the query matrix, key matrix and value matrix respectively, is the feature dimension. In the multi-head self-attention mechanism, through multiple groups of different , , The matrix calculates multiple sets of attention weights to capture the characteristic patterns of different time scales and obtain the behavior feature vector At the same time, the graph adjacency matrix The input graph convolutional network layer extracts spatial dimension features, aggregates node neighborhood information through the message passing mechanism, and obtains the transaction relationship representation. In the graph convolutional network, the feature representation of each node is updated by aggregating the features of itself and its neighboring nodes. For example, the node feature update formula is: in It is Layer node feature matrix, is the weight matrix, is the activation function (such as ReLU), It is the graph relationship adjacency matrix. Through multi-layer graph convolution operations, the direct and indirect relationships of nodes are integrated into the feature representation to obtain the final transaction relationship representation. . The behavior feature vector and transaction relationship Input the hybrid attention layer, and fuse the temporal and spatial features through an adaptive weight allocation mechanism to achieve the characterization of user behavior. In the hybrid attention layer, weights are assigned to temporal features and spatial features respectively. and , by weighted sum calculation of user behavior representation vector : in , dynamically adjust these two weights through learning to adapt to the changes in the importance of spatiotemporal features in different scenarios. The input anomaly detection module calculates the behavior deviation through the contrastive learning framework, and standardizes it with historical statistical distribution to obtain risk quantification indicators. In the contrastive learning framework, the behavior deviation is calculated by calculating the difference between the current behavior and the normal behavior pattern, such as by cosine similarity. : in is the reference behavior vector (normal behavior mode), Represents the norm of the vector. The calculated behavior deviation Compare with historical statistical distribution and calculate risk quantification indicators through standardization methods (such as Z-score standardization) : in and are the mean and standard deviation of historical behavior deviations. Through this method, the risk changes of user behavior can be dynamically monitored and the risk level can be quantitatively evaluated.

[0031] In a specific embodiment, the execution step performs time series decomposition and topology structure construction processing on multi-dimensional user data, converts transaction records into time series features, maps related transactions into graph structure relationships, and obtains a time series feature matrix and a graph relationship adjacency matrix. The process may specifically include the following steps: Sort and segment the transaction records in the multi-dimensional user data to obtain a time-series segmented data set; Based on the time series segmented data set, the transaction frequency, transaction amount change rate and cumulative transaction volume are calculated, and standardized transformation is performed to obtain the initial time series feature set; Perform multi-scale decomposition on the initial time series feature set, extract the time series patterns at different frequencies, and extract the periodic information in combination with Fourier features to obtain the time series feature matrix; Extract transaction entity nodes and transaction relationship edges based on transaction records in multi-dimensional user data, and construct an initial transaction relationship network based on the transaction entity nodes and transaction relationship edges; The nodes of the initial transaction relationship network are clustered, and the centrality index, structural similarity and connectivity characteristics are calculated to obtain the graph relationship adjacency matrix.

