Search model establishment method and system based on heterogeneous graph neural network and application

Abstract

The application discloses a search model establishment method and system based on a heterogeneous graph neural network and application, and belongs to the technical field of artificial intelligence and graph neural networks. The search model establishment method comprises the following steps: collecting meta-paths related to a target node to form an initial meta-path set; generating a meta-path redundancy graph by using meta-paths in the initial meta-path set, which have a dependency higher than a preset threshold; finding an optimal non-adjacent meta-path combination from the meta-path redundancy graph as an initial non-redundant meta-path combination; optimizing the architecture parameters and network weights of the initial non-redundant meta-path combination according to the meta-paths and meta-path strengths in the initial non-redundant meta-path combination to obtain an optimal non-redundant meta-path combination; establishing a meta-path aggregation search space for the optimal non-redundant meta-path combination as a search model; and optimizing and training the search model by using a differentiable modular architecture search strategy. The application improves the performance and efficiency of a heterogeneous graph representation learning task.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and graph neural network technology, and more specifically, relates to a method, system and application for establishing a search model based on heterogeneous graph neural networks. Background Technology

[0002] Heterogeneous graph neural networks (HNNs), as important tools for processing complex heterogeneous information networks, have demonstrated significant value in fields such as social network analysis and knowledge graph reasoning by mining high-order semantic relationships between different types of nodes. The core mechanism of HNNs relies on the structured semantic modeling capability of meta-paths, which can transform multi-hop interactions between heterogeneous nodes into computable graph topological features. However, existing technologies face multi-dimensional challenges in the design and optimization of key components.

[0003] In recommender systems, user and product node types constitute a complex e-commerce network, which can be abstracted as a heterogeneous graph. In this graph, each node has a corresponding type, and each edge has a clearly defined relationship type, fully representing the complex relationships between users, orders, categories, and products in an e-commerce scenario. Heterogeneous graphs contain various complex meta-paths, such as user-order-category-product-user, user-product-user, and user-product-type. This complexity requires recommender systems to handle massive amounts of product information and diverse user behaviors. However, existing technologies are inefficient and computationally expensive in handling this complexity. On one hand, traditional meta-path selection methods, such as random sampling or exhaustive screening, face an exponentially growing meta-path combination space, leading to rapidly increasing computational resource consumption and difficulty in efficiently finding the optimal meta-path combination. On the other hand, a large number of invalid or low-value meta-paths introduce semantic noise, interfering with the accurate extraction of user preferences and product association features. This not only makes it difficult to further improve the performance of recommender systems but also directly affects the accuracy and recall of recommendations, thereby reducing user experience and commercial value.

[0004] Similarly, other fields face similar challenges. Take, for example, a heterogeneous graph composed of four node types: author-paper-term-author. Researchers aim to uncover relationships such as academic collaboration and knowledge transfer. However, current feature aggregation frameworks typically employ static weight allocation to integrate multi-path features. This fixed aggregation strategy ignores the differentiated semantic information requirements of different academic mining tasks (such as academic influence assessment and research direction prediction). In the context of interdisciplinary research integration or the dynamic evolution of academic fields, this rigid architecture is particularly inadequate. It easily leads to the loss of key academic features or the over-amplification of noisy features, making it difficult for the mining results to accurately reflect the true state of academic relationships. This not only hinders the improvement of academic knowledge graphs but also limits the in-depth development of academic research, failing to meet the increasingly complex needs of modern academic research.

[0005] In summary, existing search methods based on heterogeneous graph neural networks suffer from insufficient accuracy and low efficiency. Summary of the Invention

[0006] In view of the shortcomings of the prior art, the purpose of this invention is to provide a method, system and application for establishing a search model based on heterogeneous graph neural networks, so as to solve the problems of insufficient accuracy and low efficiency of existing search methods based on heterogeneous graph neural networks.

[0007] To achieve the above objectives, a first aspect of the present invention provides a method for establishing a search model based on a heterogeneous graph neural network, comprising: collecting meta-paths related to a target node to form an initial meta-path set; generating a meta-path redundancy graph using meta-paths in the initial meta-path set whose pairwise dependencies exceed a preset threshold; finding the optimal combination of non-adjacent meta-paths from the meta-path redundancy graph as an initial non-redundant meta-path combination; optimizing the architecture parameters and network weights of the initial non-redundant meta-path combination based on the meta-paths and their strengths to obtain an optimal non-redundant meta-path combination; establishing a meta-path aggregation search space for the optimal non-redundant meta-path combination as a search model; and optimizing and training the search model using a differentiable modular architecture search strategy.

