254 results about "Bipartite graph" patented technology

The invention provides a resource scheduling control method and system for a big data server, and the method comprises the steps: constructing a multi-dimensional resource portrait module, collecting the CPU, memory, network, storage I / O load and task queue length of each node in real time, and predicting a resource demand trend through a time sequence algorithm; extracting characteristics such as calculation intensity, data dependence, memory requirements, network transmission quantity and the like; adjusting the weight coefficients of the resource utilization rate, the task completion time and the energy consumption efficiency according to the system load and the historical effect; establishing a bipartite graph model by taking a resource trend as a node feature and a task vector as an edge feature, and calculating a matching score through graph convolution and a multi-objective optimization function; the scheduling scheme is synchronized by adopting a consistency algorithm; automatic rollback and reallocation are carried out when resources are detected to be insufficient; and optimizing a weight coefficient and a network parameter through reinforcement learning. Through the method, the system resource utilization rate can be improved, the task execution efficiency is improved, the overall scheduling effect stability is improved, and the system fault recovery time is shortened.

The invention discloses a multi-agent cooperation enhancement method, system and equipment based on a knowledge graph, and the method comprises the steps: obtaining original data in an external environment, carrying out the preprocessing and feature extraction of the original data, generating a knowledge triple, storing the knowledge triple in a local knowledge graph, and submitting the knowledge triple to a shared knowledge graph for knowledge updating; when a to-be-executed task is received, decomposing the to-be-executed task by utilizing the large language model and querying global knowledge in the shared knowledge graph and local knowledge in the local knowledge graph to obtain a plurality of sub-tasks; a bipartite graph minimum cost matching algorithm is adopted to match a plurality of sub-tasks with the capability and availability of each agent to generate a preliminary task allocation scheme, and a large language model is utilized to optimize the preliminary task allocation scheme to generate an optimal task allocation scheme; and sending each task allocation knowledge fragment in the optimal task allocation scheme to a corresponding agent for collaborative execution through a semantic communication protocol.

The invention discloses a three-dimensional shielding target tracking method based on multi-modal space-time interaction, and relates to the technical field of target tracking. The method comprises the following steps: acquiring a point cloud and an image and preprocessing to obtain global fusion features; obtaining an initial detection frame and region-of-interest features through region proposal network processing; projecting the non-empty voxel point cloud to the image features, and reconstructing shielded target features; convolution and neural network processing are utilized to obtain a refined detection frame; screening legal detection frames through distance calculation and legality judgment; the bipartite graph and the self-adaptive channel graph are adopted for convolution, and appearance correlation scores are calculated; and matching the detection frame and the trajectory based on a Hungary algorithm to realize whole-course tracking. The target identification accuracy and robustness are improved, the shielding problem is solved, the accuracy of the detection frame is ensured, the correlation accuracy is improved by using the bipartite graph and the adaptive convolution, the nodes are matched in combination with the geometric cost matrix, and whole-course tracking and error calibration are realized.

The invention provides a self-adaptive intent perception enhanced graph contrast learning recommendation method, which comprises the following steps of: constructing a user-article bipartite graph, carrying out multi-layer embedded coding on a user and an article by adopting a graph neural network, and obtaining a multi-intention embedded representation of Gaussian distribution through a variational auto-encoder in combination with a multi-intention hypothesis; noise disturbance is adaptively added to intention embedding so as to enhance feature robustness, and the method respectively implements comparative learning of isomorphic and heterogeneous nodes in a node interaction space domain and an intention perception domain, so that the problems of data sparsity and intention entanglement are effectively relieved; and finally, carrying out joint optimization on recommendation task loss, KL divergence loss and double-domain comparison loss, and realizing accurate modeling and recommendation of the personalized preference of the user. Experimental results show that the method is superior to a mainstream recommendation system on a plurality of real data sets, and has strong generalization ability and robustness.

The invention relates to the technical field of multi-view clustering, in particular to a multi-view clustering method and device, and can solve the problem that the overall effect of an existing method in a large-scale clustering task is limited due to the fact that the existing method has problems in the aspects of calculation efficiency, robustness and multi-view information integration to a certain extent. The method comprises the following steps: dynamically learning an anchor matrix and a projection matrix for each view, and constructing a bipartite graph to generate a similarity matrix; calculating a graph Laplacian matrix based on the similarity matrix of each view, and extracting spectrum embedding; the spectrums of multiple views are embedded and stacked into a third-order tensor, and cross-view shared information is extracted by using a low-rank tensor constraint; multi-view atlas embedding is aligned through a spectrum rotation technology, and a discrete clustering indication matrix is directly output.