[0032] Specifically, the transaction records in the multi-dimensional user data are sorted and segmented to obtain a time-series segmented data set. In the form of It is The time when a transaction occurred. Sort the transaction records in chronological order to ensure the logical continuity of the data in the time dimension. Sorted transaction data Include timestamp and the corresponding transaction amount In order to extract time series features, the sliding window method is used to segment the sorted transaction records, and the length of each segment is , represents the transaction data within a time window, such as daily, weekly or monthly transaction records. The time series segmented data set obtained by segmentation processing is expressed as ,in It is A subset of transactions in a time window. Based on the time series segmented dataset , calculate the core time series features such as transaction frequency, transaction amount change rate and cumulative transaction volume to obtain the initial time series feature set. Indicates in segment The number of transactions in is obtained by counting method: in Indicates segmentation The number of transaction records in the transaction amount. By calculating the change rate between the total transaction amount of the current segment and the total transaction amount of the previous segment, we can get: in It is The amount of the transaction, Represents the data set of the previous time segment. Cumulative transaction volume The historical transaction amounts are accumulated to indicate the long-term transaction trend of the user in the current time window: The calculated time series characteristics , and Standardization is performed and the Z-score standardization method is used to transform the eigenvalues ​​into a standard normal distribution form: in is the original eigenvalue, and are the mean and standard deviation of the feature data respectively. The standardized feature data constitute the initial time series feature set Based on the initial time series feature set, multi-scale decomposition is performed to extract time series patterns at different frequencies, and the periodic information is extracted in combination with Fourier features to generate a time series feature matrix. The multi-scale decomposition process uses methods such as wavelet transform empirical mode decomposition to decompose the time series into components of different frequencies, for example: in is a timing signal, and are the basis functions of the low-frequency and high-frequency components, respectively. and is the corresponding component coefficient. Combined with Fourier transform to calculate frequency domain features, extract the periodic patterns in trading behavior, and obtain the main periodic features through spectrum analysis. : Where FFT stands for Fast Fourier Transform, It is the intensity of the frequency component. The frequency distribution can be used to identify the periodicity of user transaction behavior, such as daily, weekend or seasonal consumption habits. The results of multi-scale decomposition and frequency domain analysis constitute the time series feature matrix , each row of the matrix represents the multi-frequency and multi-period features of a time window. At the same time, based on the transaction records in the multi-dimensional user data, the transaction entity nodes and transaction relationship edges are extracted to construct the initial transaction relationship network. The transaction relationship network is shown in Figure In the form of It is a collection of nodes, representing transaction entities such as users, products, and stores. It is a set of edges, representing transaction behaviors, evaluation relationships, or product association relationships. Through transaction records, we analyze the interaction between users and products, products and stores, and users and users. For example, for user nodes and product nodes , if the user purchases a product, an edge is created in the graph , the edge weight Calculated based on transaction frequency or amount, for example: in is the number of purchases, is the cumulative transaction amount, is the adjustment coefficient. After constructing the initial transaction relationship network, node clustering is performed, and centrality indicators, structural similarity, and connectivity characteristics are calculated to obtain the graph relationship adjacency matrix. The node clustering method divides the nodes in the transaction network into several groups through community discovery algorithms (such as Louvain algorithm, spectral clustering). Centrality indicators (such as degree centrality, betweenness centrality, eigenvector centrality) are used to evaluate the importance of nodes. Structural similarity is calculated by calculating the similarity of node neighborhoods (such as Jaccard coefficient). Connectivity characteristics are analyzed by analyzing the number of paths between different nodes in the network, clustering coefficients and other characteristics. The graph relationship adjacency matrix constructed by these network analysis methods is Contains node connections and their weights.

[0033] In a specific embodiment, the process of executing step 400 may specifically include the following steps: Perform multi-layer threshold division on risk quantification indicators, build an initial rule set based on historical approval data, and obtain initial decision rules containing conditional judgment statements and execution actions; A multi-layer decision tree structure is constructed based on the initial decision rules, and the path of the multi-layer decision tree structure is optimized and pruned to obtain a basic decision tree model; The node conditions and threshold parameters of the basic decision tree model are encoded into gene sequences, and a genetic algorithm is used to initialize the population and evaluate the fitness of the gene sequences to obtain an initial population set; Perform crossover mutation and selection operations on the initial population set, verify the accuracy of the calculation rules through historical cases, and iteratively optimize the threshold parameters to obtain the optimal decision rule; The risk quantification indicators are classified into multiple levels according to the optimal decision-making rules, and hierarchical approval resolutions including automatic approval, manual review and automatic rejection are obtained.

