An ai-dag pre-experiment pre-checking and automatic reconstruction method based on heterogeneous graph linkage

Abstract

This invention discloses a pre-testing and automatic reconstruction method for AI-DAG experiments based on heterogeneous graph linkage. It abstracts the AI ​​modeling experiment process into a DAG, constructing a static theoretical feature graph and a dynamic topological interaction graph to link the theoretical knowledge domain with the practical topological domain. Isolated logical subgraphs are generated based on the dynamic topological interaction graph. During the topology building phase, incremental topology connection legality local checks are performed, and layered compliance checks are conducted based on branch stability and normalized topology deviation values ​​to identify various topology defects and multi-level risks. Abnormal data streams are intercepted based on the check results, and topology regularization and node parameter optimization are completed in stages. The graph is incrementally updated based on experimental data to optimize risk assessment capabilities. This invention achieves pre-testing risk identification throughout the entire topology building process, intercepting illegal connections and topology errors, and automatically repairing defects, thus improving the standardization and operational efficiency of AI modeling experiment topology building.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence modeling and experimental technology, specifically to a method for pre-experiment verification and automatic reconstruction of AI-DAG based on heterogeneous graph linkage. Background Technology

[0002] Artificial intelligence modeling processes typically employ directed acyclic graphs (DAGs) to organize tasks such as dataset loading, feature processing, operator operations, model training, and performance evaluation. These DAGs are visualized and drag-and-drop, allowing users to customize and build their AI-DAG experimental topologies. While existing AI experimental platforms support graphical editing of topologies, they generally lack comprehensive pre-verification and automated error correction capabilities, exhibiting numerous technical shortcomings.

[0003] First, traditional platforms can only simply verify the data type matching relationship between upstream and downstream nodes, lacking a systematic, layered risk identification mechanism. They cannot intercept illegal data flow connections in real time during topology drag-and-drop construction, and it is difficult to constrain the problem of arbitrary cross-connection of nodes between different standard topologies; they can only detect explicit link errors, and cannot identify implicit operational risks in topology branches by combining historical operation data, nor can they quantify the degree of distortion of the entire topology relative to the standard process, making it easy to miss structural defects such as topology disorder and missing parts.

[0004] Secondly, theoretical constraints and practical topology are disconnected, lacking an effective cross-domain linkage and mapping mechanism. Prior knowledge in areas such as operator adaptation scenarios, reasonable parameter ranges, operator mutual exclusion constraints, and process pre-ordering is mostly statically stored in documents and cannot be automatically linked to the topology nodes built by the user. Constraint verification relies heavily on human experience, is highly subjective and inefficient, and is prone to problems such as parameter out-of-bounds errors and operator combination exclusion within modules. Furthermore, it cannot distinguish between serious errors that disrupt data flow and internal module vulnerabilities that only affect stability, lacking the ability for refined risk classification and control.

[0005] Furthermore, existing solutions have weak capabilities for correcting topology defects. The rectification of topology structure and parameter configurations relies on manual adjustments, lacking automatic standardization based on standard topology templates, illegal element cleanup, and intelligent link completion mechanisms. It also struggles to automatically generate compliant configuration schemes for non-compliant parameters. Submitting defective topologies directly for execution easily leads to task interruptions, experimental errors, and a continuous waste of computing resources, increasing the cost of AI modeling and trial-and-error.

[0006] In addition, most topology verification rules are static and fixed, making it impossible to continuously update the judgment criteria using experimental success and failure data, and unable to adaptively optimize the risk identification accuracy as experimental cases accumulate.

[0007] In summary, existing AI-DAG visualization experimental platforms struggle to achieve automated pre-construction risk management throughout the entire topology building process. Theoretical constraints and user-operated topologies are independent of each other, lacking an effective linkage mechanism. They lack a multi-layered screening architecture that combines incremental real-time continuation verification, branch stability assessment, and global structural compliance verification, making it easy to miss issues such as illegal data flow splicing across topologies, hidden operational defects, and topology errors or omissions. Furthermore, they cannot differentiate between different levels of topology risk and implement differentiated handling, relying primarily on manual rectification for topology defects and parameter misconfigurations. Verification rules cannot be dynamically iterated and optimized based on historical experimental data, and directly running defective topologies results in ineffective consumption of computing resources. Summary of the Invention

[0008] To overcome the shortcomings of the existing technologies, this invention provides a method for pre-experiment verification and automatic reconstruction of AI-DAG based on heterogeneous graph linkage. This method aims to achieve hierarchical pre-experiment risk verification, structural defect identification, and automatic topology correction and optimization during the topology construction process of artificial intelligence modeling experiments. It can intercept various hidden risks such as illegal data stream continuation and topology disorder before experimental operation. Furthermore, based on the theoretical knowledge of heterogeneous graphs and the linkage mechanism of practical data, it dynamically optimizes the topology risk judgment and topology compliance reconstruction effect, effectively improving the standardization, stability, and efficiency of AI modeling experiments.

[0009] The technical solution to achieve the purpose of this invention is: a method for pre-experiment verification and automatic reconstruction of AI-DAG based on heterogeneous spectrum linkage, comprising the following steps:

[0010] S1. The complete experimental process of artificial intelligence modeling, training, parameter optimization, and effect evaluation is abstracted and modeled as a directed acyclic graph structured data flow model; the topological nodes of the directed acyclic graph correspond to various functional modules of AI modeling, and the directed edges of the directed acyclic graph represent the directional flow dependencies of data tensors, model parameters, and feature information;

[0011] S2. Construct mutually independent static theoretical feature maps and dynamic topological interaction maps, and realize the data association between the theoretical knowledge domain and the practical topological domain through a cross-domain feature alignment mechanism;

[0012] The static theoretical feature map is a pre-constructed offline domain knowledge directed graph model, which solidifies the domain prior knowledge, constraint rules and knowledge association relationships, and is independent of the topology of a single experiment and is not dynamically updated.

[0013] The dynamic topology interaction graph is a dynamic directed graph model for the business practice layer. The dynamic topology interaction graph uses the AI ​​modeling function module as the topology function node and has two types of topology directed edges built in: standard topology edges and user-defined topology edges. The dynamic topology interaction graph generates experimental-specific logical subgraphs without physical storage as needed. All topology retrieval, traversal and quantization calculations are completed within the logical subgraphs.

[0014] S3. During the user topology construction process, the prefix topology sequence is extracted in real time based on the experimental dedicated logical subgraph and the parent and child topology branches are obtained. Candidate standard topology structures are screened by matching all standard topology structures in the logical subgraph. Based on the set of compliant successor nodes of the candidate standard topology structures, the legality of the data flow continuation of this topology connection is determined, and the local illegal data flow continuation generated by a single operation is intercepted in real time.