[0008] Optionally, the dependency is:

[0009] ;

[0010] in, Metapath Harmony Path Dependence between them , Metapath Metapath The set of covered nodes , Metapath Metapath The edge set covered.

[0011] Optionally, the meta-path strength is:

[0012] ;

[0013] in, Metapath Metapath strength, , They are respectively the metapath Metapath Related architectural parameters, This represents the total number of meta-paths in the initial non-redundant meta-path combination. , Metapath Metapath The size of the covered node set, , Metapath Metapath The size of the covered edge set, This refers to the temperature parameter.

[0014] Optionally, the architecture parameters and network weights of the initial non-redundant meta-path combination can be optimized by alternately updating the architecture parameters and network weights until the optimal architecture parameters and network weights are obtained.

[0015] Optionally, the meta-path aggregation search space can realize one or more of message passing, feature aggregation, and meta-path fusion; the message passing uses direct aggregation, average aggregation, and / or multilayer perceptron transformation to pass node information in the meta-path; the feature aggregation uses average pooling, max pooling, and / or attention mechanisms to aggregate the features after message passing; the meta-path fusion uses average pooling, single-head attention mechanisms, and / or multi-head attention mechanisms to fuse the features after feature aggregation, and outputs the meta-path fused features.

[0016] Optionally, the search model can be optimized using a differentiable modular architecture search strategy, specifically including: dividing the meta-path aggregation process in the search model into multiple modules, each module containing a set of candidate operations; using a Softmax weighted supernetwork to transform discrete architecture selection into continuous probabilities, and jointly optimizing the weight parameters and model weights in the search model through gradient descent.

[0017] A second aspect of the present invention provides a search method based on heterogeneous graph neural networks, comprising: using the optimized search model in the method described above to search for comprehensive feature information related to the target node.

[0018] A third aspect of the present invention provides a search model building system based on heterogeneous graph neural networks, comprising: a collection module for collecting meta-paths related to a target node to form an initial meta-path set; a generation and search module for generating a meta-path redundancy graph using meta-paths in the initial meta-path set whose pairwise dependencies are higher than a preset threshold; and searching for the optimal combination of non-adjacent meta-paths from the meta-path redundancy graph as an initial non-redundant meta-path combination; a first optimization module for optimizing the architecture parameters and network weights of the initial non-redundant meta-path combination based on the meta-paths and meta-path strengths in the initial non-redundant meta-path combination to obtain the optimal non-redundant meta-path combination; a building module for building a meta-path aggregation search space for the optimal non-redundant meta-path combination as a search model; and a second optimization module for optimizing and training the search model using a differentiable modular architecture search strategy.

[0019] A fourth aspect of the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described above.

[0020] A fifth aspect of the present invention provides a non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform the method described above.

[0021] Compared with existing technologies, the advantages of this invention include: providing a method for establishing a search model based on heterogeneous graph neural networks; generating a meta-path redundancy graph by utilizing meta-paths in the initial meta-path set whose pairwise dependencies exceed a preset threshold, effectively reducing information redundancy and enabling precise selection of the most suitable meta-path combination; leveraging a differentiable modular architecture search strategy to comprehensively mine the structural and semantic information of heterogeneous graphs, significantly improving the performance and applicability of the search model, thereby improving the accuracy and efficiency of search results; experimental results show that this method outperforms existing state-of-the-art methods, dynamically adjusting the meta-path aggregation strategy to adapt to different task requirements, thus achieving higher accuracy and efficiency in heterogeneous graph representation learning tasks. Attached Figure Description

[0022] Figure 1 A flowchart illustrating the method for establishing a search model based on a heterogeneous graph neural network, as provided in an embodiment of the present invention.

[0023] Figure 2 Implementation provided for embodiments of the present invention Figure 1The system framework diagram of the method is shown.

[0024] Figure 3 A block diagram of the system for establishing a search model based on a heterogeneous graph neural network provided in an embodiment of the present invention.

[0025] Figure 4 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0026] In view of the shortcomings of the prior art, the inventors of this invention, through long-term research and extensive practice, have proposed the technical solution of this invention. The following will further explain and illustrate this technical solution, its implementation process, and its principles.

