A multi-view based news theme mining method

Abstract

This invention discloses a news topic mining method based on multiple views, belonging to the technical field of text analysis and data mining. First, the invention constructs an LT-MSC model, then builds manifold regularization, sparse constraint, and diversity regularization terms. By introducing the manifold regularization term, geometric information in the multiple-view news data is mined; by using the sparse constraint term, the block diagonal structure of the subspace representation matrix is ​​enhanced; and by incorporating the diversity regularization term, complementary information between different views in the news data is captured. Finally, spectral clustering is used to cluster the data. Compared to existing single-view and multi-view mining schemes, the method of this invention achieves the best clustering effect and recognition performance in news topic mining scenarios, thereby effectively improving the efficiency and accuracy of news topic mining.

Description

Technical Field

[0001] This invention relates to a news topic mining method based on multiple views, belonging to the fields of news recommendation and public opinion analysis. Background Technology

[0002] With the rapid development of internet technology, a massive amount of news data is generated online every day. However, real-life news data often contains a certain amount of noise, making the effective extraction of essential information from this data a highly challenging task. Mining thematic content from news data is also a significant problem. Mining effective thematic content from news data has crucial value in practical applications, such as detecting trending topics on social media and providing early warnings of public opinion. However, several challenges remain to be addressed in the effective mining of news data: firstly, existing methods still suffer from low accuracy in mining thematic content; secondly, the extraction of feature information from news data is not deep enough, as current methods for extracting news features often do not consider the essential information within the news data.

[0003] Xu Shuang and Yu Li proposed a multi-view clustering method using regularized matrix decomposition technology (MRMD). The method uses multivariate nonnegative matrix decomposition technology and norm to describe the loss function of matrix decomposition, and adopts multi-view manifold regularization to regularize the matrix decomposition to improve the effectiveness of the identification information in the mined data (Xu Shuang, Yu Li. Multi-view clustering method using regularized matrix decomposition technology [J]. Computer Engineering and Applications, 2019, 55(14):142-147.).

[0004] Wang et al. proposed a graph-based multi-view clustering (GMC) method. This method uses multi-view fusion technology to merge the data graph matrices of all views and automatically weights each data graph matrix to generate a unified graph matrix. Without introducing tuning parameters, a rank constraint is imposed on the graph Laplacian matrix of the unified matrix, which is more conducive to naturally dividing the data points into the required number of clusters (Wang H, Yang Y, Liu B. GMC: Graph-based multi-view clustering[J]. IEEE Transactions on Knowledge and Data Engineering, 2019, 32(6):1116-1129.).

[0005] While the methods described above are effective in mining news topics, they cannot solve the problem of noisy data when the original news data contains noise. Therefore, their performance in mining news topics will be significantly affected. Summary of the Invention

[0006] To address the aforementioned problems, this invention proposes a multi-view-based news topic mining method. To solve the problem of ineffective news topic mining caused by noise in the original news data, this invention captures the sparse representation of news data through sparse constraints, mines complementary information between different views in the news data through diversity regularization, and combines a low-rank constraint multi-view subspace clustering method to construct a model for a multi-view-based news topic mining method. The method includes:

[0007] Step 1: Transform the multi-view news data matrix A (1) A (2) A (V) Decomposed into matrix A (v) F (v) S and matrix S (v) ,Right now:

[0008] A (v) =A (v) F (v) +S (v) v = 1, 2, ..., V

[0009] Among them, A (v) F (v) F (v) and S (v) Let A represent the low-rank part, subspace representation part, and reconstruction error part corresponding to the v-th view, respectively. (V) ∈R d×n d represents the dimension of the sample in the news data, n represents the number of news data, and V represents the number of views;

[0010] Step 2: Based on the subspace representation part F of each single view (v) Construct an N-order tensor

[0011] F = Ψ(F (1) ,F (2) ,...,F (V) )

[0012] Where Ψ(·) represents the subspace representation matrix F of different views. (v) After merging, an N-order tensor F is constructed;

[0013] Expanding the N-order tensor F over the nth mode yields the matrix F. (n) The expression is:

[0014]

[0015] Among them, I n Represents matrix F (n) dimensionality;

[0016] Step 3: Calculate the nuclear norm of the Nth-order tensor F, using the following formula:

[0017]

[0018] in,‖·‖ * ζ represents the nuclear norm of a tensor. n It is a constant greater than 0, and satisfies

[0019] Step 4: Construct a low-rank tensor-constrained multi-view subspace clustering LT-MSC model:

[0020]

[0021] Where S = [S (1) S (2) ;...;S (V) The symbol represents the matrix formed by concatenating the reconstruction error matrices of each view. 2，1 L represents the matrix 2,1 Norm;

[0022] Step 5: Calculate the hypergraph Laplacian matrix for each view.

[0023] Step 6: Construct the manifold regularization term Where tr(·) represents the trace of the matrix;

[0024] Step 7: Construct sparse constraint terms ||F T F||1; where ‖·‖1 denotes the L1 norm of the matrix;

[0025] Step 8: Construct diversity regularization terms

[0026] Step 9: Based on the multi-view subspace clustering LT-MSC model and the manifold regularization term Sparse constraint term ||F T F||1, Diversity Regularization Term Construct a news topic mining model based on multiple views. The model's expression is:

[0027]

[0028] Where λ1, λ2, λ3 and λ4 are equilibrium parameters;

[0029] Step 10: Update variable μ, the update formula is:

[0030] μ = min(ρμ, μ) max )

[0031] Where ρ represents a given positive penalty parameter, and μ represents a penalty parameter greater than 0.

