Multi-view clustering method, system, device and medium based on non-negative matrix decomposition

Abstract

The present application discloses a multi-view clustering method, system, device and medium based on non-negative matrix decomposition. The method obtains the original data matrix of multiple views; performs multiple non-negative matrix decompositions on the original data matrix to obtain a clustering target matrix and a spatial basis matrix; constructs a first target loss function and a second target loss function based on the original data matrix, the clustering target matrix and the spatial basis matrix, wherein the first target loss function is used to indicate denoising of the original data matrix, and the second target loss function is used to indicate feature importance evaluation of the original data matrix to select features according to feature importance; and uses an alternating direction multiplier method to iterate the first target loss function and the second target loss function to obtain a target clustering result. The present application can improve the accuracy of clustering results, thereby improving the clustering effect.

Description

Technical Field

[0001] The present application relates to the technical field of multi-view clustering, and in particular to a multi-view clustering method, system, device and medium based on non-negative matrix decomposition. Background Art

[0002] As multi-view data becomes more and more common in many real-world applications, learning from multi-view data is an emerging problem in machine learning research. For example, multi-class data is now ubiquitous, and different types of information can be obtained from the same object, such as texture information, shape information, color information, etc. In these scenarios, the same set of objects has different views collected by different measurement methods or ways, where any specific single view data may not be sufficient to fully describe the information of all objects. Therefore, a major goal of multi-view unsupervised learning is to find consensus clustering across views so that similar objects are grouped into the same cluster, while dissimilar objects are divided into different clusters.

[0003] Existing related methods address this problem through different objective functions driven by different applications. However, most existing multi-view NMF related methods either only define view weights and ignore the effectiveness of different features in a single view, or a priori know the weights of specific features when deriving consensus clustering. In practice, such assumptions may not be valid because datasets are often feature-rich and have many noisy features. Since these noisy features may lead to clustering errors, existing related methods have poor clustering effects. Summary of the invention

[0004] The present application aims to propose a multi-view clustering method, system, device and medium based on non-negative matrix decomposition, which can improve the accuracy of clustering results and thus improve the clustering effect.

[0005] In a first aspect, an embodiment of the present application provides a multi-view clustering method based on non-negative matrix decomposition, the method comprising:

[0006] Get the original data matrix of multiple views;

[0007] Performing multiple non-negative matrix decompositions on the original data matrix to obtain a clustering target matrix and a spatial basis matrix;

[0008] Based on the original data matrix, the clustering target matrix and the spatial basis matrix, construct a first objective loss function and a second objective loss function, wherein the first objective loss function is used to indicate denoising of the original data matrix, and the second objective loss function is used to indicate feature importance evaluation of the original data matrix to perform feature selection according to the feature importance;

[0009] The first objective loss function and the second objective loss function are iterated by using an alternating direction multiplier method to obtain a target clustering result.

[0010] Compared with the prior art, the first aspect of the present application has the following beneficial effects:

[0011] This method obtains the clustering target matrix and the spatial basis matrix by performing multiple non-negative matrix decompositions on the original data matrix; based on the original data matrix, the clustering target matrix and the spatial basis matrix, the first target loss function and the second target loss function are constructed, wherein the first target loss function is used to indicate the denoising of the original data matrix, and the second target loss function is used to indicate the feature importance evaluation of the original data matrix, so as to select features according to the feature importance. Since the original data matrix often contains various nonlinear and non-Gaussian noises and outliers, the denoising of the original data matrix by the first target loss function can well reduce the influence of the complex noise of high-dimensional data on the clustering results, and enhance the robustness of the clustering process; at the same time, the feature importance of the original data matrix is ​​evaluated by the second target loss function, and then the feature selection is performed according to the feature importance, which can improve the correctness of clustering. The alternating direction multiplier method is used to iterate the first target loss function and the second target loss function. The alternating direction multiplier method can improve the convergence speed and has good convergence performance. After convergence, the clustering results can be directly obtained according to the weighting, which improves the clustering effect.

[0012] In some implementations, constructing a first objective loss function and a second objective loss function based on the original data matrix, the clustering target matrix, and the spatial basis matrix includes:

[0013] Construct feature scale factor matrix;

[0014] Constraining the relationship between the clustering target matrix and the original data matrix through the spatial basis matrix, and constructing a first objective loss function introducing the feature scale factor matrix by constraining the non-negativity of the spatial basis matrix, the clustering target matrix and the original data matrix;

[0015] pass The norm constrains the original data matrix and the clustering target matrix, and constructs a second objective loss function that introduces the feature scale factor matrix.

