164 results about "Network representation learning" patented technology

The invention discloses a method and device for obtaining a network representation learning vector, equipment and a storage medium, belongs to the technical field of computers, and is used for enabling the representation learning vector to express more multilevel information on the basis that the calculation cost is as low as possible. The method comprises the following steps: constructing a social network association subgraph set according to social network data; obtaining a network representation learning initial vector of a non-central node included in the social network association subgraph set; for each sub-graph, obtaining a first attention weight from each non-central node to the forward edge of the central node, and obtaining an attention summary vector of the sub-graph according to the network representation learning initial vector of each non-central node and the first attention weight corresponding to each non-central node; and for each non-central node, obtaining a networkrepresentation learning adjustment vector of the non-central node according to the attention summary vector of the subgraph existing from the central node to the reverse edge of each non-central node.

The invention belongs to the field of short-time intelligent traffic control, and particularly relates to a short-time traffic flow control method based on deep learning and spatio-temporal data fusion. The method comprises the steps of: obtaining a data source; analyzing the time and space relevance of the historical data of an intersection; respectively establishing a time dimension GRU model and a space dimension CNN regression model according to the historical data of a prediction checkpoint and the related intersection thereof; fusing output results of the time dimension GRU model and thespace dimension CNN regression model to obtain an adaptive spatio-temporal data fusion model; and counting the prediction result of the adaptive spatio-temporal data fusion model, and sending the prediction result to a traffic department. According to the method, on the space-time level, the space-time dependence of a road traffic flow in a complex road network is analyzed, the network representation learning is used for screening the traffic data, and the screened data is used for inputting the model, so that the accuracy of the traffic flow prediction result of the intersection is improved.

The invention discloses a drug target interaction prediction method based on multilayer network representation learning, and mainly solves the problem of low prediction accuracy in the prior art. Themethod comprises the following steps: downloading data from a drug and protein database, and respectively constructing multilayer similarity networks of drugs and proteins; calculating diffusion states of the two similarity networks respectively, and integrating the diffusion states respectively to obtain feature vectors of drugs and proteins; taking known drug target interaction data as supervision information, putting the drug and protein feature vectors into the same drug target space, and respectively obtaining projection matrixes of the drug and the protein by using a bilinear function; obtaining a prediction score matrix of drug target interaction according to the two projection matrixes and ranking the prediction score matrix; regarding eight top-ranked unknown drug target pairs aspotential drug target interactions. According to the method, the prediction accuracy of drug target interaction is improved, and the method can be used for predicting candidates of drug target pairs.

The invention relates to a personalized information recommendation method by information fusion, belonging to the technical field of Internet information recommendation. The method includes first pre-processing a data set to extract item shape information: determining a relationship type between items, constructing a relationship network between items, determining item text information, and determining item image information; using a network representation learning method to extract network relationship features, extracting text features by using a text representation learning method, and extracting image features by using an image feature extraction method; then calculating a user's preference feature value for each item in each dimension; finally inputting preference features into a sorting model, and recommending items with scores TOP-N in an alternative set to the user. Compared with the prior art, the method improves the accuracy of recommendation results by excavating and utilizing attribute information of the items to supplement sparse user active interaction data; and the integration of the item attribute information can make the recommendation not only rely on rating data,but also help solve the problem of cold boot of new items.

The present invention discloses a microblog-based network user enhancement representation method. The method belongs to the field of microblog data mining, and particularly to a network representation learning method targeted at microblog data. The method comprises: in consideration of colloquial features of microblogging short texts, performing preprocessing on the texts in a targeted way, so as to reduce effects of noise data; by using an LDA theme model, generating feature representation of a historical blog text of user, and calculating cosine similarity between blog text features of any two users, so as to construct a potential friend relation network; and integrating structure information of an original network, and fusing a potential friend relation into the original network, to obtain a modified network structure. According to the method disclosed by the present invention, by using the potential friend relation network extracted from the user-generated text, the original network topological structure is modified, so as to obtain more accurate feature representation of microblog user nodes. Compared with the network representation learning method that only considers network structures, the method obviously improves accuracy of gender and age inference.