[0032] Step 11: Update the subspace representation matrix F (v) The reconstruction error matrix S, the auxiliary variable G n The Lagrange multiplier B v The vectorization f of the Nth-order tensor F, and the auxiliary variable G n vectorization of g n and Lagrange multipliers α n until ||A is satisfied (v) -A v F (v) -S (v) || ∞ <ε, and ||P n fg n || ∞ <ε, where ε represents a very small positive number that satisfies the convergence condition;

[0033] Wherein, the updated subspace representation matrix F (v) The update process includes:

[0034] According to the formula The multi-view-based news topic mining model is further represented in the following form:

[0035]

[0036] stA (v) =A (v) F (v) +S (v) v = 1, 2, ..., V

[0037] F = Ψ(F (1) ,F (2) ,...,F (V) ),

[0038] S = [S (1) S (2) ;...;S (V) ],

[0039] diag(F (v) ) = 0

[0040] in,

[0041] Optimizing the above representation of the present invention can be accomplished using the Augmented Lagrange Multiplier Method (ALM). When solving the problem using the ALM method, the objective function of the present invention needs to be separable; therefore, an auxiliary variable G is introduced. n To replace F (n) Then we have:

[0042]

[0043] stP n f = g n n = 1, 2, ..., N

[0044] A (v) =A (v) F (v) +S (v) v = 1, 2, ..., V

[0045] F = Ψ(F (1) ,F (2) ,...,F (V) ),

[0046] S = [S (1) S (2) ;...;S (V) ],

[0047] diag(F (v) ) = 0

[0048] Among them, f and g n These are the Nth-order tensor F and the auxiliary variable G, respectively. n Vectorization, P n It is a permutation matrix used to calibrate F (n) and the auxiliary variable G n The corresponding elements between them;

[0049] By using the augmented Lagrange multiplier method, the multi-view-based news topic mining model is further represented as follows:

[0050]

[0051] Where, α n and B v It is a Lagrange multiplier, μ is the equilibrium parameter, and L μ>0 Represents the augmented Lagrangian function of a multi-view-based news topic mining model;

[0052] When the reconstruction error S of each view(v) and auxiliary variable G n When fixed, the subspace representation matrix F (v) The update formula is:

[0053]

[0054] Among them, Ω v (·) indicates selecting the element corresponding to the v-th view, and then reconstructing these elements into a matrix, where I1 represents a matrix of all 1s;

[0055] The updated subspace representation matrix is ​​then obtained. as follows:

[0056]

[0057] in,

[0058] Step 12: Calculate the similarity matrix

[0059] Step 13: Based on the similarity matrix obtained in Step 12, classify the news topics using spectral clustering and output the final classification results.

[0060] Optionally, the formula for calculating the hypergraph Laplacian matrix of each view in step five is:

[0061]

[0062] Among them, D C H is a diagonal matrix whose diagonal elements are the degree of each vertex; G ∈R |C|×|T| The hypergraph's incidence matrix represents the relationships between vertices and hyperedges; W represents the set of hyperedge weights; D T A diagonal matrix representing hyperedge degree.

[0063] Optionally, the process of constructing the manifold regularization term in step six includes:

[0064] The manifold regularization is hypergraph regularization, where the hypergraph G = (C, T, W) consists of a set of vertices C, a set of hyperedges T, and a set of weights W for each hyperedge. Each hyperedge t is assigned a positive weight w(t), and the elements in the weight matrix W are related to each hyperedge t. T and C satisfy U t∈T =C, U represents the union operation;

[0065] The correlation matrix H of the hypergraph G G ∈R |C|×|T|The elements of the incidence matrix, used to represent the relationships between vertices and hyperedges, are expressed as follows:

[0066]

[0067] in, Let c represent the degree of each hyperedge, and c represent the vertex.

[0068] The manifold regularization term is:

[0069] Optionally, the process of constructing the diversity regularization term in step nine includes:

[0070]

[0071] in, H represents a matrix with a mean of 0.

[0072] Optionally, the process of updating and reconstructing the error matrix S includes:

[0073] When the subspace representation matrix F of the single view (v) and the auxiliary variable G n When fixed, the update formula for the reconstruction error matrix S can be obtained as follows:

[0074]

[0075] Where Z is matrix A (v) -A (v) F (v) +B (v) The solution to the above equation is obtained by connecting the columns vertically together to form a matrix:

[0076]

[0077] Among them, Z :,i Let represent the i-th column of matrix Z.

[0078] Optionally, the auxiliary variable G n The iterative update formula is as follows:

[0079]

[0080] in, Ω n (P n f+a n ) is based on the nth pattern of the expansion, which will be used to expand P n f+a n Reshape into the corresponding matrix. This represents the spectral soft thresholding operation of a matrix, where the spectrum of a matrix refers to the maximum value of the modulus of its eigenvalues.

[0081] Optionally, the Lagrange multiplier B v The update formula is as follows:

[0082]

[0083] Optionally, the method for updating the vectorized f of the tensor F is: directly replacing the updated F. (v) Update f using the corresponding elements in the table, i.e., f * ←F (v) .

