55 results about "Conditional mutual information" patented technology

The invention discloses a word vector model based on point mutual information and a text classification method based on CNN. The method comprises the following steps: (1) training a word vector modelthrough a global word vector method based on point mutual information; (2) determining a word vector matrix of the text according to the trained word vector model; (3) extracting features from word vector matrix by CNN and training classification model; (4) extracting input text features according to the trained word vector model and CNN feature extraction model; (5) according to the text featuresextracted from CNN feature extraction model, calculating the mapping distance between text and preset categories by softmax and the cross entropy method, wherein the nearest one is the correspondingcategory of text. This method overcomes the shortcomings of Glove word vector in the semantic capture and statistical co-occurrence matrix, reduces the training complexity of the model, can accuratelymine the text classification features, is suitable for text classification in various fields, and has great practical value.

The invention provides a remote sensing hyperspectral image band selection method based on conditional mutual information, which comprises the following steps of: A, opening a hyperspectral remote sensing image and labeling a sample to be classified by man-machine interaction; B, grouping: carrying out band grouping by utilizing the conditional mutual information among all adjacent bands under given class condition according to the sample to be classified obtained from the step A; and C, searching: carrying out search calculation on the grouped bands obtained from the step B by utilizing a search algorithm combining a support vector machine and a genetic algorithm so as to find out an optimal band combination; and on this basis, pruning by using a self-adaptation branch and bound algorithm. Through the combinational use of the band grouping based on the conditional mutual information and the pruning based on the self-adaptation branch and bound algorithm, redundancy and noise grouping which are caused by noise perturbation are avoided, the frequency of band grouping is reduced and the classification accuracy of the band combination is improved.

The invention provides an unsupervised feature selecting method based on conditional mutual information and K-means. Multiple times of clustering of unclassified labels is carried out by adopting K-means algorithms having different initial conditions, and on the basis of each time of clustering, a modularization metric value of every feature and the conditional mutual information between among the features are considered comprehensively, and related independence indexes among the features are used to select feature subsets having high relevancy and small redundancy. The feature subsets acquired by the clustering of the different K-means are gathered together to acquire a final feature subset. The unsupervised feature selecting method is effectively used for the imbalanced data sets having no labels, and the acquired feature subsets have the high relevancy and the small redundancy.

The present invention relates to methods and arrangements in a multi-antenna radio communication system, in particular to methods and arrangements for improved multiple HARQ transmission in such systems. While HARQ transmission schemes, as known in the art, only can consider the fact whether or not a transmission attempt has been successful the present invention provides a HARQ retransmission scheme that considers the reception quality for already performed transmissions of a same data packet when selecting a resource allocation for necessary re-transmissions. Resource allocation for retransmissions is based on a pre-defined metric indicating a quality of the reception of the previous transmission attempts. Such a metric can be derived from a quality measure derived in the receiver unit, e.g. a CQI or CSI-based value, or an appropriate measure of the mutual information, e.g. the accumulated conditional mutual information (ACMI).

The invention provides a financial field term recognizing method based on information entropy and term credibility. Only simple characteristics are selected, and financial terms are recognized through a CRF model; candidate terms belonging to the specific error type are screened out by setting a threshold according to an information entropy formula based on marginal probability in a recognition result, and the candidate terms are processed in a more targeted mode; words are converted into word vectors with rich semantic information when the candidate terms are filtered, and a large number of financial field terms can be obtained through filtering since a similarity calculation method and a traditional mutual information method complement each other. The too complicated characteristic selection process of an existing robot learning model can be effectively avoided, post-processing part is flexible and not limited to specific linguistic data, the recall rate can be easily increased, the term structure integrity can be improved, and the method can be used as a universal term recognizing method.

The invention provides a multicarrier radar system maximum likelihood distance estimation algorithm for estimating the distance information of the target. Supposing that the reflection coefficient isa constant for single target detection under the complex additive white Gaussian noise (CAWGN), the probability density distribution and the distance mutual information of the target distance under different signal-to-noise ratios are obtained by using the Zadoff-Chu multicarrier signal, and the Cramer-Rao bound (CRB) of the distance variance and the analytical expression of the distance mutual information under the condition of high signal-to-noise ratio can be obtained. The simulation result indicates that the distance mutual information of the target and the signal-to-noise ratio have a linear relation under the condition of high signal-to-noise ratio, and the time bandwidth product (TBP) of the radar detection system is doubled and the distance mutual information is increased for 1.5 bits. The simulation result proves the correctness of theoretical analysis. The conclusion has important theoretical guidance significance for actual design of the radar detection system.