[0034] Specifically, the risk quantification index is divided into multiple layers of thresholds, and the initial rule set is constructed in combination with the historical approval data to form the initial decision rule containing conditional judgment statements and execution actions. ,in Indicates risk factors, such as user credit score, transaction behavior abnormality, historical overdue rate, etc. By analyzing the risk value distribution in the historical approval data, each risk factor is divided into multiple threshold intervals, such as low risk, medium risk and high risk. The division rules are expressed as: in and Respectively represent The lower and upper limits of the threshold interval, is the corresponding classification label (such as passed, reviewed, rejected). Build the initial rule set based on the threshold rule , where each rule Represents a specific combination of conditional judgment and execution action. A multi-layer decision tree structure is constructed based on the initial rule set. By recursively splitting the feature space of risk quantification indicators, the complex decision logic is represented as a decision tree model. Each node in the decision tree represents a conditional judgment, each branch represents a decision path under different conditions, and the leaf node corresponds to the final execution action (such as automatic approval, manual review, and automatic rejection). The optimal split point is selected by information gain or Gini coefficient. The information gain calculation formula is: in represents information gain, is the initial entropy, It is in the characteristics Conditional entropy under the condition. The system selects the feature with the largest information gain for splitting, and obtains a multi-layer decision tree structure. ,in is a collection of nodes, is a set of branches. Path optimization and pruning are performed on multi-layer decision trees to improve the generalization ability of the model. Pruning methods include pre-pruning and post-pruning. In the pre-pruning process, the maximum tree depth is set. , minimum number of samples Conditions such as these prevent the decision tree from growing too much. Subsequent pruning is done by calculating the error rate of the subtree. And the error rate after pruning When making a judgment Pruning is performed when , the formula is: in is the number of misclassified samples, is the total number of samples. Through pruning, a more concise and efficient basic decision tree model is obtained. The node conditions and threshold parameters in the basic decision tree model are encoded as gene sequences, which facilitates rule optimization through genetic algorithms. The gene sequence is represented as , each gene Corresponding to a node condition or threshold. Generate the initial population set through random initialization, regular sampling or historical data-based methods ,in is the population size, and each gene sequence represents a specific rule combination. The initial population set is subjected to crossover, mutation, and selection operations, the accuracy of the calculation rules is verified through historical cases, and the threshold parameters are iteratively optimized to obtain the optimal decision rule. In the crossover operation, two gene sequences are randomly selected. and Exchange some fragments to generate new gene sequences , the crossover formula is: in is the index of the intersection. The mutation operation increases the diversity of the population by randomly modifying some conditions or thresholds in the gene sequence. In the selection operation, the gene sequence with high fitness is selected based on the fitness evaluation. The fitness calculation formula is: in It is a gene sequence The fitness of is the number of samples in historical data where the rule is correctly judged. is the total number of samples. After multiple iterations, through the cycle of selection-crossover-mutation-evaluation, the node conditions and thresholds in the decision rule are continuously optimized to obtain the optimal decision rule According to the optimal decision rule Carry out multi-level classification of risk quantification indicators to realize automated approval decision output. The decision tree model is input, and through layer-by-layer condition judgment and path selection, after reaching the leaf node, a hierarchical approval resolution (such as automatic approval, manual review, and automatic rejection) is output based on the node's execution action.

[0035] In a specific embodiment, the process of executing step 500 may specifically include the following steps: Generate an original instruction data structure containing operation type, execution parameters and business identification based on the hierarchical approval resolution, and perform integrity check on the original instruction data structure to obtain a standardized execution instruction; Input standardized execution instructions into the permission control module, allocate the required number of signatures and signature permission levels according to the operation type and risk level, and obtain the multi-signature requirement configuration; Generate a signature request task based on the multi-signature requirement configuration, distribute the signature request to relevant authority holders through a distributed message queue, and obtain a multi-party signature response set; Verify the validity and number of signatures of the multi-party signature response set, and combine the multi-party private key fragments through the threshold signature algorithm to obtain digital signature authentication information; The standardized execution instructions and digital signature authentication information are encrypted and encapsulated, and the instruction content is protected by an asymmetric encryption algorithm to obtain safe operation instructions. The safe operation instructions are distributed and the execution status is tracked to obtain the execution status information of the safe operation instructions.