[0084] Optionally, the Lagrange multiplier α n The update rules are as follows:

[0085]

[0086] This invention also provides a news topic mining system based on multiple views, comprising:

[0087] The data acquisition module is used to acquire news data;

[0088] The data classification module is used to classify news topics based on the news data using the aforementioned multi-view-based news topic mining method.

[0089] The output display module is used to output the categorization results of news topics.

[0090] The beneficial effects of this invention are:

[0091] This invention presents a multi-view-based news topic mining method. Building upon the LT-MSC framework, it first introduces manifold regularization and sparse constraints, then utilizes HSIC as a diversity regularization term to construct a multi-view-based news topic mining model. Spectral clustering is then employed for clustering. Compared to existing technologies, this invention considers low-rank principles to mine essential identification information from news data, uses tensors to explore higher-order correlations between different views in news data, uses manifold regularization to mine geometric information in multi-view news data, uses sparse constraints to enhance the block diagonal structure of the subspace representation matrix, and utilizes HISC as a diversity regularization term to enhance the complementarity between different views in news data. Simulation results demonstrate that, compared to existing single-view and multi-view mining methods, this invention effectively addresses the problem of noisy data and achieves optimal clustering and recognition performance in news topic mining scenarios, thereby significantly improving the efficiency and accuracy of news topic mining. Attached Figure Description

[0092] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0093] Figure 1 This is an overall flowchart of the news topic mining method based on multiple views according to the present invention.

[0094] Figure 2 This is an experimental result diagram of Embodiment 2 of the present invention on the news dataset 3Sources.

[0095] Figure 3 This is a graph showing the experimental results of Embodiment 2 of the present invention on the BBC news dataset.

[0096] Figure 4 This is an experimental result diagram of Embodiment 2 of the present invention on the news dataset BBCSport2view.

[0097] Figure 5 This is an experimental result diagram of Embodiment 2 of the present invention on the news dataset WebKB. Detailed Implementation

[0098] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0099] The basic theoretical knowledge involved in this invention is introduced as follows:

[0100] 1. K-means clustering method

[0101] The basic idea of ​​the K-means method is to cluster objects based on k points in space, grouping them into categories that are closest to these k points. The values ​​of each cluster center are updated iteratively until the best clustering result is obtained.

[0102] 2. Spectral Clustering Method

[0103] Spectral clustering is an unsupervised learning method based on graph theory. This method divides a weighted undirected graph into two or more optimal subgraphs, making the subgraphs as similar as possible internally and as far apart as possible between them, in order to achieve the purpose of clustering.

[0104] 3. Evaluation Criteria for the Model

[0105] The evaluation criteria employ four methods: accuracy (AC), normalized mutual information (NMI), precision, and adjusted Rand index (AR).

[0106] 4. HSIC (Hilbert-Schmidt independence criterion)

[0107] HSIC is often used to measure the correlation between two sets of data. ab It is the cross covariance, where a and b are two random variables, and Φ(a) is a mapping from a∈A to the kernel space F, where the inner product between vectors is defined by a kernel function k1(a). i ,a j )=<Φ(a i ),Φ(a j The following is given. Let S be the second kernel space over B, and the specific kernel function be k2(b i ,b j )=<Φ(b i ),Φ(b j Cross covariance C ab It is a function of the variances of two random variables, and its expression is as follows:

[0108]

[0109] in, Let E(·) represent the tensor product, E(·) represent the expectation of the variable, and η represent the tensor product. a =E[(Φ(a)],η b =E[(Φ(b)];

[0110] F and S are two given separable regenerable nucleus Hilbert spaces (RKHS), p ab Let be a joint probability distribution over spaces A and B. The expression for HSIC is as follows:

[0111]

[0112] Among them, ||C ab || HS Representing matrix C ab Hilbert-Schmidt norm,

[0113] HSIC expresses the degree of independence between samples in space A and space B. In reality, the joint probability distribution p on spaces A and B... abIt is often difficult to estimate or is unknown, therefore we consider starting from p. ab n independent observations extracted An estimate of HSIC is written as HSIC(P,F,S), then HSIC(P,F,S)=(n-1) -2 tr(K1HK2H);

[0114] Where K1 and K2 represent Gram matrices, and k 1,ij =k1(a i ,a j ),k 2,ij =k2(b i ,b j H ij =δ ij -1 / n is a matrix with a mean of 0.

[0115] 5. The Augmented Lagrange Multiplier (ALM) method is based on the Lagrange multiplier method, incorporating a penalty function to construct a more suitable objective function. This method obtains an approximate optimal solution to the objective function by continuously adjusting the Lagrange multipliers and penalty parameters during the search for the optimal solution.