The invention discloses a large language model semantic enhancement collaborative filtering recommendation method based on comparative learning, and belongs to the field of recommendation systems.The method comprises the steps that a user-item bigraph is constructed, user-item semantic information collaborative information is provided through GNN, user-item semantic information is extracted through cue words, a deterministic topology view enhancement strategy is adopted, and user-item semantic information collaborative filtering recommendation is achieved. According to the method, a semantic neighbor extension view is generated, a semantic neighbor reconstruction view is generated, collaboration and semantic information alignment are performed, double-view structure alignment is performed, and a loss function is integrally trained, so that the accuracy and the cold start capability of a recommendation system are remarkably improved through collaborative graph structure learning and semantic enhancement.

The invention relates to an identification method for matching scene behaviors by using multi-modal features, and belongs to the technical field of scene behavior matching. The method comprises the following steps: acquiring multi-modal scene behavior data, and carrying out noise self-adaptive purification processing on the multi-modal scene behavior data to obtain a preprocessed scene image, scene audio data and a scene label text; secondly, performing feature collaborative extraction on the preprocessed data to obtain visual features, audio features and text features, inputting the cooperatively extracted features into a scene behavior matching network, and performing scene feature fusion and behavior feature fusion respectively to obtain a scene feature vector and a behavior feature vector; and constructing a bipartite graph according to the scene feature vector and the behavior feature vector, calculating the semantic similarity between nodes of the bipartite graph, dynamically updating the edge weight of the bipartite graph according to the semantic similarity between the nodes, and normalizing the updated bipartite graph to obtain a scene behavior matching result. According to the method, the association degree of the scene and the behavior can be accurately quantified, and the accuracy of a matching result is greatly improved.

The invention provides a method for detecting the quality of a foundation steel bar welded joint, and belongs to the technical field of steel bar welded joint quality detection.The method comprises the steps that a stress distribution matrix is obtained through ultrasonic detection equipment, and a fatigue distribution matrix is established in combination with a fatigue analysis algorithm; a stress fatigue correlation analysis method and a bipartite graph maximum matching algorithm are used to construct a stress fatigue correlation matrix, a damage evaluation deep learning model based on a Transform architecture is used to generate a damage distribution matrix, and a time sequence prediction algorithm is used to consider a fatigue center change effect to establish a damage expansion trend matrix. A damage prediction matrix is deduced through a damage prediction deep learning model of a bidirectional long-short-term memory network, and finally, a quality detection result matrix is comprehensively evaluated and output by utilizing a game model, a multi-objective optimization algorithm and a gating weighting function, so that the technical problem that the quality detection accuracy of the steel bar welded joint is not enough is solved.

The invention discloses a recommendation system noise trimming and long tail enhancement method based on two-stage graph optimization. In order to solve the problems of noise interaction and long-tail user data sparsity in an implicit feedback recommendation system, the method comprises the following steps: firstly, constructing a user-article interaction bipartite graph, and initializing a graph convolutional network model to generate a preliminary embedded representation; in the first stage, the reliability of an interaction edge is evaluated through a node similarity index (Nsim), a noise edge is trimmed in combination with a dynamic threshold strategy, and a de-noised subgraph is generated to improve the embedding quality. And in the second stage, for the long-tail user, a probability sampling mechanism is adopted to add a high-confidence potential interaction edge, and an enhanced sub-graph is generated to improve the long-tail recommendation effect. Finally, Bayesian personalized ranking (BPR) loss is optimized through iterative training, and an accurate personalized recommendation result is generated. The accuracy and fairness of the recommendation system are remarkably improved, and the method is suitable for application scenes such as e-commerce, social media and content recommendation.

The invention provides a multi-feature factor matching and interrupt processing method for storm tracking. The method comprises the following steps: calculating storm features required by tracking; carrying out discretization processing and coding on the storm features, and constructing a sparse matrix; constructing an auto-encoder model, performing data reconstruction on the sparse matrix based on the auto-encoder model, and calculating main features of the single storm; on the basis of the main features, calculating the feature vector distance between the previous and later time secondary storms, and constructing a feature space distance matrix; a bipartite graph algorithm is adopted, the incidence relation of the previous and later time-order storms is obtained based on the feature space distance matrix, and a storm matching result is constructed; identifying an interrupted storm based on a storm matching result and carrying out secondary matching; and constructing a storm trajectory based on a secondary matching result and removing a repeated storm trajectory. According to the method, the movement track of the storm can be accurately tracked from formation to extinction of the storm, reliable data support is provided for meteorological monitoring and forecasting, and the timeliness and accuracy of severe convective weather early warning are improved.

