65 results about "Graph encoding" patented technology

The application discloses a kind of based on graph coding's diagnosis and surgical matching check method and system, the method includes: obtaining the diagnosis data and surgical plan data of target object;Based on feature extraction algorithm, extract the diagnosis change feature and surgical step feature in the diagnosis data and the surgical plan data;Based on graph network coding algorithm, the diagnosis change feature and surgical step feature are encoded into corresponding node relationship graph;The node relationship graph is input into trained graph neural network to obtain the matching degree parameter of output diagnosis and surgical plan.It can be seen that the present application can realize accurate medical diagnosis and surgery matching evaluation based on graph structure feature association, improve the scientificity and adaptability of surgical plan selection, reduce the surgical risk and treatment deviation caused by mismatch between scheme and diagnosis.

This invention relates to the field of bioinformatics, and particularly to an artificial intelligence-based drug relocation method, apparatus, device, and storage medium. The AI-based drug relocation method of this invention first constructs a heterogeneous network, and then builds a relocation model including a multi-relational heterogeneous graph encoder, a multi-head attention mechanism module, and an interactive decoder. Based on the heterogeneous network, it predicts the treatment score of a drug for a disease, achieving better test metrics compared to other methods. Furthermore, this invention interprets the prediction results, identifying key functional nodes and explaining the pharmacological mechanism of drug treatment for diseases.

The application discloses a large model access control system and method based on calling intention analysis, relates to the technical field of natural language data processing, and standardizes, cleans and unifies the format of user input text, and obtains an intention vector based on an intention encoder constructed by a large model. Then, the intention vector continuously generated is arranged in sequence to form a discrete point sequence with a dialogue round as a discrete time axis, and the end points are connected to form a dialogue moving track through sliding window maintenance. The curvature of each point of the track and the maximum curvature in the sliding window are calculated by processing the obtained digital data. A history normal curvature library is established for each user, a personalized curvature threshold is obtained, and an abnormal curvature segment is marked, a first-level early warning is triggered through a continuous abnormality counter, a second-level early warning is triggered and a graded response is executed through accumulated abnormal curvature integration and variance determination. The application achieves the purposes of reducing the false positive rate and improving the accuracy and reliability of large language model security detection.

The application provides a kind of full-body control method and system of leg type operating robot based on graph model, belongs to robot control field, method includes: with the perception unit of robot and actuating mechanism establish node set, with the physical structure connection relationship of robot establish edge set, with the tensor splicing of multi-source perception data obtain the feature vector of node, the feature vector of node is made into the feature matrix of node, graph model is constructed with node set, edge set and the feature matrix of node;Graph model is input into permutation invariant graph encoder and is encoded, to obtain fusion feature;The task to be controlled is decomposed into subtask using high-level planner in strategy network, in each subtask, determine trigger condition according to fusion feature, according to trigger condition, output corresponding instruction;According to corresponding instruction, using low-level controller outputs corresponding action to control robot.The application realizes the internal collaborative optimization of movement and operation by dynamically aggregating the heterogeneous features of each node.

This invention discloses an AI-based integrated system for intelligent classification, storage, and retrieval of archives. The system includes: an archive classification module that inputs archive text into a MacBERT model for classification; a category template generation module that determines the target entity field identifier set based on archive category labels; a named entity recognition module that generates structured archive metadata; a multi-view encoding module that generates view pooling vectors; a multi-view interactive sentence vector generation module that generates sentence vector representations of target archives; a vector index library module that stores sentence vector representations of target archives; a structured database module that stores structured archive metadata; and an archive retrieval module that outputs retrieval results by combining structured archive metadata during the retrieval phase. This invention is applicable to automated management and semantic-level intelligent retrieval scenarios for large-scale government and enterprise archives.

The application relates to a heterogeneous network key node identification method and device based on decoupling causal interaction, which comprises the following steps: designing an intention decoupling graph encoder, introducing a generation mechanism of a variational graph autoencoder, mapping node features of a heterogeneous network into a plurality of mutually independent Gaussian distributions, performing differentiable sampling on the node features through reparameterization sampling, and physically separating the node features; based on a contrast learning mechanism of mutual information maximization, regarding each layer of the heterogeneous network as a view, using the contrast learning mechanism to compare and learn mutual information of decoupled representations of the same node under different views, and filtering random noise edges appearing only in a single layer; constructing a structural causal model, identifying mixed factors based on the structural causal model, estimating the weight of the edge by adopting an inverse propensity weighting strategy, and performing causal readjustment on the weight of the edge to identify key nodes that truly have structural control power. The application has the effects of realizing semantic decoupling, structural denoising and causal correction, and improving the accuracy and robustness of network key node identification.