The invention discloses an epistasis locus mining method based on a genetic tabu and Bayesian network, and the method comprises the following steps: 1, converting genotype data into Boolean data in binary representation; 2, using the logic and the operation to quickly calculate the mutual information between arbitrary SNP locus pairs and phenotype, extracting a top-N node pair, and constructing aninitial network map comprising the SNP loci; 3, generating new individuals based on an initial network individual through randomly adding edges, deleting edges and reversing the edges until the number of network individuals reaches the size of the population; 4, evolving a Bayesian network structure through three operations of the genetic algorithm and a scoring mechanism of the Bayesian network,finding an optimal solution of the network structure, and quickly and accurately obtaining the epistasis locus affecting the phenotypic traits. The method can help biological researchers to obtain epistatic gene loci affecting specific phenotypic traits, thereby assisting in gene function mining, and providing reference for genetic basis analysis of complex quantitative traits of different species.

The method includes steps: (1) training phase: (a) collecting known determined whether dialog events are intruded, and extracting features as training set; (b) pretreating the training set; (c) obtaining trained bilaminar half lazy Bayes classifier based on conditional mutual information; (d) ending; (2) classifying phase: (e) pretreating dialog events to be tested; (f) using classifier obtained from step (1)-(c) to classify pretreated dialog events; (g) returning back classified result; (h) ending. Keeping low time complexity in application phase, the invention raises performance of classified precision so as to raise intrusion detection performance of intrusion detection system.

The invention relates to the field of cement strength prediction, and particularly discloses a construction method of a cement strength prediction model and a cement strength prediction method, and the method comprises the steps: collecting a plurality of cement sample quality inspection data sets, each quality inspection data set comprising a plurality of feature parameter values; sorting the plurality of characteristic parameters from large to small according to the relevancy with the cement strength through characteristic selection based on conditional mutual information, calling the valuesof the first m parameters sorted in the quality inspection data of each cement sample to form a characteristic set of the cement sample, and training an auxiliary prediction model based on the characteristic set of all the cement samples; and determining an m value corresponding to the auxiliary prediction model with the highest prediction precision obtained by training, and synchronously optimizing a plurality of parameters of the to-be-trained model in each training iteration by adopting GA based on all feature sets corresponding to the m value to obtain a cement strength prediction model.The training sample for training the cement strength prediction model is reasonable, the training efficiency is high, and the prediction precision of the model obtained through training is high.

The invention relates to the technical field of bioinformatics and provides a method for mining epistasis loci of an artificial bee colony optimized Bayesian network, including four steps S1 to S4. The method for mining epistasis loci of the artificial bee colony optimized Bayesian network, comprises firstly using three stages of expansion, contraction, and symmetry detection to calculate the Markov blanket of nodes through conditional mutual information so as to construct an initial nectar source network structure; then, based on the initial nectar source, randomly adding, subtracting and reversing edges to generate new nectar source until the maximum number of initial nectar sources is reached. The three operations (collecting bees, observing bees, and reconnoitering bees) of artificialbee colony algorithm and the BIC and MIT scoring method of the Bayesian network are configured to evolve the structure of the Bayesian network, find the optimal network structure, quickly and accurately obtain the epistasis gene loci that affect phenotypic traits, and assist the gene function mining.

The invention belongs to the technical field of text topic clustering, in particular to a pointwise mutual information-based text semantic similarity measurement method. The method comprises steps ofbased on a co-occurrence latent semantic vector space model, further extracting a potential semantic similarity relationship among the keywords by utilizing the pointwise mutual information so that two keywords which do not have a co-occurrence relation originally are enabled; by constructing the keyword co-occurrence vector, further extracting and mining the potential semantic similarity relationship between the keywords, so that semantic extraction is more sufficient, a text semantic similarity measurement method based on point mutual information is established, and the application of the method can effectively improve the text clustering and information retrieval precision and reduce the retrieval cost.

The invention discloses a process reliability evaluation method based on nonlinear correlation analysis. The method comprises the following steps: firstly, performing failure mechanism analysis and FMEA (Failure Mode and Effects Analysis) analysis on products, determining a product characteristic that affects the inherent reliability of the products and a process characteristic that affects each product characteristic, using a partial mutual information estimation method based on a Clayton copula entropy to select a key product characteristic and a key process characteristic; secondly, giving a dependence structure between the process characteristics by means of a Clayton copula function; finally, giving a product inherent reliability prediction method based on a support vector machine. The process reliability evaluation method based on the nonlinear correlation analysis provided by the invention uses the partial mutual information to effectively measure a nonlinear relation between variables, and uses a relation between the partial mutual information and the Clayton Copula entropy to avoid from estimating the joint probability density function, and improve the accuracy of the partial mutual information estimation; besides, the input variables are effectively selected, thus the prediction accuracy and efficiency of a model are improved.