[0016] In some embodiments, the feature scale factor matrix is ​​constructed as follows:

[0017]

[0018] in, represents the characteristic scale factor in the jth row and jth column of the characteristic scale factor matrix, Represents the characteristic scale factor matrix The element in the jth column, represents the element of the jth row of the clustering target matrix, Represents the total number of rows of the clustering target matrix, express Norm.

[0019] In some embodiments, the first objective loss function introducing the feature scale factor matrix is ​​constructed as follows:

[0020]

[0021] in, represents the original data matrix, represents the feature scale factor matrix, represents the clustering target matrix, represents the spatial basis matrix, represents transpose, represents the hyperparameter, express Norm.

[0022] In some embodiments, the second objective loss function introducing the feature scale factor matrix is ​​constructed as follows:

[0023]

[0024] in, and represents an adjustable hyperparameter, represents the original data matrix, represents the feature scale factor matrix, represents the clustering target matrix, represents the loss term, represents a vector whose elements are all 1, express Norm.

[0025] In some implementations, the step of iterating the first objective loss function and the second objective loss function using an alternating direction multiplier method to obtain a target clustering result includes:

[0026] Adding the first objective loss function and the second objective loss function to obtain a total loss function;

[0027] The feature scale factor matrix, the clustering target matrix and the original data matrix are input into the total loss function, and the total loss function is iterated by using the alternating direction multiplication method to obtain the target clustering result.

[0028] In some embodiments, the inputting the feature scale factor matrix, the clustering target matrix, and the original data matrix into a total loss function, and iterating the total loss function using an alternating direction multiplier method to obtain a target clustering result includes:

[0029] Taking any one parameter among the feature scale factor matrix, the cluster target matrix and the original data matrix as a variable, and the other two parameters as constants;

[0030] The variables and the constants are input into the total loss function, and the total loss function is iterated cyclically using the alternating direction multiplier method until a preset condition is met to obtain a target clustering result.

[0031] In a second aspect, an embodiment of the present application further provides a multi-view clustering system based on non-negative matrix decomposition, the system comprising:

[0032] A data matrix acquisition unit, used for acquiring the original data matrix of multiple views;

[0033] A non-negative matrix decomposition unit, used for performing multiple non-negative matrix decompositions on the original data matrix to obtain a clustering target matrix and a spatial basis matrix;

[0034] A loss function construction unit, configured to construct a first target loss function and a second target loss function based on the original data matrix, the clustering target matrix and the spatial basis matrix, wherein the first target loss function is used to indicate denoising of the original data matrix, and the second target loss function is used to indicate feature importance evaluation of the original data matrix, so as to perform feature selection according to the feature importance;

[0035] The clustering result obtaining unit is used to iterate the first target loss function and the second target loss function by using an alternating direction multiplier method to obtain a target clustering result.

[0036] In the third aspect, an embodiment of the present application also provides an electronic device, comprising at least one control processor and a memory for communicating with the at least one control processor; the memory stores instructions that can be executed by the at least one control processor, and the instructions are executed by the at least one control processor so that the at least one control processor can execute a multi-view clustering method based on non-negative matrix decomposition as described above.

[0037] In a fourth aspect, an embodiment of the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute a multi-view clustering method based on non-negative matrix decomposition as described above.

[0038] It can be understood that the beneficial effects of the second to fourth aspects compared with the related art are the same as the beneficial effects of the first aspect compared with the related art. Please refer to the relevant description in the first aspect, and no further details will be given here. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0040] Figure 1 It is a flowchart of an embodiment of a multi-view clustering method based on non-negative matrix decomposition provided by the present application;

[0041] Figure 2 is a schematic diagram of a related entropy constraint process in an embodiment of a multi-view clustering method based on non-negative matrix decomposition provided by the present application;

[0042] Figure 3 is a schematic diagram of a feature selection constraint flow in an embodiment of a multi-view clustering method based on non-negative matrix decomposition provided by the present application;

[0043] Figure 4 It is a schematic diagram of an alternating iteration process of an alternating direction multiplier method in an embodiment of a multi-view clustering method based on non-negative matrix decomposition provided by the present application;

[0044] Figure 5 is a structural schematic diagram of an embodiment of a multi-view clustering system based on non-negative matrix decomposition provided by the present application;

[0045] Figure 6 It is a schematic diagram of the structure of an embodiment of the electronic device provided by the present application. DETAILED DESCRIPTION

[0046] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be understood as limiting the present application.

