35 results about "Graph sampling" patented technology

The invention relates to an unsupervised graph representation learning method and device on a large-scale attribute graph based on subgraph sampling. The method comprises the steps of performing sub-graph sampling on an attribute graph according to structure information and node attribute information of the attribute graph to generate a plurality of sub-graphs; and performing graph learning on anauto-encoder on each sub-graph by utilizing the structure information, the node attribute information and the community information of the attribute graph to obtain low-dimensional vector representation of nodes in the attribute graph. The graph auto-encoder comprises an encoder and a decoder; the encoder adopts a graph convolutional neural network; the decoder includes a graph structure loss reconstruction decoder, a graph content loss reconstruction decoder, and a graph community loss reconstruction decoder. A user is supported to learn low-dimensional vector representations of nodes in a large-scale attribute graph in an unsupervised mode, topological structure information and node attribute information on the graph can be reserved as much as possible through the vector representations,and the vectors serve as input to be applied to different downstream tasks to conduct data mining tasks on the graph.

The invention discloses a generalized maximum degree random walk graph sampling algorithm. The generalized maximum degree random walk graph sampling algorithm comprises the following steps of enabling a sample to walk on a graph randomly; and performing unbiased estimation according to the sample. A 'large deviation problem' of an RW algorithm and a 'repeated sample problem' of an MD algorithm can be solved effectively, so that the overall efficiency on sample points acquired from the internet is improved.

The invention belongs to the technical field of recommendation systems, and relates to a heterogeneous graph neural network-based recommendation method, which comprises the following steps of: collecting a data set with social relationships among users, user-commodity interaction historical data and commodity category information, filtering invalid data and carrying out negative sampling; randomly selecting a user set and a related commodity set, and carrying out multi-order graph sampling and mapping; node feature extraction: inputting the constructed graph into a heterogeneous graph neural network for processing to obtain a fusion node embedding vector of the nodes, wherein for the commodity nodes which do not need to be subjected to the re-calibration step, the fusion node embedding vector of the commodity nodes is the commodity fusion embedding vector; re-calibration: re-calibrating the user fusion node embedding vector to obtain a user final expression embedding vector; and performing preference prediction by using the user final representation embedding vector and the commodity fusion embedding vector, and obtaining a recommendation sequence. The method solves the problems of data sparsity and data missing, and has the advantages of being accurate in recommendation and the like.

The invention discloses a large graph sampling visualization method based on graph representation learning, and belongs to the field of graph visualization and graph sampling. According to the method,nodes in an original network are converted into high-dimensional vectors through a node2vec algorithm, then the high-dimensional vectors of the nodes are projected to a low-dimensional space througha dimension reduction algorithm, and the semantic structure similarity of the corresponding nodes in the network space can be effectively expressed through the distance between projection points. Secondly, a multi-target sampling model of adaptive blue noise sampling is designed to effectively maintain a topological structure of an original network; and measurement indexes based on network attribute characteristics are proposed, quantitative evaluation is carried out on different sampling algorithms to obtain graph sampling result evaluation, and the graph sampling result evaluation is presented by utilizing a visualization method. According to the method, the nodes are sampled in the representation space, the context structure of the original network is well simplified and reserved, and the topological structure of the network is effectively kept while the node scale is reduced.

The invention discloses a cold start user-oriented recommended meta-learning method, and the method comprises the steps: carrying out the sampling of a data set in a mode of dynamic subgraph sampling,and enabling the data obtained through sampling to serve as training data, wherein the data set comprises interaction records between a plurality of users and different articles; training a collaborative filtering model by using the training data set, the training comprising an inner cycle and an outer cycle: in the inner cycle, performing recommendation prediction on each user, and updating model parameters of the users based on a prediction result; and in the outer loop, updating the overall model parameters by utilizing the model parameters of all the users. The method can be suitable forany differentiable collaborative filtering-based model, personalized recommendation can be better carried out for new users, and the model performance is improved.