The invention relates to a bus peak load prediction method considering complex meteorological influence, and belongs to the technical field of bus peak load prediction. According to the method, the feature importance degree result of the condition mutual information on the to-be-selected features in the original feature set is used as the basis, and an IPSO-ELM serves as a predictor, forward feature selection is performed, an optimal characteristic set of bus peak load prediction is determined; the influence of characteristic redundancy on the prediction precision during bus peak load prediction is reduced; and optimal prediction models are constructed for different buses, so that the prediction precision of different buses is effectively improved, an improved particle swarm optimization extreme learning machine is introduced to be combined with a linear method, peak load prediction in different scenes is carried out, and the prediction requirements in small-sample or sample-free scenes are met.

The invention discloses a method for selecting hyperspectral image bands based on extraction of all kinds of important bands. The hyperspectral data of each training sample of the bands are used as a time sequence; smooth denoising processing is carried out on each time sequence through wavelet transformation; the important band sets of the training samples are ensured by extracting the important points of the smoothed time sequences, wherein the important points correspond to the important bands respectively; the important band sets of the training samples are collected to form an initial band set; the final band combination is selected on the basis of the initial band set through a branch and bound method. According to the method, conditional mutual information grouping is introduced in the branch and bound method as constraint conditions, and compared with the search efficiency of the existing fast branch-and-bound search algorithm, search efficiency is improved by about one half.

The invention relates to a news automatic labeling method based on an LDA model. The news automatic labeling method includes the steps: extracting text data features at a semantic level, and having better effect in practical application; proposing improvements for an LDA model, utilizing point mutual information for quantizing the subject term relation, obtaining the co-occurrence relation betweensubject terms by calculating the weights of the subject terms, and setting a threshold value to select the optimal subject term. For the news automatic labeling method, keywords with high accuracy are selected according to the strength of the representation ability of vocabularies corresponding to different topics, and mutual information can be introduced to improve a topic-lexical item matrix, so that the accuracy of an LDA model in news document automatic label application is improved, and the correlation between subject terms is better described.

The invention discloses a dynamic feature selection method based on conditional mutual information, and the method specifically comprises the following steps: 1, carrying out the preprocessing of a data set, and obtaining a preprocessed data set; step 2, discretization processing is performed on the preprocessed data set, and all features in the preprocessed data set are divided into different feature levels; 3, calculating the importance degree between all the features X and the class variable Y in the data set subjected to discretization processing in the step 2; and step 4, according to theimportance I (X, Y) between the features and the classes calculated in the step 3, selecting the feature with the maximum importance as an important feature, deleting the important feature from the original feature set, adding the important feature into the candidate feature set to serve as a first candidate feature selected into the candidate feature set, and then calculating other candidate features. According to the invention, by improving the direct correlation between the features and the classes, the redundancy between the features is reduced, so that the accuracy and efficiency of feature selection are improved.

The invention discloses a refrigeration system fault diagnosis method and a refrigeration device, and belongs to the field of fault diagnosis and artificial intelligence, and the method comprises the following steps: (1) constructing a BN network model; (2) obtaining prior probability values of a target signal node and a fault signal node; (3) collecting BN network information; (4) performing relaxation operation on the data of the conditional mutual information matrix, and constructing a TAN classifier model matched with the fault features; (5) recalculating the conditional probability matrix; (6) calculating the posterior probability between the fault signal node and the characteristic signal node; (7) checking a posterior probability value; and (8) sorting the posterior probability values in each state from large to small, and taking the state corresponding to the maximum posterior probability value as a priority diagnosis / prediction classification result of the target signal node. The method is simple in network model construction, stable in classification efficiency, capable of accurately processing various types of tasks, insensitive to missing data, high in diagnosis speed and high in efficiency.

The invention discloses a cellular network base station state time-varying model establishing method based on a Bayesian network. The cellular network base station state time-varying model establishing method comprises the following steps of (1) using the existing actual cellular network as a scene, sensing states of a base station switch in a system model by using secondary sensing equipment in a cellular network, collecting sensing data and forming an observation sequence; (2) creating a Bayesian network model by using the observation sequence and learning the model according to a Bayesian structure learning algorithm of a totally connected graph and condition mutual information to obtain a value of a dependency relation between a conditional probability chart and nodes; and (3) establishing a time-varying statistic model of the states of the cellular network base station by using the value of the dependency relation between the conditional probability chart and the nodes. By the cellular network base station state time-varying model establishing method, the problem that the existing method is high in complexity and cannot be adaptively adjusted along with change of the nodes of the network is solved, by the base station state time-varying model with low complexity, data business collision probability of master mobile users of a cellular network is reduced effectively, and data transmission efficiency in the network is improved.