[0015] S4. Based on the statistical data of topology branch operation and the theoretical adaptation weighting factor, calculate the branch topology stability and identify the hidden topology operation risks according to the legal branch stability benchmark.

[0016] Traverse all standard topologies within the logical subgraph, calculate the normalized topology deviation value corresponding to the user topology, select the minimum value as the global optimal structure deviation value, implement topology risk stratification based on a critical threshold of 0.5, and distinguish the types of topology defects.

[0017] S5. Based on the verification results, intercept the data flow of abnormal topologies; combine the set of compliant successor nodes from the same source with the set of theoretical constraint attributes to complete the topology structure compliance data reconstruction, topology operation risk verification and node parameter configuration optimization.

[0018] S6. Based on the results of the topology branch experiment, update the running statistics parameters of the dynamic topology interaction graph in real time, refresh the stability judgment benchmark, and iteratively optimize the accuracy of topology risk judgment and structural deviation quantification.

[0019] Compared with the prior art, the present invention has the following advantages:

[0020] First, a heterogeneous architecture is constructed, consisting of a static theoretical feature graph and a dynamic topological interaction graph. This architecture decouples the theoretical knowledge domain from the practical topological domain and establishes connections through cross-domain semantic alignment. The static theoretical feature graph solidifies prior constraints in the domain, while the dynamic topological interaction graph collects topology building behavior and operational statistics. Based on logical subgraph slicing, it enables isolated computation across multiple experimental scenarios, eliminating the drawbacks of the separation between theoretical constraints and practical topology, and providing unified theoretical support for risk quantification and assessment.

[0021] Second, a three-tiered verification system is established, comprising incremental connection legality verification, local branch stability verification, and global structure compliance verification. Incremental local verification is conducted based on the current change location of the topology, and illegal data flow connections generated by single operations are intercepted in real time using memory-level logical subgraphs. Branch stability is quantified by combining historical operational data to identify hidden operational risks. Global distortion is quantified through topology structure deviation values ​​to distinguish between three types of topology defects: disordered, missing, and mixed.

[0022] Third, a set of compliant successor nodes from the same source is set up to achieve refined constraints on the continuation of the same source. At the same time, the candidate standard topology range is dynamically switched to adapt to different stages of topology construction, reducing misjudgments in short-link scenarios in the early stage of construction. Hierarchical handling is implemented for blocking adaptation risks and performance optimization risks. Structural defects are automatically cleaned up with abnormal links and missing topologies are completed with reference to the optimal standard topology. Parameter-related risks are automatically generated with compliant configuration schemes, realizing differentiated intelligent reconstruction of topology structure and internal parameters.

[0023] Fourth, based on the experimental results, the statistical indicators of the dynamic topology interaction graph are incrementally updated to continuously refresh the stability judgment benchmark and achieve adaptive iterative optimization of risk judgment rules. The entire process completes pre-experiment risk screening before the experiment starts, intercepts various topology defects in advance, and avoids the waste of computing resources caused by invalid experiments; the solution is compatible with visual drag-and-drop DAG construction scenarios, has a high degree of automation, and strong engineering feasibility.

[0024] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. Attached Figure Description

[0025] Figure 1 This is a flowchart of the overall process of the AI-DAG experimental pre-verification and automatic reconstruction method based on heterogeneous graph linkage according to the present invention.

[0026] Figure 2 This is a schematic diagram of the three-layer pre-verification and automatic reconstruction process of the present invention. Detailed Implementation

[0027] A method for pre-experiment verification and automatic reconstruction of AI-DAG based on heterogeneous graph linkage includes the following steps:

[0028] S1, AI-DAG topology structure modeling.

[0029] The complete experimental process of artificial intelligence modeling, training, parameter optimization, and performance evaluation is abstracted and modeled as a directed acyclic graph (DAG) structured data flow model. The DAG topology nodes correspond to various functional modules of general artificial intelligence modeling, such as dataset loading, feature processing, various machine learning operators, deep learning network operators, parameter optimization, loss evaluation, model training, and model evaluation. These modules respectively implement the basic functions of the entire AI modeling process, including dataset reading, feature transformation, operator operations, parameter optimization, loss calculation, model fitting, and model generalization metric detection. The directed edges of the DAG represent the directional flow dependencies between data tensors, model parameters, and feature information. User-defined topological structures and data flow logic are recorded, such as the complete record of topological features and data flow logic formed by user-defined drag-and-drop and connection methods, serving as the structured data foundation for the subsequent construction of two types of heterogeneous graphs.

[0030] S2. Construct two types of heterogeneous maps and complete cross-domain feature alignment.

[0031] This step constructs two types of heterogeneous graphs with independent functions, storage content, structure, and update mechanisms: a static theoretical feature graph and a dynamic topological interaction graph. These two types of graphs achieve data association between the theoretical knowledge domain and the practical topological domain through a cross-domain feature alignment mechanism.

[0032] 2.1 Constructing a static theoretical feature map.

[0033] This graph is a pre-built, domain-specific directed graph model with a fixed underlying knowledge framework. It uses knowledge points from the field of artificial intelligence as feature nodes and is divided into theoretical nodes and practical nodes. All knowledge content is standardized, labeled, and stored offline in the database based on prior knowledge of the domain during the graph initialization phase. It is independent of any single experimental topology and does not dynamically update with the experimental process.

[0034] Among them, theoretical nodes are basic knowledge nodes of the discipline, used to support the underlying principle tracing of topological defects; practical nodes are engineering application knowledge nodes for topology construction, which are structured extraction and summarization by domain experts based on official operator documents, standardized modeling documents, compliant topology cases and historical experience. They are pre-defined offline and marked with fixed category tags, specifically including three categories: functional module knowledge points, general constraint knowledge points, and combination specification knowledge points. Functional module knowledge points include algorithm principles, general parameter adaptation rules, and module applicable scenarios. General constraint knowledge points include topology hierarchical subordinate constraints and general rules for data flow pre-dependency.

[0035] Furthermore, the static theoretical feature map pre-configures standardized constraint labels for all constraint-related knowledge points, which are used as constraint rules for subsequent automated topology verification and risk differentiation. The labels are divided into three categories: data flow matching constraints, module adaptation boundary constraints, and process pre-constraints. Among them, data flow matching constraints represent the data types and tensor dimension adaptation conditions of upstream and downstream node inputs and outputs; module adaptation boundary constraints represent the applicable scenarios of functional modules, input object constraints, operator combination exclusion relationships, and reasonable parameter ranges; and process pre-constraints represent the pre-dependency order constraints between topology nodes.

