A machine learning based information system integration method

Abstract

The application discloses a kind of information system integration method based on machine learning, the present application relates to machine learning and information system integration field, this method detects whether the field structure and semantic change of unannounced of source system interface occurs, the structure and semantic features of changed field are extracted using the deep map encoder obtained by meta-learning training, candidate mapping is generated and comprehensive confidence is calculated, candidate mapping is translated into first-order logic expression, abstract syntax tree is structured in combination with industry ontology knowledge base, formal verification is carried out using satisfiability module theory solver, for the candidate mapping that does not satisfy preset constraint condition in verification, according to the constraint condition that is violated, penalty feedback value is calculated, and the candidate mapping is regenerated and verified again using reinforcement learning strategy network guidance, form generation, verification, feedback technical closed loop, the present application can be used for the automatic, safe data integration under the scene of implicit change of interface between enterprise-level information systems.

Description

Technical Field

[0001] This invention relates to the field of machine learning and information system integration technology, specifically to a machine learning-based information system integration method. Background Technology

[0002] Traditional technical methods mainly rely on ETL tools and ESB architecture. Integration engineers manually define static field mapping rules in advance according to the interface specifications of the source system and the target system, and hardcode the mapping logic into the integration script. Data extraction, transformation and loading are completed according to fixed rules. This type of solution is highly dependent on the manual input of integration engineers and has high operation and maintenance costs. When the source system interface undergoes unnotified implicit structural changes or semantic changes, the hard-coded mapping logic cannot recognize the new structure, and the integration link will fail to resolve. Manual intervention is required to troubleshoot and repair, and business continuity is difficult to guarantee.

[0003] The closest existing technology to this invention is an automatic interface mapping scheme based on a machine learning model. This scheme consists of an interface document parsing module, a feature vector encoding module, a similarity calculation and matching module, and a mapping code generation module, which are connected in sequence. This scheme calculates the semantic similarity between the source field and the target field through a pre-trained language model or graph neural network, selects the field pair with the highest similarity as the mapping result, and directly generates a data conversion script for integration. This scheme has a certain generalization ability, but its decision-making process is essentially a probability distribution output, and there is no logical correctness verification link independent of the neural network. When faced with new field combinations that have not appeared in the training data, it is easy to output semantically incorrect mappings and execute them directly. Moreover, there is no feedback loop between the machine learning module and the execution link. The model cannot learn and improve autonomously from the cases of verification failure. It can only rely on manual re-labeling of data and offline retraining, resulting in a long iteration cycle. Summary of the Invention

[0004] The technical problem this invention aims to solve is that existing enterprise-level information system integration technologies lack an automated data mapping method that combines dynamic adaptability with logical security when dealing with unannounced changes in field structure and semantics at the source system interface. Specifically, static rule-based integration methods cannot handle dynamic changes in the interface. Once the field name, nesting level, or data type of the source system is implicitly changed, the integration link will fail to resolve the issue, requiring manual intervention for troubleshooting and repair.

[0005] The decision-making process of interface mapping schemes based on machine learning models is essentially a probability distribution output, lacking verifiable logical guarantees. It is prone to semantic errors when faced with new and unseen field combinations. Once such a mapping is directly deployed to the production environment, there is a risk of data inconsistency.

[0006] In existing technologies, there is a lack of feedback loop between the machine learning module and the verification execution stage. The model cannot use verification failure information to self-correct and can only rely on manual re-labeling of data for offline retraining.

[0007] To address the aforementioned technical problems, this invention provides an information system integration method based on machine learning, the method comprising: detecting whether unnotified field structure changes and field semantic changes have occurred at the source system interface;

[0008] When the change is detected, the changed field and its tree-like context structure are input into the deep map encoder trained by meta-learning to obtain the hidden layer embedding vector of the field. The same encoding is performed on each candidate field in the candidate field set of the target system to obtain the corresponding hidden layer embedding vector.

[0009] The semantic similarity between the source field and each candidate field is calculated based on the hidden layer embedding vector. The semantic similarity is then weighted and fused by combining the similarity of field name strings and the consistency of data types to obtain the comprehensive confidence score corresponding to each candidate field. Field pairs with comprehensive confidence scores not lower than a preset threshold are determined as the candidate mapping set.