[0116] Example 1:

[0117] This embodiment provides a news topic mining method based on multiple views. See [link to relevant documentation]. Figure 1 The method includes:

[0118] Step 1: Transform the multi-view news data matrix A (1) A (2) A (V) Decomposed into matrix A (v) F (v) S and matrix S (v) ,Right now:

[0119] A (v) =A (v) F (v) +S (v) v = 1, 2, ..., V

[0120] Among them, A (v) F (v) F (v) and S (v) Let A represent the low-rank part, subspace representation part, and reconstruction error part corresponding to the v-th view, respectively. (V) ∈R d×nd represents the dimension of the sample in the news data, n represents the number of news data, and V represents the number of views;

[0121] Step 2: Based on the subspace representation part F of each single view (v) Construct an N-order tensor

[0122] F = Ψ(F (1) ,F (2) ,...,F (V) )

[0123] Where Ψ(·) represents the subspace representation matrix F of different views. (v) After merging, an N-order tensor F is constructed;

[0124] Expanding the N-order tensor F on the nth mode yields the matrix F. (n) The expression is:

[0125]

[0126] Among them, I n Represents matrix F (n) dimensionality;

[0127] Step 3: Calculate the nuclear norm of the Nth-order tensor F, using the following formula:

[0128]

[0129] in,‖·‖ * Let ζ represent the nuclear norm of a tensor. n It is a constant greater than 0, and satisfies

[0130] Step 4: Construct a low-rank tensor-constrained multi-view subspace clustering LT-MSC model:

[0131]

[0132] Where S = [S (1) S (2) ;...;S (V) The symbol represents the matrix formed by concatenating the reconstruction error matrices of each view. 2，1 L represents the matrix 2,1 Norm;

[0133] Step 5: Calculate the hypergraph Laplacian matrix for each view.

[0134] Step 6: Construct the manifold regularization term Where tr(·) represents the trace of the matrix;

[0135] Step 7: Construct sparse constraint terms ||F T F||1; where ‖·‖1 denotes the L1 norm of the matrix;

[0136] Step 8: Construct diversity regularization terms

[0137] Step 9: Based on the multi-view subspace clustering LT-MSC model and the manifold regularization term Sparse constraint term ||F T F||1, Diversity Regularization Term Construct a news topic mining model based on multiple views. The model's expression is:

[0138]

[0139] Where λ1, λ2, λ3 and λ4 are equilibrium parameters;

[0140] Step 10: Update variable μ, the update formula is:

[0141] μ = min(ρμ, μ) max )

[0142] Where ρ represents a given positive penalty parameter, and μ represents a penalty parameter greater than 0.

[0143] Step 11: Update the subspace representation matrix F (v) The reconstruction error matrix S, the auxiliary variable G n The Lagrange multiplier B v The vectorization f of the Nth-order tensor F, and the auxiliary variable G n vectorization of g n and Lagrange multipliers α n until ||A is satisfied (v) -A v F (v) -S (v) || ∞ <ε, and ||P n fg n || ∞ <ε, where ε represents a very small positive number that satisfies the convergence condition;

[0144] Wherein, the updated subspace representation matrix F (v) The update process includes:

[0145] According to the formula The multi-view-based news topic mining model is further represented in the following form:

[0146]

[0147] stA (v) =A (v) F (v) +S (v) v = 1, 2, ..., V

[0148] F = Ψ(F (1) ,F (2) ,...,F (V) ),

[0149] S = [S (1) S (2) ;...;S (V) ],

[0150] diag(F (v) ) = 0

[0151] in,

[0152] Optimizing the above representation of the present invention can be accomplished using the Augmented Lagrange Multiplier Method (ALM). When solving the problem using the ALM method, the objective function of the present invention needs to be separable; therefore, an auxiliary variable G is introduced. n To replace F (n) Then we have:

[0153]

[0154] stP n f = g n n = 1, 2, ..., N

[0155] A (v) =A (v) F (v) +S (v) v = 1, 2, ..., V

[0156] F = Ψ(F (1) ,F (2) ,...,F (V) ),

[0157] S = [S (1) S (2) ;...;S (V) ],

[0158] diag(F (v) ) = 0

[0159] Among them, f and g n These are the Nth-order tensor F and the auxiliary variable G, respectively. n Vectorization, P n It is a permutation matrix used to calibrate F(n) and the auxiliary variable G n The corresponding elements between;

[0160] By using the augmented Lagrange multiplier method, the multi-view-based news topic mining model is further represented as follows:

[0161]

[0162] Where, α n and B v It is a Lagrange multiplier, μ is the equilibrium parameter, and L μ>0 This represents the augmented Lagrangian function of the model in this embodiment;

[0163] When the reconstruction error S of each view (v) and auxiliary variable G n When fixed, the subspace representation matrix F (v) The update formula is:

[0164]

[0165] Among them, Ω v (·) indicates selecting the element corresponding to the v-th view, and then reconstructing these elements into a matrix, where I1 represents a matrix of all 1s;

[0166] The updated subspace representation matrix is ​​then obtained. as follows:

[0167]

[0168] in,

[0169] Step 12: Calculate the similarity matrix

[0170] Step 13: Based on the similarity matrix obtained in Step 12, classify the news topics using spectral clustering and output the final classification results.

[0171] Example 2:

[0172] This embodiment provides a news topic mining method based on multiple views. See [link to relevant documentation]. Figure 1 The method includes:

[0173] (1) Given a multi-view news data matrix A (1) A (2) A (V) Here A (V) ∈R d×nWhere d represents the dimension of the sample in the news data, n represents the number of news data, and V represents the number of views;

[0174] (2) Calculate the matrix A corresponding to the v-th view in the multi-view news data. (v) Decomposed into matrix A (v) F (v) S and matrix S (v) ,Right now:

[0175] A (v) =A (v) F (v) +S (v) v = 1, 2, ..., V

[0176] Among them, A (v) F (v) F (v) and S (v) These represent the low-rank part, subspace representation part, and reconstruction error part (or sparse noise) corresponding to the v-th view, respectively.