[0036] The static theoretical feature map contains three types of directed topological edges: pre-dependency edges, hierarchical subordinate edges, and similarity-related edges, forming a complete network of disciplinary knowledge connections. Among them, pre-dependency edges are used to represent the theoretical sequence relationship between knowledge points, hierarchical subordinate edges are used to represent the hierarchical inclusion relationship between knowledge points, and similarity-related edges are used to represent the approximate relationship between the principles and application scopes of knowledge points.

[0037] 2.2 Constructing a dynamic topological interaction graph.

[0038] This graph is a dynamic directed graph model for practical business operations. It uses AI modeling modules as topology nodes to record user topology building behavior, node data flow dependencies, and experimental running status data for each topology branch. It does not store theoretical knowledge or general theoretical constraints and is heterogeneously isolated from the static theoretical feature graph. The dynamic topology interaction graph has two types of directed edges: standard topology edges and user-defined topology edges. Standard topology edges represent the inherent standard data flow continuity constraints of the experimental process; a complete feasible experimental scheme connected by standard topology edges is called a standard topology structure. User-defined topology edges represent personalized topology connections built and dynamically adjusted by the user; an experimental scheme to be verified connected by user-defined topology edges is called a user topology structure. The standard topology structure has two attribute labels for topology nodes: required nodes and optional nodes. Required nodes are core functional modules needed to form a complete experimental closed loop and cannot be omitted; optional nodes are optional modules for experimental optimization and can be legally bypassed without being considered as missing topology structures.

[0039] The dynamic topology interaction graph adopts a unified persistent architecture for the global graph, collecting DAG topology data from each experimental template. Each topology functional node has built-in text metadata, including the functional definition, input / output constraints, and task adaptation description of the corresponding module, which is entered and stored during the AI ​​functional module initialization phase. Each topology edge in the graph is configured with independent structured attribute fields, bound to the unique identifier of the experimental template, the total branch run count, and the branch successful run count. With a single template and a single topology branch as the smallest statistical unit, independent accumulation and physical isolation of topology data from multiple scenarios are achieved.

[0040] The dynamic topology interaction graph construction runtime dynamic slicing mechanism can filter the global graph based on the unique identifier of the experiment template, and dynamically generate memory-level, isolated experiment-specific logical subgraphs. The logical subgraphs are temporary computing views without independent physical storage. All topology searches, branch traversals and quantization calculations are completed independently within the subgraphs, realizing a decoupled architecture of globally unified physical storage and on-demand isolated logical calculations.

[0041] The dynamic topology interaction graph addresses scenarios where multiple standard topologies coexist. Based on complete and feasible standard topology samples within the logical subgraph, it pre-constructs a set of compliant, same-origin successor nodes for each node according to a single standard topology. This set of compliant, same-origin successor nodes represents the set of immediately adjacent downstream nodes that can legally continue the current node within its respective standard topology. The successor sets corresponding to different standard topologies are independent of each other, preventing cross-connection of nodes between different standard topologies and forming a refined topology continuity verification mechanism with same-origin constraints. The same-origin constraint is a topology verification rule that restricts a node to completing data flow continuity only within its own standard topology.

[0042] 2.3 Cross-domain feature alignment mapping.

[0043] The text metadata embedded in the topological functional nodes in the dynamic topological interaction graph is extracted and preprocessed with the knowledge point entity text in the static theoretical feature graph by word segmentation and stop word filtering. Then, a general pre-trained text representation model is used to achieve unified semantic encoding and convert it into equal-dimensional semantic vectors. Based on the obtained semantic vectors, the cosine similarity algorithm is used to complete the cross-domain association mapping between practical topological nodes and theoretical knowledge points, and generate node theoretical adaptation scores.

[0044] The node theoretical adaptation scores are normalized by a maximum-minimum value, mapping all scores to a unified interval to obtain a theoretical adaptation weighting factor with a value range of [0,1]. This factor represents the degree of matching between the topological functional node and the theoretical constraints, providing a theoretical dimension weighting benchmark for subsequent branch topology stability calculations and realizing data association between the theoretical knowledge domain and the practical topology domain. Simultaneously, based on this cross-domain association mapping relationship, the matched knowledge points and constraint labels are collected to generate a static theoretical feature node attribute set corresponding to the topological node, providing theoretical constraint data support for subsequent topology operation risk verification, constraint rule comparison, and parameter optimization.

[0045] S3. Implement continuation validity verification based on incremental prefix topological sequence retrieval using dynamic logical subgraphs.

[0046] During the topology building process, which involves users dragging and dropping nodes, adding nodes, and adjusting connections, the currently completed prefix topology sequence is captured in real time based on the experimentally specific isolated logic subgraph and the user's dynamic topology building sequence. Using the data flow connection corresponding to this topology change as the dividing boundary, the prefix topology sequence is split into consecutive parent-child topology branches. For topology forking scenarios, multiple independent parent-child topology branches can be generated simultaneously. The prefix topology sequence refers to the local topology subgraph formed by the topology starting node to all current end nodes during the user's real-time building process, which may contain multiple parallel branches and forking structures. The parent-child topology branches are upstream and downstream local link segments obtained by dividing the prefix topology sequence along the data flow direction. The upstream segment closer to the topology starting end is the parent topology branch, and the newly built downstream end segment is the child topology branch.

[0047] Traverse all standard topologies within the logical subgraph, comparing the node arrangement, data flow continuity, and branching structure of the standard topologies with those of their parent topologies. If the continuity of a parent topology branch from its upstream starting point to its end corresponds to a continuous link within a standard topology, the branch is considered a successful match; otherwise, the match is considered a failure. If the parent topology branch is a successful match, all successfully matched standard topologies are considered candidate standard topologies. Otherwise, if the parent topology branch does not match any standard topologies, it indicates that the current topology fragment is too short and has not yet formed a locatable standardized link; in this case, all standard topologies within the logical subgraph are considered candidate standard topologies.

[0048] The terminal node of the parent topology branch is taken as the parent node, and the starting node of the child topology branch is taken as the child node. The legality of the data flow continuation of adjacent nodes at the current construction position is judged based on the set of compliant successor nodes of the candidate standard topology structure. If there is no candidate standard topology structure that allows the current child node to be included in the set of compliant successor nodes of the corresponding parent node, the verification is deemed to have failed, and the current segment is marked as having a local data flow connection anomaly. This segment of data flow cannot form a legal continuation, thus realizing the pre-construction structural risk screening of the entire topology construction process.

[0049] The aforementioned incremental connection legality verification only performs local verification on the parent and child topology branches corresponding to this topology change, and is used to intercept illegal data flow connections generated by a single operation; for cross-topology structural defects accumulated from multiple builds, they are identified by subsequent global structural compliance verification.

[0050] S4. Local single-step branch stability verification and global structure compliance verification.

[0051] 4.1 Local single-step branch stability verification.