[0047] In the description of this application, if there is a description of first, second, etc., it is only for the purpose of distinguishing technical features, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features.

[0048] In the description of the present application, it should be understood that the descriptions involving orientation, such as the orientation or positional relationship indicated as up, down, etc., are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.

[0049] In the description of this application, it should be noted that, unless otherwise clearly defined, terms such as setting, installing, connecting, etc. should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meaning of the above terms in this application based on the specific content of the technical solution.

[0050] As multi-view data becomes more and more common in many real-world applications, learning from multi-view data is an emerging problem in machine learning research. For example, multi-class data is now ubiquitous, and different types of information can be obtained from the same object, such as texture information, shape information, color information, etc. In these scenarios, the same set of objects has different views collected by different measurement methods or ways, where any specific single view data may not be sufficient to fully describe the information of all objects. Therefore, a major goal of multi-view unsupervised learning is to find consensus clustering across views so that similar objects are grouped into the same cluster, while dissimilar objects are divided into different clusters. In the literature, such a learning problem is called multi-view clustering. Multi-view clustering algorithms are actually widely used in a wide range of scenarios, including big data, large model processing, user portrait analysis, vision and other fields.

[0051] There are two main categories of methods in the existing literature: generative (model-based) and discriminative (similarity and dimensionality reduction-based). For generative methods, we usually use mixture models and regression-based matrix factorization. The idea is to model the probability distribution of each data view and obtain a common clustering result by allowing all views to share the same prior or derive from shared latent factors. The advantage of generative methods is that it can well explain the basis for building clusters, but in the context of multi-view learning, this method is more computationally expensive. Discriminative methods focus on optimizing the objective functions of average similarity within clusters and differences between clusters. Different objective functions lead to different methods, such as multi-view clustering algorithms, multi-view subspace clustering algorithms, and multi-task multi-view clustering algorithms. Discriminative methods usually involve non-convex objective functions, and it may be difficult to find a good solution. In the field of computer vision and pattern recognition, the high dimensionality of input data increases the computational complexity and makes it difficult to process data.

[0052] Existing related methods address this problem through different objective functions driven by different applications. However, most existing multi-view NMF related methods either only define view weights and ignore the effectiveness of different features in a single view, or a priori know the weights of specific features when deriving consensus clustering. In practice, such assumptions may not be valid because datasets are often feature-rich and have many noisy features. A small number of outlier features and a large amount of high-dimensional data make traditional multi-view clustering algorithms challenging. Since these noisy features may lead to clustering errors, existing related methods have poor clustering effects.

[0053] In order to solve the problems of the above method being complex in calculation, prone to producing erroneous clustering results and poor clustering effect, the present application proposes a multi-view clustering method, system, device and medium based on non-negative matrix decomposition.

[0054] Reference Figure 1 The embodiment of the present application provides a multi-view clustering method based on non-negative matrix decomposition, and the method comprises the following steps:

[0055] Step S100, obtaining the original data matrix of multiple views;

[0056] Step S200, performing multiple non-negative matrix decompositions on the original data matrix to obtain a clustering target matrix and a spatial basis matrix;

[0057] Step S300, constructing a first objective loss function and a second objective loss function based on the original data matrix, the clustering target matrix and the spatial basis matrix, wherein the first objective loss function is used to indicate denoising of the original data matrix, and the second objective loss function is used to indicate feature importance evaluation of the original data matrix, so as to perform feature selection according to feature importance;

[0058] Step S400: Adopting the alternating direction multiplier method to iterate the first objective loss function and the second objective loss function to obtain a target clustering result.