The invention discloses an optimization solution system for a large-scale mixed integer linear programming problem, and the system is characterized in that the system comprises a problem decomposition module which is used for modeling the large-scale mixed integer linear programming problem into a bipartite graph for representation, and dividing the bipartite graph into a plurality of sub-graphs through a graph decomposition algorithm; the problem reduction module is used for obtaining a feature code of each sub-graph node, inputting the feature code into a pre-training neural network to obtain a variable prediction value, reducing the scale of the large-scale mixed integer linear programming problem through a redundancy constraint removal and coefficient priority variable fixing strategy based on the variable prediction value, and then outputting the problem; and the efficient solving module is used for solving the large-scale mixed integer linear programming problem based on the result output by the problem reduction module and weighted subgraph division. According to the method, the high-quality solution of the large-scale mixed integer linear programming problem can be quickly given, so that the high-quality feasible solution can be obtained within the acceptable time.

The invention provides a personalized recommendation method and system based on a large model. According to the method, the content data stream is disassembled into independent feature packets in different modals through the multi-channel acquisition server; and analyzing a cross-modal semantic relationship among the feature packets by using a large model, and establishing a cross-modal similarity corresponding table. And generating a dynamic compensation coefficient through deviation analysis in combination with a cross-modal conversion record in a historical behavior. And constructing a bipartite graph structure of the content and the user, converting a compensation coefficient into a connection strength adjustment factor, and integrating weights of adjacent nodes to generate an expression vector. And finally, extracting a user purpose, content association and aging characteristic value of the vector, constructing a dynamic weight model in combination with a preference change track, integrating multiple characteristics to generate a decision vector, realizing cross-modal matching sorting, and outputting a personalized recommendation result. According to the method, the cross-modal data deviation is dynamically compensated and the user track is fused, so that high-precision matching recommendation of the multi-modal content and the user demand is realized.

The invention discloses a medical image clustering method based on correlation entropy and an adaptive bipartite graph, and the method comprises the steps: collecting image data of medical imaging equipment, and forming a sample data matrix; selecting anchor point samples from the sample data matrix, and establishing an original sample low-dimensional representation-anchor point low-dimensional representation graph for capturing a local geometric structure of an embedded space; introducing an orthogonal basis matrix; carrying out joint modeling on the sample reconstruction error, the structure retentivity, the adaptive graph constraint and the discriminative features by adopting a joint objective function; iteratively updating the joint objective function; and inputting the finally obtained low-dimensional representation of the original sample as an embedded representation into a clustering device of K-means, and completing final clustering label distribution. According to the method, more stable, efficient and interpretable unsupervised clustering is realized, the clustering precision and generalization ability are improved, and the method is particularly suitable for medical image data analysis tasks of high-dimensional and complex structures.

The invention discloses a drug recommendation model construction method for relieving entity sparsity based on a hierarchical algorithm, and belongs to the technical field of drug recommendation, and the method comprises the steps: building a pre-training module, and obtaining the health representation of a patient; performing correlation layering; constructing a dual-attribute graph network representation stage; aggregating the drug graph representing the safety and the bipartite graph used for representing the drug disease / operation accuracy into dual-attribute representation in a graph network; recommending related medicine combinations to adapt to long-term health conditions; constructing a loss function, wherein the loss function comprises a prediction probability based on recommended drugs and cross entropy loss and DDI loss of a real label; and training the deep learning model to obtain predicted drug recommendation representation. By layering historical record data of a patient, capturing a sequence relation and forming an accurate graph structure and a safe graph structure for drugs, diseases and operations, the drug recommendation performance can be improved, and the calculation amount is reduced.

