Cloud platform access path optimization method based on graph convolutional network

Abstract

The invention discloses a cloud platform access path optimization method based on a graph convolutional network, and the method comprises the following steps: S1, collecting user access behavior data and resource state data in a cloud platform, and constructing an original access data set; s2, constructing an access path graph; s3, graph structure features, access behavior features and resource state features are extracted, and node input vectors are generated; s4, inputting to an improved GraphSAGE model, and outputting a path scoring result; s5, identifying bottleneck nodes and high-risk path segments; s6, constructing an optimized path candidate set, and generating an optimal access path combination; and S7, collecting an actual access result of the user to construct access feedback data, updating graph edge attributes and graph nodes, and adjusting aggregation parameters and scoring weights to realize adaptive updating of the path scoring network. The resource scheduling efficiency of the cloud platform can be effectively improved, the access conflict risk is reduced, and the system response performance is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence and cloud computing, and particularly relates to a cloud platform access path optimization method based on a graph convolution network. BACKGROUND

[0002] With the popularity of cloud computing services and the rapid growth of multi-tenant applications, access path optimization for user requests has become a key task to ensure service quality and resource efficiency. The current cloud platform generally uses a path selection method based on static strategies or heuristic rules, which usually selects paths based on pre-defined resource load thresholds, network delays or service accessibility indicators. However, in a complex and heterogeneous cloud platform environment, resource states change frequently and the dependency structure between nodes is complex, which leads to problems such as path congestion, increased response time and uneven resource load in the path optimization method based on static scoring mechanism when facing high concurrency requests or resource fluctuation scenarios.

[0003] The existing method often ignores the topology structure information between the resource nodes of the cloud platform in the path modeling process, lacks sufficient modeling ability for the correlation characteristics between nodes, and is difficult to capture potential resource bottlenecks and access conflict relationships. In addition, the processing of resource state information in the path scoring process mostly adopts a simple weighted superposition or linear fusion method, which lacks a modeling mechanism for the complex nonlinear relationship between multi-dimensional state scores, resulting in poor stability and weak discrimination ability of the final path score result in the case of uneven resource distribution or dramatic connection state fluctuations. For the screening process of the optimized path candidate set, the traditional method also fails to introduce a structured semantic or global scoring mechanism for effective judgment, lacks fine screening and sorting ability, and is difficult to dynamically adapt to the access requirements of service types and target resources.

[0004] Therefore, how to provide a cloud platform access path optimization method based on a graph convolution network is a problem that those skilled in the art need to solve. SUMMARY

[0005] One purpose of the present application is to propose a cloud platform access path optimization method based on graph convolution network, which combines the access record relationship and service call relationship between resource nodes to construct an access path atlas, extracts the graph structure features, access behavior features and resource state features of the resource nodes, inputs them into an improved GraphSAGE model, performs multiple rounds of graph convolution and feature aggregation operations, and outputs the path score results; based on the path score, the bottleneck nodes and high-risk path segments are identified, an optimized path candidate set is constructed, and the access path combination with the highest path score is selected in combination with the user access target and resource state information to generate the optimal access path combination; by collecting user access feedback data, the access path atlas and the graph convolution network model parameters are dynamically updated to form a self-adaptive optimization closed loop. The present application can effectively improve the access path optimization capability and service response efficiency of the cloud platform in a complex load environment.

[0006] According to an embodiment of the present application, a cloud platform access path optimization method based on graph convolution network comprises the following steps: S1, collect user access behavior data and resource state data in the cloud platform, extract an access record relationship set, a service call relationship set and a resource node set, and construct an original access data set; S2, construct an access path atlas based on the original access data set; the access path atlas takes resource nodes as graph nodes and the connection relationship in the access record relationship set and the service call relationship set as graph edges, and the graph edge attributes include access frequency, response delay and failure times; S3, extract graph structure features, access behavior features and resource state features of the resource nodes in the access path atlas to generate node input vectors; S4, input the access path atlas and the node input vectors into an improved GraphSAGE model to perform multiple rounds of graph convolution and feature aggregation operations to output path score results; the improved GraphSAGE model includes a node feature fusion encoder, a time-aware aggregation module, a dynamic neighbor sampling mechanism and a path score output layer; S5, identify bottleneck nodes and high-risk path segments according to the path score results; if the resource load score of a resource node in the access path is greater than a first preset threshold, the resource node is marked as a bottleneck node; if the conflict score of an access path segment is greater than the average value and the historical failure rate is greater than a second preset threshold, the access path segment is marked as a high-risk path segment; S6, based on the path score results and the identification labels, construct an optimized path candidate set, combine the user access target and resource state information, select the access path combination with the highest path score from the optimized path candidate set, and generate the optimal access path combination; S7, collecting user actual access results to construct access feedback data, updating graph edge attributes and graph nodes, and adjusting aggregation parameters and scoring weights of the improved GraphSAGE model to realize adaptive update of the path scoring network.

[0007] Preferably, S1 is specifically: Setting a data collection period and a data access interface to collect user access behavior data and resource state data in the cloud platform; the user access behavior data includes user identification, access time, access resource identification, access entry, access result, access time consumption, and access path sequence; the resource state data includes resource identification, resource type, service unit where the resource is located, current resource load, resource connection state, and resource response time delay; Based on the user access behavior data, an access record relationship set is constructed, each record in the access record relationship set including user identification, starting resource identification, target resource identification, access time, and access path sequence; Based on the resource state data and the service call log, a service call relationship set is constructed, each relationship in the service call relationship set including starting resource identification, called resource identification, call frequency, average response time, and failure number; From the access record relationship set and the service call relationship set, all appearing resource identifications are extracted to construct a resource node set; the resource node includes resource identification, resource type, service unit, and historical load record; the access record relationship set, the service call relationship set, and the resource node set are unified and arranged into a structured original access data set.

