30 results about "Level structure" patented technology

The invention discloses a cluster job planning method and system based on a dual time window mechanism. The method comprises the steps that jobs are divided into a multi-level structure with an execution sequence, and a directed acyclic graph is constructed to form a task topological structure; constructing a dual time window mechanism; designing a multi-objective optimization function; modeling cluster job scheduling into a multi-agent Markov decision process, setting a hierarchical action space, and designing a reward function reflecting scheduling balance; the method comprises the following steps: constructing a hybrid neural network architecture, extracting a time sequence correlation feature of a task, encoding a heterogeneous feature between the task and a server by using a heterogeneous graph neural network, and generating an intelligent matching score between the task and the server; a multi-agent near-end strategy optimization algorithm is applied, and agent collaborative optimization is achieved through a collaborative mechanism updated by a centralized value function and a distributed strategy. According to the method, fine modeling and state dynamic perception of complex operation topology are realized, and scientificity and controllability of scheduling decision are improved.

The invention relates to the technical field of block chains, in particular to a smart contract vulnerability detection method and system based on a multi-modal large language model.The detection method comprises the steps that S1, the system receives source code input of a smart contract; s2, executing semantic branches and graph structure branches in parallel; executing semantic branch processing, calling an annotation agent, and generating an annotation from the source code of the smart contract; calling a vectorization agent, coding the annotation and the source code, and converting the annotation and the source code into a high-dimensional vector; executing graph structure branch processing, compiling a source code into a byte code by the system, and generating a control flow graph; the vectorization agent generates high-dimensional embedded representation of the nodes based on the control flow graph; and S3, integrating features extracted from the semantic branches and the graph structure branches through a multi-modal feature fusion strategy, and inputting the integrated features into a classifier to detect vulnerabilities. The method is based on a multi-modal large language model agent, and potential vulnerabilities in the smart contract are comprehensively analyzed by integrating high-level semantic information and low-level structured data.

The invention discloses a quaternary logic driven polymorphic structure coding method and application thereof. The technical problem to be solved is to express state, direction, hierarchy and identification integration of a nested structure, a direction path and a graph structure of data. The method comprises the following steps of mapping, identifying, encoding and forming structural unit encoding, and is applied to structural identity card encoding, graph neural network GNN structural path input encoding, storage and calculation integrated chip path encoding instructions, data authority and / or experience sharing structural identification or chip circuit encoding. In the field of artificial intelligence underlying data coding and processing, structural unit coding, direction, hierarchy and unique identification are adopted, unified coding and reversible analysis of a direction path, a nested structure and a graph topological relation of data are achieved, and the method can be widely applied to the fields of artificial intelligence, graph neural networks, chip paths, identity systems and multi-hierarchy structure modeling.

A knowledge graph construction method based on an ontology and a large language model belongs to the field of artificial intelligence and knowledge graph construction, and comprises the following steps: step 1, initial ontology construction for a knowledge extraction task: step (1.1), field demand analysis and range definition; (1.2) reusing the body; (1.3) carrying out concept classification and hierarchical structure construction; (1.4) relation definition and attribute establishment; step (1.5), designing an attribute system; the second step is an ontology extension method based on the large language model, and the method comprises the following sub-steps: (2.1) an extension candidate set is automatically generated by the LLM; (2.2) verifying the consistency of the expansion candidates; (2.3) updating the domain ontology; thirdly, triple automatic extraction is carried out in combination with a self-adaptive chain type thinking mechanism; and step 4, triple fusion and knowledge graph construction oriented to heterogeneous data. According to the method, the structure normalization and semantic consistency of knowledge graph construction are effectively improved.

The invention discloses a layout parameterization hierarchical parallel mesh generation method based on Net. The method comprises the following steps: inputting a layout design file, analyzing hierarchical structure identification ports, generating a port set P and creating an accessed port set V (initial V =); a graph G = (VG, EG) is constructed, VG comprises key vertexes or geometric centers of ports and conductor fragments, an electrically independent Net is generated through traversal of a breadth-first search algorithm or a depth-first search algorithm, and iteration is performed until V is completely equal to P; configuring independent subdivision parameters such as a grid size and a grid algorithm for each Net; net is used as an independent task, grids are divided through parallel computing frameworks such as multithreading, and a complete grid file for subsequent simulation is generated after merging. According to the method, Net automatic identification is realized through graph theory analysis; different Net precision requirements are adapted by differential subdivision parameters, so that waste of computing resources is avoided; through parallel processing, the time consumption of layout grid division is shortened to the time consumption of processing the maximum Net, and the efficiency and the flexibility of simulation pre-processing are remarkably improved.