[0059] In this embodiment, the clustering target matrix and the spatial basis matrix are obtained by performing non-negative matrix decomposition on the original data matrix for multiple times; based on the original data matrix, the clustering target matrix and the spatial basis matrix, the first target loss function and the second target loss function are constructed, wherein the first target loss function is used to indicate that the original data matrix is ​​denoised, and the second target loss function is used to indicate that the original data matrix is ​​evaluated for feature importance, so as to select features according to feature importance. Since the original data matrix often contains various nonlinear and non-Gaussian noises and outliers, the original data matrix is ​​denoised by the first target loss function, which can well reduce the influence of complex noise of high-dimensional data on clustering results and enhance the robustness of the clustering process; at the same time, the original data matrix is ​​evaluated for feature importance by the second target loss function, and then feature selection is performed according to feature importance, which can improve the correctness of clustering. The alternating direction multiplier method is used to iterate the first target loss function and the second target loss function, and the alternating direction multiplier method can improve the convergence speed and has good convergence performance, and can directly obtain the clustering results according to weighting after convergence, thereby improving the clustering effect.

[0060] The above non-negative matrix decomposition can make all the decomposed components non-negative (requiring a purely additive description) and simultaneously achieve non-linear dimensionality reduction.

[0061] The above-mentioned alternating direction multiplier method (i.e., ADMM algorithm) can be a simple method to solve decomposable convex optimization problems, especially effective in solving large-scale problems. Using the ADMM algorithm, the objective function of the original problem can be equivalently decomposed into several solvable sub-problems, and then each sub-problem is solved in parallel, and finally the solutions of the sub-problems are coordinated to obtain the global solution of the original problem.

[0062] In some embodiments, based on the original data matrix, the clustering target matrix and the spatial basis matrix, a first objective loss function and a second objective loss function are constructed, including:

[0063] Construct feature scale factor matrix;

[0064] By constraining the relationship between the clustering target matrix and the original data matrix through the spatial basis matrix, and by constraining the non-negativity of the spatial basis matrix, the clustering target matrix and the original data matrix, a first objective loss function introducing a feature scale factor matrix is ​​constructed;

[0065] pass The norm constrains the original data matrix and the clustering target matrix, and constructs the second objective loss function that introduces the feature scale factor matrix.

[0066] In this embodiment, the regression coefficients are rescaled by the feature scale factor matrix to evaluate the importance of the features. Therefore, by constructing the first objective loss function introducing the feature scale factor matrix and the second objective loss function introducing the feature scale factor matrix, the noise of important features can be better removed (i.e., many noisy features are removed), and suitable features can be better selected, thereby laying a good data foundation for later clustering.

[0067] Above Norms include Norm and Norm.

[0068] The above feature scaling factor matrix is ​​obtained by considering the addition of an implicit Regularization, a sparse feature scale factor matrix is ​​learned to achieve the purpose of selecting appropriate features in high-dimensional data.

[0069] In some embodiments, the feature scale factor matrix is ​​constructed as follows:

[0070]

[0071] in, represents the characteristic scale factor in the jth row and jth column of the characteristic scale factor matrix, Represents the characteristic scale factor matrix The element in the jth column, represents the element of the jth row of the clustering target matrix, Represents the total number of rows of the clustering target matrix, express Norm.

[0072] In this embodiment, by considering adding an implicit Regularization, a sparse feature scale factor matrix is ​​learned to achieve the purpose of selecting appropriate features in high-dimensional data.

[0073] In some embodiments, the first objective loss function introducing the feature scale factor matrix is ​​constructed as follows:

[0074]

[0075] in, represents the original data matrix, represents the feature scale factor matrix, represents the clustering target matrix, represents the spatial basis matrix, represents transpose, represents the hyperparameter, express Norm.

[0076] In this embodiment, by constructing a first objective loss function that introduces a feature scale factor matrix for denoising, nonlinear noise in the original data matrix can be better removed.

[0077] In some embodiments, the second objective loss function introducing the feature scale factor matrix is ​​constructed as follows:

[0078]

[0079] in, and represents an adjustable hyperparameter, represents the original data matrix, represents the feature scale factor matrix, represents the clustering target matrix, represents the loss term, represents a vector whose elements are all 1, express Norm.

[0080] In this embodiment, by constructing a second objective loss function that introduces a feature scale factor matrix, the importance of features can be evaluated through the feature scale factor matrix, so that appropriate features can be selected according to the importance of the features.

[0081] In some implementations, the first objective loss function and the second objective loss function are iterated using an alternating direction multiplier method to obtain a target clustering result, including:

[0082] Add the first objective loss function and the second objective loss function to obtain the total loss function;

[0083] The feature scale factor matrix, clustering target matrix and original data matrix are input into the total loss function, and the total loss function is iterated using the alternating direction multiplication method to obtain the target clustering result.