The invention provides a network representation learning method. Learning is performed by comprehensively considering text information and a network structure; for the text information part, different types of continuous bag-of-words-based and convolutional neural network-based text encoding models are designed; the network structure information of nodes in a network is used for predicting neighbor nodes of current nodes, and the text information of the nodes are used for predicting representation vectors of the text information of the current nodes; and by use of the method, the text information and the network structure information of the nodes can be effectively encoded in the representation vectors, and the classification accuracy is remarkably improved in a node classification task. Meanwhile, the method fully considers effective information such as the text information in a practical network, achieves an excellent effect in different types of information network data, and has high practicality.

The invention provides a large-scale network disintegration method based on deep reinforcement learning, a storage device and a computer readable storage medium, and the method comprises the followingsteps: training a network representation learning model which is a neural network model mapped from nodes of the network to corresponding feature vectors of the network; training a network disintegration model according to the network representation learning model and a reinforcement learning algorithm, the network disintegration model being a neural network model fitting a reinforcement learningQ value function; and performing network disruption on the target network through the network disruption model. According to the method, a conventional network disintegration problem and a generalized network disintegration problem can be systematically solved. The method has the advantages that the method is simple, the expansibility is high, the required priori knowledge is less, the network disintegration strategy can be efficiently learned only by taking the network as input and defining corresponding reward functions according to different problems, the problem solving scale can be expanded to more than ten thousands of nodes, and the application scene is very wide.

The embodiment of the invention provides a network anomaly detection method and device, relates to the field of AI, and can improve the accuracy of a network anomaly detection result. The method comprises the following steps: layering a target network into at least one stage of sub-network; for each stage of sub-network, determining a similar matrix of the stage of sub-network through a network representation learning algorithm, the similar matrix being used for indicating topological information of the stage of sub-network; training according to the similarity matrix of the at least one stageof sub-network and the historical performance data of the target network to obtain an anomaly detection model; and inputting the current performance data of the target network into the anomaly detection model, and outputting a detection result. The embodiment of the invention is applied to a scene of performing anomaly detection on networks such as a telecommunication network, a cable televisionnetwork or a computer network.

The invention discloses an energy flow distribution prediction method and a system for a regional integrated energy system. The energy flow distribution prediction method comprises the following steps: collecting multi-energy flow network data; constructing a multi-energy flow network data representation model; mining multi-source spatial and temporal features related to energy flow distribution;and outputting an energy flow distribution prediction result. The system comprises a multi-energy flow network data representation module, an energy flow distribution spatial feature extraction module, an energy flow distribution timing feature extraction module, an energy flow distribution external factor feature extraction module, an energy flow distribution feature fusion module and an energy flow distribution prediction output module. According to the energy flow distribution prediction method and the system for the regional integrated energy system, multi-energy flow network data is converted and analyzed based on a network representation learning method and a deep learning method, spatial and temporal features and external factor features related to energy flow distribution are mined, and a new method and a technical base are provided for the problem of energy flow distribution state prediction for the regional integrated energy system.

The invention discloses a network representation learning method and device based on random walk of edges, and the method comprises the steps: calculating the similarity between the edges in a networkaccording to a topic vector of each node of the network and an associated timestamp of the edge; Calculating the edge-to-edge transition probability according to the calculated edge-to-edge similarity; Based on the guidance of the meta-path, carrying out random walk according to the calculated transition probability to generate a node sequence; And performing representation learning of the node according to the obtained node sequence to obtain low-dimensional representation of the node. According to the method, semantic information and time information can be interpreted to obtain richer network content, so that potential information of a real world can be more truly and effectively mined; And the method can be used for carrying out more appropriate and practical representation on the network which changes along with the passage of time.