[0008] Preferably, S2 is specifically: According to the resource identification in the resource node set, a graph node set is constructed, each graph node corresponding to a unique resource identification and being associated with resource type, service unit, and historical load record; From the access record relationship set, access relationships between resource nodes are extracted to generate an access relationship edge set, each access relationship edge including starting resource identification, target resource identification, access frequency, average response delay, and failure number; from the service call relationship set, service call relationships between resource nodes are extracted to generate a service call edge set, each service call edge including starting resource identification, called resource identification, call frequency, average response time, and call failure number; the access relationship edge set and the service call edge set are merged to construct a graph edge set, and a de-duplication operation is performed on repeated edges and independent attribute fields of each graph edge are retained; Establish an access path graph structure, which includes a graph node set and a graph edge set. The graph node set includes all resource nodes, and the graph edge set includes access relationship edges and service call edges. Assign three attribute fields to each graph edge: access frequency, response latency, and number of failures. Standardize the edge attribute format and record it as numerical structured data.

[0009] Preferably, S3 specifically includes: For each resource node in the access path graph, graph structure features, access behavior features, and resource status features are extracted. The graph structure features include the node's in-degree, node's out-degree, the average shortest path length between the node and other nodes in the graph, the node's topological level, and the number of adjacent nodes in the graph. The access behavior features include the total number of times the resource node is accessed in the access path, the average access interval, the average response time during access, the number of access failures, and the average dwell time in the access path. The resource status features include the average CPU load, average memory usage, average disk utilization, average response latency, and network connection status corresponding to the resource node. Normalization and missing data completion are performed separately, and a unified vector representation is constructed. The normalized graph structure feature vector, access behavior feature vector, and resource status feature vector are concatenated to generate the node input vector corresponding to the resource node.

[0010] Preferably, the improved GraphSAGE model includes a node feature fusion encoder, a time-aware aggregation module, a dynamic neighbor sampling mechanism, and a path scoring output layer, specifically: A node feature fusion encoder is constructed to perform feature compression and encoding mapping on the node input vector. The node feature fusion encoder includes a linear transformation layer and a nonlinear activation layer. The linear transformation layer performs matrix multiplication on the node input vector to generate a basic encoding vector, and the nonlinear activation layer performs activation function operation on the basic encoding vector to generate a fused feature representation. A time-aware aggregation module is set up, which introduces a time weight vector in each round of graph convolution operation to perform weighted aggregation of neighbor node features. The time weight vector is calculated based on the access time interval, interaction frequency and delay change. The aggregation function includes a weighted average function and a time decay function, and outputs time-aware aggregated features. A dynamic neighbor sampling mechanism is constructed to dynamically adjust the sampling quantity and sampling order based on the access frequency and structural similarity of neighbor nodes. For each target node, a fixed number of nodes are sampled from the set of neighbor nodes according to the access frequency distribution, and neighbor filtering is performed based on the similarity matrix between nodes to generate a set of neighbor features. In each round of graph convolution computation, the fused feature representation is fused with the time-aware aggregated features, and feature concatenation and weight fusion operations are performed to update the node representation; A path scoring output layer is set up, which takes the updated node representation as input to the fully connected scoring layer and calculates the path score value of the connection between nodes. The path scoring output layer includes a weight matrix, a normalization layer and a scoring function, and outputs the path score result.

[0011] Preferably, the path scoring output layer includes a weight matrix, a normalization layer, and a scoring function, and outputs the path scoring result, specifically as follows: Based on the resource node representations generated by each round of graph convolution operations, a node representation set is constructed, where each item in the node representation set corresponds to the final representation vector of a resource node; a linear mapping operation is performed to generate an initial set of score vectors. The initial score vector set is normalized, and the path score values ​​between adjacent resource nodes are scaled proportionally according to the connection order of the resource nodes to generate a normalized score vector set. The path score is calculated based on the connection relationship and score difference between resource node pairs. The scoring function includes the Euclidean distance function, the inner product function and the weighted fusion function. The path score values ​​of resource node connections in all access paths are combined into a path score matrix; each row of the path score matrix corresponds to an access path, and each column corresponds to the path score value of a segment of resource node connection relationship, and the path score result is output.

[0012] Preferably, S5 specifically includes: Based on the resource nodes and connection relationships in the access path graph, the path score value associated with each access path is obtained. The access frequency, average response latency, and resource utilization rate of each resource node in the path are extracted according to the path score value. The access frequency, average response latency, and resource utilization rate are weighted to generate a resource load score. If there is a resource node in the access path with a resource load score greater than a first preset threshold, it is marked as a bottleneck node. For the connection relationship of each path segment in the access path graph, extract the access record relationship and service call relationship between adjacent resource nodes within the path segment, and generate a conflict score and historical failure rate based on the number of access failures and the number of conflict requests; if the conflict score of a path segment is greater than the average conflict score and the historical failure rate is greater than the second preset threshold, it is marked as a high-risk path segment.

[0013] Preferably, the step of constructing an optimized path candidate set based on the path scoring results and identification labels specifically involves: Based on the path score values ​​in the path score matrix and the identification tags of resource nodes in the access path graph, access paths with path score values ​​greater than the average score value are extracted to construct an initial path set; access paths containing bottleneck nodes or high-risk path segments are removed from the initial path set to form a path filtering set. Collect user access target and resource status information. User access target includes target resource identifier and service type. Resource status information includes resource load status, response capability status and connection availability status. Match the end resource node of the path according to the user access target. Match the status of each resource node in the path according to the resource status information. Calculate the path availability score. Perform availability score calculation and path score fusion processing on all access paths in the path filtering set, sort them according to weighted score, extract the top N access paths with the highest weighted score, and construct an optimized path candidate set.