The invention provides a remote sensing hyperspectral image band selection method based on conditional mutual information, which comprises the following steps of: A, opening a hyperspectral remote sensing image and labeling a sample to be classified by man-machine interaction; B, grouping: carrying out band grouping by utilizing the conditional mutual information among all adjacent bands under given class condition according to the sample to be classified obtained from the step A; and C, searching: carrying out search calculation on the grouped bands obtained from the step B by utilizing a search algorithm combining a support vector machine and a genetic algorithm so as to find out an optimal band combination; and on this basis, pruning by using a self-adaptation branch and bound algorithm. Through the combinational use of the band grouping based on the conditional mutual information and the pruning based on the self-adaptation branch and bound algorithm, redundancy and noise groupingwhich are caused by noise perturbation are avoided, the frequency of band grouping is reduced and the classification accuracy of the band combination is improved.

The invention belongs to the technical field of data mining, and discloses a feature selection method and device based on conditional mutual information, equipment and a storage medium, and the method comprises the steps: obtaining a data set to form a candidate feature set F; calculating mutual information of each candidate feature in the candidate feature set F and the category attribute C, and putting the selected features into a feature set S; setting a threshold value, and entering circulation until the threshold value is met; training a model for the selected feature set S through a classifier, predicting the category by using the trained model, and calculating the prediction accuracy; changing the weight coefficient, repeatedly screening the feature sets S, calculating the prediction accuracy, and selecting the feature set S with the highest accuracy as a final output feature set. According to the method, feature selection can be performed more efficiently and quickly, and the precision and efficiency of data mining are improved.

The invention discloses a hydrological dependent structure modeling method based on mutual information and vine copula. Firstly, mutual information and conditional mutual information are used to measure the correlation and uncertainty of hydrological variables, in combination with the principle of the strongest correlation and the least uncertainty, the structure of vine copula is selected, starting from the first tree, the mutual information of pairs of paired variables is calculated, the pairing mode that maximizes the sum of mutual information is selected as the edge of the tree, the conditional mutual information of possible pairing variables is calculated, and the pairing mode that maximizes the sum of conditional mutual information is selected as tree 2, and the pairing mode is repeated until the structure of the whole tree is determined. Secondly, according to the tree structure, fitting of edge distribution is carried out, a goodness-of-fit test is carried out, starting from tree 1, the AIC criterion is utilized to determine the copula type of the edge, parameters are estimated, a goodness-of-fit test is performed, then the conditional edge distribution of variables is calculated, and the determination of copula type, estimating parameters and testing steps are repeated until all the trees are determined. All trees and edges are connected to complete the modeling of hydrological dependent structures.

The present invention relates to methods and arrangements in a multi-antenna radio communication system, in particular to methods and arrangements for improved multiple HARQ transmission in such systems. While HARQ transmission schemes, as known in the art, only can consider the fact whether or not a transmission attempt has been successful the present invention provides a HARQ retransmission scheme that considers the reception quality for already performed transmissions of a same data packet when selecting a resource allocation for necessary re-transmissions. Resource allocation for retransmissions is based on a pre-defined metric indicating a quality of the reception of the previous transmission attempts. Such a metric can be derived from a quality measure derived in the receiver unit, e.g. a CQI or CSI-based value, or an appropriate measure of the mutual information, e.g. the accumulated conditional mutual information (ACMI).

The invention discloses an intelligent decision making system reduction method based on ant colony, Comprising the following steps: (1) solving the attribute core of a decision making system, and initializing mutual information and iteration time; (2) generating k ants, initializing the k ants by the attribute core, and randomly selecting certain attribute for the k ants; (3) calculating heuristic information composed of attribute importance degree and pheromone; (4) selecting next attribute for each ant according to the heuristic information; (5) if the mutual information of the current ant is equal to the initial mutual information, ending the ant, and otherwise turning to (3); (6) obtaining a local solution; and (7) if the mutual information is less than the maximum iteration time or evolutionary trend, obtaining a global solution, outputting a minimum reduction, otherwise updating the pheromone and turning to (2). According to the technical scheme disclosed by the invention, the minimum reduction of the attribute in the decision making system can be quickly and effectively obtained, and information accuracy is effectively improved.