The invention discloses an interest point recommendation method based on heterogeneous attribute network representation learning, and the method is suitable for recommending interest points to a useron a sign-in data set with abundant description information of the interest points and strong seriality. The method comprises the following steps of firstly, constructing a directed weighted heterogeneous attribute network based on a social network of the user and the sign-in data set; secondly, acquiring node attribute embedding information in the heterogeneous network based on text description of interest points, and acquiring meta-path embedding information of nodes based on random walk of multiple meta-paths in combination with a self-attention mechanism; then, fusing the attribute embedding of the node and the embedding information of the plurality of meta-paths, and performing representation vector learning of the node based on a heterogeneous skip_gram; and finally, based on the similarity of the representation vectors, performing accurate recommendation of the next interest point according to the time and place of the target user.

The invention provides a knowledge graph architecture construction and application method and system and a medium. The method comprises the steps: 1, completing knowledge modeling by defining entitiesin the academic field and constructing an ontology of an academic knowledge graph; 2, carrying out entity alignment, that is, for each entity in the heterogeneous data source knowledge base, the sameentity belonging to the real world is found out; 3, enriching the knowledge graph by utilizing a rule-based knowledge graph reasoning method; step 4, evaluating several most advanced methods of knowledge graph architecture-AceKG embedded knowledge; and step 5, evaluating several most advanced methods of knowledge graph architecture-AceKG network representation learning. According to the method, pure academic information is provided, a large-scale reference data set is provided for researchers, a basis is provided for evaluating knowledge embedding and network representation learning methods,and a knowledge graph framework is provided for enriching the proposed knowledge graph framework.

The invention provides an enhanced network representation learning method based on community perception and relationship attention, and the method comprises the following steps: firstly obtaining a network topology structure and the text information of a node; secondly, acquiring community information by adopting a community discovery algorithm, marking a community where each node is located, andcombining the community with the topological structure to generate a community structure network; then, learning structure embedding of nodes on the community structure network by adopting a communityperception module; then, learning text embedding of each pair of adjacent nodes by adopting a relation attention module; and finally, performing model training in combination with structure embeddingand text embedding to obtain embedded representation of each node. According to the invention, local and global structures of the network and rich semantic relationships among nodes can be completelycaptured.

The invention realizes a set of differential privacy recommendation method based on heterogeneous information network embedding. The differential privacy recommendation method comprises the followingfour steps of: performing network representation learning by using HAN, and calculating heterogeneous attention sensitivity by using characterizations of HAN and an attention weight result; based on adifferential privacy definition, using the heterogeneous attention sensitivity to generate corresponding random noise, and generating a random noise matrix through using a heterogeneous attention random disturbance mechanism; constructing an objective function of differential privacy recommendation embedded with heterogeneous information for learning to obtain a prediction score matrix; and outputting the score matrix as a prediction score capable of keeping privacy. Therefore, the original scoring data is protected for the recommendation system scene under the heterogeneous information network, an attacker is prevented from improving the reasoning attack capability by utilizing the heterogeneous information network data acquired by other channels, and the original scoring data can be guessed or learned again with high probability by observing the recommendation result change of the score.

The invention belongs to the field of social network data mining, and particularly relates to a social network abnormal account detection method and system based on network representation learning, and the method comprises the steps: building a network G (V, E, C) by using social network data; constructing an M * M-dimensional adjacent matrix of the network G (V, E, C); and constructing a representation learning joint optimization model of the network about the topological structure and the account node attribute, and the like. According to the invention, the abnormal account detection task iscombined with the network representation learning; an abnormal factor of each account node on a topological structure and an account node attribute is determined by solving the topological structureof the network account node and a low-dimensional vector representation form corresponding to the account node attribute; furthermore, the consistency characterization forms of the two account nodes in the low-dimensional space are solved, the consistency abnormal factor of each account node is calculated, and finally, the abnormal degree of each account node in the social network is evaluated bycombining the abnormal factors, so that the detection and identification of abnormal accounts are completed.

The invention discloses a semi-supervised network representation learning algorithm LSDNE (Labeled Structural Deep Network Embedding) based on a deep compression self-encoder. The method comprises thefollowing steps: building a model, pre-training the input data with a deep belief network (DBN) to obtain the initial values of the model parameters, and taking the adjacency matrix and Laplace matrix of the network as inputs; encoding the network by a self-encoder with deep compression, and obtaining the global structure of the node; using Laplacian feature mapping, and obtaining the local structural features of nodes; using an SVM classifier to classify the known label nodes and optimize the whole model; using the Adam optimization model and obtaining a representation of the node. The invention can simultaneously use the structure information of the network and the label information of the node to carry out network representation learning, and the deep learning model is used, so that the performance of the representation of the node on the label classification task is better than the existing algorithm. Deep compression self-encoder can reduce the over-fitting phenomenon and make the model have better generalization performance.