[0014] Preferably, the step of combining user access goals and resource status information to select the access path combination with the highest path score from the optimized path candidate set to generate the optimal access path combination specifically involves: Based on the target resource identifier and service type in the user's access target, extract access paths from the optimized path candidate set that have the same destination resource node as the target resource identifier and match the service type, and construct a target matching path set; For each access path in the target matching path set, extract the resource status information of all resource nodes in the path; the resource status information includes resource load status score, response capability score and connection availability score, and perform weighted fusion according to the set weights to generate a path status score value; The path status score is fused with the path score in the path score matrix to generate a comprehensive score according to a set fusion formula, which includes a weighted average function and a maximum value preservation function. The target matching path set is sorted in descending order according to the comprehensive score, and the top K access paths are extracted to construct the optimal access path combination.

[0015] Preferably, S7 specifically includes: The system collects actual access results of user access path combinations and generates access feedback data. Based on the access feedback data, it updates the graph edge attributes and graph node information in the access path graph. The graph edge attribute updates include incremental updates of the response latency field, failure count field, and access frequency field. The graph node information updates include synchronous updates of resource load records and connection status. Based on the updated graph structure, it recalculates the node input vectors and path score values. Based on the score deviation in the access feedback data, it adjusts the aggregation parameters and score weights of the improved GraphSAGE model.

[0016] The beneficial effects of this invention are: This invention constructs an access path graph and extracts graph structure features, access behavior features, and resource status features of resource nodes. It combines a node feature fusion encoder, a time-aware aggregation module, a dynamic neighbor sampling mechanism, and a path scoring output layer. Based on an improved GraphSAGE model, it performs multiple rounds of graph convolution and feature aggregation to output path scoring results. Building upon these path scores, it identifies bottleneck nodes with resource load scores exceeding a first preset threshold and high-risk path segments with conflict scores exceeding the average and historical failure rates exceeding a second preset threshold. Furthermore, based on the path scoring results and identified labels, combined with user access goals and resource status information, it constructs an optimized path candidate set. The system integrates path score and path status score according to a set fusion formula, outputs a comprehensive score, and selects the access path combination with the highest score to construct the optimal access path combination. Finally, it collects the actual access results of user access path combinations, adjusts aggregation parameters and score weights based on score deviations, and updates the graph structure and scoring mechanism to achieve adaptive optimization of the improved GraphSAGE model. This invention realizes a graph structure representation of resource node status and connection relationships in cloud platform access paths, improves the dynamic adaptability and bottleneck identification capability of the path scoring mechanism, and enhances the availability, robustness, and comprehensive scoring accuracy of access paths during multi-path selection. Ultimately, it achieves the technical effects of intelligent access path optimization and dynamic self-updating of the scoring network. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0018] Fig. 1 This is a flowchart of a cloud platform access path optimization method based on graph convolutional networks proposed in this invention; Fig. 2 This is a schematic diagram of the structure of the improved GraphSAGE model proposed in this invention; Fig. 3 This is a data flow graph for a cloud platform access path optimization method based on graph convolutional networks proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figs. 1-3 A cloud platform access path optimization method based on graph convolutional networks includes the following steps: S1. Collect user access behavior data and resource status data from the cloud platform, extract the access record relationship set, service call relationship set, and resource node set, and construct the original access dataset; S2. Construct an access path graph based on the original access dataset; the access path graph uses resource nodes as graph nodes and the connection relationships in the access record relationship set and the service call relationship set as graph edges, and the graph edge attributes include access frequency, response latency and number of failures; S3. Extract graph structure features, access behavior features, and resource status features from the resource nodes in the access path graph to generate node input vectors; S4. Input the access path graph and the node input vector into the improved GraphSAGE model, perform multi-round graph convolution and feature aggregation operations, and output the path score result; the improved GraphSAGE model includes a node feature fusion encoder, a time-aware aggregation module, a dynamic neighbor sampling mechanism, and a path score output layer; S5. Identify bottleneck nodes and high-risk path segments based on the path scoring results; if the resource load score of a resource node in the access path is greater than the first preset threshold, it is marked as a bottleneck node; if the conflict score of an access path segment is greater than the average value and the historical failure rate is greater than the second preset threshold, it is marked as a high-risk path segment. S6. Based on the path score results and identification tags, construct an optimized path candidate set. Combine user access goals and resource status information, select the access path combination with the highest path score from the optimized path candidate set, and generate the optimal access path combination. S7. Collect actual user access results to construct access feedback data, update graph edge attributes and graph nodes, and adjust the aggregation parameters and scoring weights of the improved GraphSAGE model to achieve adaptive updating of the path scoring network.

[0021] This implementation method collects user access behavior data and resource status data from the cloud platform to construct an original access dataset containing a set of access record relationships, a set of service call relationships, and a set of resource nodes, thus achieving comprehensive integration of multi-source information in the access path. Based on the original access dataset, an access path graph is constructed, with resource nodes as graph nodes and access record relationships and service call relationships as graph edges. The graph edge attributes are set as access frequency, response latency, and number of failures, thus realizing graph structure modeling of resource relationships and status in the access path.

[0022] Furthermore, by extracting the graph structure features, access behavior features, and resource status features of resource nodes, a node input vector is constructed, enabling the path feature representation to possess structural, dynamic, and resource status information. The access path graph and node input vector are then input into an improved GraphSAGE model that includes a node feature fusion encoder, a time-aware aggregation module, a dynamic neighbor sampling mechanism, and a path scoring output layer. This model performs multiple rounds of graph convolution and feature aggregation to output path scoring results, thereby enhancing the dynamic expressiveness and temporal sensitivity of path scoring.