[0036] Specifically, an original instruction data structure including operation type, execution parameters and business identifier is generated based on the hierarchical approval resolution. Assume that the approval resolution is ,in Indicates the type of operation, such as "transfer", "authority modification" or "data query". Indicates execution parameters, such as transfer amount, target account, or query conditions; Represents a business identifier, such as an order ID, task ID, or session ID. This information is combined into a raw instruction data structure. : in Indicates the current timestamp, Represents a random number, used to prevent replay attacks. After generating the original instruction data structure, the instruction data is checked for integrity to ensure the consistency of the data during generation and transmission. The integrity check process uses a hash algorithm. , such as SHA-256, hash the instruction data to get the data summary : By verifying the data digest in subsequent operations After verification, the original instruction data structure is standardized, and all fields are formatted into a unified structure, such as the timestamp formatted into the ISO8601 standard, the amount field converted to a numeric type, etc., to obtain a standardized execution instruction. . Standardize execution instructions Input permission control module, according to the operation type and risk level Dynamically allocate the required number of signatures and signing permission levels to generate a multi-signature requirement configuration. Multi-signature requirement configuration Contains the number of signers and signing permission levels , the formula is: in is the distribution rule function. Number of signers Determines the minimum number of signatures and signature permission levels required for command execution Used to specify the role or permission level of the signer. Configure based on multi-signature requirements , generate a signature request task, distribute the signature request to relevant authority holders through a distributed message queue, and obtain a multi-party signature response set. Assume that there is Signatories , the system will generate a signature request for each signer , the signature request contains the instruction summary , Signer Identity and signing permission levels : The signature request is asynchronously pushed to the signer's approval system through a distributed message queue, such as Katka or RabbitMQ. When the signer receives the signature request, he verifies his own authority and digitally signs the instruction data. The signature operation is performed through the private key To encrypt, the formula is: Where Sign represents the signature function, It is a digital signature. After the signature is completed, the signature response Returned to the system through a secure channel, ultimately forming a multi-party signature response set ,in After receiving the multi-party signature response set, the validity and number of signatures are verified for the signature set, and the multi-party private key fragments are combined through the threshold signature algorithm to obtain the digital signature authentication information. Verify the signature Whether to use data summary match: Signature quantity verification confirms the number of valid signatures in the signature response set Does it meet the multi-signature requirement? , only when The system will continue to operate only when the threshold signature algorithm is used to generate valid digital signature authentication information without the participation of all signatories. , the combination process is expressed as: Combine is the combination function of the threshold signature algorithm. Generate a unique digital signature when the standard execution instruction is Authenticate information with a digital signature Encrypt and encapsulate, protect the instruction content through asymmetric encryption algorithms (such as RSA, ECC), and obtain secure operation instructions The encryption operation is performed using the recipient's public key. To ensure the security of the instruction content during transmission: The system will encrypt the security operation instructions Distribute to specific execution nodes, such as fund transfer system, permission management system or data query system. During the execution process, the execution status of the instruction is monitored in real time through the status tracking module, including "received", "executing", "completed" or "failed" status. Execution status information Indicated by instruction ID and status flag: in Is the business logo, It is the current instruction execution status. The execution status information is fed back to the approval system to achieve closed-loop management of the approval process, and all operations are recorded in the log system to support subsequent auditing and tracing, ensuring the security, transparency and compliance of the entire process.