[0177] (3) Based on the subspace representation part F of each single view (v) Construct an N-order tensor

[0178] F = Ψ(F (1) ,F (2) ,...,F (V) )

[0179] Where Ψ(·) represents the subspace representation matrix F of different views. (v) After merging, construct an N-order tensor F.

[0180] Expanding the N-order tensor F on the nth mode yields the matrix F. (n) The expression is:

[0181]

[0182] Among them, I n Represents matrix F (n) The dimension of.

[0183] (4) According to the formula Calculate the nuclear norm of the Nth-order tensor F;

[0184] in,‖·‖ * Let ζ represent the nuclear norm of a tensor. n It is a constant greater than 0, and satisfies

[0185] (5) Construct a low-rank tensor-constrained multi-view subspace clustering model (LT-MSC):

[0186]

[0187] Where S = [S (1) S (2) ;...;S (V) The matrix formed by concatenating the reconstruction error matrices of each view is ‖·‖ 2，1 L represents the matrix 2,1 Norm;

[0188] (6) Calculate the Laplacian matrix L of the hypergraph of a single view. h :

[0189] A hypergraph G = (C, T, W) consists of a set of vertices C, a set of hyperedges T, and a set of weights W for each hyperedge. Each hyperedge t is assigned a positive weight w(t), and the elements in the weight matrix W are related to each hyperedge t. T and C satisfy U t∈T =C, U represents the union operation;

[0190] The incidence matrix H of the hypergraph G G ∈R |C|×|T| The elements of the incidence matrix, used to represent the relationships between vertices and hyperedges, are defined as follows:

[0191]

[0192] in, This represents the degree of each hyperedge.

[0193] The Laplace matrix L of the hypergraph G h for:

[0194]

[0195] Among them, D C It is a diagonal matrix whose diagonal elements are the degree of each vertex, D. T A diagonal matrix representing the degree of a hyperedge.

[0196] (7) Constructing manifold regularization terms Here, tr(·) represents the trace of the matrix.

[0197] In this embodiment, manifold regularization refers to hypergraph regularization, and the manifold regularization term is:

[0198] (8) Construct sparse constraint terms ||F T F||1; where ‖·‖1 represents the L1 norm of the matrix.

[0199] (9) Constructing diversity regularization terms

[0200] HSIC is often used to measure the correlation between two sets of data. ab It is the cross covariance, where a and b are two random variables, and Φ(a) is a mapping from a∈A to the kernel space F, where the inner product between vectors is defined by a kernel function k1(a). i ,a j )=<Φ(a i ),Φ(a j )>Given.

[0201] Let S be the second kernel space over B, and let the specific kernel function be k2(b i ,b j )=<Φ(b i ),Φ(b j Cross covariance C ab It is a function of the variances of two random variables, and its expression is as follows:

[0202]

[0203] in, Let E(·) represent the tensor product, E(·) represent the expectation of the variable, and η represent the tensor product. a =E[(Φ(a)],η b =E[(Φ(b)];

[0204] F and S are two given separable regenerable nucleus Hilbert spaces (RKHS), p ab Let be a joint probability distribution over spaces A and B. The HSIC expression is as follows:

[0205]

[0206] Among them, ||C ab || HS Representing matrix C ab Hilbert-Schmidt norm,

[0207] HSIC expresses the degree of independence between samples in space A and space B. In reality, the joint probability distribution p on spaces A and B... ab Often, these are difficult to estimate or unknown, therefore, we should consider starting from p. ab n independent observations extracted An estimate of HSIC is written as HSIC(P,F,S), then HSIC(P,F,S)=(n-1) -2 tr(K1HK2H);

[0208] Where K1 and K2 represent Gram matrices, and k 1,ij =k1(ai ,a j ),k 2,ij =k2(b i ,b j H ij =δ ij -1 / n is a matrix with a mean of 0.

[0209] For a given view, when the HSIC scaling factor (n-1) is ignored... -2 When using HSIC as the regularization term for diversity, the expression is as follows:

[0210]

[0211] in, H represents a matrix with a mean of 0.

[0212] (10) Construct a model for a news topic mining method based on multiple views;

[0213] Based on the LT-MSC framework, the aforementioned manifold regularization term and sparsity constraint term are introduced. The sparsity constraint applies to the product of the transpose of the subspace representation matrix and itself. In addition, a diversity regularization term is incorporated. The formula for the news topic mining method based on multiple views in this embodiment is as follows:

[0214]

[0215] Where λ1, λ2, λ3, and λ4 are equilibrium parameters. Indicates the first A matrix representing the subspace of each view. It is based on F (v) The Laplacian matrix of the hypergraph of the concrete view on the graph, with the manifold regularization term being: The sparse constraint term is Diversity regularization term is

[0216] (11) Update variable μ;

[0217] μ = min(ρμ, μ) max )

[0218] Where ρ represents a given positive penalty parameter, and μ represents a penalty parameter greater than 0.