[0084] In this embodiment, the total loss function is iterated by the alternating direction multiplier method until the clustering result converges. The alternating direction multiplier method is usually used to solve optimization problems with only equality constraints and two optimization variables. The use of the alternating direction multiplier method can increase the convergence speed and has good convergence performance. After convergence, the clustering result is directly obtained based on the weighting, thereby improving the clustering effect.

[0085] In some embodiments, the feature scale factor matrix, the clustering target matrix, and the original data matrix are input into the total loss function, and the total loss function is iterated using the alternating direction multiplier method to obtain the target clustering result, including:

[0086] Take any one parameter among the feature scale factor matrix, cluster target matrix and original data matrix as a variable and the other two parameters as constants;

[0087] The variables and constants are input into the total loss function, and the total loss function is iterated cyclically using the alternating direction multiplier method until the preset conditions are met and the target clustering result is obtained.

[0088] In this embodiment, any one of the characteristic scale factor matrix, the clustering target matrix and the original data matrix is ​​used as a variable, and the other two parameters are used as constants. That is, the alternating direction multiplier method reduces the scale of the problem by dividing a large problem into two sub-problems, thereby accelerating the convergence speed.

[0089] The above preset condition can be that the total loss function is minimized, that the total loss function converges, or that a preset number of iterations is reached.

[0090] The above preset number of iterations can be changed according to actual conditions and is not specifically limited in this embodiment.

[0091] To facilitate understanding by those skilled in the art, a set of best embodiments is provided below:

[0092] Existing related methods address this problem through different objective functions driven by different applications. However, most existing multi-view NMF related methods either only define view weights and ignore the effectiveness of different features in a single view, or a priori know the weights of specific features when deriving consensus clustering. In practice, such assumptions may not be valid because datasets are often feature-rich and have many noisy features. A small number of outlier features and a large amount of high-dimensional data make traditional multi-view clustering algorithms challenging.

[0093] Usually, when we deal with multi-view datasets, each view in the multi-view data usually involves a large number of features. Different feature components in different views have different dimensions. Therefore, during the learning process, the different importance of feature components in each view needs to be considered. If there are irrelevant features in the clustering process, the clustering algorithm will take longer to calculate and may even produce wrong clustering results, especially for multi-view data.

[0094] To solve the above problems, this embodiment proposes a multi-view clustering method based on non-negative matrix decomposition. This embodiment can clean (i.e., denoise) and cluster analyze the original data matrix, and finally obtain the specific category of the corresponding sample. Basic data mining and processing problems such as clustering have wide applications in the fields of big data, user portrait analysis, etc. Therefore, this application has research significance and practical significance. The solution proposed in this embodiment specifically includes the following steps:

[0095] Step 1. Perform certain preprocessing on the given original data set, including standardization and normalization, etc. The given original data set may be a text data set.

[0096] Step 2. After the preprocessing is completed, the processed original data set (i.e., the original data matrix) is subjected to multiple non-negative matrix factorizations to separate it into a clustering target matrix W and a spatial basis matrix V.

[0097] Step 3. Use the Gaussian kernel function to increase the dimension of the clustering target matrix W and the spatial basis matrix V. After the increase in dimension, perform relevant entropy constraints to identify outliers in the low-dimensional part.

[0098] Specifically, refer to Figure 2 , the original data matrix is ​​divided into to After inputting several views together, the characteristic factor matrix is ​​used to influence them and perform non-negative decomposition. The two matrices of non-negative decomposition are defined as the clustering target matrix W, which has further constraints in the subsequent process to obtain the clustering results, and the other is the spatial basis matrix V to constrain the relationship between W and the original matrix X and its non-negativity. The outer layer performs the Gaussian kernel correlation entropy constraint (including Gaussian kernel function and correlation entropy function) process, which can map the original data matrix to a high-dimensional space, thereby mining the nonlinear noise in the original data matrix, and denoising in the iteration in the form of correlation entropy constraints.

[0099] In practice, only a limited number of samples are usually available, and The joint distribution function of is often unknown. In this case, the relevant entropy can be approximated as:

[0100] (1)

[0101] in, represents the relevant entropy estimator, represents the hyperparameter, represents the total number of samples, represents the clustering target matrix after high-dimensional mapping, represents the spatial basis matrix after high-dimensional mapping, Represents input The probability density distribution of Represents input The probability density distribution of represents the probability density function, Represents an exponential function.