[0027] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0028] Furthermore, in the description of this invention, it should be understood that the terms "upper," "lower," "inner," "outer," "horizontal," "vertical," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0029] In the description of this specification, the references to terms such as "an embodiment," "a particular embodiment," or "the embodiment" indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0030] This invention provides a method for establishing a search model based on a heterogeneous graph neural network. (See also...) Figure 1 The method for establishing a search model based on heterogeneous graph neural networks includes steps S100-S500.

[0031] Step S100: Collect meta-paths related to the target node to form an initial meta-path set.

[0032] Collect all possible meta-paths related to the target node to form an initial meta-path set. The target node is the node type for which classification and prediction tasks are required. Each meta-path consists of different types of nodes and relationships combined in a certain order, such as a path like author-paper-term-author.

[0033] Step S200: Using the meta-paths in the initial meta-path set whose pairwise dependencies are higher than a preset threshold, generate a meta-path redundancy graph; find the optimal combination of non-adjacent meta-paths from the meta-path redundancy graph as the initial non-redundant meta-path combination.

[0034] For each metapath in the initial metapath set, calculate the set of nodes and edges it covers. For any two metapaths, calculate the dependency between them by comparing the set of nodes and edges they cover. Specifically, calculate the sum of the proportions of the intersection and union of the two metapaths to quantify the dependencies between different metapaths and identify semantically similar or redundant metapaths.

[0035] The initial metapath set can be represented as Preferably, for any two metapaths... and The degree of dependence between them is:

[0036] ;

[0037] in, Metapath Harmony Path Dependence between them , Metapath Metapath The set of covered nodes , Metapath Metapath The edge set covered.

[0038] Preferably, the process of generating a meta-path redundancy graph is as follows: If the meta-path and Dependence between If the value is higher than the preset threshold δ, in the meta-path redundancy graph Add an edge connecting them, marking highly dependent meta-path combinations; use a greedy algorithm to select unselected nodes, add them to the meta-path combination, and delete the node and its adjacent nodes, until all nodes have been processed. This is a set of nodes, representing the set of all meta-paths; Let be the edge set, representing the redundant relationships between metapaths.

[0039] A further preferred approach is to use a greedy algorithm to find the optimal combination of non-adjacent meta-paths from the meta-path redundancy graph, that is, to select a set of meta-paths without direct dependencies as the initial non-redundant meta-path combination, thereby reducing resource consumption in subsequent calculations.

[0040] Step S300: Based on the meta-paths and meta-path strengths in the initial non-redundant meta-path combination, optimize the architecture parameters and network weights of the initial non-redundant meta-path combination to obtain the optimal non-redundant meta-path combination.

[0041] Preferably, the meta-path strength is:

[0042] ;

[0043] in, Metapath Metapath strength, , They are respectively the metapath Metapath Related architectural parameters, This represents the total number of meta-paths in the initial non-redundant meta-path combination. , Metapath Metapath The size of the covered node set, , Metapath Metapath The size of the covered edge set, Temperature parameters. Architecture parameters. .

[0044] Preferably, in one embodiment of the present invention, optimizing the architecture parameters and network weights of the initial non-redundant meta-path combination specifically includes: alternately updating the architecture parameters and network weights until the optimal architecture parameters and network weights are obtained.

[0045] The following example illustrates the specific process of the optimization operation in step S300.

[0046] Using a multilayer perceptron to combine the initial non-redundant metapaths The features of the mid-path are transformed and aggregated to obtain the updated node representation. The update formula is as follows:

[0047] ;

[0048] in, For the target node The update indicates; For projection functions; Represents the projection function The weight parameters in the code are the network weights. Metapath The characteristic matrix of .

[0049] Use an alternating update strategy to optimize architecture parameters and network weight That is, fixing one parameter and optimizing another, alternating between the two until the most suitable network architecture is found. This process can be represented as:

[0050] ;

[0051] ;

[0052] in, and Let represent the training set loss and the validation set loss, respectively, and t∈{1,2,…,T} represent the number of iterations. In each iteration, the architecture parameters are first fixed to optimize the network weights, and then the updated network weights are fixed to optimize the architecture parameters.

[0053] Without sacrificing generality, the training set loss and validation set loss can be expressed as follows when handling node classification tasks:

[0054] ;

[0055] in, The number of target nodes. The number of classes. For real labels, for The probability corresponding to the c-th category. For the target node The predicted probability vector, The probability that the c-th element corresponds to the c-th category. . The output layer's softmax function is used to... The probability distributions obtained by transforming the data into different categories are as follows:

[0056] ;

[0057] in, Let T represent a vector where all elements are 1, and let T denote the transpose. yes One component, representing the target node Feature strength or prediction score across different categories.