A method for dynamic execution of sub-graphs within executable graph-based models is provided. Processing circuitry obtains an executable graph-based model comprising a plurality of sub-graphs and an overlay structure comprising processing logic associated with the plurality of sub-graphs. Each sub-graph defines a hierarchical structure of related nodes. The processing circuitry receives a stimulus and a context associated with the stimulus. In response to the stimulus being received and based on the context, the processing circuitry maps the stimulus to a first sub-graph of the executable graph-based model. The processing circuitry causes execution of processing logic within the overlay structure based on the mapping. The processing logic is associated with one or more nodes of the first sub-graph.

The invention relates to the technical field of evaluation and judgment, in particular to a typical task system performance evaluation method based on an entropy weight method and a graph convolutional network, which comprises the following steps of: weighting evaluation index data with a hierarchical structure by using the entropy weight method, and converting the traditional hierarchical index data into a data structure of a non-Euclidean space; and constructing a depth map convolutional network model, extracting deep features in an input sample, mining a coupling relationship between sensor data and task system performance, and realizing evaluation of the task system performance. The typical task system performance is preliminarily evaluated through the information entropy, and the relation between bottom layer single items can be well processed. According to the method, the graph network is applied to typical task system performance evaluation, the model can fully extract shallow and deep feature information of an input sample, the graph attention network is adopted to fuse and enhance the features, and it is guaranteed that the model has higher robustness and accuracy.

A knowledge graph construction method based on ontology and large language model belongs to the field of artificial intelligence and knowledge graph construction, comprising the following steps: first, initial ontology construction for knowledge extraction task: step (1.1) domain requirement analysis and scope definition; step (1.2) ontology reuse; step (1.3) concept classification and hierarchical structure construction; step (1.4) relationship definition and attribute construction; step (1.5) attribute system design; second, ontology extension method based on large language model, comprising the following substeps: step (2.1) LLM automatically generates an extension candidate set; step (2.2) consistency verification of extension candidates; step (2.3) update domain ontology; third, automatic extraction of triples combined with adaptive chain thinking mechanism; fourth, triple fusion and knowledge graph construction oriented to heterogeneous data. The application effectively improves the structural standardization and semantic consistency of knowledge graph construction.

The application discloses a sequence construction method and system. The method constructs a switching matrix based on the number of various types of elements in the sequence, establishes a unified constraint system, and obtains a numerical solution of the switching matrix through optimization. Based on the numerical solution, a graph structure is constructed and connectivity is determined to determine whether a complete sequence can be formed; when the connectivity is satisfied, a target sequence that meets the switching relationship and quantity requirements is generated based on the graph structure. The method supports segment-level structure expression, modeling of the head-tail relationship of the augmented matrix, and sub-sequence combination expression, and is suitable for multiple sequence construction tasks. Accordingly, the application also provides a sequence construction system, which includes a matrix modeling module, a constraint construction module, a solving module, a connectivity determination module, and a sequence generation module.

The present application relates to a long text retrieval method, apparatus, and computer device based on a graph structure. The method comprises: dividing an input long text into blocks, guiding a large language model to analyze the sub-texts after the blocks are divided, obtaining text information of each sub-text, connecting the nodes of the sub-texts with similar core elements to obtain a first structure graph, processing the first structure graph using a community clustering algorithm to obtain a second structure graph with a hierarchical structure of a graph community, receiving a question input by a user, processing and optimizing the question based on a large language model, confirming whether the question is a specific question or an abstract question, if the question is a specific question, retrieving the question from the first structure graph to obtain a reading queue, generating notebook content based on the reading queue, if the question is an abstract question, retrieving the question from the second structure graph to obtain notebook content containing answers and scores, and summarizing and outputting the notebook content through the large language model.

Polygonal shapes can be generated automatically from an example using a graph grammar. The graph grammars can then be used to generate graphs that are locally similar to a given example. An input shape can be disassembled into small pieces called primitives and then the primitives can be reassembled into new graphs. All possible locally similar graphs can be organized into a graph hierarchy and then matching graphs can be found within the graph hierarchy. These matches are used to create a graph grammar that can construct every locally similar graph. The graphs can be generated using the graph grammar and then converted into a planar graph drawing to produce the final shape.