The invention discloses an automatic optical detection laser welding system which is composed of an automatic optical detection system and a laser welding system. The automatic optical detection system comprises a microcomputer, a graph sampling system, a data conversion system, a revision system and an automatic working platform; and the laser welding system comprises a laser, a coupling lens and a focusing lens assembly. The automatic optical detection system is integrated with the laser welding system, and bugs in the printed circuit board (PCB) welding and detection process are reduced by matching the high-automation high-accurate automatic optical detection system with the laser welding technology which is small in influenced area and high in energy concentration. Therefore, working efficiency is improved, multiple undesirable phenomena during welding are reduced, the welding devices are ensured to be free of damage during welding, and simultaneously detection speed, efficiency and stability are ensured.

The invention provides a face clustering method and device based on structure perception, and the method comprises the steps: obtaining a plurality of to-be-processed face images, extracting the face features of each to-be-processed face image based on a pre-trained convolutional neural network model, and constructing a K-nearest neighbor graph according to the face features of each to-be-processed face image; inputting the K neighbor graph into a pre-trained edge score prediction model to obtain the score of each edge in the K neighbor graph, wherein the edge score prediction model is obtained by sampling a K neighbor graph by using a structure retention sub-graph sampling strategy and training a graph convolutional neural network by using a sub-graph obtained by sampling; and according to the score of each edge in the K neighbor graph, performing a first pruning operation on the K neighbor graph to obtain a face cluster for the plurality of to-be-processed face images. The technical problem of insufficient face clustering accuracy in the prior art is solved.

The invention discloses a hyperspectral image active learning method based on graph signal sampling. The method comprises the following steps: after reading three-dimensional hyperspectral image data,rearranging, taking its category label as graph signal, using the hyperspectral image data to construct a weight matrix, and characterizing the connection relationship between graph signal points; preserving 8 nearest neighbor connections and sparse weight matrices; calculating a degree matrix, a normalized weight matrix, a normalized graph Laplace matrix, a second-order graph Laplace matrix; acquiring an initial training sample as an initial sampling point of a graph signal; using the graph sampling method, the weakest pixel points are connected in the non-sampling set of the graph signal; adding the sampling pixel points into the sampling set of the graph signal; judging whether the sampling pixel points belong to a test set, if the sampling pixel points belong to the test set, giving an expert tag and adding a training set to remove the test set; classification accuracy being verified by using graph reconstruction classification method; whether the number of training samples reaches the set value or not being judged; if the training samples do not reach the set value, the active learning process being withdrawn.

The invention relates to a graph neural network representation learning-based structure graph alignment method and a multi-graph joint data mining method. The method comprises the following steps: performing sub-graph sampling on a graph in training data; learning low-dimensional vector representation of nodes in the sub-graphs by using a graph neural network through marked aligned node pairs; calculating the similarity among the nodes according to the low-dimensional vector representation of the nodes in the sub-graphs, aligning the graphs based on the similarity, and finally obtaining a graph neural network with trained parameters; in the speculation stage, obtaining low-dimensional vector representation of each node of two to-be-aligned graphs through the trained graph neural network, then calculating the similarity between the nodes, aligning the two graphs on the basis of the similarity, and then performing joint data mining through aligned multi-graph data. According to the invention, under the supervised setting, the expression performance of the model, the loss function setting, the spatial constraint of the representation vector and the expandability are considered, and the improvement of the existing method is realized.

The invention discloses a wireless network data missing attribute recovery method and device based on a graph neural network. The method comprises the following steps: mapping wireless network data into a corresponding topological graph structure, and sequentially mapping sample data with missing attributes into attribute vectors of nodes in the topological graph structure; obtaining an adjacent matrix of the topological graph structure according to the attribute vector of the node; simplifying the topological graph structure by using a graph sampling algorithm to obtain a sparse adjacent matrix; based on the attribute vector and the sparse adjacent matrix, learning by using a graph neural network model, and outputting the attribute vector recovered after reconstruction. According to the method, an attribute recovery framework based on a graph automatic encoder is used, a graph neural network learning algorithm based on strategy gradient is adopted, modeling and learning are directly carried out on the attribute recovery problem of the wireless network data, the correlation in the wireless network data is fully utilized, and therefore, the attribute performance of wireless network data recovery is improved.