[0102] The maximum value of the correlation entropy in formula (1) is called the maximum correlation entropy criterion. Figure 2, construct the loss function of the Gaussian kernel correlation entropy process. The input here is the clustering target matrix W and the spatial basis matrix V obtained by non-negative decomposition, and the feature part is added for constraint. In summary, the loss function of this part (that is, the first target loss function) is as follows:

[0103] (2)

[0104] in, represents the original data matrix, represents the feature scale factor matrix, represents the clustering target matrix, represents the spatial basis matrix, represents transpose, express Norm.

[0105] Step 4. The feature selection constraint process constructs an objective loss function for feature selection (i.e., the second objective loss function), which mainly includes a feature weight scaling matrix (i.e., the feature scale factor matrix, which includes multiple scale factors ), an original data matrix, a loss fitting term and a tunable hyperparameter.

[0106] Specifically, refer to Figure 3 , the objective loss function of feature selection is constructed as follows:

[0107] (3)

[0108] in, and represents an adjustable hyperparameter, represents the original data matrix, represents the feature scale factor matrix, represents the clustering target matrix, represents the loss term, represents a vector whose elements are all 1, express Norm.

[0109] Here is the scaling matrix of the diagonal matrix, by considering adding the implicit Regularization to learn sparse , so as to achieve the purpose of selecting appropriate features in high-dimensional data, so there are:

[0110] (4)

[0111] in, represents the characteristic scale factor in the jth row and jth column of the characteristic scale factor matrix, Represents the characteristic scale factor matrix The element in the jth column, Represents the element of the jth row of the clustering target matrix. Here, it is assumed that each row in the original data matrix X is a sample. Represents the total number of rows of the clustering target matrix, express Norm.

[0112] Step 5. Then constrain the entire module (i.e. construct a total loss function) to avoid overfitting of the model. Figure 4 , the data is iterated by alternating direction multiplication method until the clustering target matrix converges, that is, the change in iteration does not exceed 5%. Then the clustering result of the original data is directly obtained by weight.

[0113] Specifically, add formula (2) and formula (3) to construct the total loss function for:

[0114] (5)

[0115] in, represents the first objective loss function, represents the second objective loss function, represents the original data matrix.

[0116] The total loss function is iterated by alternating direction multiplication method, and variable data and constant data are input into the total loss function for iteration until the clustering target matrix converges or the total loss function is minimized as a whole, that is, the change in iteration does not exceed 5%. Then the clustering result of the original data is directly obtained by weight.

[0117] For example, when the alternating direction multiplication method iterates the clustering target matrix W, it treats the clustering target matrix W as a variable and treats all other matrices (i.e. and ) are all regarded as constants, that is, invariants. At this time, the extreme value of W is found by derivation or other methods, and then W is updated. This is done for all variable matrices in each round of iteration.

[0118] In this embodiment, the present application is robust to high-dimensional data, and after introducing feature selection, the clustering performance is improved while making the clustering results more interpretable. Therefore, the present application has good practical significance and research significance.

[0119] Reference Figure 5 The embodiment of the present application also provides a multi-view clustering system based on non-negative matrix decomposition, the system includes a data matrix acquisition unit 100, a non-negative matrix decomposition unit 200, a loss function construction unit 300 and a clustering result acquisition unit 400, wherein:

[0120] The data matrix acquisition unit 100 is used to acquire the original data matrix of multiple views;

[0121] A non-negative matrix decomposition unit 200 is used to perform multiple non-negative matrix decompositions on the original data matrix to obtain a clustering target matrix and a spatial basis matrix;

[0122] A loss function construction unit 300 is used to construct a first target loss function and a second target loss function based on the original data matrix, the clustering target matrix and the spatial basis matrix, wherein the first target loss function is used to indicate denoising of the original data matrix, and the second target loss function is used to indicate feature importance evaluation of the original data matrix to perform feature selection according to feature importance;

[0123] The clustering result obtaining unit 400 is used to iterate the first target loss function and the second target loss function by using an alternating direction multiplier method to obtain a target clustering result.

[0124] It should be noted that since the multi-view clustering system based on non-negative matrix decomposition in this embodiment and the above-mentioned multi-view clustering method based on non-negative matrix decomposition are based on the same inventive concept, the corresponding contents in the method embodiment are also applicable to the system embodiment and will not be described in detail here.