Method for analyzing a mechatronic system, wherein the mechatronic system has one or more functions, the functions comprising one or more hardware and / or software functions, and in the method - a first network (20) in the form of a tree graph with several hierarchy levels and nodes (201, 202, 203, 204, 205, 206) arranged in the hierarchy levels is provided, wherein the nodes (201-206) are each representative of one of the functions of the mechatronic system and are linked to each other via one or more logical operators, such that the tree graph represents dependencies between the functions represented by the nodes (201-206), wherein the top hierarchy level has a single node (201) as the initial node, - a node of the tree graph of the first network (20) is specified as the first trigger node, - Software network data is provided that is representative of the software functions of the mechatronic system and their dependencies, - depending on the first trigger node and the software network data, a second network (30) in the form of a tree graph with several hierarchy levels and nodes (301, 302, 303, 304, 305) arranged in the hierarchy levels is determined, wherein the nodes (301-305) are each representative of one of the software functions of the mechatronic system and are linked to each other via one or more logical operators, such that the tree graph represents dependencies between the software functions represented by the nodes (301-305), wherein the top hierarchy level has a single node (301) as the initial node, and - depending on the first network of action (20) and the second network of action (30), the mechatronic system is analyzed, wherein an analysis of the mechatronic system includes that - the first network (20) is extended by the second network (30) such that the first trigger node of the first network (20) corresponds to the initial node of the second network (30), and - the mechatronic system is analyzed depending on the extended first network of action (40), wherein the mechatronic system has one or more diagnostic functions that are representative of one or more software functions for diagnosing the mechatronic system, in which - a node of the tree graph of the extended first network (40) is specified as the second trigger node, - one of the diagnostic functions is specified and assigned to the second trigger node, - Diagnostic network data is provided that is representative of the diagnostic functions of the mechatronic system and their dependencies, - depending on the second trigger node and the diagnostic network data, a third network (50) is determined in the form of a tree graph with several hierarchy levels and nodes (501, 502, 503, 504, 505) arranged in the hierarchy levels, wherein the nodes (501-505) are each representative of one of the diagnostic functions of the mechatronic system and are linked to each other via one or more logical operators, such that the tree graph represents dependencies between the diagnostic functions represented by the nodes (501-505), wherein the top hierarchy level has a single node (501) as the initial node, such that the second trigger node of the extended first network (40) corresponds to the initial node of the third network (50), and - depending on the extended first network (40) and the third network (50) the mechatronic system is analyzed.

A method for making predictions includes identifying, for each node in the heterogeneous graph structure, a set of node-target paths that connect the node to a target node; assigning, to each of the node-target paths, a path type identifier indicative of a number of edges and corresponding edge types in the associated node-target path; and extracting a semantic tree from the heterogeneous graph structure. The semantic tree includes the target node as a root node and defines a hierarchy of metapaths that each correspond to a subset of the node-target paths in the heterogeneous graph structure assigned to a same path type identifier. The semantic tree is encoded by generating a metapath embedding corresponding to each metapath in the semantic tree. A label is predicted for the target node in the heterogeneous graph structure based on the set of metapath embeddings.

The embodiment of the invention provides a structured information processing method, equipment and device, a storage medium and a program product. A hierarchical clustering mechanism with coordinated structure and semantics is provided, a plurality of nodes in a graph structure are coded into a plurality of feature vectors through a target coding model, and the feature vectors comprise text semantic information and hierarchical structure information between the nodes to lay a foundation for subsequent clustering. According to the method, clustering is carried out on multiple feature vectors and clustering clusters by adopting two-layer clustering operation, and in the hierarchical clustering process, category labels of the clustering clusters are generated through a language model, and the category labels have cross-hierarchical semantic interpretability. The hierarchical category label is generated based on the category label of each cluster, the semantic abstract hierarchical structure from fine granularity to coarse granularity is constructed, the hierarchical structure of the hierarchical category label is clear, the semantic interpretability is high, the context consistency and knowledge controllability of the text information retrieved from the graph structure are improved, and the text information retrieval efficiency is improved. And the accuracy of downstream tasks is improved.