[0023] Furthermore, based on the path scoring results, bottleneck nodes with resource load scores exceeding a first preset threshold and high-risk path segments with conflict scores exceeding the average and historical failure rates exceeding a second preset threshold are identified, achieving precise path bottleneck location and risk identification. Simultaneously, based on the path scoring results and identified tags, combined with user access goals and resource status information, the highest-scoring access path combination is selected from the optimized path candidate set to construct the optimal access path combination, enhancing the robustness and diversity of access path selection. Finally, by collecting actual user access results to construct access feedback data, updating graph edge attributes and graph node information, and adjusting the aggregation parameters and scoring weights of the improved GraphSAGE model, adaptive updates and continuous optimization of the path scoring network are achieved, thereby effectively improving the scoring accuracy, path availability, and access success rate of cloud platform access paths.

[0024] In this embodiment, S1 specifically refers to: Set the data collection cycle and data access interface to collect user access behavior data and resource status data from the cloud platform; the user access behavior data includes user identifier, access time, access resource identifier, access entry point, access result, access time, and access path sequence; the resource status data includes resource identifier, resource type, service unit where the resource is located, current resource load, resource connection status, and resource response latency; A set of access record relationships is constructed based on user access behavior data. Each record in the set of access record relationships includes user identifier, starting resource identifier, target resource identifier, access time, and access path order. A service call relationship set is constructed based on resource status data and service call logs. Each relationship in the service call relationship set includes the starting resource identifier, the called resource identifier, the call frequency, the average response time, and the number of failures. Extract all resource identifiers from the access record relationship set and the service call relationship set to construct a resource node set; the resource node includes resource identifier, resource type, service unit and historical load record; and organize the access record relationship set, the service call relationship set and the resource node set into a structured raw access dataset.

[0025] In this embodiment, S2 specifically refers to: A graph node set is constructed based on the resource identifiers in the resource node set. Each graph node corresponds to a unique resource identifier and is associated with the resource type, service unit, and historical load record. Extract access relationships between resource nodes from the access record relationship set to generate an access relationship edge set. Each access relationship edge includes the starting resource identifier, the target resource identifier, the access frequency, the average response latency, and the number of failures. Extract service call relationships between resource nodes from the service call relationship set to generate a service call edge set. Each service call edge includes the starting resource identifier, the called resource identifier, the call frequency, the average response time, and the number of call failures. Merge the access relationship edge set and the service call edge set to construct a graph edge set. Perform deduplication on duplicate edges and retain the independent attribute fields of each graph edge. Establish an access path graph structure, which includes a graph node set and a graph edge set. The graph node set includes all resource nodes, and the graph edge set includes access relationship edges and service call edges. Assign three attribute fields to each graph edge: access frequency, response latency, and number of failures. Standardize the edge attribute format and record it as numerical structured data.

[0026] In this embodiment, S3 specifically refers to: For each resource node in the access path graph, graph structure features, access behavior features, and resource status features are extracted. The graph structure features include the node's in-degree, node's out-degree, the average shortest path length between the node and other nodes in the graph, the node's topological level, and the number of adjacent nodes in the graph. The access behavior features include the total number of times the resource node is accessed in the access path, the average access interval, the average response time during access, the number of access failures, and the average dwell time in the access path. The resource status features include the average CPU load, average memory usage, average disk utilization, average response latency, and network connection status corresponding to the resource node. Normalization and missing data completion are performed separately, and a unified vector representation is constructed. The normalized graph structure feature vector, access behavior feature vector, and resource status feature vector are concatenated to generate the node input vector corresponding to the resource node.

[0027] In this embodiment, the improved GraphSAGE model includes a node feature fusion encoder, a time-aware aggregation module, a dynamic neighbor sampling mechanism, and a path scoring output layer, specifically: A node feature fusion encoder is constructed to compress and encode the node input vector. The node input vector includes four basic features: resource utilization rate, connection stability, request response time, and average task consumption. A basic encoded vector is generated by performing a linear matrix multiplication operation on the node input vector through a transformation matrix. The basic encoded vector is then input into a nonlinear activation layer, and the ReLU activation function is called to perform element-wise mapping to generate the fused feature representation vector of the node. The linear transformation weights and activation parameters are fitted through a backpropagation training process to enhance the representation ability and discriminative power of the original node features. A time-aware aggregation module is set up to introduce a time weight vector in each round of graph convolution calculation. The time weight vector is constructed based on the access time interval, interaction frequency, and response latency between the target node and its neighboring nodes. The access time interval is processed using an exponential time decay function, and the frequency and latency features are normalized and used as weighting factors. A weighted average function is used to generate time-aware weights. The time-aware weights are then used to perform a weighted summation operation on the fused feature representations of neighboring nodes, and the time-aware aggregated feature vector of the neighbors is output. The aggregation result can enhance the model's attention to recently active nodes and improve its responsiveness to changes in access patterns. A dynamic neighbor sampling mechanism is constructed, which dynamically determines the set of sampling nodes in each round of graph convolution based on the access frequency and structural similarity of neighbor nodes. During the sampling process, probabilistic sampling is first performed from the neighbor nodes according to the access frequency distribution, with a fixed number N. Then, the structural similarity matrix is ​​called to calculate the similarity score between each sampling node and the target node in terms of resource status, connection relationship, and response performance. Neighbor nodes with scores below the average level are removed, and the remaining nodes form the neighbor feature set. This mechanism can avoid the feature redundancy problem caused by fixed neighbor sampling and improve the ability to focus on access hotspots. In each round of graph convolution operation, the node fusion feature representation is fused with the neighbor time-aware aggregated feature. The two types of feature vectors are connected by a feature concatenation operation, and a weighted fusion function is called to perform a linear combination of weights to generate an updated node representation vector. The weighted fusion weights are obtained by backfitting by minimizing the path scoring error to ensure a reasonable proportion of neighbor information and node-owned information in different access scenarios. A path scoring output layer is set up, and the updated node representations in each round are input into the path scoring layer. The scoring layer includes a fully connected weight matrix, normalization processing, and a scoring function. The path scoring function receives any pair of node representation vectors, calculates the access path score value of their connecting edges, and includes a vector inner product function and a normalization function. The output score value is mapped to the interval [0, 1]. The path score value can comprehensively reflect the resource load status, connection stability, and historical access performance of the path, providing a scoring basis for subsequent path optimization.