This invention provides a method, system, device, and storage medium for smart contract vulnerability detection, belonging to the fields of blockchain security and artificial intelligence technology. It includes extracting an abstract syntax tree, control flow graph, and data dependency graph from the original smart contract code; obtaining a composite graph object based on the abstract syntax tree, control flow graph, and data dependency graph; processing the composite graph object based on node semantics and function call relationships to obtain a hierarchical heterogeneous graph; constructing a heterogeneous graph Transformer network; using a heterogeneous graph Transformer encoder to process the hierarchical heterogeneous graph for encoding node and edge information, obtaining node representation vectors that fuse semantic and structural information; extracting multi-scale features from the node representation vectors; fusing the extracted multi-scale features using learnable weights; and processing the fused multi-scale features using a base classifier cluster to obtain vulnerability type prediction results and confidence levels. This invention achieves full-process optimization from structural modeling and semantic fusion to decision reasoning, significantly improving the accuracy and interpretability of smart contract vulnerability detection.

The application discloses a main bearing damage diagnosis method based on flow-heat coupling dynamic topology and a physical information graph autoencoder, and comprises the following steps: one-dimensional residual feature extraction is performed on multi-source SCADA time series data to obtain time series feature vectors of nodes; a Nusselt number is calculated, edge weights of a directed adjacency matrix are dynamically calculated based on the Nusselt number, a temperature gradient gating function and a thermal resistance parameter, and a dynamic flow-heat topology graph is constructed in combination with the time series feature vectors; a graph encoder is used to perform spatial aggregation on neighbor node features along a directed edge to obtain a predicted temperature of the main bearing at the next moment; a total loss function is calculated based on mean square errors of the predicted temperature and an actual temperature and physical consistency loss, and training is constrained to obtain a state reconstruction model; a residual error between the predicted temperature output by the model and the actual temperature, and a decay factor calculated based on a bearing fatigue life equation are used to dynamically adjust a warning threshold, and damage warning is triggered when the residual error exceeds the threshold.

The application discloses an e-commerce platform consumer behavior prediction method based on edge computing, comprising the following steps: collecting the behavior event flow of consumers in the e-commerce platform at the corresponding edge node of user access, and performing preprocessing; constructing a session state feature vector; constructing a local behavior transition graph for a micro-session segment and performing timing diagram coding on the local behavior transition graph; aligning and calculating the local behavior representation vector with a long-term preference anchor point set; generating a state digest based on an edge side prediction network and a collaborative routing indication value, and sending the state digest to a cloud node; the cloud node performs collaborative prediction in combination with the historical behavior characteristics stored in the cloud; collecting real behavior feedback of the edge side prediction result or the cloud prediction result to obtain difficult samples, generate a parameter correction amount or an anchor point correction amount, and write back to the edge node. The edge collaborative prediction realizes accurate identification of consumer behavior.

This application relates to a method and system for reconstructing watertight meshes from building point clouds based on multi-scale graph networks, belonging to the field of 3D mesh reconstruction technology. The method for reconstructing watertight meshes from building point clouds includes: acquiring and processing the original building point cloud and combining it with a radius adaptive mechanism to generate a building point cloud map; analyzing the building point cloud map according to a multi-scale graph encoder and outputting a composite code for building features; generating initial triangular patches based on a joint classifier calculation and the composite code for topological building features, and determining the type of holes; parsing and repairing all holes according to a hole repair mechanism, projecting the building boundary to close the bottom surface, and generating an initial watertight mesh for the building; optimizing the initial watertight mesh for the building based on a mixture of curvature diffusion terms and sparse constraint terms to obtain the optimal watertight mesh for the building; verifying the optimal watertight mesh for the building through topology, and generating a material property table and mesh partition labels; and solving the problems of low efficiency, numerous holes, and poor watertightness in existing technologies by using an end-to-end fusion of geometric repair and semantic optimization algorithms.