[0125] Reference Figure 6 The embodiment of the present application further provides an electronic device, the electronic device comprising:

[0126] at least one memory;

[0127] at least one processor;

[0128] at least one program;

[0129] The programs are stored in the memory, and the processor executes at least one program to implement the multi-view clustering method based on non-negative matrix decomposition in the present disclosure.

[0130] The electronic device may be any intelligent terminal including a mobile phone, a tablet computer, a personal digital assistant (PDA), a car computer, etc.

[0131] The electronic device according to the embodiment of the present application is described in detail below.

[0132] The processor 1600 may be implemented by a general-purpose central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided by the embodiments of the present disclosure;

[0133] The memory 1700 may be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 1700 may store an operating system and other application programs. When the technical solution provided in the embodiment of this specification is implemented by software or firmware, the relevant program code is stored in the memory 1700, and the processor 1600 calls and executes the multi-view clustering method based on non-negative matrix decomposition of the embodiment of the present disclosure.

[0134] Input / output interface 1800, used to implement information input and output;

[0135] Communication interface 1900, used to realize communication interaction between the device and other devices, which can be realized through wired mode (such as USB, network cable, etc.) or wireless mode (such as mobile network, WIFI, Bluetooth, etc.);

[0136] Bus 2000 , which transmits information between various components of the device (e.g., processor 1600 , memory 1700 , input / output interface 1800 , and communication interface 1900 );

[0137] The processor 1600 , the memory 1700 , the input / output interface 1800 , and the communication interface 1900 are connected to each other in communication within the device via the bus 2000 .

[0138] The embodiment of the present disclosure also provides a storage medium, which is a computer-readable storage medium. The computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute the above-mentioned multi-view clustering method based on non-negative matrix decomposition.

[0139] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0140] The embodiments described in the embodiments of the present disclosure are intended to more clearly illustrate the technical solutions of the embodiments of the present disclosure and do not constitute a limitation on the technical solutions provided by the embodiments of the present disclosure. Those skilled in the art will appreciate that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of the present disclosure are also applicable to similar technical problems.

[0141] Those skilled in the art will appreciate that the technical solutions shown in the figures do not limit the embodiments of the present disclosure and may include more or fewer steps than shown in the figures, or a combination of certain steps, or different steps.

[0142] The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separated, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0143] Those skilled in the art will appreciate that all or some of the steps in the methods disclosed above, and the functional modules / units in the systems and devices may be implemented as software, firmware, hardware, or a suitable combination thereof.

[0144] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0145] It should be understood that in the present application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0146] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0147] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0148] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0149] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including multiple instructions to enable an electronic device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), disk or optical disk and other media that can store programs. The above is a detailed description of the embodiments of the present application in conjunction with the accompanying drawings, but the present application is not limited to the above embodiments. Within the scope of knowledge possessed by ordinary technicians in the relevant technical field, various changes can be made without departing from the purpose of the present application.

Claims

1. A multi-view clustering method based on non-negative matrix decomposition, characterized in that: The method comprises: Acquire an original data matrix of multiple views, wherein the original data matrix is ​​different types of information obtained from multiple objects, including texture information, shape information, and color information; Performing multiple non-negative matrix decompositions on the original data matrix to obtain a clustering target matrix and a spatial basis matrix; Based on the original data matrix, the clustering target matrix and the spatial basis matrix, a first objective loss function and a second objective loss function are constructed, wherein the first objective loss function is used to indicate that the original data matrix is ​​denoised, and the second objective loss function is used to indicate that the feature importance of the original data matrix is ​​evaluated to perform feature selection according to the feature importance; specifically comprising: Construct feature scale factor matrix; Constraining the relationship between the clustering target matrix and the original data matrix through the spatial basis matrix, and constructing a first objective loss function introducing the feature scale factor matrix by constraining the non-negativity of the spatial basis matrix, the clustering target matrix and the original data matrix; pass The norm constrains the original data matrix and the clustering target matrix, and constructs a second objective loss function that introduces the feature scale factor matrix; The feature scale factor matrix is ​​constructed as follows: ； in, represents the characteristic scale factor in the jth row and jth column of the characteristic scale factor matrix, Represents the characteristic scale factor matrix The element in the jth column, represents the element of the jth row of the clustering target matrix, Represents the total number of rows of the clustering target matrix, express norm; The first objective loss function introducing the feature scale factor matrix is ​​constructed as follows: ； in, represents the original data matrix, represents the feature scale factor matrix, represents the clustering target matrix, represents the spatial basis matrix, represents transpose, represents the hyperparameter, express norm; The second objective loss function introducing the feature scale factor matrix is ​​constructed as follows: ； in, and represents an adjustable hyperparameter, represents the original data matrix, represents the feature scale factor matrix, represents the clustering target matrix, represents the loss term, represents a vector whose elements are all 1, express norm; The first objective loss function and the second objective loss function are iterated by using an alternating direction multiplier method to obtain a target clustering result.