The invention belongs to the technical field of information technology and security. The invention relates to the technical field of transaction, in particular to a bitcoin transaction station addressidentification method, system and device, and aims to solve the problem that whether input address information is a transaction station address is judged based on bitcoin transaction information, andthe method comprises the following steps: obtaining to-be-identified transaction data comprising a bitcoin address identifier and bitcoin flow direction data as input information; constructing a bitcoin transaction network based on the input information; A network representation learning method is used for obtaining feature vectors of nodes in a bitcoin transaction network to form a feature space, and whether a bitcoin address identifier in input information is a real bitcoin transaction address or not is further identified through an address classifier. The address classifier is obtained through training on the basis of a transaction data sample and a label sample, and is a combination of a classifier model based on a plurality of mapping functions. The method is less in dependence on resources, can directly identify the address of the transaction, and achieves a better identification rate.

The present application relates to a recommendation method, system and electronic device for learning based on network representation. The method comprises the following steps: step a: constructing auser-article co-occurrence network based on a bipartite graph network and a single projection image; step b: For the user-article co-occurrence network, defining search strategy to get neighbor nodesof each user node and item node. C, according to each user node, article node and respective neighbor node, obtaining vector representation of each user node and article node by using representation learning on network; Step d: according to the vector representation of each user node and the article node, obtaining the most relevant article node of each user node through vector calculation, and recommending the most relevant article to each user according to the calculation result. The present application alleviates the problem of sparsity of collaborative filtering, makes the recommendation system more interpretable, and greatly alleviates the problem of scalability in collaborative filtering.

The invention discloses a multi-behavior migration recommendation method based on deep learning, and the method comprises the steps: firstly obtaining and processing a plurality of implicit feedback data sets of a user; constructing a base network Gb and a plurality of behavior networks G (k), and learning low-dimensional embedded representations of users and article nodes in each network by usinga network representation learning method; then, based on different influence of multiple implicit behavior feedbacks of the user on user preference modeling, using an attention mechanism for automatically learning the weight of each behavior, and acquiring fused low-dimensional embedded representation of the user and the object finally, naturally splicing and sending low-dimensional embedding vectors of the user and the articles and to a full-connection embedding layer, adopting and feeding back a preference learning method based on a deep neural network to a feedforward neural network witha hidden layer, wherein the preference of the user for articles is learned on an output layer. The method can better capture the preference of the user and realize personalized recommendation, and has the advantages of high recommendation accuracy, strong generalization ability, easiness in realization and the like.

The invention provides a cross-social-network user identity recognition method based on a neural tensor network. The method comprises the following steps: step 1, based on network representation learning of Random Walks and Skipgram models, mapping network structure spaces of a source network Gs and a target network Gt to vector spaces respectively; step 2, based on the vector space obtained in the step 1, modeling an association relationship between the user nodes in the source network Gs and the target network Gt by using a neural tensor network model; and 3, inputting the incidence relationvector obtained by modeling in the step 2 into a multi-layer perceptron model for dichotomy, and judging whether the user node pairs between the source network Gs and the target network Gt point to the same real user or not according to a classification result. According to the method, the neural tensor network model is adopted to replace a standard neural network model, the model has stronger capability of expressing the relationship among cross-network users, and two user vectors can be associated in multiple dimensions.

The invention discloses a high-order neighborhood hybrid network representation learning method and a device. A self-attention mechanism and a cascade aggregation layer are added on the basis of an original graph convolution layer, and the method comprises the following steps: converting a Laplace matrix of a graph into a node pair graph attention matrix by using the self-attention mechanism, andtraining weight parameters to learn different attention coefficients; gathering information flows with different distances through the cascade aggregation layer, and taking the output of the previousorder as the input of the next order to control the calculation complexity; and determining that the embedded vector is output to a downstream machine learning task, or outputting a classification result. According to the method, real end-to-end training can be realized, the training speed of the model is effectively improved, and the proposed idea of network high-order and low-order information hybrid learning has field expandability and is simple and easy to realize.