This invention discloses a collision calculation acceleration method combining parametric geometry and collision algorithms, belonging to the field of physical simulation technology. The method includes: traversing a 3D scene to classify and label geometric objects as parametric entities or discrete mesh objects; constructing a hybrid bounding volume hierarchical structure tree, with leaf nodes storing algebraic equations or vertex data respectively; performing generalized collision detection to screen potential collision pairs; calculating a collision urgency index based on overlapping volume and complexity, and prioritizing collision pairs; using analytical geometry for parametric entities and primitive intersection testing for discrete meshes, and dynamically allocating computational resources according to priority; finally, using pose hash caching and a warm-start strategy to accelerate convergence. This invention effectively solves the problems of high computational load and inefficient resource allocation in traditional discretized collision detection by combining hybrid geometric representation and dynamic scheduling mechanisms, significantly improving the real-time performance of physical simulation.

The invention discloses an extensible and comprehensive graph converter graph representation learning method, system and device and a storage medium, and relates to the technical field of machine learning. The method comprises the steps of obtaining target graph data, determining corresponding node features, edge features and an adjacent matrix according to the target graph data, calculating a node embedding matrix and an edge embedding matrix of a graph, and generating an initial graph structure; based on the initial graph structure, transmitting an adjacent matrix into a graph coarsening module, generating a graph hierarchical structure by using a graph coarsening algorithm, and calculating an absolute position coding matrix and a relative position coding matrix of each graph layer; inputting each coding matrix into a comprehensive position coding module for merging to obtain a comprehensive position code; and inputting the comprehensive position code and the node embedding matrix into a focused graph linear attention module, calculating focused linear attention, and outputting final node representation. The efficiency and expression ability of graph representation learning can be improved.

The application discloses a topic classification method based on a hyperbolic graph convolutional network and hierarchical clustering, and belongs to the technical field of text topic classification in natural language processing. The method comprises the following steps: preprocessing microblog text data, including data cleaning, Chinese word segmentation and removal of stop words; and calculating a feature vector by using a TF-IDF method. The preprocessed text data is mapped to a hyperbolic space, information is converted between Euclidean space and hyperbolic space by using exponential mapping and logarithmic mapping, and the hierarchical structure characteristics of the hyperbolic space are retained. The hyperbolic distance between nodes is calculated, and a hierarchical clustering algorithm is combined to capture the relationship in the hierarchical structure data. Similar or related nodes can obtain a higher weight in the information aggregation process of the hyperbolic graph convolutional network, the hierarchical structure of the text data is mined, the hierarchical structure relationship of the microblog text data is captured by training the hyperbolic graph convolutional neural network, and high-quality topic classification is realized by using the trained hyperbolic graph convolutional neural network.

The invention belongs to the field of graph data processing, and discloses a multi-modal graph representation learning method and system based on a dynamic hyperbolic hypergraph Transform model.The method comprises the steps that multi-modal data are obtained and preprocessed, and multi-modal graph data are obtained; generating hyperedges in a self-adaptive manner on the basis of the multi-modal graph data, and constructing a sparse incidence matrix between nodes of the graph data and the hyperedges; nodes and hyperedges of graph data are mapped to a hyperbolic geometric space, and bidirectional feature propagation and aggregation between the nodes and the hyperedges are executed according to a sparse incidence matrix to update node representation; performing global dependency modeling on the node representation based on the updated node representation to obtain a fused node representation; and based on a category weighted loss function and a multi-label threshold search strategy, performing classification prediction on the fusion node representation, and completing a graph representation learning task. According to the method, the problems of insufficient high-order relation modeling, hierarchical structure representation distortion, high calculation complexity, poor category imbalance adaptability and the like of a graph representation learning method in the prior art are solved.

The application discloses a kind of multi-dimensional graph tensor fusion representation and embedding of code and application, belong to artificial intelligence field.It includes: extracting the syntax information and hierarchical structure information of source code file and binary file;Abstract Syntax Tree abstract syntax tree (AST), Data Dependence Graph data dependence graph (DDG), Control Flow Graph control flow graph (CFG), Natural Code Sequence natural language sequence (NCS) four different heterogeneous code graph structures are generated simultaneously to source code file and binary file;Four kinds of heterogeneous code graph structures are combined to generate high-dimensional graph tensor;Using Graph Tensor Convolution Network interpretable graph tensor convolution neural network (GTCN) to generate accurate code semantic embedding and capture code internal features, and related technology is applied in various downstream tasks, such as malicious code identification, in detection efficiency and accuracy rate aspect, good balance is achieved.