[0219] (12) Update the subspace representation matrix F (v) ;

[0220] According to the formula The model of the news topic mining method based on multiple views in this embodiment can be further represented in the following form:

[0221]

[0222] stA (v) =A (v) F (v) +S (v) v = 1, 2, ..., V

[0223] F = Ψ(F (1) ,F (2) ,...,F (V) ),

[0224] S = [S (1) S (2) ;...;S (V) ],

[0225] diag(F (v) ) = 0

[0226] in, γ n Indicates the penalty parameter;

[0227] Optimizing the above representation of the present invention can be accomplished using the Augmented Lagrange Multiplier Method (ALM). When solving the problem using the ALM method in this embodiment of the invention, it is necessary to make the objective function of this embodiment separable. Therefore, an auxiliary variable G is introduced into the model of this embodiment. n To replace F (n) Then we have:

[0228]

[0229] stP n f = g n n = 1, 2, ..., N

[0230] A (v) =A (v) F (v) +S (v) v = 1, 2, ..., V

[0231] F = Ψ(F (1) ,F (2) ,...,F (V) ),

[0232] S = [S (1) S (2) ;...;S (V) ],

[0233] diag(F (v) ) = 0

[0234] Among them, f and gn They are tensor F and matrix G, respectively. n Vectorization, P n It is a permutation matrix used to calibrate F (n) and G n The corresponding elements between;

[0235] By using the Augmented Lagrange Multiplier (ALM) method, the model in this embodiment can be further represented as follows:

[0236]

[0237] Where, α n and B v It is a Lagrange multiplier, μ is the equilibrium parameter, and L μ>0 This represents the augmented Lagrangian function of the model in this invention;

[0238] F (v) When S (v) and G n When fixed, F (v) This can be obtained by solving the following subproblem:

[0239]

[0240] Among them, Ω v (·) indicates selecting the element corresponding to the v-th view, and then reconstructing these elements into a matrix. I1 represents a matrix of all 1s;

[0241] It can be obtained as follows:

[0242]

[0243] in,

[0244] (13) Update and reconstruct the error matrix S;

[0245] S: When F (v) and G n When fixed, we can obtain S * as follows:

[0246]

[0247] Where Z is matrix A (v) -A (v) F (v) +B (v) The solution to the above equation can be obtained by connecting the columns vertically together.

[0248]

[0249] Among them, Z :,i Let represent the i-th column of matrix Z.

[0250] (14) Update auxiliary variable G n ;

[0251] G n Auxiliary variable G n The iterative formula is as follows:

[0252]

[0253] in, Ω n (P n f+a n ) is based on the nth pattern of the expansion, which will be used to expand P n f+a n Reshape into the corresponding matrix. This represents the spectral soft thresholding operation of a matrix, where the spectrum of a matrix refers to the maximum value of the modulus of its eigenvalues.

[0254] (15) Update Lagrange multiplier B v ;

[0255] B v Lagrange multiplier B v The update rules are as follows:

[0256]

[0257] (16) Update the vectorization of tensor F;

[0258] Regarding the aforementioned updated F (v) You can directly replace F (v) Update f using the corresponding elements in the table, i.e., f * ←F (v) .

[0259] (17) Update auxiliary variable G n vectorization of g n ;

[0260] g n :g n The update is similar to the update of f, and you can obtain... as follows:

[0261]

[0262] (18) Update the Lagrange multiplier α n ;

[0263] α n : Lagrange multiplier α n Updates and B v Similarly, α n The update rules are as follows:

[0264]

[0265] Until A is satisfied (v) -A v F (v) -S (v) || ∞ <ε, and ||P n fg n || ∞ <ε, where ε represents a very small positive number that satisfies the convergence condition;

[0266] (19) Calculate the similarity matrix

[0267] (20) Use spectral clustering to classify news topics based on the similarity matrix E mentioned above;

[0268] (21) Output news topic categories.

[0269] The performance of this invention can be further illustrated by the following simulation experiments.

[0270] 1. Simulation conditions and parameters

[0271] To verify the effectiveness of the multi-view-based news topic mining method of the present invention, simulation experiments were conducted on four multi-view news datasets: 3Sources, BBC, BBCSport2view, and WebKB.

[0272] The 3Sources news dataset consists of 169 news articles from three news organizations: BBC, Reuters, and The Guardian. Their dimensions are 3560, 3631, and 3068, respectively, covering six themes: business, entertainment, health, politics, sports, and technology.

[0273] The BBC News dataset consists of 685 documents, each divided into four parts with dimensions of 4659, 4633, 4665, and 4684, covering five themes: business, entertainment, politics, sports, and technology.

[0274] The BBCSport2view news dataset consists of 544 news articles from the BBC Sport website. It contains two views with dimensions of 3183 and 3203, respectively, and the news articles belong to five sports: athletics, cricket, football, rugby, and tennis.

[0275] The WebKB News Dataset has three views with dimensions of 1703, 230, and 230 respectively. It comes from the web pages of four universities and is divided into four categories: faculty, students, projects, and courses.

[0276] In the simulation experiment, the parameter γ of the method of the present invention... n The value range of is {0.01, 0.1, 1, 10, 100}, the value range of parameter γ is {0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9}, the value range of parameter δ is [0.0001, 0.5], the value range of parameter τ is {0.0001, 0.0005, 0.001, 0.005, 0.01, 0.05, 0.1, 0.2, 0.3, 0.4, 0.5}, and the nearest neighbor parameter is set to 5.

[0277] The number of times all method experiments are repeated is set to 50.