[0037] The above describes the business approval processing method in the embodiment of the present application. The following describes the business approval processing device 10 in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the business approval processing device 10 includes: Verification module 11, used to receive a user authorization request, verify the user authorization request, and obtain a sub-account credential with read-only permission; A collection module 12, for collecting data from multiple e-commerce platforms in real time based on the sub-account credentials, and obtaining multi-dimensional user data including transaction records, operation data and evaluation information; The dynamic evaluation module 13 is used to input multi-dimensional user data into a preset spatiotemporal sequence attention network and graph convolution hybrid model to dynamically evaluate user behavior patterns and obtain risk quantification indicators; The automatic decision module 14 is used to input the risk quantification index into the rule engine integrating the adaptive decision tree and the genetic algorithm to make an automatic decision and obtain a hierarchical approval resolution; The encryption module 15 is used to perform multi-signature encryption on the execution instruction according to the hierarchical approval resolution to obtain the security operation instruction and execution status information.

[0038] Through the collaboration of the above components, the sub-account credential mechanism with only read permissions is adopted to realize the non-intrusive and secure collection of e-commerce platform data, and the data collection efficiency is greatly improved by combining distributed task scheduling and multi-threaded parallel processing technology. The combination of spatiotemporal sequence attention network and graph convolution hybrid model enables risk assessment to capture the temporal dynamic changes of user behavior, analyze the spatial topological structure of transaction relationships, and identify complex risk patterns that are difficult to find in traditional models. The rule engine that integrates adaptive decision tree and genetic algorithm can automatically adjust decision rules and threshold parameters according to changes in business environment, avoiding the lag of traditional fixed rule model. In terms of security, multi-signature encryption and threshold signature algorithm ensure that execution instructions cannot be forged or tampered with, and the application of blockchain technology enables the approval process to form a complete and traceable operation chain. The present invention realizes the automation of the whole process from data collection, risk assessment to decision execution, supports all-weather uninterrupted operation, eliminates the time delay of manual review, and reasonably allocates resources through a multi-level hierarchical approval mechanism, so that low-risk businesses are automatically passed, and high-risk or borderline cases are guided to manual review, optimizing resource allocation and overall system performance.

[0039] See also Figure 3 , Figure 3This is a schematic block diagram of the structure of an electronic device 300 provided in an embodiment of the present application. The electronic device 300 includes a processor 301 and a memory 302. The processor 301 and the memory 302 are connected via a device bus 303, wherein the memory 302 may include a non-volatile storage medium and an internal memory.

[0040] The non-volatile storage medium can store a computer program. The computer program includes program instructions, and when the program instructions are executed by the processor 301, the processor 301 can execute any of the above-mentioned business approval processing methods.

[0041] The processor 301 is used to provide computing and control capabilities to support the operation of the entire electronic device 300 .

[0042] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor 301, the processor 301 can execute any of the above-mentioned business approval processing methods.

[0043] Those skilled in the art will understand that Figure 3 The structure shown in the figure is only a block diagram of a partial structure related to the present application scheme, and does not constitute a limitation on the electronic device 300 involved in the present application scheme. The specific electronic device 300 may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0044] It should be understood that the processor 301 may be a central processing unit (CPU), and the processor 301 may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0045] It should be noted that those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the electronic device 300 described above can refer to the corresponding process of the aforementioned business approval processing method, and will not be repeated here.

[0046] An embodiment of the present application also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by one or more processors, the one or more processors implement the business approval processing method provided in the embodiment of the present application.

[0047] The computer-readable storage medium may be an internal storage unit of the electronic device 300 in the aforementioned embodiment, such as a hard disk or memory of the electronic device 300. The computer-readable storage medium may also be an external storage device of the electronic device 300, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped with the electronic device 300.

[0048] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0049] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable an electronic device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.

