182 results about "Matrix factorisation" patented technology

In a recommender method, Bayesian Matrix Factorization (BMF) is performed on a matrix having user and item dimensions and matrix elements containing user ratings for items made by users in order to train a probabilistic collaborative filtering model. A recommendation is generated for a user using the probabilistic collaborative filtering model. The recommendation may comprise a predicted item rating, or an identification of one or more recommended items. The recommender method is suitably performed by an electronic data processing device. The BMF may employ non-Gaussian priors, such as Student-t priors. The BMF may additionally or alternatively employ a heteroscedastic noise model comprising priors that include (1) a row dependent variance component that depends upon the matrix row and (2) a column dependent variance component that depends upon the matrix column.

System and method for performing Collaborative Filtering while preserving complete user anonymity are provided. Each of a group of client devices sends a rating vector anonymously to a server. The cells in each rating vector correspond to a set of items, and selected cells have ratings provided by the user associated with the corresponding client device for the corresponding items. The server aggregates all the rating vectors into a rating matrix, and factorizes the rating matrix into a user feature matrix and an item feature matrix through approximation, such that the rating matrix equals the product of the user feature matrix and the item feature matrix. The item feature matrix is sent to the client devices. Each of the client devices calculates its own user feature vector based on its rating vector and the item feature matrix, and provides personalized recommendations on selected items based on the client's user feature vector and the item feature matrix.

Documents of a set of documents are represented by bag-of-words (BOW) vectors. L labeled topics are provided, each labeled with a word list comprising words of a vocabulary that are representative of the labeled topic and possibly a list of relevant documents. Probabilistic classification of the documents generates for each labeled topic a document vector whose elements store scores of the documents for the labeled topic and a word vector whose elements store scores of the words of the vocabulary for the labeled topic. Non-negative matrix factorization (NMF) is performed to generate a document-topic model that clusters the documents into k topics where k>L. NMF factors representing L topics of the k topics are initialized to the document and word vectors for the L labeled topics. In some embodiments the NMF factors representing the L topics initialized to the document and word vectors are frozen, that is, are not updated by the NMF after the initialization.

The invention belongs to the technical field of e-commerce information processing, and discloses a cross-domain recommendation data processing method with multiple auxiliary domains and a cross-domainrecommendation data processing system. The method comprises obtaining a scoring matrix of an auxiliary domain, calculating the scoring reliability of a user, carrying out equal-proportion segmented mapping on a threshold value, and emptying scores with the number of scores lower than the threshold value in the auxiliary domain; obtaining clustering level scoring matrix of all domains by using a K-means clustering algorithm, and carrying out matrix decomposition; meanwhile, decomposing the target domain scoring matrix to learn a feature mapping function for the cold start user; evaluating predicted scoring matrix by using an average absolute error. Compared with the prior art, the method has the advantages that K-means clustering algorithm is used in the data processing process for obtaining clustering-level user project scoring matrix combined with all domains, the data sparsity of the cold start user is reduced. The problem that the recommendation effect is not ideal due to poor prediction accuracy of a traditional single auxiliary domain matrix factorization model is solved, the recommendation effect of the recommendation system is improved, and the method has higher universality.

The invention discloses a voice conversion method based on convolutive nonnegative matrix factorization. The method comprises the following steps: (1) training a transformation model through training data: carrying out time calibration and parameter decomposition of training voice data, analyzing STRAIGHT spectrum by using a convolutive nonnegative matrix factorization method, and analyzing pitch frequency of source voice and object voice; (2) converting new input voice based on a training model: carrying out parameter decomposition on source voice data A[c] to be converted by employing a STRAIGHT model, realizing sound channel frequency spectrum parameter conversion based on convolutive nonnegative matrix factorization, realizing conversion of the pitch frequency based on obtained mean value and variance in a training phase, and synthesizing voice after conversion, wherein the voice is voice after synthesis and conversion of the STRAIGHT spectrum S[Bc] which is obtained through conversion, the pitch frequency f[Bc] and original aperiodic component ap[Ac]. According to the invention, training effect of voice conversion is improved, and voice quality of conversion voice is improved.

