31 results about "Graph inference" patented technology

Disclosed is a method and system for detecting malicious entities and malicious behavior in a time evolving network via a graph framework by modeling activity in a network graph representing associations between entities. The system utilizes classification methods to give score predictions indicative of a degree of suspected maliciousness, and presents a unified graph inference method for surfacing previously undetected malicious entities that utilizes both the structure and behavioral features to detect malicious entities.

The invention discloses a data parallelization processing method, system, and device and a storage medium. The method comprises the following steps: defining at least three first computing nodes witha logical relationship as a first parallel node group, the first parallel node group comprising a first front node and at least two first rear nodes; obtaining a first input data model of a first front node and generating a first input tensor of the first front node; calculating a first output tensor of the first front node according to the first input data model and the first input tensor; obtaining second input data models of at least two first back nodes and taking the first output tensor as a second input tensor; and respectively calculating second output tensors of the at least two firstrear nodes according to the second input data model and the second input tensor to obtain a first calculation result of the first parallel node group. According to the method, the technical effect offully utilizing CPU and operation card resources is achieved through overlapping operation of multiple input graph reasoning.

The invention relates to a question-answering method based on knowledge graph completion, and belongs to the field of natural language processing, and the method comprises the following steps: S1, dividing input Q into words or phrases; s2, representing the words as vectors by using a word vector model BERT to obtain a matrix as model input; s3, identifying entities in the Q by using an entity identification technology to obtain a candidate entity set; s4, querying the category of the eKGs, and replacing the entities in the Q with c; s5, constructing a declarative query cyber, and obtaining acandidate triple set so as to obtain a candidate relationship set; s6, linking based on the relationship between Qc and rij; s7, in the KGs, if the relationship between the eKGs and the rij is absent;s8, learning new vector representation of entities eKGs and entities in neighborhoods of the eKGs; s9, estimating the importance of entities in the neighborhood of the central entity; s10, executingrelation prediction on the basis of existing related triples; and S11, obtaining an answer A based on knowledge graph reasoning of entities and relationships.

The invention discloses a method for identifying the social relationship of a person in an image based on a graph inference model, which comprises the following steps: step S1, modeling the coexistence relationship between the social relationship and a semantic object to form a priori knowledge map; S2, combine high-level knowledge map and depth learning to form a graph reasoning network capable of end-to-end learn, and utilizing that propagation mechanism of the graph reasoning network to propagate and learn the information of nodes in the prior knowledge map to obtain weighted graph node output; Step S3, combining the output of the weighted graph nodes and outputting the distribution of the social relation labels of the target person pair through the classification network. By adopting the invention, the efficiency of identifying the social relation of the person in the image can be improved.

The embodiment of the invention discloses a scheduling method and a device based on deep learning node calculation and a storage medium. The scheduling method comprises the following steps: obtainingto-be-calculated nodes of a preset neural network calculation graph; judging the node type of the node to be calculated, wherein the node type comprises a hardware calculation node and a software calculation node; scheduling the hardware computing node to a first queue under the condition that the node type is the hardware computing node, and judging whether a hardware computing power module corresponding to the hardware computing node is occupied or not; and inputting the hardware computing node into the hardware computing power module for computing under the condition that the hardware computing power module is not occupied. According to the embodiment of the invention, asynchronization of a graph reasoning process is realized, and resources of software and hardware are fully utilized.

The invention discloses a method for solving a polymorphic statement video positioning task through a space-time graph reasoning network, and belongs to the field of natural language visual positioning. According to the method, firstly, a video is parsed into a space-time region graph, and the space-time region graph not only has implicit and explicit space sub-graphs of each frame, but also has across-frame time dynamic sub-graph; next, a text clue is added into the space-time region graph, and multi-step cross-modal graph reasoning is established; the multi-step process may support multi-order relational modeling; and thereafter, a temporal boundary of the pipeline is determined using a temporal locator, then the object is located in each frame using a spatial locator having a dynamic selection method, and a smooth pipeline is generated. According to the method, the video does not need to be trimmed when the natural language is positioned, so that the video positioning cost is reduced; and question sentences and declaration sentences can be effectively processed, technical support is provided for higher-level natural language processing and computational vision combined research(such as video questions and answers), and the application prospect is wide.