[0010] Each candidate mapping in the candidate mapping set is translated into a first-order logical predicate expression. Relevant constraint axioms are retrieved from the pre-built industry ontology knowledge base according to the business concepts to which the source field and target field belong. The first-order logical predicate expression is combined with the retrieved constraint axioms to construct an abstract syntax tree.

[0011] The abstract syntax tree is converted into a constraint satisfaction problem formula that can be recognized by the satisfiability modular theory solver, and the satisfiability of the constraint satisfaction problem formula is determined by the satisfiability modular theory solver.

[0012] When the determination result meets the preset constraints, the corresponding data transformation and integration execution code is generated and put into the integration link for execution;

[0013] When the determination result is that the preset constraints are not met, the subset of the violated constraints that caused the non-compliance is output. The penalty feedback value is calculated based on the subset of the violated constraints. The reinforcement learning policy network is used to generate an adjustment action for the candidate mapping scheme based on the penalty feedback value and the current state. The adjusted candidate mapping is obtained and re-verified until the determination result is that the preset constraints are met.

[0014] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0015] 1. This invention trains a deep graph encoder through meta-learning, enabling it to learn general meta-knowledge that aligns structural and semantic features, rather than memorizing static mapping samples from a specific business system. This allows the system to extract the semantic and structural features of the new field and generate candidate mappings in real time, even when there are no labeled samples for that field, thus improving the system's ability to adapt to dynamic interface changes.

[0016] 2. This invention translates candidate mappings into first-order logical expressions and performs formal verification of satisfiability using an industry ontology knowledge base. Only mapping operations that meet preset constraints after logical judgment are allowed. This adds an independent logical correctness verification step before the machine learning model output results are put into production execution, reducing the risk of data consistency problems caused by semantically incorrect mappings being directly executed.

[0017] 3. This invention transforms the violated constraints output in the verification process into a penalty feedback signal for reinforcement learning, which in turn guides the iterative update of the candidate mapping generation model's strategy. The verified mapping samples are then fed back into the meta-learning training task set, enabling the system to automatically adjust candidate mappings and re-verify them when verification fails. This reduces the reliance on manual re-labeling of data for offline retraining and shortens the system's response cycle to interface changes. Attached Figure Description

[0018] Figure 1 This is a diagram illustrating the overall system architecture in an embodiment of the present invention.

[0019] Figure 2 This is a flowchart of the closed-loop algorithm in the embodiments of the present invention;

[0020] Figure 3 This is a diagram showing the structure of the depth map encoder and candidate mapping generation in an embodiment of the present invention. Detailed Implementation

[0021] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0022] The technical solution of the present invention will be further described below with reference to specific embodiments. The information system integration method based on machine learning provided by the present invention can be executed by a computing device with a processor and a memory. The memory stores a computer program that can be loaded and executed by the processor. When the processor executes the program, it implements the various processing steps described in this embodiment.

[0023] The method consists of an interface awareness and feature extraction stage, a neural symbol security verification stage, and a reinforcement learning closed-loop feedback stage working together to achieve a complete technical closed loop of candidate mapping generation, verification, and feedback regeneration.

[0024] I. System Overall Architecture

[0025] The source system interface layer and the target system interface layer correspond to the interface specification documents provided by the third-party business systems to be integrated. The interface parsing and graph structure construction module parses the interface specification documents and transforms information such as field names, field types, field descriptions, and nesting levels into a tree structure graph composed of nodes and edges. Nodes represent interface fields, and edges represent nesting or subordinate relationships between fields. This tree structure graph is input to the meta-learning feature extraction and candidate mapping generation module, where the internal deep graph encoder encodes the nodes in the structure graph to obtain a set of candidate mappings, which is then output to the neural symbol security verification engine. The verification engine consists of a logic translation submodule, an industry ontology knowledge base submodule, an abstract syntax tree construction submodule, and a satisfiability modular theory solver submodule, which work together to formally verify the candidate mappings. Candidate mappings whose verification results meet the preset constraints are output to the integration execution layer, which generates the corresponding data conversion and integration execution code and puts it into actual integration chain operation.