The invention discloses a graph neural network data sampling method and device, equipment and a storage medium. According to the scheme, vertexes of an original image data set are clustered, and the training vertexes are sorted according to the clustering categories of the training vertexes, so that a batch of training vertexes can be sampled in the same cluster at the same time during sampling in the sampling process, and the data locality of sampling is improved; moreover, as the training vertexes in the same cluster generally have more similar attributes and are closely connected parts, and connection among different clusters is little, the neighborhood vertexes expanded in the same cluster are concentrated in the same cluster, and the vertexes in the same cluster are close in storage, so that the sampling data locality can be improved, the range of neighborhood expansion is limited, and the efficiency of sub-graph sampling is improved.

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 fingerprint image noise reduction method and a fingerprint image noise reduction device, which are used for reducing noise interference. Comprising the following steps: extracting a first graph with the height of h, the width of w and the number of channels na; sampling to obtain a second graph with the height of h / 2, the width of w / 2 and the number of channels na; extracting a third graph with the height of h / 2, the width of w / 2 and the number of channels nb; sampling to obtain a fourth graph with the height of h / 4, the width of w / 4 and the number of channels nb, extracting a fifth graph with the height of h / 4, the width of w / 4 and the number of channels nc, and sampling to obtain a sixth graph with the height of h / 8, the width of w / 8 and the number of channels nc; extracting a seventh image with the height of h / 8, the width of w / 8 and the number of channels nd, and sampling to obtain an eighth image with the height of h / 4, the width of w / 4 and the number of channels nd; extracting a ninth image with the height of h / 4, the width of w / 4 and the number of channels nc from the fifth image and the eighth image, and sampling to obtain a tenth image with the height of h / 2, the width of w / 2 and the number of channels nc; extracting an eleventh graph with the height of h / 2, the width of w / 2 and the number of channels nb from the third graph and the tenth graph, and sampling to obtain a twelfth graph with the height of h, the width of w and the number of channels nb; a thirteenth graph with the height h, the width w and the channel number na is extracted from the first graph and the twelfth graph.

The invention discloses a graph attention network inductive learning method based on graph sampling, which mainly comprises two parts, namely a graph sampling process and a graph training process, and is characterized in that a plurality of sub-graphs are sampled from an original data set large graph by using a random walk sampler to form mini batch, and then the mini batch is input into a graph attention network for training; and according to the method, a big data set is split into small data sets, and the number of training rounds is increased, so that the performance of the method is remarkably improved, and the method is ensured to have good robustness. The method can also serve as the basis of the technical thought, and has reference value and deployment significance for researchers in the industry to design related algorithms in the future.

A pre-training method based on a dynamic graph neural network learns node representation from three aspects of time, structure and semantics, and comprises the following steps: determining the size of a sampling sub-graph according to actual requirements and system performance, and performing sub-graph sampling on large-scale dynamic graph data by using a time sensitive sampling algorithm to obtain a sub-graph; for the sub-graph, performing covering processing on the sub-graph by using a time-sensitive edge covering algorithm and a node feature covering algorithm to obtain a processed new sub-graph; a dynamic graph generation algorithm is combined with a GNN model to predict covering edges and covering node features of the sub-graphs, optimal parameters are stored, and the pre-training process is ended; and loading the optimal parameters, and performing fine adjustment on the predicted graph data according to different downstream tasks to obtain a final result. According to the method, large-scale dynamic graph data can be processed, and compared with other pre-training methods, the method is higher in expression ability, learned nodes are more accurate in expression, and the method can be better applied to various downstream tasks.

The present invention discloses a social network sampling method based on hybrid jump. In the social network sampling, the method is based on a current classic MHRW (Metropolis-Hasting Random Walk) sampling method and employs a random jump strategy to prevent sampling from falling into a local sub network to effectively solve the large-scale complex social network sampling problem so as to obtainunbiased social network sample set data. The method provided by the invention combines a BFS (Breath-first Search) method at the first time to have fast sampling speed and have no duplicate nodes in the samples, and employs a cubic spline interpolation method to establish a three-dimensional average degree distribution model to determine a jump parameter optimization value of a social network sampling method. The method can better provide guidance for selection of graph sampling setting parameters to allow the sampling method to achieve the best sampling effect. The social network sampling method based on hybrid jump provides a new idea for a social network sampling method so as to facilitate performance research of the large-scale complex social network.