[0052] For each topology branch built by the user in real time, local single-step branch stability verification is performed. During the user's topology building process, the total run count and successful run count of the corresponding topology branch in the dynamic topology interaction graph are updated in real time. The stability value of the current topology branch to be verified is calculated using the branch topology stability calculation formula. At the same time, the stability of all legal topology branches in the current logical subgraph is counted offline, and their average value is calculated as the stability judgment benchmark for local topology anomaly detection. The stability value of the current single branch is compared with the stability judgment benchmark. If the stability value of the current single branch is less than the stability judgment benchmark, the current single-step topology is determined to have operational risks, triggering a local risk marker to accurately identify hidden operational defects in the topology branch.

[0053] The formula for calculating the stability of the branch topology is:

[0054]

[0055] in, The branch topology stability is represented by a value in the range [0,1]. A larger value indicates higher stability and theoretical compliance of the data flow operation of the corresponding topology branch. The total number of executions of the topology branch is the cumulative number of executions of the branch as counted by the dynamic topology interaction graph, and the value is a non-negative integer. The success count of the successful execution of the topology branch is the number of times the branch has successfully completed the experiment without errors as counted by the dynamic topology interaction graph, and the value is a non-negative integer. The weighting factor, which is theoretically adapted, has a value range of [0,1]. It is generated by the normalization of the aforementioned cross-domain feature alignment mapping, and its value follows the constraint rule: when the node function and constraints completely match the domain theoretical constraints, =1; when the node part adapts to theoretical constraints and there are slight scene deviations, w k Linear mapping based on semantic similarity to the (0,1) interval; when node functions, adaptation logic, and domain theory constraints are completely contradictory. =0.

[0056] 4.2 Global structural compliance verification.

[0057] For any user topology structure built in real time, a global structural compliance check is performed. During the user topology building process, all standard topologies within the current experimental logic subgraph are traversed, and the normalized topology deviation value corresponding to each standard topology structure for the user topology structure is calculated one by one. The minimum value is selected as the global optimal structural deviation value, and the corresponding standard topology structure is the optimal standard topology structure.

[0058] The formula for calculating the normalized topology deviation is as follows:

[0059]

[0060] In the formula, The normalized topology deviation value is in the range of [0,1). The larger the value, the higher the degree of deviation from the standard topology. When the user topology perfectly matches the same source standard topology, the normalized topology deviation value is 0; when the user topology is an empty topology, the normalized topology deviation value is 0.5.

[0061] The total cost of the standard topology represents the inherent structural cost of a single standard topology. It is calculated offline and stored during the experimental template initialization phase. This cost is calculated by using a legal bypass to skip optional nodes and their corresponding directed edges, and its value is greater than 0. The legal bypass refers to a data flow continuation method that allows direct connection to upstream and downstream nodes without accessing optional nodes, based on the set of compliant successor nodes from the same source within the standard topology.

[0062] The total cost of user topology distortion represents the cost of structural distortion relative to the standard topology, including two categories: missing cost and disorder cost. Missing cost corresponds to structural defects caused by missing topology, which includes topological nodes and edges present in the standard topology but not constructed in the user topology. Disorder cost corresponds to disorder defects caused by disordered topology, which includes topological nodes and edges present in the user topology but absent in the standard topology. When the standard topology contains optional nodes or directed edges, the calculation of missing and disorder costs is determined based on whether the user topology contains such optional nodes or directed edges. If it does, the optional node or directed edge is included in the cost calculation; otherwise, it is not.

[0063] The Based on the inherent structure of the topology, the node structure weight is 2, and the link structure weight is 1. The calculation formula is as follows:

[0064] ,

[0065] in, The number of nodes in a standard topology. This represents the number of directed edges in the standard topology.

[0066] The Based on the structural distortion calculation of the user topology relative to the standard topology, the calculation formula is as follows:

[0067] ,

[0068] in, The cost of missing user topology is calculated by weighted summation of the missing nodes and edges. , For the number of missing topology nodes, This represents the number of missing directed edges in the topology. , representing the cost of user topology disorder, is calculated by weighted summation of disordered topology nodes and edges. For the number of disordered topology nodes, This represents the number of directed edges in the disordered topology.

[0069] The inherent critical value of 0.5, corresponding to the normalized topology deviation calculation formula, is selected as the unified judgment threshold to adapt to the special scenario of empty topology, balance the evaluation of the costs of missing and disordered distortions, and perform hierarchical judgment:

[0070] (1) When the deviation value of the global optimal structure is greater than or equal to 0.5, it is in a high-risk state. It is determined that the current topology structure deviates from the standard and there are many disordered or invalid structures. This triggers the global risk mark and completes the global topology compliance pre-judgment.

[0071] (2) When the deviation value of the global optimal structure is less than 0.5 and the cost of user topology disorder is greater than 0, it is in a medium-risk state. It is determined that the current topology has a higher effective construction ratio and there is a risk, triggering a medium-level global risk marker.

[0072] (3) When the deviation value of the global optimal structure is less than 0.5 and the cost of user topology structure disorder is equal to 0, it is in a low-risk state. It is determined that the current topology only has unbuilt structures and no topology errors, and is in the normal building process.

[0073] Under different levels of risk, a larger normalized topology deviation value indicates a greater degree of deviation between the user's topology and the optimal standard topology. Based on the structural distortion cost of the user's topology relative to the standard topology, defect types are classified as follows:

[0074] When the cost of missing user topology is 0 and the cost of disorder is not 0, it is a disorder-type defect.

[0075] When the cost of missing user topology structure is not 0 and the cost of disorder is 0, it is a missing defect.

[0076] When the cost of missing user topology structure is not zero and the cost of disorder is not zero, it is a hybrid defect.

[0077] S5. Based on the hierarchical verification results, perform data stream interception and automatic reconstruction of topology structured data.

[0078] After completing three-layer hierarchical verifications—local continuation legality verification, local single-step branch stability verification, and global structural compliance verification—if a local data flow connection anomaly is detected, or if local or global risk markers exist, a data flow interception mechanism is automatically triggered to block the flow of the current abnormal topology structure to downstream experimental modules, thus avoiding invalid experimental runs and wasted computing resources. Simultaneously, based on the cross-domain feature alignment mapping results and the pre-built set of compliant successor nodes from the same source, standard compliant topology structured data is automatically generated according to the hierarchy of topology defects from topology structure to internal configuration. This enables the correction of topology defects and compliance heuristic guidance. The specific generation process is as follows:

[0079] (1) Topology Compliance Data Reconstruction: This stage focuses on the overall compliance and integrity of the topology structure, without verifying theoretical constraints or operational correctness. When a user's topology structure has defects, all standard topologies within the logical subgraph are traversed, and the standard topology structure corresponding to the global optimal structure deviation value is selected as the reference topology structure. For global disorder defects and mixed defects, the graph data corresponding to illegal custom topology edges and abnormal nodes in the user's topology structure is automatically removed; for missing defects and mixed defects that have completed the cleanup of abnormal nodes and topology edges, the missing graph data is completed by calibrating the topology hierarchy and data flow pre-dependency order according to the reference topology structure. Starting from the initial node of the reference topology structure, the set of compliant successor nodes corresponding to each node in the topology is extracted sequentially. If the successor node of the node belongs to a node in the current topology structure that has not yet established a connection, the existing node in the current topology structure and its corresponding connecting link are selected; otherwise, the branch stability of the topology branch formed by the current node extending to each successor node is obtained, and the downstream node with the best branch stability and its corresponding connecting link are selected. Based on the selected downstream nodes and their corresponding connecting links, the missing node and data flow link graph metadata in the reference topology is automatically extracted to generate regularized topology structured data, thus completing the cleaning and completion of topology nodes and data flow links.