[0026] Candidate mappings whose verification results do not meet the preset constraints, along with the violated constraints, are output to the reinforcement learning closed-loop feedback module. This module, in collaboration with the state construction submodule, policy network submodule, and reward and penalty calculation submodule, adjusts the candidate mapping scheme. The adjusted candidate mapping is then re-input into the verification engine for verification. This process is repeated until the verification results meet the preset constraints, or the number of repeated verifications reaches the preset iteration threshold, at which point the system is transferred to the manual processing channel for fallback processing.

[0027] II. Meta-learning and Zero-Shot Candidate Map Generation

[0028] The deep graph encoder adopts a graph neural network structure, and its training process is implemented using a meta-learning approach. Specifically, multiple meta-training tasks are constructed by sampling from the interface mapping corpus of multiple historical business systems. Each meta-training task corresponds to a set of samples consisting of source interfaces, target interfaces, and known correct mappings. Each meta-training task is divided into a support set for rapid adaptation and a query set for evaluating the adaptation effect according to the grouping method.

[0029] For each sampled meta-training task, based on the current meta-parameters, the task loss is calculated using the support set data of that task, and gradient descent is performed along the gradient direction to obtain the corresponding adaptation parameters for that task. This process is represented as follows:

[0030] ;

[0031] in, This represents the initial shared parameters of the meta-learning model;

[0032] Indicates the task Task-specific adaptation parameters obtained after gradient descent;

[0033] This represents the learning rate used to control the step size of gradient descent;

[0034] Indicates the parameter Operators for calculating gradients;

[0035] Representation Model In the mission The loss function value calculated on the support set.

[0036] Subsequently, the generalization loss of the model after adaptive parameter updates is evaluated using the query set data of each task, and the gradient of the meta-parameters is updated by summing the losses of multiple tasks. This process is represented as follows:

[0037] ;

[0038] in, This represents the parameters of the meta-learning model obtained after iterative optimization.

[0039] This represents the parameter value that minimizes the summation term;

[0040] Indicates the distribution of historical tasks. This represents the first sample obtained from this distribution. One training task;

[0041] Indicates the use of tasks The query set data is used to update the model with adapted parameters. The calculated loss function value is obtained through the above iterative training. These are the meta-knowledge model parameters possessed by the depth map encoder, which enable the alignment of structural and semantic features.

[0042] When unnotified field structure and semantic changes are detected at the source system interface, the changed fields and their tree-structured context are input to the system. For a depth map encoder with parameters, output the hidden embedding vector of this field. ;

[0043] Simultaneously, each candidate field in the candidate field set of the target system is also encoded to obtain the corresponding hidden layer embedding vector. The semantic similarity between the source field and the candidate field is calculated as follows:

[0044] ;

[0045] in, Indicates source field With candidate fields Semantic similarity score between them;

[0046] , Representing fields respectively , Hidden layer embedding vectors output by the depth map encoder;

[0047] This represents the inner product operation of two vectors;

[0048] , Representing vectors respectively , The Euclidean norm is used to weight and fuse semantic similarity with field name string similarity and data type consistency features to obtain a comprehensive confidence score, and the comprehensive confidence score is not lower than a preset threshold. The field pairs are included in the candidate mapping set, which is represented as:

[0049] ;

[0050] in, This represents the set of candidate maps for the final output.

[0051] This indicates the pre-set confidence threshold for filtering. Sort the candidate mappings by overall confidence level and output the top-ranked ones for processing by the neural symbolic security verification engine.

[0052] III. Neural Symbol Security Verification

[0053] The neural symbol security verification engine receives a set of candidate maps. Then, the logic translation submodule maps each of the candidate lines. Translate into first-order logical predicate expressions, formally describe the field type assertions, value range assertions, and business relationship assertions between fields implicit in the mapping, and generate a set of mapping logical expressions.

[0054] The industry ontology knowledge base submodule retrieves relevant constraint axioms from the pre-built industry ontology knowledge base based on the business concepts to which the source and target fields belong, denoted as... Each This represents a specific business constraint axiom, including the dimensional consistency axiom, the numerical boundary constraint axiom, and the strong business constraint axiom. The strong business constraint axiom is used to characterize constraints such as the consistency of units before and after currency settlement and the non-nullability and uniqueness of primary key fields.