The invention relates to a knowledge graph intelligent question-answering method based on relation prediction, and belongs to the field of natural language processing. The method comprises the following steps: S1, inputting a problem Q, and preprocessing the problem; S2, identifying an entity equalization in the problem by utilizing an entity identification technology, and mapping the entity equalization to a corresponding entity eKGs in the KGs; S3, querying the category c of the entity eKGs in the KGs, replacing the entity equest in the problem Q with the category c, and marking the categoryc as Qc; S4, mapping a relationship r from the Qc; S5, in the KGs, if there is no relation between the entity eKGs and the relation r; S6, learning new vector representation of the center entity eKGs; S7, deducing a hidden relationship in the KGs based on the existing related triples; and S8, obtaining an answer A based on knowledge graph reasoning of entities and relationships. According to themethod, the corresponding relation between the question entity and the knowledge graph entity and the corresponding relation between the question natural language description and the knowledge graph semantic relation can be found.

The invention discloses a spatial pyramid graph convolution network implementation method, which comprises the following steps: firstly, generating an incidence matrix aiming at a deep network characteristic graph, and exporting an efficient calculation method by decomposing the incidence matrix and using a multiplication combination law, so that graph convolution can be directly carried out in anoriginal characteristic space. As a lightweight network, the method breaks through the limitation that graph convolution needs to be carried out in semantic nodes in previous methods, and further improves the expression ability of the network by carrying out graph reasoning on multiple scales. According to the method, the problem of insufficient perception domain of the full convolutional networkcan be effectively solved, and the performance of the full convolutional network is remarkably improved; according to the graph convolution scheme provided by the invention, significant performance is obtained in a density prediction task in computer vision, and examples and experiments fully verify the effectiveness and potential application value of the method.

The invention discloses an answer prediction method and device based on a graph reasoning model. The method comprises the following steps of receiving questions and a support document set; screening out irrelevant documents through text slimming, and performing semantic coding on all texts; semantic interaction of various texts and initialization of graph nodes are carried out by using various attention mechanisms; performing multi-hop reasoning by using a GNN information transfer algorithm based on the constructed graph; calculating a score for predicting an answer for each candidate based on the updated graph node representation; and predicting an unknown right entity, namely an answer, according to the score distribution result of the answer. A new graph is provided, multiple types of elements are regarded as graph nodes, and reasoning is more comprehensive. Meanwhile, due to the fact that sentence nodes are adopted, reasoning becomes more accurate and specific, multiple attention mechanisms are fused to carry out multiple semantic representation, the influence of relative correctness between candidates on reasoning is innovatively considered, and answer prediction is more accurate.

The invention discloses a video-paragraph retrieval method and a video-paragraph retrieval system based on a local-overall graph inference network, which belong to the field of cross-modal retrieval. The method mainly comprise the following steps of: 1) firstly preprocessing videos and texts (paragraph); 2) encoding given videos and texts by using a local-overall graph inference network to obtain final video features and text features; 3) calculating the similarity between the video features and the text features by using cosine similarity; 4) performing retrieval according to a similarity measurement result. Compared with a traditional video-paragraph retrieval method, the video-paragraph retrieval method has the advantages that the video and the text are decomposed into four layers of semantic structures respectively, the four layers of semantic structures are constructed into the local graph and the overall graph respectively, then the graph convolution network is used for conducting graph reasoning operation, and the result obtained in video-paragraph retrieval is better than that obtained in the traditional method.