The application relates to the technical field of dialogue sentiment computing and natural language processing, and particularly discloses a dialogue sentiment-reason pair extraction method based on a double-channel graph encoder network. The method comprises the following steps: preprocessing an input dialogue to construct a dialogue directed graph; inputting dialogue features and an adjacency matrix into a global structure learning channel and a local feature extraction channel in parallel, and modeling long-term dependence and local implicit clues respectively; integrating double-channel features through an adaptive gating fusion module; calculating matching scores of all candidate sentiment-reason dialogue pairs based on the integrated features, and outputting final sentiment-reason pairs.

This invention relates to the fields of automated machine learning and neural network technology, specifically to a method and system for searching neural network structures based on dynamic collaborative graph encoding. This method leverages the collaborative evolution mechanism of maintainer, sterile, and restorer lines in breeding optimization algorithms, achieving inter-population structural migration through dynamic hybridization probability control. A support vector machine surrogate model is constructed to replace time-consuming performance evaluation, rapidly predicting the accuracy of offspring structures and screening high-potential candidate structures. Gradient fine-tuning is applied to candidate structures to output the optimal network structure. Modular recombination and topological evolution of the neural network structure are achieved through graph encoding. Utilizing the collaborative mechanism of maintaining the stability of the maintainer line, exploring the diversity of the sterile line, and accelerating the convergence of the restorer line, combined with a two-layer dynamic regulation of population-level migration control and individual-level variation optimization, this method efficiently obtains high-performance deep neural network structures.

This invention relates to a method and system for rapid prediction of 3D flow fields of aircraft based on graph encoder networks, belonging to the interdisciplinary field of computational fluid dynamics and deep learning. This invention extracts unstructured meshes into point cloud data, defines the features and weights of graph nodes and edges, and constructs a graph structure using the K-nearest neighbor algorithm. It designs a fusion network containing a graph encoder and decoder, introducing normalization, pooling aggregators, and DropEdge techniques to deconstruct flow field information and avoid overfitting. Through redundant channel identification, model pruning, and hierarchical point-by-point decoder design, it achieves network lightweighting, ensuring prediction accuracy while improving computational efficiency, adapting to practical engineering needs. This invention also provides a system for implementing the above method, including electronic devices and computer-readable storage media, which can be deployed on conventional computing platforms to achieve rapid prediction of 3D flow fields of aircraft.

The application discloses a knowledge graph enhanced retrieval method based on a value subgraph and structure perception, and comprises the following steps: 1) designing a whole framework of a graph enhanced retrieval method based on a value subgraph and structure perception; 2) retrieving a value subgraph, first obtaining a globally optimal value subgraph under the constraints of reward, cost trade-off and connectivity by using a PCST algorithm, and then performing propagation expansion based on similarity from a high correlation seed by using an AdaProp algorithm to obtain a final value subgraph; 3) encoding the value subgraph, encoding the value subgraph into a vector representation containing text semantic information and graph structure information at the same time to generate a graph-level representation; and 4) generating an answer of a large model: taking the graph-level representation as a soft prompt to guide the large model to generate a final answer. The knowledge graph enhanced retrieval technology disclosed by the application can significantly improve the accuracy of a large model knowledge question and answer.

This invention provides a method for constructing a logic optimization generation system based on reinforcement learning. The method includes: step S1, obtaining a training dataset; step S2, constructing an initial generation model, which includes a graph encoder, an average pooling layer, a word segmenter, and a sequence decoder; step S3, using the training dataset to iteratively train the initial generation model multiple times until convergence; step S4, using the converged initial generation model as a reference network and initializing a policy network and a value network, and using a reinforcement learning mechanism with human feedback to fine-tune the policy network and the value network, and using the fine-tuned policy network as a logic optimization generation system. The fine-tuning process includes multiple rounds of iterative optimization. In each round of iterative optimization, the graph structure representation of the logic circuit in the training dataset is used as the initial state input reference network, the policy network optimized in the previous round, and the value network to perform multiple interactions.

Some embodiments of a method may include: decoding features representing Region-Adaptive Hierarchical Transform (RAHT) coefficients corresponding to a point cloud frame at a current resolution; reconstructing the RAHT coefficients corresponding to the point cloud frame at the current resolution; and performing a RAHT inverse transform based on the reconstructed RAHT coefficients. Some embodiments of a method may include: computing RAHT coefficients based on attributes from a child octree level; performing feature extraction based on the RAHT coefficients using a neural network module; generating a feature map; and encoding the current feature map into a bitstream.