[0050] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A business approval processing method, characterized in that: The business approval processing method includes: Receiving a user authorization request, verifying the user authorization request, and obtaining a sub-account credential with read-only permission; Based on the sub-account credentials, real-time data collection is performed on multiple e-commerce platforms to obtain multi-dimensional user data including transaction records, operation data and evaluation information; Inputting the multi-dimensional user data into a preset spatiotemporal sequence attention network and graph convolution hybrid model to dynamically evaluate the user behavior pattern and obtain a risk quantification indicator; Input the risk quantification index into a rule engine that integrates an adaptive decision tree and a genetic algorithm to make an automatic decision and obtain a hierarchical approval resolution; The execution instructions are encrypted with multiple signatures according to the hierarchical approval resolution to obtain security operation instructions and execution status information.

2. The business approval processing method according to claim 1, characterized in that: The receiving a user authorization request, verifying the user authorization request, and obtaining a sub-account credential with read-only permission includes: Receive a user authorization request and obtain platform store identification information, and normalize the platform store identification information to obtain a standardized store identity; Constructing an authorization request data packet based on the standardized store identity identifier, and transmitting the authorization request data packet through an e-commerce platform API interface to obtain platform response data; Perform digital signature verification and authorization scope matching analysis on the platform response data to obtain authorization validity verification results and permission level markings; Determine the authorization status according to the authorization validity verification result, and when the authorization is valid, perform permission configuration on the permission level mark to obtain sub-account configuration information; The sub-account configuration information is written into a security credential database for timestamp marking and access control list setting, to obtain a sub-account credential with read-only permission.

3. The business approval processing method according to claim 1, characterized in that: The real-time data collection of multiple e-commerce platforms based on the sub-account credentials to obtain multi-dimensional user data including transaction records, operation data and evaluation information includes: Constructing an API call parameter set based on the sub-account credentials, and encrypting and encapsulating the API call parameter set to obtain a platform interface access request package; The platform interface access request packet is distributed to a multi-threaded data engine through a distributed task scheduler, and parallel data extraction operations are performed on multiple e-commerce platforms to obtain original platform data in heterogeneous formats; Performing field mapping and type conversion processing on the original platform data in the heterogeneous format to obtain standardized structured data; Building a multidimensional association index based on the standardized structured data, cross-validating and checking the integrity of the multidimensional association index, and obtaining a user data set updated in real time; Dimension expansion and feature derivation calculation are performed on the real-time updated user data set to obtain multi-dimensional user data including transaction records, operation data and evaluation information.

4. The business approval processing method according to claim 1, characterized in that: The multi-dimensional user data is input into a preset spatiotemporal sequence attention network and graph convolution hybrid model to dynamically evaluate the user behavior pattern to obtain risk quantification indicators, including: Performing time series decomposition and topological structure construction on the multi-dimensional user data, converting transaction records into time series features, mapping related transactions into graph structure relationships, and obtaining a time series feature matrix and a graph relationship adjacency matrix; The temporal feature matrix is ​​input into the multi-head self-attention network layer to extract the time dimension features, and the long-term and short-term dependencies are captured through the position encoding and masking mechanism to obtain the behavior feature vector; The graph relationship adjacency matrix is ​​input into the graph convolutional network layer to extract spatial dimension features, and the node neighborhood information is aggregated through the message passing mechanism to obtain the transaction relationship representation; Input the behavior feature vector and the transaction relationship representation into a hybrid attention layer, perform adaptive weight allocation on the spatiotemporal features, and obtain a user behavior representation vector; The user behavior representation vector is input into the anomaly detection module, the behavior deviation is calculated through a comparative learning framework, and standardized processing is performed in combination with historical statistical distribution to obtain a risk quantification indicator.

5. The business approval processing method according to claim 4 is characterized in that: The multi-dimensional user data is subjected to time series decomposition and topological structure construction processing, transaction records are converted into time series features, related transactions are mapped into graph structure relationships, and a time series feature matrix and a graph relationship adjacency matrix are obtained, including: Sorting and segmenting the transaction records in the multi-dimensional user data to obtain a time-series segmented data set; Calculating the transaction frequency, transaction amount change rate and cumulative transaction volume based on the time series segmented data set, and performing standardized transformation to obtain an initial time series feature set; Performing multi-scale decomposition on the initial time series feature set, extracting time series patterns at different frequencies, and extracting periodic information in combination with Fourier features to obtain a time series feature matrix; Extracting transaction entity nodes and transaction relationship edges based on the transaction records in the multi-dimensional user data, and constructing an initial transaction relationship network based on the transaction entity nodes and the transaction relationship edges; The initial transaction relationship network is clustered, and centrality indexes, structural similarity and connectivity characteristics are calculated to obtain a graph relationship adjacency matrix.