[0058] Step S400: Establish a meta-path aggregation search space as a search model for the optimal non-redundant meta-path combination.

[0059] The design incorporates a meta-path aggregation search space with multiple functionalities. Preferably, the meta-path aggregation search space can implement one or more of the following: message passing, feature aggregation, and meta-path fusion. The message passing function is used to pass information between nodes in a meta-path; the feature aggregation function is used to integrate features from different instances under the same meta-path; and the meta-path fusion function is used to merge features from different meta-paths.

[0060] Preferably, message passing utilizes direct aggregation, average aggregation, and / or multilayer perceptron transformation to transmit node information in the meta-path, thereby obtaining the features after message passing. .

[0061] Direct aggregation only considers the start and end points of the path, ignoring information about intermediate nodes. For the k-th meta-path... Features after direct aggregation for:

[0062] ;

[0063] in, Indicates about and The corresponding adjacency matrix, Metapath Upper i Each node type Metapath The nodes on This represents the optimal combination of non-redundant metapaths. Indicates the length of the corresponding metapath. This indicates a direct aggregation operation.

[0064] The average aggregation method generates new aggregated features by averaging the features of all nodes along the path element by element. for:

[0065] ;

[0066] The multilayer perceptron transform method performs multilayer nonlinear transformations on node features through multilayer perceptron transformation, improving the model's expressive power and ability to capture complex semantic relationships. The aggregated features... for:

[0067] ;

[0068] Preferably, feature aggregation utilizes average pooling, max pooling, and / or attention mechanisms to aggregate the features after message passing, resulting in aggregated features. .

[0069] Average pooling performs element-wise averaging of features across all instances along the same meta-path, effectively integrating information from all instances to obtain a comprehensive feature representation. for:

[0070] ;

[0071] in, The number of metapath instances. Let i be the feature of the i-th metapath instance.

[0072] Max pooling extracts the maximum value from all instance features along the same meta-path to highlight the most important features. The aggregated features... for:

[0073] ;

[0074] Attention mechanisms learn the normalized importance weights of each metapath instance, perform weighted summation to model and flexibly capture the different contributions of different instances to the target node, and aggregate the features. for:

[0075] ;

[0076] in, Represents neighbors based on metapaths. Indicates node u in the metapath The attention weight for node v is then determined.

[0077] Preferably, meta-path fusion utilizes average pooling, single-head attention, and / or multi-head attention mechanisms to fuse the aggregated features, outputting the meta-path fused features. .

[0078] Average pooling takes into account information from different meta-paths and obtains a comprehensive feature representation by averaging element by element. The fused features for:

[0079] ;

[0080] Where K is the number of meta-paths. It is the feature representation of the target node v after metapath fusion.

[0081] The single-head attention mechanism dynamically adjusts the feature aggregation method based on the importance of meta-paths. It considers the different contributions of different meta-paths to the target node, weighs and sums the relevant meta-paths, and then fuses the resulting features. for:

[0082] ;

[0083] in, This is the normalized attention score.

[0084] The multi-head attention mechanism is extended to a multi-head version to address the high variance problem caused by heterogeneity in large-scale heterogeneous graphs. Specifically, it executes independent attention mechanisms in R, then concatenates their outputs to fuse the features. for:

[0085] ;

[0086] in, The Softmax activation function is used. For the first A head of attention for Normalized importance weights For the first feature after aggregation Individual path features, This represents the total number of meta-paths in the optimal non-redundant meta-path combination. The number of independent attention mechanisms.

[0087] Step S500: Optimize the training search model using a differentiable modular architecture search strategy.

[0088] Preferably, in one embodiment of the present invention, step S500 specifically includes: dividing the meta-path aggregation process in the search model into multiple modules, each module containing a set of candidate operations; using a Softmax weighted hypernetwork to transform discrete architecture selection into continuous probabilities, and jointly optimizing the weight parameters and model weights in the search model through gradient descent.