The invention provides a personalized semantic understanding learning method, medium and system under an AI platform, and belongs to the technical field of semantic understanding. According to the technical scheme, the personalized semantic understanding learning method comprises the steps that a personalized semantic file recording user question and answer habits and personal terms is constructed, and dual time decay values are set; a semantic understanding optimization model is established based on an attention mechanism, personalized feature vectors and input semantic vectors are fused, a concept drift detection algorithm adopting bipartite graph maximum matching and a Hungary algorithm is designed to monitor semantic habit changes in real time, and an incremental learning mechanism based on a variable sliding window is established to calculate a semantic difference matrix. A forgetting function based on a Gaussian kernel is applied to adjust the historical semantic feature weight according to the time distance and the use frequency, and a reinforcement learning feedback module is constructed to collect user satisfaction evaluation and generate reward signal optimization model parameters; and executing a self-iterative optimization process to periodically update semantic archives and models so as to continuously adapt to personalized semantic requirements of users.

The invention provides a confidence-guided adaptive graph representation reinforcement learning method, equipment and a medium, and relates to the technical field of label prediction and electric data processing. The method comprises the following steps: acquiring a crowdsourcing data set composed of multi-source heterogeneous electronic data, and inputting an adaptive graph representation learning model; mapping the crowdsourcing data set to construct an original bipartite graph; based on the original bipartite graph, generating a trimmed view and an enhanced view through dynamic confidence trimming and conservative edge addition; performing feature extraction, adaptive information enhancement and learning on the worker node and the task node through a dual-channel heterogeneous adaptive network, and performing prediction through a predictor to obtain an initial prediction label; and in combination with the original bipartite graph, performing label correction on the initial prediction label to obtain a final prediction label. The method can effectively suppress noise interference, capture local features and adapt to a dynamic environment, is remarkably superior to the prior art in the aspects of accuracy, robustness and calculation efficiency, and is suitable for the fields of multi-source heterogeneous intelligent labeling and the like.

The invention provides a method and a system for accurately converting a large-model natural language into an SQL (Structured Query Language) based on multi-modal fusion. The method comprises the following steps: S1, performing named entity recognition and dependency syntax analysis on a natural language text, performing vision-text dual-path processing on image data, and constructing a bipartite graph structure for table data so as to obtain a vision text composite feature and a table feature; s2, designing a modal attention controller and calculating the weight of each modal through a gating network; s3, constructing a knowledge graph, clearly recording all legal table names and column names in a database, and retrieving the knowledge graph in real time when a decoder generates the table names or the column names; s4, generating an SQL draft; s5, performing bidirectional verification correction on the SQL statement; and S6, outputting the legal SQL statement. According to the method, multi-modal information can be integrated, information loss and semantic deviation are reduced, illegal statements are avoided through grammar and semantic dual verification, and the overall reliability of SQL statement generation is improved.

The present application discloses a risk account determination method and apparatus, a device, a storage medium and a program product. The risk account determination method comprises: constructing a transfer network graph on the basis of transfer information among a plurality of accounts, wherein the plurality of accounts comprise at least one known risk account; converting the transfer network graph into a corresponding bipartite graph, and constructing a quantum walk space on the basis of the bipartite graph; using a plurality of quantum bits to perform quantum state preparation on each node among a plurality of nodes, so as to obtain an initial quantum state corresponding to each node; on the basis of the initial quantum state, taking each node as a starting point respectively to perform a quantum walk in the quantum walk space, so as to obtain a final quantum state corresponding to each node; on the basis of the final quantum state, determining embedded information corresponding to each node; and on the basis of the embedded information, determining, from among the plurality of accounts, an account similar to the known risk account as a risk account.

The present disclosure provides methods and systems for modeling cellular behavior. A method for generating a model of a biological system may include obtaining sample data including records derived from samples of the biological system. The records may indicate the presence, absence, and / or expression levels of entities in respective samples of the biological system. The method may further include dividing the sample data into a training set and a validation set, providing biological system data as input to a machine learning model to initialize the model, training the model to model dynamic behavior of the biological system based on the training set, and validating the trained model using the validation set. The biological system data may include a bipartite graph representing the biological system and structured as an optimal control loop.

The application discloses an intelligent decision support system based on big data and artificial intelligence, and relates to the technical field of decision support, comprising: a decision data acquisition module, which is used for acquiring decision data, and performing feature extraction and anchoring effect feature marking; a propagation path identification module, which constructs a decision participant opinion influence bipartite graph based on a graph neural network, identifies a core anchor source and a strong cascading propagation path thereof, and outputs an anchoring effect strength value; a weight adjustment module, which is used for generating a weight adjustment factor based on the anchoring effect strength value, generating a weight through iterative adjustment and correlation correction, and generating a decision scheme according to weighted statistics; and a decision correction module, which is used for decomposing a historical decision implementation effect signal to identify a deviation component, generating an counterfactual decision scheme based on causal inference, and correcting a deviation of an actual decision scheme; and effectively solves the information cascade and collective irrationality problems caused by over-amplification of a few opinions or authority influence in traditional group decision-making.