[0278] The method of this invention compares K-means and NMF. best SVD best PCA best Four single-view methods and three multi-view methods: LT-MSC, LMSC (Zhang C, Hu Q, Fu H, et al. Latent multi-view subspace clustering[C] / / Proceedings of the IEEE conference on computer vision and pattern recognition.2017:4279-4287.), and FMR (Li R, Zhang C, Hu Q, et al. Flexible Multi-View Representation Learning for Subspace Clustering[C] / / IJCAI.2019:2916-2922.).

[0279] 2. Simulation Content and Result Analysis

[0280] (1) Clustering experiments conducted on a multi-view news dataset

[0281] The clustering results of each method are given in the table below, where AC represents the clustering accuracy, NMI represents the normalized information, AR represents the adjusted Land coefficient, and Precision represents the clustering precision. The higher the value of the above four evaluation indicators, the better the clustering effect of the method.

[0282] Experiment 1: Comparative experiment conducted on the 3Sources news dataset. The experimental results of different methods on the 3Sources news dataset are shown in Table 1.

[0283] Table 1. Clustering results (%) of each method on 3Sources

[0284]

[0285]

[0286] As can be seen from Table 1, the present invention achieves the highest scores in all four evaluation indicators compared to other methods, indicating that the method of the present invention exhibits the best recognition performance.

[0287] Experiment 2: A comparative experiment was conducted on the BBC News dataset. The experimental results of different methods on the BBC News dataset are shown in Table 2.

[0288] Table 2. Clustering results (%) of each method on BBC

[0289]

[0290] The experimental results in Table 2 demonstrate that the method of the present invention has the highest clustering accuracy compared to other methods.

[0291] Experiment 3: Comparative experiment on the BBCSport2view news dataset. The experimental results of different methods on the BBCSport2view news dataset are shown in Table 3.

[0292] Table 3. Clustering results (%) of each method on BBCSport2view

[0293]

[0294]

[0295] As can be seen from Table 3, the method of the present invention still shows the best recognition performance in all four evaluation metrics, whether compared with single-view clustering methods or multi-view clustering methods.

[0296] Experiment 4: Comparative experiments were conducted on the WebKB News Dataset. The experimental results of different methods on the WebKB News Dataset are shown in Table 4.

[0297] Table 4. Clustering results (%) of each method on WebKB

[0298]

[0299] The experimental results in Table 4 demonstrate that the method of the present invention achieves the best clustering results.

[0300] (2) Experiments in exploring news topics

[0301] from Figure 2 As can be seen, there are certain differences in the sector area displayed by different news topics. The larger the sector area, the larger the data volume of the news topic and the higher the degree of attention people pay to the content of the topic. Among all news topics, business occupies the largest area, which indicates that people pay more attention to business content.

[0302] Figure 3 The area corresponding to politics is the largest, indicating that people are more concerned about political content.

[0303] Figure 4 The area corresponding to the football in the middle is the largest, indicating that people pay more attention to football-related content in sports.

[0304] Figure 5 The project with the largest area indicates that this project has the largest amount of data.

[0305] Based on the above experimental results, the news topic mining method based on multiple views of the present invention outperforms other single-view clustering methods and multiple-view clustering methods in terms of evaluation metrics on four multiple-view news datasets, including 3Sources, BBC, BBCSport2view and WebKB, and achieves the best clustering effect. At the same time, the experimental results of mining news topics show that the method of the present invention can effectively mine the corresponding news topics.

[0306] Some steps in the embodiments of the present invention can be implemented using software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk.