The invention discloses an efficient forgetting method for hierarchical clustering data, relates to the technical field of hierarchical clustering, solves the problem that forgetting nodes are not determined by analyzing global features, and comprises the following steps: recording distance parameters of father nodes of each layer through a feature example graph to form a distance mapping relationship from bottom to top; when data is updated, affected father nodes can be quickly positioned based on the example graph, and global traversal is avoided; according to the method, initial child nodes at the bottommost layer are recognized from a feature example graph as undetermined nodes, low-distance nodes are preferentially selected as candidate forgetting objects through similar distance sorting of father nodes of the initial child nodes, it is ensured that the influence of elimination operation on an upper-layer clustering structure is minimum, and the nodes with the low similar distance generally belong to local dense clusters; the disturbance to the overall hierarchical structure is smaller after elimination, the clustering feature change of each undetermined node after elimination is quantified by calculating a feature difference value, and the node with the minimum difference value is selected as a forgotten node to ensure that the change quantity of the original clustering feature is reduced to the minimum.

The invention belongs to the field of software testing, and relates to a function-oriented knowledge graph construction and test scene automatic generation method and system based on a user manual. The method comprises the following steps: preprocessing a user manual of a to-be-tested application, including format unification and structured processing, and extracting a hierarchical structure and a content unit of a document; extracting knowledge from the preprocessed user manual through a large language model and constructing an initial knowledge graph; the initial knowledge graph is optimized through a graph-based reasoning technology, inconsistency caused by illusion of a large language model is relieved, knowledge implied in a user manual is deduced, and a function-oriented knowledge graph is obtained; and automatically generating a diversified test scene set based on the optimized function-oriented knowledge graph. According to the method, the high-quality knowledge graph can be automatically constructed from the user manual, and the test scene comprehensively covering normal functions, abnormal conditions and complex combinations is generated based on the high-quality knowledge graph.

The invention provides a sensitive data detection method and device based on hierarchical information enhancement and a graph convolutional network, and relates to the technical field of data processing. According to the method, text sequence semantics (BERT), an entity hierarchical structure (hierarchical embedding) and an entity association relationship (graph convolution) are combined, so that the modeling capability for a complex sensitive scene (such as multi-entity co-occurrence) is improved, and the detection accuracy of sensitive data in the complex sensitive scene is further improved; moreover, a tree-shaped hierarchical structure is constructed through the risk levels of the suspected sensitive entity types, so that risk weights associated with entities in different dimensions can be flexibly adjusted, and diversified sensitive data detection requirements are met; besides, based on a pre-trained BERT model, a graph structure comprising nodes and edges can be quickly migrated to sensitive data detection tasks in different fields (such as medical treatment and finance).

The application provides a sensitive data detection method and device based on hierarchical information enhancement and a graph convolutional network, and relates to the technical field of data processing. The method combines text sequence semantics (BERT), entity hierarchy (hierarchical embedding), and entity correlation (graph convolution) to improve the modeling capability for complex sensitive scenarios (such as multi-entity co-occurrence), thereby improving the detection accuracy of sensitive data in complex sensitive scenarios. Furthermore, by constructing a tree-like hierarchical structure based on the risk level of suspected sensitive entity types, the risk weights of different dimensional entity correlations can be flexibly adjusted to adapt to diversified sensitive data detection requirements. In addition, based on the pre-trained BERT model and the graph structure containing nodes and edges, the sensitive data detection task can be quickly migrated to different fields (such as medical care, finance, etc.).

A computer-implemented method includes obtaining a first kinematic graph that represents a first articulated object. The method includes generating a set of primitive structures by decomposing the first kinematic graph. The method includes generating a graph hierarchy by enumerating one or more valid combinations of the set of primitives. The method includes determining the production rules based on the graph hierarchy. The method includes determining kinematic graph grammar based on the set of primitive structures and the production rules. The method includes determining the continuous parameters for the discrete kinematic graph structure based on the production rules using a generative flow network. The method includes generating a second kinematic graph to represent a second articulated object based on the kinematic graph grammar and the continuous parameters.

The application discloses a knowledge graph embedding method and system based on quaternions, comprising the following steps: S1, constructing a projection matrix for the hierarchical structure of a triple; S2, constructing a knowledge graph embedding model based on quaternions, using the knowledge graph embedding model based on quaternions to realize the rotation between entities, using the modulus of the quaternion to realize the ability of constructing the hierarchical structure; obtaining a score function of the knowledge graph embedding model based on quaternions, and constructing a final loss function through the score function; S3, using the knowledge graph embedding model based on quaternions to model a main relationship mode; S4, learning the problem of entity and relationship representation in the knowledge graph through the final loss function of the knowledge graph embedding model based on quaternions, predicting the connection, so as to obtain the missing and real triple facts in the knowledge graph, realize the completion of the knowledge graph, and solve the multiple representations of entities in different scenes and improve the entity classification performance.