Methods, systems, and computer program products are provided for bipartite graph pre-trained dual transformers. Graph data for a bipartite graph is received, the bipartite graph including a first node / entity, a second node / entity, and an edge. The first entity context of each first entity includes a second entity connected to the first entity through the edge. The second entity context of each second entity / node includes the first entity connected to the second entity through the edge. A first embedding of each first entity is generated. A first encoded representation of each first entity is generated based on the first entity context and a first transformer encoder. A second embedding is generated for each second entity. A second encoded representation of each second entity is generated based on the second entity context and a second transformer encoder. The first embedding and the first transformer encoder are adjusted based on a first contrast loss. The second embedding and the second transformer encoder are adjusted based on a second contrast loss.

The invention discloses a large language model alignment method based on multi-user preference dynamic balance, and the method comprises the steps: collecting user interaction data, preference comparison data and public policy question and answer data in a livelihood consultation scene, and dividing the data into structured preference data and unstructured data; preprocessing the divided data to generate a user feature vector and a response feature vector; constructing a user-response bipartite graph, performing multi-hop preference propagation through a graph neural network, capturing potential association among users, and outputting a user preference embedding vector; embedding a vector selection expert path according to user preference through a dynamic gating function, and generating a personalized response; and through an optimization-free embedding aggregation strategy, similar users are retrieved based on a graph structure, and the embedding of the similar users is weighted and aggregated, so that rapid cold start of new users is realized. The precision and efficiency of livelihood consultation can be remarkably improved, and the method is particularly suitable for smart city scenes with high-frequency policy updating and frequent user flow.

This application discloses a method, apparatus, electronic device, and storage medium for determining anomaly node sets. By constructing a target bipartite graph using the frequency of similar transactions occurring between similar nodes as the weight of the edges, the target bipartite graph can carry more information. Furthermore, by dividing the target bipartite graph into communities, the set of anomaly nodes can be determined by combining the characteristics of different transaction initiating nodes and service providing nodes, effectively improving the accuracy of anomaly node determination. Moreover, anomaly nodes are determined through the first distribution characteristics of service providing nodes in the target node set. On the one hand, service providing nodes play a passive role in transactions, reducing the imitability of transactions; on the other hand, the first distribution characteristics can accurately reflect the transaction characteristics of service providing nodes, thereby further improving the accuracy of anomaly node determination. This method can be widely applied in technologies such as cloud computing and artificial intelligence.

The application relates to a recognition method for matching scene behaviors by using multi-modal features, and belongs to the technical field of scene behavior matching. The method acquires multi-modal scene behavior data, and carries out noise self-adaptive purification processing on the multi-modal scene behavior data, so as to obtain preprocessed scene image data, scene audio data and scene label text. Then, feature collaborative extraction is carried out on the preprocessed data, so as to obtain visual features, audio features and text features. The collaboratively extracted features are input into a scene behavior matching network, scene feature fusion and behavior feature fusion are respectively carried out, scene feature vectors and behavior feature vectors are obtained, a bipartite graph is constructed according to the scene feature vectors and the behavior feature vectors, the semantic similarity between nodes of the bipartite graph is calculated, the edge weight of the bipartite graph is dynamically updated according to the semantic similarity between the nodes, the updated bipartite graph is normalized, and a scene behavior matching result is obtained. The application can accurately quantify the correlation degree of scenes and behaviors, and greatly improves the accuracy of the matching result.

The present invention belongs to the field of data clustering technology, and provides a method and system for industrial coal-fired boiler operating condition division based on data integration clustering to obtain the operating status data of the industrial coal-fired boiler; the present invention performs integrated clustering operations on the industrial coal-fired boiler data through three steps of mixed representative nearest neighbor similarity, bipartite graph segmentation and third-order tensor integration, and realizes effective operating condition division of the boiler data; specifically, by constructing a sparse affinity submatrix through mixed representative nearest neighbor similarity, it can solve the problem that the traditional clustering method has too high computational time complexity and cannot effectively construct the affinity matrix of the coal-fired boiler data; by dividing the bipartite graph, the time for solving the characteristic problem is reduced; and by integrating the multi-base clustering results into a unified integrated clustering framework, the accuracy and robustness of the clustering are further improved while maintaining high efficiency.