[0089] The differentiable modular architecture search strategy relaxes discrete choices using Softmax weighting, transforming architecture selection into continuous probabilities. Gradient descent is then used to jointly optimize the weight parameters in the search model. γ and model weights are used to determine the weight of each operation. After training, the operation with the highest weight is selected to construct the deterministic architecture. The formula can be expressed as:

[0090] ;

[0091] in, This indicates the message passing module ( =1) or feature aggregation module ( =2), This indicates the corresponding operation. This is the weight vector corresponding to this operation. Indicates different metapath types, This represents the result of the weighted combination of operations after the softmax function. In the meta-path fusion module, the corresponding operation weight calculation formula is as follows:

[0092] ;

[0093] in, This defines the weight vector corresponding to the operation at this module. The weight parameters are adjusted using a modular differentiable search and gradient descent. , The weights of the model are jointly optimized to determine the weight of each operation.

[0094] Select the optimal operation for the three stages: message passing, feature aggregation, and meta-path fusion. , , This forms a deterministic architecture, represented as:

[0095] ;

[0096] in, It contains comprehensive feature information of the target node v. Taking a recommendation system scenario as an example, The information contained therein is obtained by modeling and aggregating nodes such as users, orders, categories, and products and the complex relationships between them. It can be used for various downstream tasks, such as node classification (predicting the category of nodes, such as the category of a user's interest in a certain product) and link prediction (predicting whether a user is likely to buy a certain product).

[0097] This invention provides a method for establishing a search model based on heterogeneous graph neural networks. It constructs a search strategy for identifying meta-path dependencies; generates an initial meta-path set; constructs a meta-path redundancy graph based on the dependencies between meta-paths to redundantly mark highly dependent meta-paths; uses a greedy algorithm to find an optimal set of non-adjacent nodes, i.e., a non-redundant meta-path combination, in the meta-path redundancy graph; defines meta-path strength; aggregates node information based on the selected meta-paths and their strengths to update node representations; dynamically finds the most suitable network architecture by alternately optimizing architecture parameters and network weights; designs a meta-path aggregation search space with multiple functions; and adopts a differentiable modular architecture search strategy to achieve personalized meta-path information aggregation. This invention optimizes the meta-path sampling strategy by constructing a meta-path redundancy graph, identifies adaptive meta-path aggregation strategies using a differentiable modular architecture search strategy, comprehensively mines the structural and semantic information of heterogeneous graphs, improves information aggregation efficiency, and reduces computational redundancy. This effectively improves the performance and efficiency of heterogeneous graph representation learning tasks and is applicable to scenarios such as social network analysis, recommendation systems, and academic relationship mining.

[0098] For recommendation system applications, this invention addresses the dual challenges of computational efficiency and semantic noise in existing automated meta-path selection mechanisms. By sampling non-redundant meta-path combinations using a greedy algorithm and optimizing the network architecture, computational resource consumption is reduced while semantic noise interference is minimized. This significantly improves the model's ability to identify user preferences and product association features, enhancing recommendation accuracy and recall, especially when handling large-scale user and product data and multimodal user behavior data. In the field of academic relationship mining, this invention overcomes the lack of dynamic adaptability in existing feature aggregation frameworks. Depending on the needs of different academic mining tasks, such as academic influence assessment, research direction prediction, and academic collaboration recommendations, the dynamic architecture search strategy of this invention can flexibly adjust the feature aggregation method, avoiding the loss of key academic features or excessive amplification of noisy features, thus enabling the mining results to more accurately reflect the true state of academic relationships.

[0099] The advantages of the search model establishment method based on heterogeneous graph neural networks of the present invention are illustrated below through specific examples. The examples used in the following examples are as follows: Figure 2 The network architecture shown in the figure illustrates the interaction relationships and data flow direction of the redundancy analysis module, path sampling module, and multi-level aggregation module.

[0100] In this example, a specific heterogeneous graph neural network architecture is used as the base model, trained with the Adam optimizer, a weight decay rate of 1e-4, a learning rate of 0.005, and a batch size of 16. Simultaneously, a meta-path dependency matrix and a meta-path redundancy graph are constructed to optimize meta-path sampling, and a differentiable modular architecture search strategy is employed to identify adaptive meta-path aggregation strategies.

[0101] I. Verification of the accuracy of the training method

[0102] The training method in this invention is compared with existing hypergraph neural network (HGNN) model training methods, and validated on the Digital Library and Documentation Project (DBLP) dataset, the Internet Movie Database (IMDB) dataset, and the Association for Computing Machinery (ACM) dataset. State-of-the-art heterogeneous graph neural network model training methods include Semantic Edge-Enhanced Heterogeneous Graph Neural Network (SeHGNN) and Long Distant Meta-Path Search Network (LMSPS) based on Progressive Sampling.