Described is a system for sensor data fusion and reconstruction. The system extracts slices from a tensor having multiple tensor modes. Each tensor mode represents a different sensor data stream of incomplete sensor signals. The tensor slices are processed into demixed outputs. The demixed outputs are converted back into tensor slices, and the tensor slices are decomposed into mode factors using matrix decomposition. Mode factors are determined for all of the tensor modes, and the mode factors are assigned to tensor factors by matching mode factors common to two or more demixings. Tensor weight factors are determined and used for fusing the sensor data streams for sensor data reconstruction. Based on the sensor data reconstruction, hidden sensor data is extracted.

This invention generally relates to methods, apparatus and computer program for processing digital data using non-negative matrix factorisation. A method of digitally processing data in a data array defining a target matrix (X) using non-negative matrix factorisation to determine a pair of matrices (F, G), a first matrix of said pair determining a set of features for representing said data, a second matrix of said pair determining weights of said features, such that a product of said first and second matrices approximates said target matrix, the method comprising: inputting said target matrix data (X); selecting a row of said one of said first and second matrices and a column of the other of said first and second matrices; determining a target contribution (R) of said selected row and column to said target matrix; determining, subject to a non- negativity constraint, updated values for said selected row and column from said target contribution; and repeating said selecting and determining for the other rows and columns of said first and second matrices until all said rows and columns have been updated.

The invention discloses an integrated method used for a finance product recommending system. Based on a data smooth collaborative filtering algorithm, filling can be conducted on sparse data, and theproblem of sparsity of the data is solved. Based on a recommendation algorithm of population statistics, historic data is not needed, dependency on attributes of objects is not needed, and the problemof cold start can be solved for a user; the two kinds of algorithms are integrated with a recommendation algorithm which is good in expressive performance and is based on item clustering and matrix decomposition, the range of application scenes is enlarged for the recommendation algorithm, and the self-adaptability of the recommendation algorithm is improved. The integrated method can effectivelyreduce the sparsity of the data and solve the problem of cold start, and promotes recommendation performance for each user.

The invention discloses a deep learning method fusing scoring data and comment texts. The influence of scoring data sparsity is relieved by introducing auxiliary information. The preference information and the item characteristics of the user can be obtained by utilizing the comment text, and the score data also contains potential association between the user and the item. Most of existing fusionmodels adopt a matrix decomposition method to process score data, in order to better utilize effective information in the score data, a convolutional neural network is utilized to process comment texts, an attention mechanism is introduced to extract representative comments in comment information, and therefore, the score data fusion accuracy is improved. Therefore, user preferences and project characteristics can be better represented. By processing the score data by using a deep neural network to extract deep features in the score data and fusing the features, the method can predict the score of the project by the user.

The invention relates to a satellite magnetic field data seismic abnormality detection method based on non-negative matrix decomposition. The method comprises a step of removing invalid data accordingto a flag bit and subtracting CHAOS-6 magnetic field model data, and solving a first-order difference of a result to obtain differential data, a step of performing short-time Fourier transform on thedifferential data to construct a non-negative time-frequency amplitude matrix corresponding to magnetic field data, a step of decomposing the non-negative time-frequency amplitude matrix by using a non-negative matrix decomposition method and separating a local influence component generated by an earthquake from a global influence component generated by a sun activity and a geomagnetic activity,a step of selecting the local influence component generated by the earthquake according to an energy ratio and carrying out abnormal track judgment on the component through an over-limit rate method,and a step of accumulating the number of abnormal orbits every day and detecting earthquake abnormality according to the degree of deviation from a background fitting straight line. According to the invention, all data obtained by measurement can be reserved and utilized to research the earthquake, and at the same time, components more related to the earthquake activity can be obtained to effectively perform earthquake abnormality detection.

The invention discloses a social collaborative filtering recommendation method based on federal learning. The social collaborative filtering recommendation method comprises the following steps that 1, a central server selects user clients participating in local training; step 2, E rounds of local training are carried out on the user client, and in each round of local training process, a model loss function is calculated through a matrix decomposition module, a user comparison module and an article comparison module; 3, after the E rounds of local training are completed, a gradient which finally needs to be uploaded is obtained through a gradient protection module, and the gradient is uploaded to a central server for gradient aggregation; step 4, repeating the step 1 to the step 3 to obtain a fully trained user embedding vector and an article embedding vector; and 5, taking the user embedded vector and the article embedded vector as input of a score prediction module to obtain an article sequence recommended to the user. According to the method, the influence of the data non-independent identical distribution problem in federated social recommendation on the model recommendation performance can be effectively relieved.