The invention provides a network representation learning method based on a meta-structure and a graph neural network. According to the method, the information of the neighbor nodes is aggregated through the graph neural network and the weighted attention mechanism, the candidate meta-structure set is generated through the hierarchical search algorithm, meta-structures do not need to be defined inadvance, and compared with a previous meta-path, the method can consider more complex structure information between the nodes. According to the method, the strong learning ability of the graph neuralnetwork and the rich semantics of the meta-structure are fused, and the problems that an existing meta-path-based method is single in consideration structure and needs to depend on experience to specify the meta-structure are effectively solved. Moreover, the introduction of a weighted attention mechanism can explicitly consider the quantity information in the meta-structure. And a final node which is more accurate than a result of a traditional representation classification mode is generated, and the final node can be used as vector representation to be used in subsequent other machine learning figures.

The invention discloses a rapid network representation learning algorithm based on a width learning system. The method comprises the following steps of S1, importing a network graph module based on atext, parsing and storing a network topological structure in a dictionary format, wherein keys in the dictionary represent network nodes, values corresponding to the keys form a list and represent a node sequence at the other end of the edge where the nodes are located; S2, performing random walk on the network nodes to generate a walk sequence; S3, constructing a network representation learning model based on a width learning system, taking the walking sequence generated in the step S2 and a representation vector with the dimension of K as input, generating a feature vector of a network nodein a feature vector layer, and enhancing the nonlinear classification capability of a network representation learning model by introducing an activation function in an enhancement vector layer to finally realize text-based network multi-label classification. A width learning system model is adopted in the algorithm, and representation learning of network nodes can be rapidly completed.

The invention provides a dynamic network abnormal link behavior detection method and a system thereof. The method comprises the following steps of: determining, according to the first T time slice networks {G1, G2, ..., GT-1, GT}, a historical network Ghistory of the current time slice network GT+1; performing network representation learning on the historical network Ghistory by using a network representation learning method to determine a new historical network G'history; determining a similarity adj (i, j) between the network node pairs (i, j) according to the distance dij between the network node pairs (i, j) in the new historical network G'history; and if it is determined that the proximity adj(i, j) is less than the preset abnormality determination threshold, determining that the link(i, j) between the node pairs (i, j) in the current time slice network GT+1 is an abnormal link behavior. The system includes a historical network construction module, a network representation learning module, a proximity determination module, and an abnormal link determination module. The invention may obtain the degree of similarity between nodes, and is further capable of more accurately identifying the abnormal link behavior of the dynamic network according to the similarity between the nodes.

The invention provides a method for detecting an intranet lateral movement attack. The method comprises the following steps: collecting flow and log data of intranet equipment; extracting all nodes inthe data, connecting the two nodes subjected to network communication, and constructing an inter-host communication graph; extracting and combinding flow data between every two connected nodes and extracting and combinding data on the nodes to serve as feature values to be assigned to edges and points of the inter-host communication graph; performing dimension reduction on the inter-host communication graph with the features by using a network representation learning method, and extracting a low-dimensional feature vector by using an auto-encoder; and classifying the low-dimensional feature vectors by using a semi-supervised classification learning algorithm, and distinguishing hosts suspected to be attacked.

The invention discloses a network representation learning-based multiplication-sharing matching method. According to the relation between the starting point and the end point of the passenger and theoriginal path of the driver, the ride sharing is divided into two types, the first type is end point sharing, the other type is path sharing, the passenger needs to walk from the starting point to theboarding point, then sharing is achieved, then the passenger walks from the boarding point to the destination, and the path track of the sharing is one part of the path track of the passenger; a heterogeneous multiplication network is constructed, and representation learning is performed on the heterogeneous multiplication network by using a network representation learning model to obtain low-dimensional vector representation of the user node; and the cosine similarity between the driver and the passenger nodes is calculated, the calculated cosine similarity values are sorted from large to small, and the first k passengers with the highest similarity values with the driver are returned as passengers capable of sharing the passengers to achieve sharing. Compared with a traditional method which only uses distance recommendation, the ride sharing recommendation method provided by the invention is more reliable, the semantic comprehensiveness is visual, potential carpooling users can be accurately discovered, and faster and more convenient services are provided for the ride sharing users.