The invention discloses a key target reasoning and predicting system based on situation data. The system comprises a target situation evaluation module, a reasoning model management module, a target intention reasoning module and a reasoning result evaluation and prediction module; the target situation evaluation module predicts the situation trend of the target in a future period of time according to the key situation elements of the real-time situation information of the key target by using a typical target activity rule model; the reasoning model management module is used for realizing management of a key target intention reasoning model; the target intention reasoning module takes the real-time situation data of the key target as input, screens proper key situation factors, and then generates the intention and action purpose of the key target through reasoning by using an activity rule model and a reasoning model, so as to realize action understanding and prediction of the national and regional ship aircraft targets of the sea battlefield object; and the reasoning result evaluation and prediction module supports the user to carry out manual evaluation on the situation prediction result and the key target intention reasoning result, and the evaluation result is fed back to the system.

The invention discloses a graph network cold start recommendation method, which comprises the following steps of: inputting pre-scored user-article data into a trained graph network to obtain a recommendation result, the training comprising the following steps: acquiring a sampling local graph of a node set or a local sub-graph to be trained; performing distance re-marking on the sampling local graph to obtain a re-marked label; distributing initial features to nodes of the sampling local graph; obtaining a prediction label and a prediction score of the initial feature; calculating a node classification error by using the prediction label and the remarking label, and calculating a score prediction error by using the prediction score and the real score of the article-user; and performing calculating by using the node classification error and the score prediction error to obtain an overall error, and updating parameters of the graph network by using the overall error. According to the method, through local graph sampling and double-task learning, inductive node reasoning and connection prediction capabilities are further realized on the basis of a deductive graph reasoning task, and the method has a feature representation capability for out-of-graph nodes.

The invention discloses a node storage method and system based on a neural network, a server and a storage medium. The method comprises the following steps: receiving a plurality of computing nodes with logical relationships sent by an off-chip memory; traversing a logical relationship among the plurality of computing nodes to confirm a first computing path and a second computing path, wherein thefirst computing path and the second computing path at least comprise one same starting node; judging whether the first calculation path and the second grounding path have the same ending node or not;if the same ending node exists in the first calculation path and the second calculation path, storing all nodes in the first calculation path and the second calculation path according to a first sequence; and if the same ending node does not exist in the first calculation path and the second calculation path, storing all nodes in the first calculation path and the second calculation path according to a second sequence. By optimizing the node storage sequence, the technical effects of improving the storage space utilization rate and accelerating the graph reasoning process are achieved.

The invention relates to the field of intellectual property technology resource conversion, in particular to a method for establishing a semantic network of intellectual property technology resourcesand providing an intellectual property technology resource recommendation service for enterprise users based on an intellectual map construction technology. The construction technology and method comprise the following steps: step 1, establishing an intellectual property technology resource technology concept label portrait; step 2, establishing a scientific research institution technology conceptlabel portrait; 3, establishing an enterprise technology concept label portrait; 4, establishing a semantic network of an intellectual property technology resource domain knowledge system; and step 5, performing graph reasoning in the intellectual property technology resource semantic network, and providing an intellectual property technology resource recommendation service for the enterprise incombination with a recommendation algorithm. According to the construction technology and method, the matching accuracy of enterprise requirements and intellectual property technology resources is improved, the working difficulty and operation cost of operators are reduced, and the intellectual property conversion rate is increased.

The invention discloses a photovoltaic module hot spot defect detection method based on a multi-scale feature map reasoning network, and the method comprises the steps: adding a self-designed multi-scale feature map reasoning module suitable for the hot spot defect detection of a photovoltaic power station into a feature fusion part in an original model based on an improved YOLOv4 neural network model; and multi-scale feature fusion is guided and complex background features in defect position area suppression images are highlighted, so that the identification capability of the hot spot defects of the photovoltaic module is effectively improved. According to the detection method, the deep learning technology and the image processing technology are combined, the low efficiency and uncertainty of traditional manual feature extraction are avoided, meanwhile, the detection process has high robustness, the detection precision is obviously improved, and the detection speed is increased.