According to the graph sampling and random walk acceleration method and system based on the graphics processor, graph data are read from a storage medium through a CPU, converted into a CSR format and then output to the GPU, and the GPU generates an alias table in real time according to a set working mode and samples the alias table, or off-line judges whether a pre-generated alias table exists or not and sampling. According to the method, the alias method can be executed efficiently and concurrently, and the performance of graph data processing can be remarkably improved on the same hardware platform, including improvement of the sampling throughput and reduction of the overall running time.

The invention relates to the field of integrated circuit mask design, in particular to a photoetching mask optimization method and device for graphic image joint optimization and electronic equipment.The method comprises the following steps: inputting a main graph; segmenting the edge of the main graph to obtain a short edge, and taking the short edge as a first variable for optimizing the main graph; generatingsame or similar auxiliary graph sampling points around the same or similar main graphs, and taking the auxiliary graph sampling points as second variables for main graph optimization;providinga target function taking a first variable and a second variable as optimization variables, wherein rules for generating auxiliary graph sampling points around each main graph are the same andare not limited by specific positions of the main graphs, so thatconsistency of optimization results of each main graph is guaranteed finally.

The invention relates to a method for adjusting graph sampling frequency, and belongs to the technical field of image acquisition. The method comprises the following steps: acquiring a falling image of an object with a certain density through a linear array camera, identifying feature points of a current image through feature extraction, calculating a ratio of an object movement speed to a sensorsampling rate according to position deviations of the same feature points corresponding to different images, and calculating the movement speed of the current object according to the ratio; Therefore,the sampling frequency of the sensor is adjusted to eliminate deformation caused by mismatching of the falling speed of the object and the sampling rate of the sensor. According to the method, the distortion and expansion of images in a color selection image acquisition system are improved, and a good effect on restoring the shape of an imaging object is achieved. The flexibility of the color selection image acquisition system for color selection is improved. The image correction accuracy of the color selection image acquisition system is improved; The method has good distinction degree for imaging objects with different materials and densities, and is widely applied to identification and screening of various materials in a color selection image acquisition system.

The invention discloses a group quantity diversity estimation method based on an information moment, belonging to the technical field of internet-of-things data analysis. According to the invention, arecursive sampling process is combined to optimize a sampling process more reasonably; characteristics like diversity in a large group can be greatly reflected by utilizing a population information moment; and an important effect is exerted on group decision and observation of group characteristics. The group quantity diversity estimation method specifically comprises recursive sampling and common group identification, wherein the step of recursive sampling is to stratify the sampling process to obtain a sampling probability; each layer of sampled groups is scanned; a byte bitmap composed oftime slots is given according to a hash function; threshold-based common group identification is performed after sampling; common groups and non-common groups in one ALOHA time slot are identified; and moment estimation of all labels, namely quantity characteristics of all labels are obtained by calculation.

The invention relates to a local-vertex-based spanning tree graph reconnection method. The method comprises steps of carrying out isomorphic function determination, carrying out frame reconnection, and carrying out ergodic verification. To be specific, a mechanism is defined to determine whether a spanning tree graph has an isomorphic nature, so that structural variation generation during the learning graph process is avoided; with a reconnection frame based on a Monte Carlo method, a head node and a tail node that are connected at an edge are split and reconnected, an adjacent end point and an adjacent node are updated, and iteration with the limited number of times is carried out on the process on a certain condition; and for the graph generated by the reconnection frame, ergodic property verification is provided to cover a leak rate caused by randomness, wherein the operation includes irreducibility, aperiodicity, and positive recursiveness. According to the invention, splitting andreconstruction are carried out by using the spanning tree graph based on the Monte Carlo method; a verification processed based on the Markov chain property is provided; and a spanning tree graph sampling method based on the fixed node number is led out creatively.

The invention discloses an OFDM (Orthogonal Frequency Division Multiplexing) system channel estimation system and method based on a graph signal method. The method is a non-blind estimation algorithm based on pilot frequency, in the algorithm, each resource block of an OFDM time-frequency doubly selective channel with structural characteristics is regarded as a node of a graph signal, and the topological structure of the graph signal is not only determined by a space structure, but also influenced by time selective fading and frequency selective fading. Smoothness constraint is used for modeling, and a signal recovery problem is mathematically modeled into an optimization problem to be solved. Meanwhile, the pilot frequency position is designed by using a graph sampling method, and a better pilot frequency placement position is found out, so that the accuracy of channel estimation is improved.