[0055] The abstract syntax tree construction submodule combines the mapping logic expression with the retrieved relevant constraint axioms to construct a unified abstract syntax tree, which is then converted into a constraint satisfaction problem formula recognizable by the satisfiability modular theory solver submodule, denoted as […]. The expression is:

[0056] ;

[0057] in, Representing candidate mapping schemes A global logical conjunction expression under all relevant constraint axioms;

[0058] Indicates subscript from arrive of Logical conjunction operation is performed using the constraint axioms;

[0059] Representation of mapping scheme In the The result of determining the value under the constraint axiom.

[0060] Satisfiability Modulus Theory Solver Submodule Satisfaction determination: If the determination result is that the preset constraints are met, it means that the candidate mapping does not have logical conflicts under the given constraint axiom system, and can be determined as a safe mapping. The corresponding data transformation and integration execution code is generated by the integration execution layer and put into actual integration link operation.

[0061] If the determination result is that the preset constraints are not met, the solver submodule will output the subset of the violated constraints that caused the non-compliance, block the execution of the candidate mapping, and pass the structured encapsulation of the subset of violated constraints to the reinforcement learning closed-loop feedback module.

[0062] For example, in a cross-system financial business field integration scenario, if a candidate mapping maps the tax-inclusive amount field of the source system to the tax-exclusive amount field of the target system, the logic translation submodule translates the mapping into a field semantic equivalence assertion. The industry ontology knowledge base submodule retrieves the strong business constraint axiom that the tax-inclusive amount and the tax-exclusive amount cannot be directly mapped equally. The abstract syntax tree construction submodule combines the above assertion with the constraint axiom. The satisfiability module theory solver submodule determines that the constraint satisfies the problem formula but does not satisfy the preset constraint conditions, and then outputs the violated constraint conditions to prevent the erroneous mapping from being put into production.

[0063] IV. Reinforce the adaptive feedback loop of learning.

[0064] For candidate mappings whose judgment result is that they do not meet the preset constraints, the state construction submodule encodes the source field embedding vector, target field embedding vector, historical mapping attempt records, and the previous satisfiability judgment result of the current interface mapping task into the state of the reinforcement learning environment. The policy network submodule is based on the current policy. In state Downsampling to obtain action This action represents one or more combinations of specific adjustment strategies for the candidate mapping scheme, such as changing the target candidate field, inserting data type conversion operators, and splitting nested paths.

[0065] The adjusted candidate mappings are resubmitted to the neural symbol security verification engine for verification. The reward and penalty calculation submodule calculates the reward value based on the verification results.

[0066] If the satisfiability determination result is that the preset constraints are met, then the reward value... ,in This represents a fixed positive reward value given upon successful verification.

[0067] If the satisfiability determination result is that the preset constraints are not met, then the reward value... ,in This represents the base penalty coefficient when the verification fails. Representation of mapping scheme The weighted metric of the number and severity of violated constraints. The larger the value, the more significant the deviation of the mapping scheme from the constraints, and the greater the corresponding penalty.

[0068] The parameter update of the policy network submodule uses the near-end policy optimization algorithm to construct the pruning objective function:

[0069] ;

[0070] in, Represents the objective function of a policy with a pruning mechanism;

[0071] Indicates time step Take the expected value;

[0072] This indicates the difference in state between the updated policy and the policy before the update. Select action The probability ratio, where The current strategy to be updated. The strategy before the update;

[0073] Indicates time step The advantage function estimate is used to measure the relative benefit of the current action relative to the average level;

[0074] This means cropping the probability ratio to an interval. Inside, The preset pruning amplitude parameter is used to limit the amplitude of a single policy update to ensure the stability of the training process. For parameters The gradient direction is updated so that the policy network submodule is more likely to generate adjustment actions that satisfy the constraints in subsequent trials.

[0075] Furthermore, candidate mapping samples that have been adjusted by the policy network submodule and whose re-verification results meet the preset constraints are added to the meta-training task set of the depth map encoder as incremental training samples for iterative training and updating of the depth map encoder. This achieves continuous closed-loop adaptive evolution of candidate generation, symbol verification, reinforcement feedback and meta-knowledge update until the verification pass rate reaches the preset convergence standard for multiple consecutive rounds, or reaches the preset iteration number threshold, after which it is transferred to the manual processing channel.