[0103] The verification results show that the training method in this invention effectively improves the accuracy of heterogeneous graph neural networks, outperforming existing heterogeneous graph neural network model training methods. In experiments, this invention achieves significant performance improvements on the Internet Movie Database dataset. Both macro-average F1 and micro-average F1 scores demonstrate accuracy in movie classification and user recommendation tasks, validating the effectiveness and advantages of this invention in recommender system applications. Experiments on the Digital Library and Document Project dataset and the Computer Association dataset show that this invention performs excellently in tasks such as author classification and research field classification, with macro-F1, micro-F1, and test accuracy scores all outperforming existing methods, fully demonstrating the superior performance and broad application prospects of this invention in academic relationship mining tasks.

[0104] II. Validation of the Meta-Path Search Algorithm

[0105] Two mainstream heterogeneous graph neural network models, semantic edge-enhanced heterogeneous graph neural networks and progressively sampled long-range meta-path search networks, were selected. The meta-paths obtained by the method of this invention were used to replace the predefined meta-path set of the HGNN model, and the performance changes were compared. Validation results show that after replacing the meta-paths, the semantic edge-enhanced heterogeneous graph neural network improved performance by 4.8%, and the progressively sampled long-range meta-path search network improved performance by 3.7%. This fully demonstrates the effectiveness and strong generalization ability of the search results obtained by the method of this invention, reflecting its significant advantages in meta-path mining.

[0106] III. Verification of the critical role of the meta-path search stage in overall model performance

[0107] This paper analyzes in depth the crucial role of the meta-path search stage in the method of this invention on the overall model performance. Through ablation experiments, the performance of the complete model and the model without the meta-path search optimization stage is compared on four datasets: DBLP, IMDB, ACM, and the Open Graph Benchmark - Microsoft Academic Graph (OGBN-MAG).

[0108] Validation results show that the complete model including the meta-path search optimization stage achieves a test accuracy of 56.20%, while the model without this stage only achieves 52.45%. This significant difference highlights the core role of the meta-path search stage in improving model performance, proving that it effectively reduces information redundancy and enhances the model's ability to express heterogeneous graph data.

[0109] Based on the above verification results, it is evident that the method provided by this invention exhibits significant advantages in the field of heterogeneous graph neural networks. Its adaptive meta-path search and aggregation strategy not only effectively improves model performance and efficiency but also verifies the superiority of the search algorithm and the necessity of key stages, providing solid technical support for the practical application of heterogeneous graph neural networks.

[0110] The method for establishing a search model based on heterogeneous graph neural networks provided by this invention, compared with existing heterogeneous graph neural network models based on meta-paths, optimizes meta-path sampling by constructing a meta-path dependency matrix and a meta-path redundancy graph, thereby avoiding redundant information in the sampling process and improving sampling efficiency. When identifying meta-path combinations, only relevant matrices and graph structures need to be constructed, without introducing a large number of additional parameters, resulting in a small data footprint. At the same time, a differentiable modular architecture search strategy is proposed, which can adaptively aggregate information for different meta-path combinations and downstream tasks, effectively improving the model's feature extraction capability for complex graph structures, and ultimately improving the efficiency and accuracy of the model in heterogeneous graph learning tasks.

[0111] Based on the same inventive concept, corresponding to any of the above embodiments, the present invention also provides a search method based on heterogeneous graph neural networks, the method comprising: using the optimized search model in the above embodiments of the search model based on heterogeneous graph neural networks to search for comprehensive feature information related to the target node.

[0112] For example, if the target node is user A, by inputting user A's identity information into the optimized search model, information such as orders, categories, products, and other related users associated with user A can be retrieved.

[0113] Based on the same inventive concept, and corresponding to any of the methods in the above embodiments, this invention also provides a search model building system based on heterogeneous graph neural networks. See also... Figure 3 The search model building system 300 based on heterogeneous graph neural networks includes a collection module 310, a generation and search module 320, a first optimization module 330, a building module 340, and a second optimization module 350.

[0114] For example, the collection module 310 performs step S100 to collect metapaths related to the target node and form an initial metapath set.

[0115] The generation and search module 320, for example, performs step S200, which uses meta-paths in the initial meta-path set whose pairwise dependencies are higher than a preset threshold to generate a meta-path redundancy graph; and searches for the optimal combination of non-adjacent meta-paths from the meta-path redundancy graph as the initial non-redundant meta-path combination.