The invention discloses a tense knowledge graph question answering method based on a subgraph, which comprises the following steps of: extracting and analyzing time information hidden in a question through knowledge in a tense knowledge graph, and replacing the question in a natural language format with a simplified question by using a regular expression; converting the problem added with the time constraint into vector representation, and obtaining candidate entities and semantic scores thereof by using a tense knowledge graph scoring function and a time sensitive function; constructing a tense neighbor sub-graph for each problem; cutting the tense neighbor sub-graph through time constraint; performing quantitative scoring on each entity in the tense neighbor subgraph by using a time activation function to obtain a subgraph reasoning score; and fusing the score semantic score and the sub-graph reasoning score to obtain a final answer. According to the method, the capability of identifying the time information contained in the problem is improved; and reasoning answers meeting time limitation through the time sub-graph to obtain reliable answers of the questions.

The invention relates to the technical field of artificial intelligence, in particular to a network model and method for face shielding detection based on graph reasoning. In order to solve the problem that the recognition precision of shielded faces is affected by the framework limitation and calculation complexity of an existing convolutional neural network, the invention provides a network model for face shielding detection based on graph reasoning. The network model comprises a feature extraction network, a graph attention reasoning module (GARM) and a decoding (deconvolution) part. Meanwhile, low-layer, high-layer and edge features of the face are extracted through a residual network and spatial pyramid pooling, similar pixel features are projected to graph nodes through graph projection, a projection data relation between the nodes is calculated to infer and analyze a possibly-shielded area, pixels are distributed to the area for detection, and finally a face shielded area is detected. Data sets such as Helen are adopted for model training and testing, and experiments prove that the detection precision and the segmentation precision of the method are superior to those of other current neural network face occlusion detection methods.

The invention discloses an intention reasoning method and device. The method comprises the steps of obtaining text data; performing emotion classification on the text data to obtain emotion classification text data; performing causal relationship extraction on the emotion classification text data to obtain causal relationship event tuple pairs; according to the causal relationship event tuple pairs, establishing a affair graph; inferring an emotion level according to the affair map; and performing intention reasoning according to the emotion level. Through the method and the device provided by the embodiment of the invention, reasons for generating the corresponding emotions of the user can be understood while the emotions of the user are calculated, so that intention reasoning can be more accurately carried out, a strategy for pacifying the user and generating common emotions can be obtained, and the method and the device can be applied to the fields of precision marketing industry, social networks, customer service quality management and the like, and have wide application prospects. The method is of great significance to final decision making of a man-machine interaction system.

The invention provides a multi-task CNN+GCN based remote sensing image building semantic segmentation edge optimization method, using CNN to extract high-level semantic features of buildings from remote sensing images, and using GCN to quickly perform graph reasoning on high-resolution original images ; Then use several times of upsampling, horizontal connection and convolution operations to remap the deep features from CNN with lower resolution to the original image, and use this to extract building edges and first semantic segmentation of buildings; combine deep features with edge The extraction results are integrated to constrain the edges of the initial building semantic segmentation results; finally, the graph feature adaptive optimization module is used to promote the GCN feature to effectively optimize the constrained building semantic segmentation results, and output the building semantic segmentation with excellent edge performance result. The beneficial effects of the invention are: adaptively optimizing the edge details of the semantic segmentation results of buildings in remote sensing images based on CNN, and improving the accuracy and application value of the results of automatic mapping of buildings.

The invention discloses a malicious user detection method based on feature learning and graph reasoning. Firstly, the suspicion degree of a product is evaluated according to a user sequence, then the suspicion degree of each user is indirectly evaluated according to the suspicion degree of the product, and then, from the product suspiciousness degree, related new features are provided, and the screened malicious users suitable for the new features are combined; and, on the other hand, the same composition hypothesis is provided, the user-user graph is constructed, feature learning of the graph neural network and a paired Markov label propagation method are fused, a unified objective function is established for iterative optimization, feature learning of graph nodes and node label reasoning are carried out, and then malicious user detection is completed.

The invention discloses a video-paragraph retrieval method and system based on a partial-whole graph reasoning network, which belongs to the field of cross-modal retrieval and mainly includes the following steps: 1) Firstly, preprocessing the video and text (paragraph). 2) For the given video and text, use the local-whole graph inference network to encode, respectively, to obtain the final video features and text features. 3) Calculate the similarity between video features and text features using cosine similarity. 4) Retrieve according to the similarity measurement results. Compared with the traditional video-paragraph retrieval method, the present invention proposes to decompose video and text into four-layer semantic structures, and construct local graphs and overall graphs respectively, and then use graph convolutional networks to perform graph reasoning operations. The results obtained in video-paragraph retrieval are better than traditional methods.