The invention provides a cloud service QoS prediction method based on multi-source feature learning. The method comprises the following steps of extracting explicit features of users and services by utilizing context data of the users and the services, based on an implicit factor embedding method combining matrix decomposition and a neural network, extracting deep implicit features of users and services from the user-service matrix, splicing the explicit features and the implicit features to obtain a multi-source feature matrix corresponding to the user-service call record, and learning to obtain a local-global feature combination of the multi-source feature matrix by using a joint deep network based on a convolutional neural network, thereby obtaining QoS prediction of the cloud service by the user. According to the method, the high-order feature combination is learned from the multi-source information, and the influence of the feature sequence on the feature combination learning is fully considered, so that the QoS prediction precision is effectively improved.

The invention discloses a method for color standardization of pathological image based on low rank embedded non-negative matrix decomposition, the data of two pathological images are expressed in theform of observation matrix, which are converted into optical density, the low rank representation matrices of the source optical density matrix and the target optical density matrix are solved respectively, the corresponding optical density matrix is decomposed into an RGB color matrix of each dyeing component and an intensity matrix of each pixel under each dyeing component, At last, that image is convert to RGB form, so as to achieve the color standardization from the source pathological image to the target pathological image, the invention can be used for color standardization of pathological images and improvement of visualization effect, Avoiding the color inconsistency of pathological images caused by different pathological scanners, different laboratory staining methods and different staining ratio has important significance for pathological image imaging, staining separation and intelligent pathological image analysis, and has broad market prospects and application value.

Aiming at a multi-channel synthetic aperture radar system under a strong clutter background, the invention provides a multi-channel synthetic aperture radar RPCA amplitude-phase joint target detectionmethod and device in combination with a robust principal component analysis (PRCA) method and an amplitude-phase joint detection method. According to the method, firstly, focused images of all channels are subjected to column vectorization and then stacked to obtain a matrix X, then the matrix X is decomposed through an improved RPCA method to obtain a low-rank matrix L and a sparse matrix S, andfinally phase information is introduced for amplitude-phase joint detection. According to the method, specific clutter and moving target information are comprehensively considered, and the moving target detection performance is effectively improved; the advantages of an RPCA and amplitude-phase joint detection method are fully combined through a two-step detection method, and the false alarm rateunder a low signal-to-clutter ratio condition is reduced; and the matrix decomposition technology adopted by the method can greatly reduce the operation complexity and improve the algorithm efficiency.

The invention discloses a sample clustering and feature recognition method based on integrated non-negative matrix factorization. The method comprises: 1, X = {X1, X2... XP} representing multi-view data composed of P different omics data matrixes of the same cancer; 2, constructing a diagonal matrix Q; 3, introducing graph regularization and sparse constraints into the integrated non-negative matrix factorization framework to obtain target functions O1 and O2; 4, solving the target function O1 to obtain a fusion feature matrix W and a coefficient matrix HI; solving the target function O2 to obtain a feature matrix WI and a fusion sample matrix H; 5, constructing an evaluation vector according to the fusion feature matrix W, and identifying common difference features according to the vector; 6, performing functional explanation on the identified common difference characteristics by using GeneCards; and 7, performing sample clustering analysis according to the fusion sample matrix. According to the method, the complementary and difference information of the multiple omics data can be fully utilized to identify the common difference characteristics, clustering analysis can be carriedout on the sample data provided by the multiple omics data, and a calculation method basis is provided for integrated research of different types of omics data.

The current methods and apparatus provide a system that learns a private attribute, such as gender, based on at least one iteration of presenting an item to a user and receiving ratings from the user for this item. In an exemplary embodiment, the system may solicit ratings for strategically selected items, such as movies for example, and then infers the user's gender. Based on the assessed confidence in the demographic selected, the system may repeat the selection, presentation and ratings of another item. The proposed system can strategically select the sequence of items that are presented to the user for a rating. By selecting the next item to be rated based on a maximum posterior probability confidence, a demographic with a certain threshold of confidence can be inferred. The inventive arrangements are based on novel usage of Bayesian matrix factorization in an active learning setting. Such a system is shown to be feasible and can be carried out using significantly fewer rated items than previously proposed static inference methods.