[0116] The first optimization module 330, for example, executes step S300, which optimizes the architecture parameters and network weights of the initial non-redundant meta-path combination based on the meta-paths and meta-path strengths in the initial non-redundant meta-path combination, to obtain the optimal non-redundant meta-path combination.

[0117] Module 340, for example, executes step S400 to establish a meta-path aggregation search space as a search model for the optimal non-redundant meta-path combination.

[0118] The second optimization module 350, for example, performs step S500 to optimize the trained search model using a differentiable modular architecture search strategy.

[0119] Through the collaborative efforts of the aforementioned components, the search model building system based on heterogeneous graph neural networks provided in this embodiment of the invention generates a meta-path redundancy graph by utilizing meta-paths in the initial meta-path set whose pairwise dependencies exceed a preset threshold. This effectively reduces information redundancy and enables precise selection of the most suitable meta-path combination. By employing a differentiable modular architecture search strategy, it comprehensively mines the structural and semantic information of the heterogeneous graph, significantly improving the performance and applicability of the search model, thereby enhancing the accuracy and efficiency of search results.

[0120] Further functional descriptions of the above modules are the same as those of the corresponding method embodiments described above, and will not be repeated here.

[0121] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the search model establishment method based on heterogeneous graph neural networks and / or the search method based on heterogeneous graph neural networks described in any of the above embodiments.

[0122] Figure 4 This illustration shows a more specific hardware structure diagram of an electronic device provided in this embodiment. The device may include: a processor 410, a memory 420, an input / output interface 430, a communication interface 440, and a bus 450. The processor 410, memory 420, input / output interface 430, and communication interface 440 are interconnected internally via the bus 450.

[0123] The processor 410 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0124] The memory 420 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 420 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 420 and is called and executed by the processor 410.

[0125] Input / output interface 430 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touch screens, microphones, various sensors, etc., and output devices may include displays, speakers, vibrators, indicator lights, etc.

[0126] The communication interface 440 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0127] Bus 450 includes a pathway for transmitting information between various components of the device, such as processor 410, memory 420, input / output interface 430, and communication interface 440.

[0128] It should be noted that although the above-described device only shows the processor 410, memory 420, input / output interface 430, communication interface 440, and bus 450, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0129] The electronic devices described above are used to implement the corresponding search model establishment method based on heterogeneous graph neural networks and / or the search method based on heterogeneous graph neural networks in any of the foregoing embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0130] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, the present invention also provides a non-transitory computer-readable storage medium storing computer instructions, which are used to cause the computer to execute the search model establishment method based on heterogeneous graph neural networks and / or the search method based on heterogeneous graph neural networks as described in any of the above embodiments.

[0131] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0132] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the search model establishment method based on heterogeneous graph neural networks and / or the search method based on heterogeneous graph neural networks as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0133] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of the invention as described above, which are not provided in the details for the sake of brevity.

[0134] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of the invention, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of the invention, and this also takes into account the fact that the details of implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of the invention will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of the invention, it will be apparent to those skilled in the art that the embodiments of the invention may be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0135] Although the invention has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0136] The embodiments of this invention are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this invention should be included within the protection scope of this invention.

Claims

1. A method for establishing a search model based on a heterogeneous graph neural network, characterized in that, include: Collect meta-paths related to the target node to form an initial meta-path set; wherein, the target node is a node type that needs to be classified or predicted. Using the meta-paths in the initial meta-path set whose pairwise dependencies are higher than a preset threshold, a meta-path redundancy graph is generated; the optimal combination of non-adjacent meta-paths is found from the meta-path redundancy graph as the initial non-redundant meta-path combination; the non-adjacent meta-path combination is a combination of meta-paths that have no direct dependencies. Based on the meta-paths and meta-path strengths in the initial non-redundant meta-path combination, the architecture parameters and network weights of the initial non-redundant meta-path combination are optimized to obtain the optimal non-redundant meta-path combination. A meta-path aggregation search space is established as a search model for the optimal non-redundant meta-path combination; The search model is optimized and trained using a differentiable modular architecture search strategy. The optimized search model is used to search for comprehensive feature information related to the target node. The target node is the user. The optimized search model is used to model and aggregate users, orders, categories, products and the relationships between them to obtain comprehensive user feature information. This information can be used to predict a user's interest category for a certain product and whether a user is likely to purchase a certain product. By inputting the user's identity information into the optimized search model, orders, categories and products associated with the user can be searched. The process of generating the meta-path redundancy graph is as follows: If the meta-path Harmony Path Dependence between If the value is higher than the preset threshold δ, in the meta-path redundancy graph Add an edge to connect the metapath Harmony Path Mark highly dependent meta-path combinations; use a greedy algorithm to select unselected nodes, add them to the meta-path combinations, and delete the node and its adjacent nodes until all nodes have been processed; This is a set of nodes, representing the set of all meta-paths; Let be the edge set, representing the redundant relationships between metapaths.