The invention discloses an interactive feature selection method based on neighborhood condition mutual information. The method comprises the steps: firstly, determining the neighborhood relation of each feature through employing an HCOM distance function for different data types, and calculating a neighborhood similarity relation matrix of each feature according to a multi-neighborhood radius set; secondly, exploring relevance between the features by utilizing neighborhood information, wherein the relevance comprises relevance between the features and classes and redundancy and interactivity between the features, and based on the relevance, establishing an evaluation function of feature importance of maximum relevance, minimum redundancy and maximum interactivity (MRmRMI). scoring the importance of the features through the evaluation function to obtain an ordered feature sequence with classification contributions from large to small; and finally, selecting a final reduction feature subset through testing on different classifiers, wherein the feature subset is a feature subset sequence corresponding to the optimal average classification performance. Compared with other six popular feature selection algorithms, the method of the invention has high classification performance and a more significant classification effect.

The invention discloses a method and system for generating a cross-modal bilateral personalized human-machine social dialogue. The invention includes encoding the dialogue context, encoding the personalized information of the robot, encoding the personalized information of the user, and encoding the output result at the last moment Perform weighted fusion to obtain a weighted fusion code, and then input the code of the output result at the previous moment into the decoder of the bilateral personalized generation model to generate the best N candidate reply list, and select the candidate reply with the largest conditional mutual information abundance value as the final output. The present invention fuses the personalized information in a cross-modal way, and at the same time considers the personalized information of the characters on both sides of the interaction, and makes full use of the personalization of the two sides of the interaction on the premise of ensuring that the reply content is reasonable, the grammar is smooth, and the logic is coherent. features that generate personalized, personalized responses.

A computer-implemented method and apparatus for supporting decisions in sequential clinical risk assessment examinations, the method comprising receiving one or more first test results and a question, both associated with a patient; and assessing by a processor associated with a computing platform, information gain provided by a second test which may be performed for the patient, as the conditional mutual information between a second test and the question, using the first test result.

The invention discloses an SAR image and visible light image registration method based on structural condition mutual information. The problem that the existing technology is unstable and the registering precision is low is mainly solved. The method comprises the following steps of 1) inputting a reference image and a to-be-registered image; 2) respectively calculating the phase consistency information processed by the non-local mean filtering algorithm of the reference image and the image to be registered; 3) respectively calculating the reference image and the phase consistency information of the image to be registered; 4) calculating the reference image and the phase consistency information of the image to be registered according to the reference image and the phase consistency information of the image to be registered; 5)recording the conversion parameters corresponding to the maximum mutual information of the structure conditions under the determination of the search space; 6)transforming the image to be registered by utilizing the conversion parameters, and obtaining a registration result. The method is stable in registration, high in registration precision and capable of being used for remote sensing image fusion and change detection.

The invention discloses a multi-scale time-frequency intermuscular coupling analysis method. The method comprises the following steps: firstly, synchronously acquiring and preprocessing multi-channelsurface electromyogram signals; and carrying out noise-assisted multivariate empirical mode decomposition on the preprocessed data to obtain useful IMF scale components; secondly, performing synchronous extraction transformation on IMF scale components, specifically, carrying out short-time Fourier transform on each IMF scale component, and then carrying out synchronous compression transform afterthe IMF scale component is multiplied by a phase factor; calculating time-frequency mutual information, time-frequency normalized mutual information and time-frequency condition mutual information; and finally, performing multi-scale time-frequency intermuscular coupling statistical analysis on the calculation result. The invention provides a new method for quantitatively researching the intermuscular nonlinear coupling strength characteristics of stroke patients under different time-frequency scales in the upper limb rehabilitation exercise process.

The invention discloses an affective computing method for a microblog emotion icon based on point mutual information. The method comprises the following steps: (1) extracting large-scale Sina microblogs and filtering and only remaining the microblogs including emotion icons and emotion words; (2) pre-processing the microblogs, combining the emotion words front connected with negative words and degree words and calculating emotion values thereof; (3) extracting 'emotion icon-emotion word' co-occurrence pairs from the pre-processed microblogs and forming a co-occurrence pair set; (4) calculating point mutual information of the emotion icons in the 'emotion icon-emotion word' co-occurrence pair set and each co-occurrence emotion word; (5) calculating an initial emotion value of each emotion icon; and (6) standardizing the initial emotion value of the emotion icon. According to the method, the point mutual information of the co-occurrence emotion word and the emotion icon is utilized to calculate and standardize the emotion values of the emotion icons; the method is simple and direct; the result is accurate.