The present development details a method and apparatus for performing channel emulation of doubly selective scenarios, where the simulation and emulation duration is arbitrarily long for a stationary or non-stationary channel, with non-separable dispersion which is achieved by combining the techniques of channel orthogonalization, decomposition of the correlation tensor in the Doppler domain into frequency-dependent correlation matrices, followed by a matrix factorization of each of the mentioned matrices and, finally, the use of the windowing method to generate arbitrarily long achievements which thereby allows the concatenation of channel realizations coming from the same or different NSSF, thus achieving reproduction of stationary or non-stationary channels, respectively.

A method, apparatus and computer-readable storage medium for determining one or more recommendations by applying efficient adaptive matrix factorization are disclosed. The method comprises causing, at least in part, an iterative performing of the following steps: a using of a current data set to optimize parameters used to adapt a current matrix factorization model by the end of the current time period, and a training of a current matrix factorization model and the current data set by the end of the current time period, based on the optimized parameters, to obtain an adapted matrix factorization model for service in a next time period.

The invention requests to protect a social recommendation method based on a heterogeneous information network, which comprises the following steps of: for the heterogeneous information network, normalizing each associated edge weight W in the network by using a mapping function to obtain a processed weight omega; and then searching all path instances under the corresponding meta-path by using a meta-path-based method, and calculating a similarity relationship s(x,y) between objects of the same type under each meta-path. In order to deeply mine the similarity of the objects, matrix decomposition is used for projecting the similarity relation of the objects to a low-dimensional feature space Feature, and each object can be subjected to characterization representation through a unique vectorin the space. After the feature information under all meta-paths is solved, all the feature data is input into the gradient boosting decision tree model, model training is carried out, the linear relation and the nonlinear relation between the features are learned, and therefore the accuracy of the recommendation model is improved.

The invention relates to the field of network structure deanonymization, in particular to a network structure deanonymization system and method based on matrix decomposition, and the method comprisesthe steps: taking a data source which comprises a topological structure of a relation between networks and a topological structure of a relation between users in the networks; dividing the data set ofthe dynamic social network into T static network data sets according to the time, and obtaining an adjacent matrix corresponding to each piece of static network structure information; inputting the adjacent matrix, performing training according to a low-rank representation model, and calculating and obtaining a target matrix; inputting the target matrix into an improved non-negative matrix decomposition model for training and prediction to obtain a de-anonymized matrix of the target matrix. According to the method, low-rank training is carried out on network data, a prediction model of non-negative matrix factorization is combined, structure information and time sequence evolution of the network are considered in a combined mode, and the prediction accuracy is improved.

The invention discloses a multi-view data missing complementing method for multi-manifold regularization non-negative matrix factorization, which comprises the following steps of: obtaining manifold and global clustering in which unmarked multi-view data tends to be consistent by utilizing a multi-manifold regularization non-negative matrix factorization algorithm through consistency assumption among multiple views; and constructing a multi-view collaborative discrimination model by adopting a view collaborative improved Gaussian mixture method. Pre-calibration of a cluster to which a sample belongs is realized by calculating the cluster relevancy level of the sample with missing data under a non-missing view angle; and establishing a missing data prediction model under a specific view angle by utilizing the consistency of multiple view angles in a low-dimensional space and adopting a multiple linear regression analysis method, thereby realizing accurate data completion under a multi-attribute missing condition. According to the method, large-scale labeled samples are not needed for training, the pre-defined category relation and related characteristics are avoided, and the understanding and discovering capacity of an existing multi-view mining technology for unlabeled multi-source data is improved.

This invention generally relates to data processing hardware, and more particularly to hardware accelerators and related methods for matrix factorisation especially non- negat ive matrix factorisation (NMF). Embodiments of the invention are particularly useful for driving electroluminescent displays such as OLED displays. A matrix factorisation hardware accelerator for determining a pair of factor matrices (R; C) which when multiplied together approximate a target matrix, the hardware accelerator comprising: an input to receive an input data matrix representing said target matrix; a first factor matrix memory for storing row and column data for a first factor matrix (R), said first factor matrix memory having a plurality of first data buses each associated with a respective block of said first factor matrix memory for accessing first factor matrix column data stored in the block; a second factor matrix memory for storing row and column data for a second factor matrix (C), said second factor matrix memory having a plurality of second data buses each associated with a respective block of said second factor matrix memory for accessing second factor matrix row data stored in the block; a matrix of processor blocks, each processor block having: a first processor block data bus coupled to one of said first data buses, a second processor block data bus coupled to one of said second data buses, and a result data output; a processor memory block for storing a portion of a matrix (Q) representing a difference between a product of said pair of factor matrices and said target matrix; and a data processor.