6. The business approval processing method according to claim 5 is characterized in that: The risk quantification index is input into a rule engine that integrates an adaptive decision tree and a genetic algorithm to make an automatic decision, and obtain a hierarchical approval resolution, including: Perform multi-layer threshold division on the risk quantification indicators, build an initial rule set in combination with historical approval data, and obtain initial decision rules including conditional judgment statements and execution actions; Constructing a multi-layer decision tree structure based on the initial decision rule, and performing path optimization and pruning processing on the multi-layer decision tree structure to obtain a basic decision tree model; The node conditions and threshold parameters of the basic decision tree model are encoded into a gene sequence, and a genetic algorithm is used to perform population initialization and fitness evaluation on the gene sequence to obtain an initial population set; Performing crossover mutation and selection operations on the initial population set, verifying the accuracy of the calculation rule through historical cases, and iteratively optimizing the threshold parameters to obtain the optimal decision rule; The risk quantification indicators are subjected to multi-level classification processing according to the optimal decision-making rule to obtain hierarchical approval resolutions including automatic approval, manual review and automatic rejection.

7. The business approval processing method according to claim 1, characterized in that: The execution instruction is encrypted with multiple signatures according to the hierarchical approval resolution to obtain the security operation instruction and execution status information, including: Based on the hierarchical approval resolution, an original instruction data structure including an operation type, execution parameters and a business identifier is generated, and integrity verification is performed on the original instruction data structure to obtain a standardized execution instruction; Input the standardized execution instruction into the authority control module, allocate the required number of signatures and signature authority level according to the operation type and risk level, and obtain the multi-signature requirement configuration; Generate a signature request task based on the multi-signature requirement configuration, distribute the signature request to relevant authority holders through a distributed message queue, and obtain a multi-party signature response set; Verifying the validity and number of signatures of the multi-party signature response set, combining the multi-party private key fragments through a threshold signature algorithm, and obtaining digital signature authentication information; The standardized execution instruction and the digital signature authentication information are encrypted and encapsulated, and the instruction content is protected by an asymmetric encryption algorithm to obtain a secure operation instruction, and the secure operation instruction is distributed and the execution status is tracked to obtain the execution status information of the secure operation instruction.

8. A business approval processing device, characterized in that: Used to execute the business approval processing method according to any one of claims 1 to 7, the business approval processing device comprises: A verification module, used to receive a user authorization request, verify the user authorization request, and obtain a sub-account credential with read-only permission; A collection module, used to collect real-time data from multiple e-commerce platforms based on the sub-account credentials to obtain multi-dimensional user data including transaction records, operation data and evaluation information; A dynamic evaluation module, used to input the multi-dimensional user data into a preset spatiotemporal sequence attention network and graph convolution hybrid model to dynamically evaluate the user behavior pattern and obtain a risk quantification indicator; An automatic decision-making module, used to input the risk quantification index into a rule engine that integrates an adaptive decision tree and a genetic algorithm to make an automatic decision and obtain a hierarchical approval resolution; The encryption module is used to perform multi-signature encryption on the execution instruction according to the hierarchical approval resolution to obtain the security operation instruction and execution status information.

9. An electronic device, characterized in that: The electronic device comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instruction in the memory so that the electronic device executes the business approval processing method according to any one of claims 1 to 7.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the business approval processing method according to any one of claims 1 to 7 is implemented.

Images