The application belongs to the technical field of Internet, and particularly relates to an industrial Internet cross-domain metadata trusted fusion method and system, comprising the following steps: collecting structured metadata, time series metadata and relationship metadata of each industrial domain node, and constructing an intra-domain metadata graph based on field attributes and call relationships; performing semantic coding and time series coding on the intra-domain metadata graph, obtaining corresponding semantic representations and time series representations, and calculating inter-field consistency; identifying stable field clusters, drift field clusters and abnormal nodes according to the inter-field consistency in each industrial domain node; and performing segmented aggregation on the stable field clusters and the drift field clusters by each industrial domain node to generate anchor field summaries and drift trajectory summaries. The application combines metadata graph coding, field clustering and federal dynamic weighting, and realizes industrial cross-domain metadata private trusted fusion through abnormal compensation correction.

The application discloses a bit parallel query push-down method based on nested document bitmap encoding, and belongs to the field of data encoding. The method comprises the following steps: acquiring a tree structure of a nested document and a mapping relationship between each parent-child node; encoding the mapping relationship between each parent-child node into a bitmap metadata column by using a bit mask mode; the bitmap metadata column is aligned with a data column; when a multi-predicate query is received, a predicate query is performed by using a post-order traversal sequence, and query load propagation is performed based on the lowest common ancestor of each two nodes positioned by the predicate. According to the scheme, the mapping relationship is encoded into the bitmap metadata column aligned with the data column by using the bit mask mode, so that the assembly overhead and intermediate materialization can be reduced; in addition, the predicate query is performed by using the post-order traversal sequence, and the query load propagation is performed based on the lowest common ancestor of each two nodes positioned by the predicate, so that the bit parallel predicate evaluation can be realized, and the query execution efficiency is greatly improved.

The application discloses a KEGG pathway constrained dynamic graph structure optimization cross-patient cell annotation method, comprising the following steps: obtaining and preprocessing single-cell RNA sequencing data of a reference patient and a query patient; calculating a functional activity score of cells based on a KEGG pathway classification system, and constructing a KEGG function semantic matrix; constructing an initial sparse kNN cell graph; using the KEGG function semantic vector as a biological constraint, dynamically adjusting the connection weight between cells through a multi-view gated similarity calculation and an iterative optimization mechanism, and generating a patient-specific cell graph containing function semantic information; constructing a shared graph encoder and a classifier, and introducing a Wasserstein discriminator for adversarial training to realize cross-domain alignment of embedding distribution of the reference patient and the query patient; and finally predicting the type of the query patient cell by using the trained model. The application solves the problem of significantly improving the accuracy, robustness and biological interpretability of cross-patient cell annotation.

Methods and apparatus are provided for unified significance map coding. An apparatus includes a video encoder (400) for encoding transform coefficients for at least a portion of a picture. The transform coefficients are obtained using a plurality of transforms. One or more context sharing maps are generated for the transform coefficients based on a unified rule. The one or more context sharing maps are for providing at least one context that is shared among at least some of the transform coefficients obtained from at least two different ones of the plurality of transforms.

This application provides a virtual interactive response system and method based on speech and semantic recognition, relating to the field of speech and semantic recognition technology. Responding to a acquired real-time speech stream, it extracts acoustic feature vectors, performs real-time speech recognition, and generates a first text sequence; it extracts word embedding features, determines nonlinear coupling weights, and generates a fused semantic vector; it performs sequence editing operations on the first text sequence based on the fused semantic vector to generate a second text sequence; it performs domain-knowledge-enhanced semantic parsing on the second text sequence to generate intent encoding and entity encoding; based on the intent encoding, it retrieves structured response data from a preset business knowledge graph; and it drives the virtual interactive object to perform corresponding action feedback based on the structured response data. This application can dynamically locate spoken language correction boundaries using acoustic paralinguistic features, and improves the robustness and accuracy of virtual interactive responses through closed-loop linkage of cross-modal information and semantic reconstruction.

A virtual power station false data injection attack cooperative defense method, in the offline stage, respectively construct and train KNN classifier containing feature construction unit, distance calculation unit, near neighbor voting unit and abnormal marking unit, and graph autoencoder containing graph signal mapping unit, node mask unit, graph encoder and graph decoder, in the real-time defense stage, through the trained KNN classifier, the measured vector polluted by false data injection attack (FDIA) is classified, the abnormal data node set is obtained, and the node set is mapped into graph signal, then through the trained graph autoencoder, data reconstruction is carried out according to the graph signal.The present application can accurately locate the aggregation unit node attacked after the system is attacked, and then restore the maliciously tampered measurement data, that is, reconstruct the operation state of the virtual power plant, effectively defend the false data injection attack, and fundamentally improve the situation awareness ability, survivability and safe operation level of the virtual power plant when it is attacked by the false data injection attack.