2. The multi-view clustering method based on non-negative matrix decomposition according to claim 1, characterized in that: The step of iterating the first objective loss function and the second objective loss function by using the alternating direction multiplier method to obtain a target clustering result includes: Adding the first objective loss function and the second objective loss function to obtain a total loss function; The feature scale factor matrix, the clustering target matrix and the original data matrix are input into the total loss function, and the total loss function is iterated by using the alternating direction multiplication method to obtain the target clustering result.

3. The multi-view clustering method based on non-negative matrix decomposition according to claim 2, characterized in that: The step of inputting the feature scale factor matrix, the clustering target matrix and the original data matrix into a total loss function, and iterating the total loss function using an alternating direction multiplier method to obtain a target clustering result includes: Taking any one parameter among the feature scale factor matrix, the cluster target matrix and the original data matrix as a variable, and the other two parameters as constants; The variables and the constants are input into the total loss function, and the total loss function is iterated cyclically using the alternating direction multiplier method until a preset condition is met to obtain a target clustering result.

4. A multi-view clustering system based on non-negative matrix factorization, characterized in that: The system comprises: A data matrix acquisition unit, used to acquire original data matrices of multiple views, wherein the original data matrices are different types of information obtained from multiple objects, including texture information, shape information, and color information; A non-negative matrix decomposition unit, used for performing multiple non-negative matrix decompositions on the original data matrix to obtain a clustering target matrix and a spatial basis matrix; A loss function construction unit is used to construct a first target loss function and a second target loss function based on the original data matrix, the clustering target matrix and the spatial basis matrix, wherein the first target loss function is used to indicate that the original data matrix is ​​denoised, and the second target loss function is used to indicate that the original data matrix is ​​evaluated for feature importance, so as to perform feature selection according to the feature importance; specifically comprising: Construct feature scale factor matrix; Constraining the relationship between the clustering target matrix and the original data matrix through the spatial basis matrix, and constructing a first objective loss function introducing the feature scale factor matrix by constraining the non-negativity of the spatial basis matrix, the clustering target matrix and the original data matrix; pass The norm constrains the original data matrix and the clustering target matrix, and constructs a second objective loss function that introduces the feature scale factor matrix; The feature scale factor matrix is ​​constructed as follows: ； in, represents the characteristic scale factor in the jth row and jth column of the characteristic scale factor matrix, Represents the characteristic scale factor matrix The element in the jth column, represents the element of the jth row of the clustering target matrix, Represents the total number of rows of the clustering target matrix, express norm; The first objective loss function introducing the feature scale factor matrix is ​​constructed as follows: ； in, represents the original data matrix, represents the feature scale factor matrix, represents the clustering target matrix, represents the spatial basis matrix, represents transpose, represents the hyperparameter, express norm; The second objective loss function introducing the feature scale factor matrix is ​​constructed as follows: ； in, and represents an adjustable hyperparameter, represents the original data matrix, represents the feature scale factor matrix, represents the clustering target matrix, represents the loss term, represents a vector whose elements are all 1, express norm; The clustering result obtaining unit is used to iterate the first target loss function and the second target loss function by using an alternating direction multiplier method to obtain a target clustering result.

5. An electronic device, characterized in that: It includes at least one control processor and a memory for communicating with the at least one control processor; the memory stores instructions that can be executed by the at least one control processor, and the instructions are executed by the at least one control processor so that the at least one control processor can execute the multi-view clustering method based on non-negative matrix decomposition as described in any one of claims 1 to 3.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute the multi-view clustering method based on non-negative matrix decomposition as described in any one of claims 1 to 3.

Images