[0028] This implementation method enhances the ability to identify resource bottlenecks and high-risk segments in access paths through the synergistic effect of node feature fusion, time-aware aggregation, dynamic neighbor sampling, and path scoring output structure, thereby achieving accurate modeling and optimization of complex access paths on cloud platforms.

[0029] In this embodiment, the path scoring output layer includes a weight matrix, a normalization layer, and a scoring function, and outputs the path scoring result, specifically as follows: A set of node representations is constructed based on the resource node representations generated by each round of graph convolution operations. Each item in the set of node representations corresponds to the final representation vector of a resource node. The set of node representations is input into the scoring calculation structure, and a linear mapping operation is performed. The linear mapping operation includes matrix multiplication of the node representation vectors and the scoring weight matrix, and outputting an initial set of scoring vectors. The scoring weight matrix is ​​obtained by fitting the path smoothness and historical performance indicators in the training samples, and can characterize the importance and stability of the connections between node pairs. The initial set of score vectors is normalized. The normalization process includes scaling each score vector of a path segment according to the sum of the scores in its path to generate a set of normalized score vectors. The scaling factor is dynamically adjusted based on the distribution of score values ​​within each access path. The normalization process maintains the relative relationship between scores and improves the comparability and stability of the score results. The path score is calculated based on the connection relationship and score difference between resource node pairs. The scoring function includes the Euclidean distance function, the inner product function, and the weighted fusion function. The Euclidean distance function is used to measure the geometric difference between node pair vectors, the inner product function is used to measure the directional consistency and similarity between nodes, and the weighted fusion function linearly weights and fuses the two types of measurement results according to the structural stability weight. The weight parameters are obtained by fitting the access success rate and resource availability rate in the historical path samples. The path score values ​​of resource node connections in all access paths are combined into a path score matrix. Each row in the path score matrix corresponds to an access path, and each column corresponds to the path score value of a resource node connection relationship. The matrix structure can support high-dimensional parallel optimization and unified modeling of node pair feature representation.

[0030] In this embodiment, S5 specifically refers to: Based on the resource nodes and connection relationships in the access path graph, obtain the path score value associated with each access path; extract the access frequency, average response latency and resource occupancy rate of each resource node in the path according to the path score value. The access frequency is the number of times a resource node is accessed per unit time, the average response latency is the average response time of a resource node to an access request, and the resource occupancy rate is the ratio of the currently accessed resources to the available resources of a resource node. Access frequency, average response latency and resource utilization are normalized separately, and the maximum and minimum scaling methods are used to unify the indicators of each dimension within the set range. The normalized access frequency, average response latency and resource utilization are weighted to generate a resource load score. The weighting process is performed by setting weight coefficients and performing a linear combination operation. The weight coefficients are obtained by fitting a set of labels constructed from the access behavior sample set and resource response records. Traverse each access path in the access path graph and determine whether there is a resource node in the path whose resource load score is greater than the first preset threshold; if so, mark the resource node as a bottleneck node. For the connection relationship of each path segment in the access path graph, extract the access record relationship and service call relationship between adjacent resource nodes within the path segment. The access record relationship includes access time, request identifier and response status, and the service call relationship includes request source, call target and call chain information. The number of access failures for path segments is extracted based on access record relationships. The number of access failures is the cumulative value of requests that were not completed, responses that timed out, or abnormal interruptions. The number of conflicting requests is extracted based on service call relationships. The number of conflicting requests is the number of times multiple access requests contend for the same resource interface per unit time. The number of failed accesses and the number of conflicting requests are linearly combined to generate a conflict score for the path segment. The weight parameters of the linear combination are obtained by fitting an anomaly annotation set constructed from service log samples and performance monitoring records. Calculate the historical failure rate of the path segment, which is the ratio of the number of failed accesses to the total number of accesses; if the conflict score of the path segment is greater than the average conflict score and the historical failure rate is greater than the second preset threshold, then the path segment is marked as a high-risk path segment.

[0031] In this embodiment, the step of constructing an optimized path candidate set based on path scoring results and identification labels specifically involves: Based on the path score value of each access path in the path score matrix and the identification label of the corresponding resource node in the access path graph, access paths with path scores greater than the average path score value are extracted to construct an initial path set; the average path score value is obtained by statistically analyzing all path scores and calculating their arithmetic mean. For each access path in the initial path set, a path structure analysis operation is performed. If the access path contains a resource node marked as a bottleneck node or a connection relationship marked as a high-risk path segment, the corresponding access path is removed from the initial path set to form a path filtering set. The marking status of bottleneck nodes and high-risk path segments is derived from the combined evaluation results of access frequency, response latency, resource utilization, conflict score, and failure rate. Collect user access target information and resource status information. User access target information includes target resource identifier and service type identifier. Resource status information includes resource load status vector, response capability status vector and connection availability status vector. Match the target resource identifier in the user access target information with the endpoint resource node of each access path in the path filtering set. If the match is successful, the path is retained. For each access path in the reserved path, a path availability score is calculated according to the state vector of all resource nodes on the path. The path availability score is obtained by constructing a path resource state matrix and performing a weighted summation operation with a preset availability weight vector. The path availability score is fused with the corresponding path score in the path score matrix to construct a weighted score. The weighted score is the weighted sum of the path score and the availability score. The weight ratio is obtained by fitting the optimal weight combination through historical path performance data. Sort all access paths in the path filtering set from high to low according to their weighted scores, and extract the top N access paths with the highest weighted scores. The number N is either a system preset value or dynamically calculated based on the number of paths and the average score distribution, and construct an optimized path candidate set.