[0080] (2) Topology operation risk verification and configuration optimization data generation: Under the premise that the regularized topology structure meets the standard topology structure, topology operation risk verification and configuration optimization data generation are carried out. Update the local single-step branch stability and risk label of the current topology structure; traverse the topology nodes and directed edges with risk labels, and obtain the knowledge points matching the risk-labeled nodes according to the cross-domain association mapping relationship between the dynamic topology interaction graph and the static theoretical feature graph, and integrate them to obtain the static theoretical feature node attribute set; use the attribute set as the theoretical constraint data for topology correction, and screen the rules with constraint label in the attribute set for verification.

[0081] Each constraint rule is compared, and the violations are automatically categorized into two types: blocking adaptation risks and performance optimization risks. Blocking adaptation risks are caused by violations of data flow matching constraints or pre-process constraints, resulting in issues such as mismatched upstream and downstream data types, incompatible tensor dimensions, or disordered topology pre-order. This will prevent the topology link from starting and running normally, and the current topology is marked as an implicit operational risk, without automatic changes to the topology structure and node internal configurations. Performance optimization risks are caused by violations of module adaptation boundary constraints. While the basic conditions for topology data flow are not disrupted, and the topology structure can flow data normally, there are potential risks such as internal operator combination exclusion conflicts and parameters exceeding theoretically reasonable ranges. While maintaining the original topology structure, violations of parameter configurations are determined based on matching theoretical knowledge, marked, and compliant parameter configuration schemes are generated, updating the internal parameters of the topology nodes. Operator combination exclusion conflicts refer to the simultaneous use of multiple logically mutually exclusive operators within the same functional module, leading to conflicts between internal operation rules.

[0082] S6: Real-time incremental updates of dynamic topological interaction graph.

[0083] After each topology branch experiment is completed, based on the success or failure status of the topology branch, the statistical parameters such as the total running count and successful running count of the corresponding topology branch in the dynamic topology interaction graph are updated; the stability judgment benchmark of each branch is refreshed simultaneously, real topology operation data are continuously accumulated, and the calculation accuracy of topology risk judgment and structural deviation quantification is dynamically optimized.

[0084] Example

[0085] This embodiment provides a method for pre-experiment verification and automatic reconstruction of AI-DAG based on heterogeneous graph linkage, deployed on a visual AI modeling experimental platform, selecting an image classification deep learning experimental topology as a typical application scenario. The standard topology process for this scenario is as follows: dataset loading, image preprocessing, feature enhancement (optional), convolutional feature extraction, fully connected classification, loss calculation, model training, and model evaluation. This standardized experimental process is pre-constructed as a standard topology structure within a logical subgraph. This invention, through a heterogeneous graph linkage mechanism, achieves incremental continuity verification, latent risk identification, structural defect determination, and automatic topology reconstruction during the AI-DAG topology construction process in this scenario. The specific implementation steps are as follows:

[0086] S1, AI-DAG topology structure modeling.

[0087] The complete experimental process of image classification is abstracted into a DAG structured data flow model. DAG topology nodes correspond to various functional modules, and directed edges in the DAG represent the directional dependencies between data tensors, feature information, and model parameters. In this embodiment, users can initially build the topology through a visual drag-and-drop interface, generating an initial topology structure including a dataset loading module, an image preprocessing module, a convolutional feature extraction module, a fully connected classification module, and a loss calculation module, thus completing the initial construction of the custom experimental topology.

[0088] S2. Construct two types of heterogeneous maps and complete cross-domain feature alignment.

[0089] Two types of heterogeneous graphs are constructed, with independent functions, storage content, structure, and update mechanisms: static theoretical feature graphs and dynamic topological interaction graphs. These two types of graphs achieve data association between the theoretical knowledge domain and the practical topological domain through a cross-domain feature alignment mechanism.

[0090] 2.1 Construction of Static Theoretical Feature Map

[0091] The static theoretical feature map is constructed offline based on prior knowledge in the deep learning domain. It includes theoretical nodes and practical nodes, and three types of constraint labels: predefined data flow matching constraints, module adaptation boundary constraints, and process pre-constraints. Among them, the module adaptation boundary constraints clearly define the mutual exclusion relationship of operators and the legal range of parameters. In this embodiment, the two specific related constraints are: the dilated convolution operator and the strided pooling operator in the convolution feature extraction module are mutually exclusive operators, and the legal range of model learning rate is [1e-6, 0.1]. The map forms a complete domain knowledge constraint network through pre-dependent edges, hierarchical dependent edges, and similar association edges.

[0092] 2.2 Constructing a dynamic topological interaction graph

[0093] The dynamic topology interaction graph records the topology structure built by the user and the experimental running data in real time. It generates a dedicated logical subgraph for image classification deep learning experiments through dynamic slicing during runtime. Based on a pre-built standard topology structure, each node is configured with an independent set of compliant successor nodes from the same source. In this embodiment, the dedicated logical subgraph for image classification deep learning experiments in the dynamic topology interaction graph has a standard topology structure. This topology structure includes nodes such as a dataset loading module, an image preprocessing module, a feature enhancement module (optional), a convolutional feature extraction module, a fully connected classification module, a loss calculation module, a model training module, and a model evaluation module. Directed edges include: dataset loading module -> image preprocessing module, image preprocessing module -> feature enhancement (optional) module, feature enhancement (optional) module -> convolutional feature extraction module, convolutional feature extraction module -> fully connected classification module, fully connected classification module -> loss calculation module, loss calculation module -> model training module, and model training module -> metric evaluation module. The legitimate downstream successor nodes of the image preprocessing module only include the feature enhancement module and the convolutional feature extraction module, restricting illegal cross-topology splicing from the topology rule level. Meanwhile, the graph provides real-time statistics on the total running count and successful running count of each topological branch, providing data support for subsequent stable quantitative calculations.