The invention discloses a text-enhancing network expression learning method, and relates to a complex network analysis technology. A new text-information-enhancing network expression learning method is put forward on the basis of non-negative matrix decomposition, and for the network structure, in combination with the first-order and second-order similarity between nodes, network expression is obtained through a decomposition similarity matrix; for the text clustering structure related to the nodes, decomposition is conducted on a text-lexical item matrix to obtain a text clustering affinity matrix, the consistency relationship is established between the network expression and the text clustering structure through the text clustering affinity matrix, and therefore network expression learning is controlled by the network structure and the text clustering structure related to the joints. By means of the method, the network structure is depicted, the text clustering structure related to the joints is also depicted, additional information beside the network structure is added to the network expression learning, the learned nodes are used for expressing more available information, and higher identifiability is achieved.

The invention discloses a dynamic network representation learning method for a social network platform, and the method comprises the following steps: numbering all appearing nodes according to input original data; constructing a dynamic network according to the numbered node sequence and the original data; obtaining an adjacency matrix, a self-loop adjacency matrix and a corresponding degree matrix of the dynamic network; taking the obtained matrix as input and sending the matrix to a deep neural network model for learning; and training a neural network model, converting the original high-dimensional sparse matrix into a low-dimensional dense vector, and embedding time sequence information carried by the network into a new vector space. According to the method, feature extraction is carried out on the network data by using the graph convolutional neural network, and potential time sequence information in the network is captured in combination with the LSTM, so that feature informationand time sequence information contained in the high-dimensional network can be captured, and the method has good universality and can be applied to all related network analysis tasks.

The invention discloses a network representation learning method under completely unbalanced tags based on approximate intra-class and inter-class constraints. The method comprises the steps that social network data is obtained, and labeled nodes and nodes belonging to unknown classes already exist in networks; modeling is conducted on network structure information; the modeling is conducted on network intra-class similarities; the modeling is conducted on network inter-class differences; an objective function of network representation learning is constructed; a solution to the objective function is obtained based on an optimization problem solving algorithm, and learned feature results are obtained. According to the method, the intra-class similarities are broadened by allowing nodes of the same tags to be on the same manifold in feature space, the inter-class differences are broadened by removing existing neighbor relationships between nodes of different tags, the method ensures tworequirements within a reasonable range, meanwhile biased results are avoided, and the method is suitable for semi-supervised network representation learning with completely unbalanced labeled data andbalanced labeled data, and is suitable for scenarios where quality of labeled information cannot be guaranteed.

The invention discloses a personalized commodity recommendation method and system based on multilayer heterogeneous attribute network representation learning, and the method comprises the steps: taking an interaction behavior between a user and a commodity as an edge, constructing a multilayer heterogeneous attribute network, and carrying out the decoupling of the multilayer heterogeneous attribute network into a plurality of simple binary networks; performing weighted accumulation on the adjacency matrixes of all the binary networks to obtain a newly combined adjacency matrix, and performingspectrogram conversion; fusing the adjacency matrix and the node attribute characteristic matrix after spectrum conversion, and finally obtaining representation vectors of all nodes by using a randomprojection method; obtaining a verification set from the historical data to perform parameter adjustment, and obtaining a representation vector of each node; and measuring the preference of the user to the commodity by utilizing cosine similarity so as to carry out personalized recommendation. According to the invention, various interaction behaviors between the user and the commodity are considered at the same time; the interactive relationship among various behaviors can be captured without human intervention; attribute information of users and commodities is effectively fused; network representation learning is carried out by using random projection so that the method efficiency is greatly improved and the recommendation performance is improved.