[0076] V. Alternative Implementation Methods

[0077] In the meta-learning candidate generation stage, in addition to the aforementioned gradient descent-based meta-learning algorithm, the training of the depth map encoder can also employ a prototype network algorithm based on metric learning or a gradient-based meta-learning algorithm based on first-order approximation to optimize parameters. This also enables the rapid generalization ability to new tasks through training on a large number of historical tasks, achieving the technical effect of zero-shot candidate mapping generation, which is an equivalent implementation of the meta-learning stage of this invention.

[0078] In the neural symbol security verification stage, in addition to using the satisfiability modular theory solver submodule, a formal reasoning engine based on logic programming or a rule engine combined with a satisfiability check library can be used to perform equivalent dimensional consistency, numerical boundary constraints and strong business constraints verification on candidate mappings, achieving the same deterministic security verification technology effect, which is an equivalent implementation method of the verification stage of this invention.

[0079] It should be noted that each processing step of the method described in this invention can be executed by a processor in a general-purpose computing device after reading the program instructions stored in the memory, or it can be implemented by hardware circuits such as application-specific integrated circuits or field-programmable logic devices. This invention does not limit this.

[0080] The computer-readable storage medium storing the above program instructions includes, but is not limited to, non-transitory storage media such as disk storage, optical disk storage, and semiconductor storage. The above embodiments are only some implementations of the present invention. All equivalent substitutions or improvements made on the basis of the technical solutions of the present invention should fall within the protection scope of the present invention.

[0081] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. A machine learning-based information system integration method, characterized in that, include: Periodically obtain the source system interface specification document, parse it into a field structure tree and extract the field semantic tags, and compare it with the field structure tree and semantic tags corresponding to the archived historical interface specification document. When the comparison results show that nodes have been added, nodes have been deleted or field semantic tags have changed, it is determined that the source system interface has undergone unnotified field structure changes and field semantic changes. When the change is detected, the changed field and its tree-like context structure are input into the deep map encoder trained by meta-learning to obtain the hidden layer embedding vector of the field. The same encoding is performed on each candidate field in the candidate field set of the target system to obtain the corresponding hidden layer embedding vector. The semantic similarity between the source field and each candidate field is calculated based on the hidden layer embedding vector. The semantic similarity is then weighted and fused by combining the similarity of field name strings and the consistency of data types to obtain the comprehensive confidence score corresponding to each candidate field. Field pairs with comprehensive confidence scores not lower than a preset threshold are determined as the candidate mapping set. Each candidate mapping in the candidate mapping set is translated into a first-order logical predicate expression. Based on the business concepts to which the source field and the target field belong, relevant constraint axioms are retrieved from the pre-built industry ontology knowledge base. The first-order logical predicate expression and the retrieved constraint axioms are combined to construct an abstract syntax tree. The abstract syntax tree is converted into a constraint satisfaction problem formula that can be recognized by the satisfiability modular theory solver, and the satisfiability of the constraint satisfaction problem formula is determined by the satisfiability modular theory solver. When the determination result meets the preset constraints, the corresponding data transformation and integration execution code is generated and put into the integration link for execution; When the determination result is that the preset constraints are not met, the subset of the violated constraints that caused the non-compliance is output. The penalty feedback value is calculated based on the subset of the violated constraints. The reinforcement learning policy network generates an adjustment action for the candidate mapping scheme based on the penalty feedback value and the current state to obtain the adjusted candidate mapping. The adjusted candidate mapping is re-translated into a first-order logical predicate expression and the verification process is repeated until the determination result is that the preset constraints are met. When the number of times the verification process is repeated reaches a preset threshold for the number of iterations, the process is switched to manual processing.

2. The method according to claim 1, characterized in that, The depth map encoder is trained as follows: multiple meta-training tasks are constructed by sampling from the interface mapping corpus of multiple historical business systems. Each meta-training task includes a source interface, a target interface, and samples consisting of known correct mappings. Each meta-training task is divided into a support set and a query set. For each meta-training task, the task loss is calculated based on the current meta-parameters and using the support set of the meta-training task. The adaptation parameters corresponding to the meta-training task are then updated along the gradient descent direction. Using the query set data of each of the meta-training tasks, the generalization loss of the model after the adaptation parameter update is calculated, and the generalization losses of multiple meta-training tasks are aggregated to perform gradient update on the current meta-parameters. The meta-parameters obtained after iterative update are determined as the model parameters of the depth map encoder.