A three-dimensional measurement method based on spatial coding includes the following steps: S1. Designing a spatially coded projection pattern in the form of a three-dimensional honeycomb structure and saving a set of intersection coding results; S2. Calibrating the intrinsic and extrinsic parameters of the camera and projector in a structured light system, projecting the spatially coded projection pattern using the projector, and synchronously capturing images using the camera; S3. Extracting an image skeleton from the captured image, extracting intersections based on the image skeleton, and classifying the intersections into two categories of intersection sets; S4. Drawing a bipartite graph based on the captured image and the intersection set using a connecting line traversal algorithm, and calculating an intersection neighborhood decoding result set based on the bipartite graph; S5. Using a matching algorithm, calculating an intersection pairing set based on the intersection neighborhood decoding result set and the intersection coding result set; and performing three-dimensional measurement of the intersection pairing set using a triangulation method, combined with the intrinsic and extrinsic parameters of the camera and projector, to generate a point cloud. This method solves the problem of non-uniform intersections caused by distortion in traditional grid coding and accelerates decoding speed.

The application relates to a drug-lncRNA relationship prediction method and system based on embedding constraints, and the method comprises the following steps: collecting lncRNA-drug association data and preprocessing, and constructing a data set; based on the data set, an lncRNA-drug association matrix is constructed, and an lncRNA-drug bipartite graph is extracted; based on the lncRNA-drug association matrix, an lncRNA similarity matrix and a drug similarity matrix are obtained, and dimension reduction processing is performed; the lncRNA-drug bipartite graph and the dimension reduction processed lncRNA similarity matrix and drug similarity matrix are input into an LDA-GNN model and an LDA-DHG model with an embedding constraint strategy, and the probability score of lncRNA and drug association is obtained; and based on the probability score, the relationship prediction of the drug-lncRNA is realized. Through the embedding constraint strategy, the prediction performance and the distinguishing ability of the model are significantly improved.

The invention belongs to the technical field of machine learning and data mining, and particularly relates to a large-scale high-dimensional image data set-oriented compressible subspace clustering method, which comprises the following steps of: firstly, designing a dictionary representation learning model based on a partitioning mechanism to select part of samples to construct a small-scale dictionary to replace the whole original data; a bipartite graph construction method is ingeniously introduced by utilizing the thought of joint clustering, the problem that the bipartite graph cannot be directly constructed due to the fact that a coefficient matrix is not a square matrix is solved, and the relevance between a dictionary sample and a new input data sample can be fully considered. Under the Laplacian matrix rank constraint of the combination graph, the method can directly learn to obtain an optimal structured bipartite graph, can directly obtain a final clustering result, and does not need any post-processing process. In addition, an efficient optimization algorithm based on alternate iteration is further designed in combination with an augmented Lagrangian multiplier method to solve the model.

A method for graph neural network-based recommendation is disclosed. The method comprises the following steps: receiving a user-project bipartite graph, wherein the user-project bipartite graph comprises a user node set representing a user set, a project node set representing a project set, and an edge set representing interaction between users and projects; obtaining user embedding of each user in the user set and project embedding of each project in the project set; performing an intra-layer aggregation process to generate an aggregated user embedding of each user in the user set and an aggregated item embedding of each item in the item set based on different weight coefficients of each user and each item; executing an interlayer propagation process to generate a propagation user embedding of each user in the user set and a propagation item embedding of each item in the item set based on different weight coefficients of each propagation layer; and making recommendations for the users in the user set based on the propagation user embedding and the propagation item embedding.

The invention discloses a classification and grading semantic enhanced dynamic graph community offset data access anomaly detection method and device. The method comprises the following steps: slicing an internal access log according to a time window, constructing a bipartite graph snapshot consisting of user nodes and data asset nodes, and classifying and grading sensitivity semantics for asset association; edge risk weights are calculated according to access statistical characteristics and sensitivity, and risk weights are introduced in graph neural network message passing to learn node embedding; community division is carried out based on embedding in each time slice, a stable community evolution trajectory is obtained through time sequence smoothing, a community offset degree and a node-community consistency deviation are calculated to form a node anomaly score, and an alarm is output. According to the scheme, structural anomalies such as collusion access, abnormal cross-community access and community migration can be identified, and interpretable clues on the community level are provided.