[0094] 2.3 Cross-domain feature alignment mapping

[0095] The metadata of topological nodes and the text of theoretical knowledge points are preprocessed and semantically encoded. A one-to-many association mapping is completed using the cosine similarity algorithm to generate a node theoretical fit score. The score is normalized to obtain a theoretical fit weighting factor with a value range of [0,1], which is used to characterize the degree of matching between the current topological node and the domain theoretical constraints. At the same time, the matching constraint labels are collected to form the theoretical attribute set corresponding to each topological node, providing theoretical support for subsequent risk verification and parameter optimization.

[0096] S3. Implementing continuation validity verification based on incremental prefix topological sequence retrieval using dynamic logical subgraphs.

[0097] During the process of users dynamically building the topology, the prefix topology sequence is divided by taking a single connection change as the dividing boundary to obtain the corresponding parent and child topology branches, and incremental connection verification is completed group by group.

[0098] In this embodiment, if the user directly connects the image preprocessing module downstream to the model evaluation module, it will result in an illegal cross-level connection. In this case, the parent topology branch corresponding to this change is the dataset loading module and the image preprocessing module, and the child topology branch is the model evaluation module. The parent topology branch is matched with the standard topology structure within the logical subgraph. If the match is successful, the image classification standard topology structure is locked as the candidate topology. The image preprocessing module, the terminal node of the parent topology branch, is taken as the parent node. Its set of compliant successor nodes, including the feature enhancement module and the convolutional feature extraction module, is queried. If the set does not contain the model evaluation module, the legal connection condition is not met, and the system determines that the connection verification fails and marks the local data stream connection as abnormal in real time. The deep-seated cross-topology defects accumulated from multiple builds are then uniformly identified by subsequent global verification.

[0099] S4. Local single-step branch stability check and global structural compliance check

[0100] 4.1 Local Single-Step Branch Stability Verification

[0101] For each topology branch built by the user in real time, a local single-step branch stability check is performed. The formula for calculating branch topology stability is:

[0102] ;

[0103] in, The branch topology stability is defined, with a value range of [0,1]; total is the total number of runs of the topology branch; success is the number of successful runs of the topology branch. The weighting factor is theoretically adapted and takes values ​​in the range [0,1].

[0104] In this embodiment, if a certain topology branch runs a total of 10 times and succeeds 5 times, the theoretical adaptation weighting factor is 0.9. Substituting this into the formula, the specific value of the branch stability, 0.45, can be accurately calculated. The average stability of all legal branches in the current logical subgraph is used as the judgment benchmark. If the stability of the current branch is lower than the benchmark value, it is determined that the branch has a hidden operational risk, triggering a local risk marker.

[0105] 4.2 Global structural compliance verification

[0106] The structural differences between user topology and standard topology are quantified by normalized topology deviation value. The formula for calculating normalized topology deviation value is as follows:

[0107] ;

[0108] ;

[0109] ,

[0110] ,

[0111] ,

[0112] In the formula, This represents the normalized topological deviation value, with a range of [0,1). This represents the total cost of the standard topology. This represents the total cost of user topology distortion. As a consequence of this deficiency, As a result of chaos, The number of nodes in a standard topology. This represents the number of directed edges in a standard topology. For the number of missing nodes, This represents the number of missing directed edges. For the number of disordered nodes, This represents the number of disordered directed edges.

[0113] In this embodiment, the total cost of the standard topology is calculated by legally bypassing the optional node feature enhancement modules and their corresponding directed edges. The number of nodes corresponding to the standard topology... The number of directed edges is 7. =6, the total cost of the standard topology is .

[0114] If the current user topology's node order is: Dataset Loading Module -> Image Preprocessing Module -> Convolutional Feature Extraction Module -> Fully Connected Classification Module -> Loss Calculation Module, then the included user topology nodes are: Dataset Loading Module, Image Preprocessing Module, Convolutional Feature Extraction Module, Fully Connected Classification Module, and Loss Calculation Module; the included user topology directed edges are: Dataset Loading Module -> Image Preprocessing Module, Image Preprocessing Module -> Convolutional Feature Extraction Module, Convolutional Feature Extraction Module -> Fully Connected Classification Module, Fully Connected Classification Module -> Loss Calculation Module; the missing topology nodes are: Model Training Module and Model Evaluation Module; the missing topology directed edges are: Loss Calculation Module -> Model Training Module and Model Training -> Model Evaluation Module. Optional feature enhancement modules are legally skipped, and these nodes and their corresponding directed edges are not included in the calculation of the total cost of user topology distortion. The number of missing topology nodes corresponding to the user topology is also considered. =2, number of missing directed edges in the topology =2, missing cost =2*2+2=6; Number of disordered topology nodes =0, number of directed edges in disordered topology =0, cost of disorder =0; Total cost of user topology distortion =6+0=6. Normalized topological deviation value. =6 / (20+6)=0.23. The inherent critical value of 0.5, corresponding to the normalized topology deviation value calculation formula, is selected as the unified judgment threshold to achieve risk classification, and the defect type is distinguished based on the missing cost and the disorder cost. When the normalized topology deviation value corresponding to the current user's topology is 0.23, which is less than 0.5 and the user's topology disorder cost is equal to 0, it is in a low-risk state. It is determined that the current topology only has unbuilt structures, no topology errors, and is in the normal building process.

[0115] If the current user topology contains unrelated image compression and clustering analysis modules, and the node order of the user topology is dataset loading module -> image compression module -> convolutional feature extraction module -> clustering analysis module -> loss calculation module, then the included user topology nodes are: dataset loading module, image compression module, convolutional feature extraction module, clustering analysis module, and loss calculation module; the included directed edges are: dataset loading module -> image compression module, image compression module -> convolutional feature extraction module, convolutional feature extraction module -> clustering analysis module, clustering analysis module -> loss calculation module; missing topology node: image preprocessing. Modules: Fully connected classification module, Model training module, Model evaluation module; Missing topological directed edges: Dataset loading module -> Image preprocessing module, Image preprocessing module -> Convolutional feature extraction module, Convolutional feature extraction module -> Fully connected classification module, Fully connected classification module -> Loss calculation module, Loss calculation module -> Model training module, Model training module -> Model evaluation module; Scrambled topological nodes: Image compression module, Cluster analysis module; Scrambled topological directed edges: Dataset loading module -> Image compression module, Image compression module -> Convolutional feature extraction module, Convolutional feature extraction module -> Cluster analysis module, Cluster analysis module -> Loss calculation module. Number of missing nodes corresponding to the user's topology structure. =4, number of missing directed edges =6, cost of missing =2*4+6=14; Number of disordered nodes =2. Number of disordered directed edges =4, Cost of Disorder =2*2+4=8; Total cost of user topology distortion =14+8=22. Normalized topological deviation value. =22 / (20+22)=0.52. The normalized topology deviation value corresponding to the current user topology is 0.52, which is greater than 0.5, indicating a high-risk state. It is determined that the current topology deviates too much overall, with many disordered or invalid structures, triggering a global risk flag and completing the pre-judgment of global topology compliance.