The invention discloses a data-to-text generation method based on fine-grained topic modeling. The data-to-text generation method comprises the following steps: learning semantic representation of each data record on a coding layer based on a bidirectional long-short-term memory network; learning topic distribution corresponding to each data record and word distribution corresponding to each topicbased on a non-negative matrix factorization method to obtain a topic word table corresponding to each data record; based on the semantic representation of each data record in a decoding layer, carrying out text generation by utilizing a long-term and short-term memory network, an attention mechanism and fine-grained topic representation in combination with a topic word table; and performing model training to obtain an optimal text generation result. According to the method, topic distribution of data and word distribution corresponding to topics are mined by utilizing a non-negative matrix factorization method, so that topic consistency between a generated text and a data table is restrained, and a model is guided to learn a more accurate word use mode; a copying mechanism is introducedin the text generation process, and it is guaranteed that the model can accurately generate numerical description.

The method of the invention definitely combines semi-supervised data representation and classification errors into a conventional non-negative matrix factorization framework for joint minimization learning, thereby applying weight coefficient construction and a label propagation process to the projection non-negative matrix factorization, avoiding the negative effects of noise, damage, or aliens included in original data on similarity measures and label prediction results. In addition, the joint minimization process can maintain the neighborhood information and the spatial structure in the projection non-negative matrix factorization process, and obtain an accurate data representation result. In addition, the weight construction and the inductive learning are integrated into a unified model so as to obtain an adaptive weight coefficient matrix, thereby avoiding the difficulty in selecting the optimal neighbor in a traditional algorithm. The method of the invention is an inductive model, can complete the induction and prediction of the data outside a sample, prevents introduction of additional reconstruction process, and has good scalability performance.

The invention discloses an implicit matrix factorization recommendation method based on differential privacy and time perception. The implicit matrix factorization recommendation method based on differential privacy and time perception can enable privacy protection and recommendation accuracy to be well balanced. The method comprises the following steps: firstly, normalizing score data of a user;before the score data of the user is sent to the recommendation system, distributing a privacy budget to each sub-scoring matrix; disturbing each score in the matrix; constructing a sub-matrix decomposition model by utilizing an improved implicit matrix decomposition algorithm; analyzing interests learned by the user from the data set at each time interval; obtaining a user characteristic matrix and a project characteristic matrix of a sub-matrix, mining the change trend of the characteristics of a user and a project along with time by utilizing an autoregression time sequence analysis model,calculating to obtain a behavior prediction value in a certain time period in the future, and selecting TOP-N projects with the maximum prediction score values to recommend the user.

The invention provides a time perception service recommendation system and method based on a self-attention factor decomposition machine, and the system comprises a PFM model which inputs a user hiding vector, a service hiding vector and a time hiding vector, input through an input layer, into an FM module for processing, and outputs the processed data to an output layer through a first full connection layer and a first activation function of a middle layer in sequence; and an SAGRU model which inputs the embedded vectors of the user and the service set on the time interval t input through theinput layer into the GRU module for processing, and outputs the processed data to the output layer through the self-attention mechanism unit, the second full connection layer and the second activation function of the middle layer in sequence. Compared with a matrix decomposition technology, the method has the advantages that the nonlinear relationship between the user and the service can be effectively learned, the dynamic behavior characteristics of the user along with time change can be captured, and the problem of service quality data sparseness in the real world can be effectively relieved.

Aiming at the clustering problem of a high-dimensional multivariate time sequence, the invention provides a multi-view clustering method based on constrained non-negative matrix factorization, in order to effectively carry out sample clustering, multivariate time sequence data is firstly projected to a multi-relation network, and then a hierarchical method is adopted to generate a plurality of independent views. The method also includes, secondly, clustering the plurality of views by adopting a constrained non-negative matrix factorization method; and finally, solving in an alternate iterativeoptimization mode, and performing k-means clustering on the representation matrix obtained by calculation of the solving result to obtain a clustering result. The correctness and effectiveness of themethod and the algorithm provided by the invention are verified through experiments, and experimental results show that the clustering method has relatively high accuracy and effectiveness.