2. The search model establishment method based on the heterogeneous graph neural network according to claim 1, characterized in that, The dependency is: ； in, Metapath Harmony Path Dependence between them , Metapath Metapath The set of covered nodes , Metapath Metapath The edge set covered. 3.The method of claim 1, wherein, The metapath strength is: ； in, Metapath Metapath strength, , They are respectively the metapath Metapath Related architectural parameters, This represents the total number of meta-paths in the initial non-redundant meta-path combination. , Metapath Metapath The size of the covered node set, , Metapath Metapath The size of the covered edge set, This refers to the temperature parameter. 4.The method of claim 1, wherein, Optimizing the architecture parameters and network weights of the initial non-redundant meta-path combination specifically includes: alternately updating the architecture parameters and network weights until the optimal architecture parameters and network weights are obtained. 5.The search model establishment method based on the heterogeneous graph neural network according to claim 1, wherein, The meta-path aggregation search space can realize one or more of message passing, feature aggregation and meta-path fusion; The message passing utilizes direct aggregation, average aggregation, and / or multilayer perceptron transformation to transmit node information in the metapath; The feature aggregation utilizes average pooling, max pooling, and / or attention mechanisms to aggregate the features after message passing; The meta-path fusion method uses average pooling, single-head attention mechanism and / or multi-head attention mechanism to fuse the features after feature aggregation, and outputs the meta-path fused features. 6.The search model establishment method based on a heterogeneous graph neural network according to claim 1, wherein, The search model is optimized and trained using a differentiable modular architecture search strategy, specifically including: The meta-path aggregation process in the search model is divided into multiple modules, each module containing a set of candidate operations; The discrete architecture selection is transformed into continuous probabilities using a Softmax weighted supernetwork, and the weight parameters and model weights in the search model are jointly optimized by gradient descent.

7. A search method based on a heterogeneous graph neural network, characterized in that, include: Using the optimized search model described in any one of claims 1-6, comprehensive feature information related to the target node is searched. 8.A system for building a search model based on a heterogeneous graph neural network, characterized in that, include: The collection module is used to collect meta-paths related to the target node to form an initial meta-path set; wherein, the target node is a node type that needs to be classified or predicted. The generation and search module is used to generate a meta-path redundancy graph by using meta-paths in the initial meta-path set whose pairwise dependencies are higher than a preset threshold; and to search for the optimal non-adjacent meta-path combination from the meta-path redundancy graph as the initial non-redundant meta-path combination; the non-adjacent meta-path combination is a meta-path with no direct dependency relationship. The first optimization module is used to optimize the architecture parameters and network weights of the initial non-redundant meta-path combination based on the meta-paths and meta-path strengths in the initial non-redundant meta-path combination, so as to obtain the optimal non-redundant meta-path combination. A module is established to create a meta-path aggregation search space as a search model for the optimal non-redundant meta-path combination. The second optimization module is used to optimize the training of the search model using a differentiable modular architecture search strategy. The optimized search model is used to search for comprehensive feature information related to the target node. The target node is a user. The optimized search model is used to model and aggregate users, orders, categories, products and the relationships between them to obtain the user's comprehensive feature information. This information can be used to predict the user's interest category for a certain product and whether the user is likely to purchase a certain product. By inputting the user's identity information into the optimized search model, orders, categories and products associated with the user can be searched. The process of generating the meta-path redundancy graph is as follows: If the meta-path Harmony Path Dependence between If the value is higher than the preset threshold δ, in the meta-path redundancy graph Add an edge to connect the metapath Harmony Path Mark highly dependent meta-path combinations; use a greedy algorithm to select unselected nodes, add them to the meta-path combinations, and delete the node and its adjacent nodes until all nodes have been processed; This is a set of nodes, representing the set of all meta-paths; Let be the edge set, representing the redundant relationships between metapaths.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method as described in any one of claims 1-7.

Images