[0116] S5. Perform data stream interception and automatic reconstruction of topology-structured data based on hierarchical verification results.

[0117] Risk interception and differentiated reconstruction are performed based on the three-layer hierarchical verification results, which are divided into two stages: structure regularization and parameter optimization. In this embodiment, the node order of the current user topology is: dataset loading module -> image compression module -> convolutional feature extraction module -> clustering analysis module -> loss calculation module.

[0118] First, topology-compliant data reconstruction. Addressing the aforementioned hybrid defects, the system automatically removes illegal image compression and clustering analysis modules, as well as corresponding redundant custom-defined directed edges, from the user's topology. Simultaneously, referencing the optimal standard topology, it automatically completes the missing image preprocessing modules, fully connected classification modules, model training modules, model evaluation modules, and corresponding directed edges, thus automatically regularizing the topology and eliminating structural errors and omissions.

[0119] Second, operational risk verification and configuration optimization. Risk types are distinguished based on static theoretical constraint labels. For blocking risks such as data flow mismatch and disordered pre-order processing, only risk marking is performed without altering the topology. For performance optimization risks such as operator conflicts and parameter out-of-bounds errors, the configuration is automatically corrected while preserving the topology. In this embodiment, if the convolution feature extraction module simultaneously uses two mutually exclusive operators—dilated convolution and strided pooling—the system identifies it as a performance optimization risk, automatically recommends disabling the conflicting operators, and corrects the parameters to a legal range, eliminating potential operational risks.

[0120] S6, Real-time incremental update of dynamic topology interaction graph

[0121] After the experiment is completed, based on the success or failure status of the topology branches, the statistical parameters such as the total running count and successful running count of the corresponding topology branches in the dynamic topology interaction graph are updated; the stability judgment benchmark of each branch is refreshed simultaneously, real topology operation data is continuously accumulated, and the calculation accuracy of topology risk judgment and structural deviation quantification is dynamically optimized, so that the system verification and reconstruction effect can be continuously improved with the accumulation of experiments.

[0122] As a preferred implementation, this method is deployed on the AI ​​visualization modeling platform server. Front-end drag-and-drop changes trigger S3 incremental continuity verification in real time to achieve immediate error interception; S4 hierarchical risk assessment is triggered when the user actively performs pre-verification; S5 risk interception and automatic topology reconstruction are uniformly executed before the experiment is submitted and run; S6 incremental map update is executed asynchronously after the experiment, forming a complete closed-loop pre-verification and adaptive optimization process.

[0123] Compared to traditional manual verification and static rule verification schemes, this embodiment achieves a deep integration of theoretical constraints and practical topology through heterogeneous graph linkage. It can not only intercept explicit illegal connection errors generated in a single construction in real time, but also quantitatively identify implicit branch operation risks and global structural distortions that traditional methods cannot detect. Through differentiated risk handling and automatic reconstruction mechanisms, it achieves intelligent topology regularization and adaptive parameter optimization, significantly reducing the cost of manual rectification and experimental trial and error. At the same time, it continuously iterates and optimizes the judgment accuracy based on experimental data, effectively improving the standardization, stability and intelligence level of AI-DAG modeling experiments.

[0124] Those skilled in the art will understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for pre-experiment verification and automatic reconstruction of AI-DAG based on heterogeneous graph linkage, characterized in that, Includes the following steps: S1. The complete experimental process of artificial intelligence modeling, training, parameter optimization, and effect evaluation is abstracted and modeled as a directed acyclic graph structured data flow model; the topological nodes of the directed acyclic graph correspond to various functional modules of AI modeling, and the directed edges of the directed acyclic graph represent the directional flow dependencies of data tensors, model parameters, and feature information; S2. Construct mutually independent static theoretical feature maps and dynamic topological interaction maps, and realize the data association between the theoretical knowledge domain and the practical topological domain through a cross-domain feature alignment mechanism; The static theoretical feature map is a pre-constructed offline domain knowledge directed graph model, which solidifies the domain prior knowledge, constraint rules and knowledge association relationships, and is independent of the topology of a single experiment and is not dynamically updated. The dynamic topology interaction graph is a dynamic directed graph model for the business practice layer. The dynamic topology interaction graph uses the AI ​​modeling function module as the topology function node and has two types of topology directed edges built in: standard topology edges and user-defined topology edges. The dynamic topology interaction graph generates experimental-specific logical subgraphs without physical storage as needed. All topology retrieval, traversal and quantization calculations are completed within the logical subgraphs. S3. During the user topology construction process, the prefix topology sequence is extracted in real time based on the experimental dedicated logical subgraph and the parent and child topology branches are obtained. Candidate standard topology structures are screened by matching all standard topology structures in the logical subgraph. Based on the set of compliant successor nodes of the candidate standard topology structures, the legality of the data flow continuation of this topology connection is determined, and the local illegal data flow continuation generated by a single operation is intercepted in real time. S4. Based on the statistical data of topology branch operation and the theoretical adaptation weighting factor, calculate the branch topology stability and identify the hidden topology operation risks according to the legal branch stability benchmark. Traverse all standard topologies within the logical subgraph, calculate the normalized topology deviation value corresponding to the user topology, select the minimum value as the global optimal structure deviation value, implement topology risk stratification based on a critical threshold of 0.5, and distinguish the types of topology defects. S5. Based on the verification results, intercept the data flow of abnormal topologies; combine the set of compliant successor nodes from the same source with the set of theoretical constraint attributes to complete the topology structure compliance data reconstruction, topology operation risk verification and node parameter configuration optimization. S6. Based on the results of the topology branch experiment, update the running statistics parameters of the dynamic topology interaction graph in real time, refresh the stability judgment benchmark, and iteratively optimize the accuracy of topology risk judgment and structural deviation quantification.

2. The AI-DAG experimental pre-verification and automatic reconstruction method based on heterogeneous graph linkage according to claim 1, characterized in that, The static theoretical feature map uses disciplinary knowledge points in the field of artificial intelligence as feature nodes. These feature nodes are divided into theoretical nodes and practical nodes. The theoretical nodes are used for tracing the origins of topological defects, while the practical nodes include functional module knowledge points, general constraint knowledge points, and combination specification knowledge points. The static theoretical feature map is pre-configured with three types of constraint labels: data flow matching constraints, module adaptation boundary constraints, and process pre-constraints. It also includes three types of topological directed edges: pre-dependent edges, hierarchical subordinate edges, and similar related edges, forming a complete disciplinary knowledge association network.