An example computer-implemented method includes: obtaining environment data descriptive of one or more traffic signal devices of a traffic control node in an environment of an autonomous vehicle; generating a control node graph based on the environment data including vertices respective to representations of the traffic signal devices and edges indicative of relationships between the representations of the traffic signal devices; providing the control node graph as input to a control node graph processing model operable to reduce the control node graph to a distilled representation of the control node graph encoding information about a state of the traffic control node; based on receipt of the control node graph as input, generating an output based on the control node graph processing model; generating a motion plan based on the output from the control node graph processing model; and controlling the autonomous vehicle based on the motion plan.

This invention provides a vulnerability assessment method for large-scale critical information infrastructure, comprising: collecting interaction event information of different large-scale critical information infrastructures; generating a temporal dynamic hypergraph based on the interaction event information; constructing a temporal hypergraph encoder, and training the temporal hypergraph encoder by combining a temporal dynamic hypergraph, a counterfactual hypergraph, and a self-supervised learning strategy; determining soft labels for node vulnerability based on dynamic propagation simulation and topological robustness analysis algorithms, wherein different soft labels represent different vulnerabilities; constructing a target inference model by combining the dynamic embedding representation of nodes output by the temporal hypergraph encoder with the soft labels, wherein the target inference model is used to determine the soft labels of each node in the temporal dynamic hypergraph, and generating a vulnerability ranking of multiple nodes based on the soft labels; and calling the target inference model to process the interaction event information of the infrastructure to be assessed to obtain the node vulnerability ranking of the corresponding target inference model.

The present application relates to the field of traffic scene license plate recognition, and discloses a wide dynamic light adaptive license plate recognition method and system, which comprises the following steps: collecting a wide dynamic scene image, generating a license plate feature enhancement graph through dynamic contrast stretching and local brightness compensation, and encoding the graph into a half-edge structure graph; running a light scene adaptive switching algorithm based on the graph, constructing a five-state hidden Markov chain to calculate state posterior probability, and outputting a scene confidence vector; activating a scene-specific AI model group to generate multi-model recognition results and construct a recognition partial order set; under the constraint of the partial order set, extracting and physically checking through the upper limit to complete license plate binding and unmanned order generation; the present application solves the problems of unclear character extraction, low recognition rate and the need for manual correction in the wide dynamic light scene, improves accuracy, reduces manual cost, and ensures smooth unmanned business.

A social e-commerce recommendation method based on multi-view network purification first constructs a user-product interaction network and a user-user social network based on user behavior and social data of the social e-commerce platform. Then, a dual-view denoising generator is used to process the user-product interaction network using implicit denoising with a variational graph autoencoder and explicit denoising based on social enhancement, resulting in two complementary purified interaction networks. Next, adaptive denoising is applied to the social network, using semantic similarity and structural similarity as metrics to eliminate false social edges. Then, a dual-graph encoder is used to learn purified user and product representations, and the representations are aligned through cross-view contrastive learning. Finally, a recommendation score is calculated based on the fused representations to achieve personalized product recommendations. This invention effectively mitigates the impact of noise on recommendation performance by integrating multi-view network denoising, graph neural networks, and contrastive learning, improving the recommendation performance of social e-commerce platforms in complex real-world environments.

This invention relates to a model training method for unsupervised graph domain adaptation tasks, belonging to the field of graph domain adaptation technology, and solves the problem of negative transfer affecting model performance in existing technologies. The method includes: acquiring labeled source domain graph data and unlabeled target domain graph data; pruning the source domain graph structure based on topological deviations; constructing an adaptive model, which includes a graph encoder, a domain discriminator, and a classifier; the graph encoder is used to encode features into the input graph data; the domain discriminator is used to determine whether the data comes from the source domain or the target domain based on the encoded features; the classifier is used to predict labels based on the encoded features; and the adaptive model is trained on the pruned source domain graph data and target domain graph data to obtain a classification model corresponding to the target domain. This significantly improves the classification accuracy of the model in the target domain.