[0032] In this embodiment, the step of combining user access goals and resource status information to select the access path combination with the highest path score from the optimized path candidate set and generating the optimal access path combination specifically involves: Based on the target resource identifier and service type in the user's access target, extract the access path whose endpoint resource node is consistent with the target resource identifier from the optimized path candidate set, and filter the path whose service type matches the service type in the user's access target to construct the target matching path set; For each access path in the target matching path set, all resource nodes in the path are traversed, and the resource load status score, response capability score, and connection availability score of the corresponding resource node are extracted from the resource status information. The resource load status score represents the current task occupancy intensity of the resource, the response capability score represents the number of service requests processed by the resource per unit time, and the connection availability score represents the stability of the connection status between the resource node and its upstream and downstream nodes. The three score values ​​are normalized to the interval [0,1], and then weighted fusion processing is performed according to the preset weight coefficients to generate the path status score value. The weighted fusion process obtains the path status score value by fitting a three-factor weighted average function. The weight coefficients are denoted as W1, W2, and W3, respectively, where the sum of W1, W2, and W3 is 1, which represents the contribution ratio of the three scoring factors. The path status score of each access path is fused with the corresponding path score in the path score matrix. The fusion process includes: setting the fusion function type; if the fusion function is a weighted average function, the path status score and the path score are linearly weighted using a set weight coefficient; if the fusion function is a maximum value preservation function, the maximum value of the two is retained as the fusion result; after fusion, a comprehensive score is generated for each access path, and a comprehensive score list is constructed; the specific selection of the fusion function is determined based on the system's preset configuration parameters, and different fusion functions reflect different strategy preferences, ensuring the flexibility and adjustability of the scoring system; The target matching path set is sorted in descending order based on the comprehensive score list. The top K access paths are extracted as output paths to construct the optimal access path combination. K is the upper limit of the optimal path capacity set by the system, which serves as the candidate path set for multi-path redundancy scheduling or load balancing. The setting of the K value needs to comprehensively consider redundancy, system complexity, and path differences.

[0033] In this embodiment, S7 specifically refers to: Collect the actual access results of users in the access path combination and construct access feedback data; the access feedback data includes access start and end time, actual response time, access result status, exception reason label and resource node response status; Update the graph edge attributes and graph node information in the access path graph based on access feedback data; graph edge attribute updates include incremental updates of response latency, failure count, and access frequency fields, and graph node information updates include overwrite updates of resource node load records and connection status. Based on the updated state of the graph nodes and edges, the node input vectors are recalculated, and the input vector set in the node feature fusion encoder is updated. Based on the updated node input vectors and graph structure, the graph convolution operation is re-executed to generate a new round of node representations and path scoring results. Based on the access anomaly types and scoring deviations recorded in the access feedback data, the scoring weights and aggregation parameters in the improved GraphSAGE model are adjusted. The scoring weight adjustment includes fine-grained tuning of the fusion weight matrix parameters, and the aggregation parameter adjustment includes adjusting the temporal coefficients of the time weight vector and dynamically changing the neighbor sampling density. The graph convolution process is re-executed in the adjusted structure, and the path scoring results are compared and analyzed to record the scoring trend and update the path scoring network structure.

[0034] Example 1: To verify the feasibility of this invention in practice, it was applied to a chip design company. In modern integrated circuit design, the EDA (Electronic Design Automation) toolchain has become a core support for chip design and verification. However, with the parallel operation of multiple tools, the increasing complexity of process collaboration, and the growing heterogeneity of data sources, data conflicts frequently occur within the EDA toolchain, such as design parameter version conflicts, incorrect file call paths, and format conversion failures. These conflicts not only reduce design efficiency but may also lead to logic verification failures, resulting in design rework losses of hundreds of thousands or even millions of dollars. This company is currently in the 28nm process chip placement and routing optimization phase, and the EDA process includes key stages such as file preprocessing, logic synthesis, static timing analysis, layout design, and formal verification. In the past, insufficient collaboration among multiple departments often led to conflicts in design parameters, environmental variables, and intermediate data within the toolchain, frequently causing project delays.

[0035] To address the aforementioned issues, this invention proposes an intelligent data conflict detection method for EDA toolchains based on graph neural networks, and has undergone a three-month implementation test in an actual integrated circuit design department. In this application, the execution flow of the enterprise's internal EDA toolchain is first modeled, extracting elements such as data files, tool versions, design modules, and script call chains to construct an initial graph. Each design module, tool script, or intermediate file is abstracted as a node in the graph, and the dependencies between nodes are represented as graph edges. Subsequently, the node information is input into an improved GraphSAGE model. The model captures the context of nodes through an adjacency sampling structure, calculates the weights of each adjacency using a weight normalization structure, retains important residual information, and adjusts the edge structure of the graph through a dynamic adjacency adjustment structure, making the graph embedding result more consistent with the semantic structure of the actual design flow. Finally, an updated graph embedding vector is output to determine whether potential conflict paths exist.

[0036] During a three-month practical application period, we used the toolchain tasks submitted daily by the EDA engineering team as the data source, with the daily data volume remaining between 3200 and 4100 entries. By introducing the method of this invention, the detection system performs graph structure analysis on task logs, scheduling scripts, and intermediate artifacts during the runtime phase, and marks potential conflict points in real time. Simultaneously, through graph embedding analysis, it automatically locates the most likely conflict path nodes and adjacent nodes that trigger conflicts.

[0037] To further verify the effectiveness of the present invention, we set up two sets of comparative experiments: "traditional manual log screening + keyword rule matching" and the method of the present invention, and evaluated them from three dimensions: conflict detection rate, false alarm rate and repair time.