3. The AI-DAG experimental pre-verification and automatic reconstruction method based on heterogeneous graph linkage according to claim 1, characterized in that, A complete and feasible experimental scheme consisting of standard topological edge connections is called a standard topology. The standard topology has two types of attribute labels for topology nodes: required nodes and optional nodes. Among them, the required nodes are the core functional modules required to form a complete experimental closed loop and cannot be omitted; the optional nodes are optional modules for experimental optimization and can be legally bypassed and are not considered as missing topology. An experimental scheme to be verified consisting of user-defined topological edge connections is called a user topology. The dynamic topology interaction graph adopts a global map persistent storage architecture, and the topology edges are bound to the experimental template identifier and branch running statistics to achieve physical isolation of topology data in multiple scenarios.

4. The AI-DAG experimental pre-verification and automatic reconstruction method based on heterogeneous graph linkage according to claim 1, characterized in that, The cross-domain feature alignment mechanism achieves the association mapping between dynamic topological nodes and static theoretical knowledge points through text preprocessing, semantic encoding, and cosine similarity matching. It also generates a theoretical adaptation weighting factor through normalization and collects the theoretical constraint attributes corresponding to the topological nodes to integrate them into a theoretical constraint attribute set.

5. The AI-DAG experimental pre-verification and automatic reconstruction method based on heterogeneous graph linkage according to claim 1, characterized in that, Step S3 specifically includes: Based on the experimental-specific isolation logic subgraph and the user dynamic topology construction sequence, the currently completed prefix topology sequence is extracted in real time; Using the data flow connection of a single topology change as the dividing boundary, the prefix topology sequence is split to obtain upstream and downstream parent and child topology branches; when the parent topology branch successfully matches the standard topology link, the matched standard topology structure is used as the candidate standard topology structure, and when it does not match, all standard topology structures are selected as candidate standard topology structures. Using the end of the parent topology branch of the candidate standard topology structure as the parent node and the beginning of the sub-topology branch as the child node, the system verifies whether the child node belongs to the set of compliant successor nodes of the parent node. If there is no valid matching relationship, the system determines that the local data flow connection is abnormal, thereby achieving real-time pre-emptive risk screening.

6. The AI-DAG experimental pre-verification and automatic reconstruction method based on heterogeneous graph linkage according to claim 1, characterized in that, The formula for calculating the branch topology stability in step S4 is: in, The branch topology stability is represented by a value of [0,1]; success is the count of successful branch executions. To adapt the weighting factors to the theory, .

7. The AI-DAG experimental pre-verification and automatic reconstruction method based on heterogeneous graph linkage according to claim 1, characterized in that, The calculation method for the normalized topology deviation value in step S4 includes: Normalized topology deviation formula: in, This is the normalized topology deviation value, taking values ​​[0,1), with a value of 0.5 in the empty topology state; This represents the total cost of the standard topology. , The number of nodes in a standard topology. This represents the number of directed edges in a standard topology. This represents the total cost of topology distortion for the user. ; As a consequence of this deficiency, As a result of chaos, For the number of missing nodes, This represents the number of missing directed edges. For the number of disordered nodes, This represents the number of disordered directed edges.

8. The AI-DAG experimental pre-verification and automatic reconstruction method based on heterogeneous graph linkage according to claim 1, characterized in that, The specific methods for determining topological risk stratification and distinguishing topological defect types based on a critical threshold of 0.5 are as follows: Topological risk stratification determination: (1) If the deviation of the global optimal structure is ≥0.5, it is judged as high risk and triggers the global risk marker; (2) If the deviation value of the global optimal structure is <0.5 and the disorder cost is >0, it is judged as medium risk and triggers a medium-level global risk marker; (3) The global optimal structure deviation value is <0.5 and the disorder cost is 0, which is judged as low risk and in the normal construction process; Distinguishing between topological defect types: When the cost of missing user topology is 0 and the cost of disorder is not 0, it is a disorder-type defect. When the cost of missing user topology structure is not 0 and the cost of disorder is 0, it is a missing defect. When the cost of missing user topology structure is not zero and the cost of disorder is not zero, it is a hybrid defect.

9. The AI-DAG experimental pre-verification and automatic reconstruction method based on heterogeneous graph linkage according to claim 8, characterized in that, The specific methods for topology-compliant data reconstruction are as follows: Traverse all standard topologies within the logical subgraph and select the standard topology corresponding to the global optimal structure deviation value as the reference topology. For global disorder defects and mixed defects, automatically remove graph data corresponding to illegal custom topology edges and abnormal nodes in the user topology. For missing defects and mixed defects that have completed the cleanup of abnormal nodes and topology edges, calibrate the topology hierarchy and data flow pre-dependency order according to the reference topology to complete the missing graph data. The process of completing missing graph data begins with the initial node of the reference topology and sequentially extracts the set of compliant successor nodes corresponding to each node in the topology. If a node's successor node belongs to a node in the current topology that has not yet established a connection, then the existing node in the current topology and its corresponding connecting link are selected; otherwise, the branch stability of the topology branch formed by the current node extending to each successor node is obtained, and the downstream node with the optimal branch stability and its corresponding connecting link are selected. Based on the selected downstream node and its corresponding connecting link, the missing node and data flow link graph data in the reference topology are automatically extracted, generating the regularized topology structured data, thus completing the cleaning and completion of topology nodes and data flow links.

10. The AI-DAG experimental pre-verification and automatic reconstruction method based on heterogeneous graph linkage according to claim 1, characterized in that, Topology operation risk verification and node parameter configuration optimization include: Based on the theoretical constraint attribute set of cross-domain mapping, we distinguish between blocking adaptation risks and performance optimization risks; for blocking risks, we only mark them without changing the topology structure, and for performance optimization risks, we generate compliant parameter schemes and update node configurations. The specific method for distinguishing between blocking adaptation risks and performance optimization risks is as follows: Update the local single-step branch stability and risk label of the current topology; traverse the topology nodes and directed edges with risk labels, and obtain the knowledge points that match the risk-labeled nodes based on the cross-domain association mapping relationship between the dynamic topology interaction graph and the static theoretical feature graph, and integrate them to obtain the attribute set of the static theoretical feature nodes; The static theoretical feature node attribute set is used as the theoretical constraint data for topology correction, and the rules carrying constraint identification labels in the attribute set are screened for verification. Each constraint rule is compared and automatically categorized into two types based on the type of violation: blocking adaptation risks and performance optimization risks. The aforementioned blocking adaptation risk is caused by violations of data flow matching constraints or process pre-constraints, resulting in issues such as data type or tensor dimension mismatch or topological order disorder; the aforementioned performance optimization risk is caused by violations of module adaptation boundary constraints, including operator combination exclusion conflicts and potential risks of parameters exceeding the theoretically reasonable range.

Images