3. The method according to claim 1, characterized in that, Calculating the semantic similarity between the source field and the candidate field includes: calculating the cosine similarity between the hidden embedding vector of the source field and the hidden embedding vector of the candidate field, and determining the cosine similarity as the semantic similarity between the source field and the candidate field.

4. The method according to claim 1, characterized in that, The constraint axioms stored in the industry ontology knowledge base include the dimensional consistency axiom, the numerical boundary constraint axiom, and the business strong constraint axiom. The business strong constraint axiom is used to characterize the consistency constraint of currency settlement and the uniqueness constraint of primary key.

5. The method according to claim 1, characterized in that, The adjustment actions for generating candidate mapping schemes include: encoding the source field embedding vector, target field embedding vector, historical mapping attempt records, and the previous satisfiability determination result of the current interface mapping task into the state of the reinforcement learning environment; the reinforcement learning policy network outputs adjustment actions for the candidate mapping schemes based on the state, such as changing the target candidate field, inserting data type conversion operators, and splitting nested paths.

6. The method according to claim 1, characterized in that, The calculation of the penalty feedback value includes: when the judgment result is that the preset constraint condition is met, a fixed positive reward value is assigned; when the judgment result is that the preset constraint condition is not met, a weighted metric value is determined according to the number and severity of the violated constraint condition, and the value obtained by multiplying the weighted metric value by the preset penalty coefficient is determined as the penalty feedback value.

7. The method according to claim 1, characterized in that, The policy parameters of the reinforcement learning policy network are updated by a proximal policy optimization algorithm. The proximal policy optimization algorithm constructs a pruning objective function based on the probability ratio of the policy before and after the update to select the same action in the same state, prunes the probability ratio to a preset pruning interval, and updates the policy parameters along the gradient direction of the pruning objective function.

8. The method according to claim 1, characterized in that, Also includes: Candidate mapping samples that have been adjusted by the reinforcement learning policy network and whose re-verification results meet the preset constraints are added to the meta-training task set of the depth map encoder as incremental training samples for iterative training and updating of the depth map encoder.

9. An information system integration device based on machine learning, characterized in that, include: The interface awareness and feature extraction module is used to detect whether there are unnotified field structure changes and field semantic changes in the source system interface. When the changes are detected, the changed field and its tree-like context structure are input into the deep graph encoder trained by meta-learning to obtain the hidden layer embedding vector of the field. The same encoding is performed on each candidate field in the candidate field set of the target system to obtain the corresponding hidden layer embedding vector. The semantic similarity between the source field and each candidate field is calculated based on the hidden layer embedding vector. The semantic similarity is weighted and fused by combining the field name string similarity and data type consistency to obtain the comprehensive confidence score corresponding to each candidate field. Field pairs with comprehensive confidence scores not lower than a preset threshold are determined as candidate mapping sets. The neural symbol security verification module is used to translate each candidate mapping in the candidate mapping set into a first-order logic predicate expression, retrieve relevant constraint axioms from a pre-built industry ontology knowledge base according to the business concepts to which the source field and target field belong, combine the first-order logic predicate expression with the retrieved constraint axioms to construct an abstract syntax tree, convert the abstract syntax tree into a constraint satisfaction problem formula that can be recognized by a satisfiability modular theory solver, and use the satisfiability modular theory solver to determine the satisfiability of the constraint satisfaction problem formula. The integrated execution module is used to generate corresponding data transformation and integrated execution code and put it into the integrated execution link when the judgment result meets the preset constraints. The reinforcement learning closed-loop feedback module is used to output a subset of the violated constraints that cause the non-compliance when the judgment result is that the preset constraints are not met. It calculates a penalty feedback value based on the subset of violated constraints, uses the reinforcement learning policy network to generate an adjustment action for the candidate mapping scheme based on the penalty feedback value and the current state, obtains the adjusted candidate mapping, and outputs the adjusted candidate mapping to the neural symbol security verification module for re-verification until the judgment result is that the preset constraints are met. When the number of repeated verifications reaches the preset iteration threshold, the process is switched to manual processing.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the machine learning-based information system integration method according to any one of claims 1 to 8.

Images