The present invention relates to a method for processing a graph. The method comprise providing (S10) the initial graph, said initial graph comprising a plurality of initial nodes and a plurality of initial edges, each initial edge connecting two initial nodes, determining (S20) a hierarchy of a plurality of communities at a plurality of levels, each community being respective to a level of said plurality of levels and comprising a respective cluster of the initial nodes, determining (S30) one or more conduits, each conduit connecting a first determined community at a first level of said plurality of levels and a second determined community at a second level of said plurality of levels, the first level being different from the second level, each conduit comprising a respective group of the initial edges and connecting the respective cluster of the initial nodes of the first community to the respective cluster of the initial nodes of the second community, and outputting (S40) a final graph based on the initial graph, the determined hierarchy of plurality of communities and the determined one or more conduits; the final graph representing said connectivity of said plurality of entities.

The invention discloses an automatic optimization method for SysML model parameters. The method comprises the following steps that S1, a SysML architecture model facing system task constraints is constructed, and design variables and target attributes are marked; s2, modeling a system performance constraint relation based on the parameter diagram and establishing a target function expression; s3, analyzing the model semantics to automatically extract an optimization problem and converting the optimization problem into a standard mathematical form; and S4, solving based on a multi-objective optimization algorithm and feeding back a result. According to the method, the hierarchical structure of the SysML model, the relation semantics and the attachment characteristics of the parameters are fully considered, deep fusion of system modeling and performance optimization can be achieved, optimization targets, design variables and constraint conditions in the model are automatically recognized and extracted, a standardized mathematical optimization model is formed, and the optimization efficiency is improved. The parameter configuration efficiency and the design decision quality in the complex system design are remarkably improved, and the method has good universality and expansibility.

The invention discloses a network security knowledge graph embedding method and device based on a mixed space, and the method comprises the following steps: S1, mapping entities and relationships of a network security knowledge graph into a hyperbolic space and a hyperspherical space at the same time; s2, respectively mapping points in the hyperbolic space and points in the hyperspherical space to a tangent space through logarithm mapping, and mapping vectors in the tangent space back to corresponding geometric spaces through exponential mapping; s3, the hyperbolic space embedding and the hyperspherical space embedding are switched to a tangent space through logarithmic mapping, the geometric information embedded in the tangent space and the geometric information in the Euclidean space are fused, and the fused geometric information is returned to a target geometric space through exponential mapping; and S4, carrying out weighted fusion on the geometric information in different geometric spaces by using an attention mechanism. According to the method, the capability of capturing complex relations and multi-layer structures of the atlas is enhanced.

A computer-implemented method for automatically determining a neural network architecture represents a neural network architecture as a data structure defining a hierarchical set of directed acyclic graphs in multiple levels. Each graph has an input, an output, and a plurality of nodes between the input and the output. At each level, a corresponding set of the nodes are connected pairwise by directed edges which indicate operations performed on outputs of one node to generate an input to another node. Each level is associated with a corresponding set of operations. At a lowest level, the operations associated with each edge are selected from a set of primitive operations. The method includes repeatedly generating new sample neural network architectures, and evaluating their fitness. The modification is performed by selecting a level, selecting two nodes at that level, and modifying, removing or adding an edge between those nodes according to operations associated with lower levels of the hierarchy.

A method for making predictions pertaining to entities represented within a heterogeneous graph includes: identifying, for each node in the heterogeneous graph structure, a set of node-target paths that connect the node to a target node; assigning, to each of the node-target paths identified for each node, a path type identifier indicative of a number of edges and corresponding edge types in the associated node-target path; and extracting a semantic tree from the heterogeneous graph structure. The semantic tree includes the target node as a root node and defines a hierarchy of metapaths that each individually correspond to a subset of the node-target paths in the heterogeneous graph structure assigned to a same path type identifier. The semantic tree is encoded, using one or more neural networks by generating a metapath embedding corresponding to each metapath in the semantic tree. Each of the resulting metapath embeddings encodes aggregated feature-label data for nodes in the heterogeneous graph structure corresponding to the path type identifier corresponding to the metapath associated with the metapath embedding. A label is predicted for the target node in the heterogeneous graph structure based on the set of metapath embeddings.