[0038] Table 1. Comparison of EDA task conflict detection accuracy over three months

[0039] As shown in Table 1, during the three-month application period, the traditional method of manual log checking combined with keyword rule matching maintained a conflict detection accuracy of 73%–75%, with a consistently high number of false positives, resulting in a significant waste of resources on ineffective checks. In contrast, the graph neural network detection method proposed in this invention consistently maintained an accuracy of over 93%, significantly outperforming the manual method. Furthermore, because the model can accurately identify conflict nodes and their associated paths, it avoids repeated and ineffective checks, reducing the average repair time from 46 minutes to less than 21 minutes, saving over 50% of the manpower costs for fault repair.

[0040] Table 2 Comparison of Task Execution Failure Rate and Conflict Delay Handling within Three Months

[0041] Table 2 further demonstrates the systematic improvements brought about by this invention. Using traditional methods, the task execution failure rate consistently remained at a high level of 2.4% to 2.6%, while conflict handling had an average delay of over 6 hours, directly impacting design progress. After introducing this invention, the failure rate dropped to around 1.1%, the delay time decreased to less than 3 hours, and the number of delayed delivery modules decreased from more than ten per month to single digits, effectively ensuring the continuity and stability of the chip design critical path.

[0042] This embodiment constructs an improved GraphSAGE graph neural network structure to accurately model the complex module dependencies and data call paths in the EDA toolchain, achieving automatic detection and intelligent localization of data conflicts. This not only significantly improves detection accuracy and reduces false alarms but also accelerates the problem-solving process, avoiding project delays. Real-world applications fully validate the practicality, stability, and advancement of this invention in the electronic design industry, demonstrating promising prospects for widespread adoption.

[0043] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A cloud platform access path optimization method based on graph convolutional networks, characterized in that, Includes the following steps: S1. Collect user access behavior data and resource status data from the cloud platform, extract the access record relationship set, service call relationship set, and resource node set, and construct the original access dataset; S2. Construct an access path graph based on the original access dataset; the access path graph uses resource nodes as graph nodes and the connection relationships in the access record relationship set and the service call relationship set as graph edges, and the graph edge attributes include access frequency, response latency and number of failures; S3. Extract graph structure features, access behavior features, and resource status features from the resource nodes in the access path graph to generate node input vectors; S4. Input the access path graph and the node input vector into the improved GraphSAGE model, perform multi-round graph convolution and feature aggregation operations, and output the path score result; the improved GraphSAGE model includes a node feature fusion encoder, a time-aware aggregation module, a dynamic neighbor sampling mechanism, and a path score output layer; S5. Identify bottleneck nodes and high-risk path segments based on the path scoring results; if the resource load score of a resource node in the access path is greater than the first preset threshold, it is marked as a bottleneck node; if the conflict score of an access path segment is greater than the average value and the historical failure rate is greater than the second preset threshold, it is marked as a high-risk path segment. S6. Based on the path score results and identification tags, construct an optimized path candidate set. Combine user access goals and resource status information, select the access path combination with the highest path score from the optimized path candidate set, and generate the optimal access path combination. S7. Collect actual user access results to construct access feedback data, update graph edge attributes and graph nodes, and adjust the aggregation parameters and scoring weights of the improved GraphSAGE model to achieve adaptive updating of the path scoring network.

2. The cloud platform access path optimization method based on graph convolutional networks according to claim 1, characterized in that, Specifically, S1 is: Set the data collection cycle and data access interface to collect user access behavior data and resource status data from the cloud platform; the user access behavior data includes user identifier, access time, access resource identifier, access entry point, access result, access time, and access path sequence; the resource status data includes resource identifier, resource type, service unit where the resource is located, current resource load, resource connection status, and resource response latency; A set of access record relationships is constructed based on user access behavior data. Each record in the set of access record relationships includes user identifier, starting resource identifier, target resource identifier, access time, and access path order. A service call relationship set is constructed based on resource status data and service call logs. Each relationship in the service call relationship set includes the starting resource identifier, the called resource identifier, the call frequency, the average response time, and the number of failures. Extract all resource identifiers from the access record relationship set and the service call relationship set to construct a resource node set; the resource node includes resource identifier, resource type, service unit and historical load record; and organize the access record relationship set, the service call relationship set and the resource node set into a structured raw access dataset.

3. The cloud platform access path optimization method based on graph convolutional networks according to claim 1, characterized in that, Specifically, S2 is: A graph node set is constructed based on the resource identifiers in the resource node set. Each graph node corresponds to a unique resource identifier and is associated with the resource type, service unit, and historical load record. Extract access relationships between resource nodes from the access record relationship set to generate an access relationship edge set. Each access relationship edge includes the starting resource identifier, the target resource identifier, the access frequency, the average response latency, and the number of failures. Extract service call relationships between resource nodes from the service call relationship set to generate a service call edge set. Each service call edge includes the starting resource identifier, the called resource identifier, the call frequency, the average response time, and the number of call failures. Merge the access relationship edge set and the service call edge set to construct a graph edge set. Perform deduplication on duplicate edges and retain the independent attribute fields of each graph edge. Establish an access path graph structure, which includes a graph node set and a graph edge set. The graph node set includes all resource nodes, and the graph edge set includes access relationship edges and service call edges. Assign three attribute fields to each graph edge: access frequency, response latency, and number of failures. Standardize the edge attribute format and record it as numerical structured data.

4. The cloud platform access path optimization method based on graph convolutional networks according to claim 1, characterized in that, Specifically, S3 is: For each resource node in the access path graph, graph structure features, access behavior features, and resource status features are extracted. The graph structure features include the node's in-degree, node's out-degree, the average shortest path length between the node and other nodes in the graph, the node's topological level, and the number of adjacent nodes in the graph. The access behavior features include the total number of times the resource node is accessed in the access path, the average access interval, the average response time during the access process, the number of access failures, and the average dwell time in the access path. The resource status characteristics include the average CPU load, average memory utilization, average disk utilization, average response latency, and network connection status corresponding to the resource nodes. Normalization and missing data completion are performed separately, and a unified vector representation is constructed. The normalized graph structure feature vector, access behavior feature vector, and resource status feature vector are concatenated to generate the node input vector corresponding to the resource node.