In various embodiments, an inference application performs tasks associated with 3D CAD objects that are represented using B-reps. A UV-net representation of a 3D CAD object that is represented using a B-rep includes a set of 2D UV-grids and a graph. In operation, the inference application maps the set of 2D UV-grids to a set of node feature vectors via a trained neural network. Based on the node feature vectors and the graph, the inference application computes a final result via a trained graph neural network. Advantageously, the UV-net representation of the 3D CAD object enabled the trained neural network and the trained graph neural network to efficiently process the 3D CAD object.

This application proposes a man-machine dialogue method, device, storage medium, and computer program product, wherein the method includes: determining the current dialogue topic and the user's current utterance information; determining the user's current utterance representation vector according to the current utterance information; combining the current utterance Information and the current discourse representation vector, perform graph reasoning calculation on the heterogeneous knowledge graph corresponding to the current dialogue topic, select the current knowledge corresponding to the current discourse information from the heterogeneous knowledge graph; according to the current discourse information and current knowledge, generate the current sentence corresponding to Reply sentences, among them, the heterogeneous knowledge graph is created based on structured knowledge and unstructured knowledge, and can generate rich reply sentences. In addition, the use of graph reasoning algorithms can improve the accuracy of knowledge selection, making the knowledge selection process It has good interpretability and good generalization ability, and at the same time, reduces the dependence of the overall solution on labeled corpus.

The invention discloses a satellite picture defect-based positioning, identifying and classifying method. The method comprises the following steps of: (1) a preprocessing step: enhancing the characteristics of noisy points and noisy lines of original pictures, and segmenting all the enhanced pictures into a training data set and a test data set; (2) a model training step: feeding the training dataset into a deep learning neural network, carrying out iterative training, evaluating the effect of the trained model by using the test data set to obtain an optimal model; and (3) a model deploying step: reasoning an updated image by using the optimal model, and realizing the positioning, identification and classification of defects in the new image in real time. According to the method, end-to-end learning is realized, image screening speed and precision are improved, and an automatic, efficient and accurate solution is provided for the problem of low production efficiency of manual image screening which puzzles people in the past.

The invention discloses a visual semantic embedding method and system based on data enhancement, and belongs to the technical field of deep learning. Performing target identification on the image by using the first network model to select a plurality of image areas; extracting fine-grained features of the image region by using a second network model, and obtaining fine-grained representation of the image in the common embedding space through fine-tuning network mapping; first semantic graph reasoning is carried out, and unified pooling operation is carried out; using a first extraction model to extract word vector representation related to the text context; performing fine-tuning mapping to the common embedding space by using a second extraction model to obtain word vector representation of the text in the common embedding space; second semantic graph reasoning is carried out, and unified pooling operation is carried out; and performing semantic alignment on the first pooling result and the second pooling result, and enhancing data in model training. According to the method, the semantic association in the modal can be well constructed, and the unified representation can be generated in the common embedded subspace.

The invention discloses an encrypted traffic identification and classification method and system based on a direct push graph. The method comprises the following steps: firstly, collecting encrypted traffic data of a known category in a known network environment and encrypted traffic data of unknown label information in a cross-network environment; then, the collected network flow data is divided into single network sessions; aggregating sessions with the same address information to form a session cluster set; then, by taking session clusters in the session cluster set as node units, calculating feature similarity among nodes, and constructing relation edges among the nodes; constructing a direct push graph according to the relation edges between the node information and the nodes; and then, predicting category information of unknown nodes through an iterative'aggregation diffusion 'graph reasoning algorithm. The method can efficiently and stably identify and classify the network application traffic collected under the general network under the condition that the diversity of the network traffic training samples is insufficient, and identifies the new class of network application traffic data not contained in the training set.