[0307] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A news topic mining method based on multiple views, characterized in that, The method includes: Step 1: Transform the multi-view news data matrix A (1) A (2) A (V) Decomposed into matrix A (v) F (v) Sum of matrix S (v) ,Right now: A (v) =A (v) F (v) +S (v) ,v=1,2,...,V Among them, A (v) F (v) F (v) and S (v) Let A represent the low-rank part, subspace representation part, and reconstruction error part corresponding to the v-th view, respectively. (V) ∈R d×n d represents the dimension of the sample in the news data, n represents the number of news data, and V represents the number of views; Step 2: Based on the subspace representation part F of each single view (v) Construct an N-order tensor F=Ψ(F (1) ,F (2) ,...,F (V) ) Where Ψ(·) represents the subspace representation matrix F of different views. (v) After merging, an N-order tensor F is constructed; Expanding the N-order tensor F on the nth mode yields the matrix F. (n) The expression is: Among them, I n Represents matrix F (n) dimensionality; Step 3: Calculate the nuclear norm of the Nth-order tensor F, using the following formula: in,‖·‖ * ζ represents the nuclear norm of a tensor. n It is a constant greater than 0, and satisfies Step 4: Construct a low-rank tensor-constrained multi-view subspace clustering LT-MSC model: Where S = [S (1) S (2) ;...;S (V) The symbol represents the matrix formed by concatenating the reconstruction error matrices of each view. 2，1 L represents the matrix 2,1 Norm; Step 5: Calculate the hypergraph Laplacian matrix for each view. Step 6: Construct the manifold regularization term Where tr(·) represents the trace of the matrix; Step 7: Construct sparse constraint terms ||F T F||1; where ‖·‖1 denotes the L1 norm of the matrix; Step 8: Construct diversity regularization terms Step 9: Based on the multi-view subspace clustering LT-MSC model and the manifold regularization term Sparse constraint term ||F T F||1, Diversity Regularization Term Construct a news topic mining model based on multiple views. The model's expression is: Where λ1, λ2, λ3 and λ4 are equilibrium parameters; Step 10: Update variable μ, the update formula is: μ=min(ρμ,μ max ) Where ρ represents the given positive penalty parameter, and μ represents the penalty parameter that is greater than 0; Step 11: Update the subspace representation matrix F (v) Reconstruction error matrix S, auxiliary variable G n The Lagrange multiplier B v The vectorization f of the Nth-order tensor F, and the auxiliary variable G n vectorization of g n and Lagrange multipliers α n until ||A is satisfied (v) -A v F (v) -S (v) || ∞ <ε, and ||P n fg n || ∞ <ε, where ε represents a very small positive number that satisfies the convergence condition; Wherein, the updated subspace representation matrix F (v) The update process includes: According to the formula The multi-view-based news topic mining model is further represented in the following form: s.t.A (v) =A (v) F (v) +S (v) ,v=1,2,...,V, F=Ψ(F (1) ,F (2) ,...,F (V) ), S=[S (1) ;S (2) ;...;S (V) ], diag(F (v) )=0 in, The above representation is optimized using the Augmented Lagrange Multiplier Method (ALM). When solving using the ALM method, the objective function needs to be separable; therefore, an auxiliary variable G is introduced. n To replace F (n) Then we have: s.t.P n f=g n ,n=1,2,...,N A (v) =A (v) F (v) +S (v) ,v=1,2,...,V, F=Ψ(F (1) ,F (2) ,...,F (V) ), S=[S (1) ;S (2) ;...;S (V) ], diag(F (v) )=0 Among them, f and g n These are the Nth-order tensor F and the auxiliary variable G, respectively. n Vectorization, P n It is a permutation matrix used to calibrate F (n) and the auxiliary variable G n The corresponding elements between them; The multi-view-based news topic mining model is further represented as follows using the ALM method: Where, α n and B v It is a Lagrange multiplier, μ is the equilibrium parameter, and L μ＞0 The augmented Lagrange function represents the news topic mining model based on multiple views. When the reconstruction error S of each view (v) and auxiliary variable G n When fixed, the subspace representation matrix F (v) The update formula is: Among them, Ω v (·) indicates selecting the element corresponding to the v-th view, and then reconstructing these elements into a matrix, where I1 represents a matrix of all 1s; The updated subspace representation matrix is ​​then obtained. as follows: in, Step 12: Calculate the similarity matrix Step 13: Based on the similarity matrix obtained in Step 12, classify the news topics using spectral clustering and output the final classification results.

2. The method according to claim 1, characterized in that, The formula for calculating the hypergraph Laplacian matrix of each view in step five is as follows: Among them, D C H is a diagonal matrix whose diagonal elements are the degree of each vertex; G ∈R |C|×|T| The hypergraph's incidence matrix represents the relationships between vertices and hyperedges; W represents the set of hyperedge weights; D T A diagonal matrix representing hyperedge degree.

3. The method according to claim 2, characterized in that, The process of constructing the manifold regularization term in step six includes: The manifold regularization is hypergraph regularization, where the hypergraph G = (C, T, W) consists of a set of vertices C, a set of hyperedges T, and a set of weights W for each hyperedge. Each hyperedge t is assigned a positive weight w(t), and the elements in the weight matrix W are related to each hyperedge t. T and C satisfy U t∈T =C, U represents the union operation; The correlation matrix H of the hypergraph G G ∈R |C|×|T| The elements of the incidence matrix, used to represent the relationships between vertices and hyperedges, are expressed as follows: in, Let c represent the degree of each hyperedge, and c represent the vertex. The manifold regularization term is:

4. The method according to claim 3, characterized in that, The process of constructing the diversity regularization term in step nine includes: in, H represents a matrix with a mean of 0.

5. The method according to claim 4, characterized in that, The process of updating and reconstructing the error matrix S includes: When the subspace representation matrix F of the single view (v) and the auxiliary variable G n When fixed, the update formula for the reconstruction error matrix S can be obtained as follows: Where Z is matrix A (v) -A (v) F (v) +B (v) The solution to the above equation is obtained by connecting the columns vertically together to form a matrix: Among them, Z :,i Let represent the i-th column of matrix Z.

6. The method according to claim 5, characterized in that, The auxiliary variable G n The iterative update formula is as follows: in, Ω n (P n f+a n ) is based on the nth pattern of the expansion, which will be used to expand P n f+a n Reshape into the corresponding matrix. This represents the spectral soft thresholding operation of a matrix.

7. The method according to claim 6, characterized in that, The Lagrange multiplier B v The update formula is as follows:

8. The method according to claim 7, characterized in that, The method for updating the vectorized f of the tensor F is: directly replace the updated F. (v) Update f using the corresponding elements in the table, i.e., f * ←F (v) .

9. The method according to claim 8, characterized in that, The Lagrange multiplier α n The update rules are as follows:

10. A news topic mining system based on multiple views, characterized in that, The system includes: The data acquisition module is used to acquire news data; The data classification module is used to classify news topics based on the news data using a multi-view-based news topic mining method as described in any one of claims 1-9. The output display module is used to output the categorization results of news topics.

Images