5. The cloud platform access path optimization method based on graph convolutional networks according to claim 1, characterized in that, The improved GraphSAGE model includes a node feature fusion encoder, a time-aware aggregation module, a dynamic neighbor sampling mechanism, and a path scoring output layer, specifically: A node feature fusion encoder is constructed to perform feature compression and encoding mapping on the node input vector. The node feature fusion encoder includes a linear transformation layer and a nonlinear activation layer. The linear transformation layer performs matrix multiplication on the node input vector to generate a basic encoding vector, and the nonlinear activation layer performs activation function operation on the basic encoding vector to generate a fused feature representation. A time-aware aggregation module is set up to introduce a time weight vector in each round of graph convolution operation to perform weighted aggregation of the features of neighboring nodes. The time weight vector is calculated based on the access time interval, interaction frequency and latency changes. The aggregation function includes a weighted average function and a time decay function, and the output is a time-aware aggregated feature. A dynamic neighbor sampling mechanism is constructed to dynamically adjust the sampling quantity and sampling order based on the access frequency and structural similarity of neighbor nodes; For each target node, a fixed number of nodes are sampled from the neighbor node set according to the access frequency distribution, and a neighbor filtering is performed based on the node similarity matrix to generate a neighbor feature set; In each round of graph convolution computation, the fused feature representation is fused with the time-aware aggregated features, and feature concatenation and weight fusion operations are performed to update the node representation; A path scoring output layer is set up, which takes the updated node representation as input to the fully connected scoring layer and calculates the path score value of the connection between nodes. The path scoring output layer includes a weight matrix, a normalization layer and a scoring function, and outputs the path score result.

6. The cloud platform access path optimization method based on graph convolutional networks according to claim 5, characterized in that, The path scoring output layer includes a weight matrix, a normalization layer, and a scoring function, and outputs the path scoring results, specifically: Based on the resource node representations generated by each round of graph convolution operations, a node representation set is constructed, where each item in the node representation set corresponds to the final representation vector of a resource node; a linear mapping operation is performed to generate an initial set of score vectors. The initial score vector set is normalized, and the path score values ​​between adjacent resource nodes are scaled proportionally according to the connection order of the resource nodes to generate a normalized score vector set. The path score is calculated based on the connection relationship and score difference between resource node pairs. The scoring function includes the Euclidean distance function, the inner product function and the weighted fusion function. The path score values ​​of resource node connections in all access paths are combined into a path score matrix; each row of the path score matrix corresponds to an access path, and each column corresponds to the path score value of a segment of resource node connection relationship, and the path score result is output.

7. The cloud platform access path optimization method based on graph convolutional networks according to claim 1, characterized in that, Specifically, S5 is: Based on the resource nodes and connection relationships in the access path graph, the path score value associated with each access path is obtained. The access frequency, average response latency, and resource utilization rate of each resource node in the path are extracted according to the path score value. The access frequency, average response latency, and resource utilization rate are weighted to generate a resource load score. If there is a resource node in the access path with a resource load score greater than a first preset threshold, it is marked as a bottleneck node. For the connection relationship of each path segment in the access path graph, extract the access record relationship and service call relationship between adjacent resource nodes within the path segment, and generate a conflict score and historical failure rate based on the number of access failures and the number of conflict requests; if the conflict score of a path segment is greater than the average conflict score and the historical failure rate is greater than the second preset threshold, it is marked as a high-risk path segment.

8. The cloud platform access path optimization method based on graph convolutional networks according to claim 1, characterized in that, The process of constructing an optimized path candidate set based on path scoring results and identified labels is as follows: Based on the path score values ​​in the path score matrix and the identification tags of resource nodes in the access path graph, access paths with path score values ​​greater than the average score value are extracted to construct an initial path set; access paths containing bottleneck nodes or high-risk path segments are removed from the initial path set to form a path filtering set. Collect user access target and resource status information. User access target includes target resource identifier and service type. Resource status information includes resource load status, response capability status and connection availability status. Match the end resource node of the path according to the user access target. Match the status of each resource node in the path according to the resource status information. Calculate the path availability score. Perform availability score calculation and path score fusion processing on all access paths in the path filtering set, sort them according to weighted score, extract the top N access paths with the highest weighted score, and construct an optimized path candidate set.

9. The cloud platform access path optimization method based on graph convolutional networks according to claim 1, characterized in that, The step of combining user access goals and resource status information to select the access path combination with the highest path score from the optimized path candidate set and generating the optimal access path combination is as follows: Based on the target resource identifier and service type in the user's access target, extract access paths from the optimized path candidate set that have the same destination resource node as the target resource identifier and match the service type, and construct a target matching path set; For each access path in the target matching path set, extract the resource status information of all resource nodes in the path; the resource status information includes resource load status score, response capability score and connection availability score, and perform weighted fusion according to the set weights to generate a path status score value; The path status score is fused with the path score in the path score matrix to generate a comprehensive score according to a set fusion formula, which includes a weighted average function and a maximum value preservation function. The target matching path set is sorted in descending order according to the comprehensive score, and the top K access paths are extracted to construct the optimal access path combination.

10. The cloud platform access path optimization method based on graph convolutional networks according to claim 1, characterized in that, Specifically, S7 is: The system collects actual access results of user access path combinations and generates access feedback data. Based on the access feedback data, it updates the graph edge attributes and graph node information in the access path graph. The graph edge attribute updates include incremental updates of the response latency field, failure count field, and access frequency field. The graph node information updates include synchronous updates of resource load records and connection status. Based on the updated graph structure, it recalculates the node input vectors and path score values. Based on the score deviation in the access feedback data, it adjusts the aggregation parameters and score weights of the improved GraphSAGE model.

Images