A uniform random anchor point generation method and anchor point adaptive multi-view clustering method

Abstract

The present invention relates to a uniform random anchor point generation method and an anchor point adaptive multi-view clustering method. The anchor point generation method includes the following steps: a first step: normalizing the view to obtain multiple sample points; a second step: mapping the multiple sample points into a coordinate system, calculating the distance of each sample point to the origin; dividing the sample points into multiple distance categories according to the distance; a third step: dividing the area where the multiple sample points are located into several blocks according to the distance category, and making each block have the same number of sample points; a fourth step: randomly extracting the same number of sample points in each block as anchor points; and merging multiple anchor points into an anchor point set. The above-mentioned uniform random anchor point generation method generates more representative anchor points with a lower running time, better represents the potential structure of the sample point set of each view, and takes into account the high efficiency and high representativeness of the anchor point generation method.

Description

Technical Field

[0001] The present invention relates to the technical field, and in particular to a uniform random anchor point generation method and an anchor point-based adaptive multi-view clustering method. Background Art

[0002] With the rapid proliferation of sensors and the internet, the same object can often be captured and represented in multiple views. For example, the same event can be reported by different news organizations; the same object can be viewed from different surveillance perspectives; or the same text can be translated into different languages. All of these are referred to as multi-view data, and they often provide complementary and compatible information. Learning and leveraging multiple views to analyze data with unknown labels is called multi-view learning.

[0003] In the process of processing multi-view data, in order to improve the processing efficiency of each view data, it is usually adopted to extract anchor points in the view to simplify the computation. The anchor points are used to represent the potential structure of the view data.

[0004] Among existing anchor point generation methods, random selection offers the best timeliness, but the anchor points it selects are not representative of multi-view data. While the K-Means anchor point generation strategy can select representative anchor points, its computational complexity makes it difficult to apply to large-scale multi-view data. Summary of the Invention

[0005] Based on this, it is necessary to address the problem that existing anchor point generation methods are difficult to ensure that the anchor points are representative while improving the efficiency of anchor point generation, and provide a uniform random anchor point generation method and an anchor point adaptive multi-view clustering method. A uniform random anchor point generation method includes the following steps:

[0006] Step 1: Normalize the view to obtain multiple sample points;

[0007] Step 2: Map multiple sample points into a coordinate system and calculate the distance from each sample point to the origin. Then, classify all sample points according to the distance to obtain multiple distance categories.

[0008] Step 3: Divide the area where multiple sample points are located into several blocks according to the distance category, and make sure that each block has the same number of sample points;

[0009] Step 4: Randomly extract the same number of sample points in each block as anchor points.

[0010] The uniform random anchor point generation method measures the relative distance of sample points and divides the area where the sample points are located according to the distance. It then randomly selects the same number of anchor points in each block, thereby generating more representative anchor points with lower runtime, better representing the potential structure of the sample point set of each view, and taking into account both high efficiency and high representativeness of the anchor point generation method.

[0011] In one embodiment, the blocks are divided as follows:

[0012] Assume that the sample point set X is: X=[x1,x2,x3,...,x n ] T ∈R n×d , the data origin x0 is: x0=[0,0,0,...,0]∈R 1×d ;

[0013] Among them, n represents the number of sample points, and d represents the feature dimension;

[0014] Let f be the sample point x in X i The distance metric function to the data origin x0, d i (i=1,2,...,m) represents x i The distance to x0 about f, d i =f(x0,x i ) ; Combine the Euclidean distances of all sample points in the sample point set to the data origin, and obtain the Euclidean distance set D as:

[0015] D = {d1, d2, ..., dn};

[0016] Sort in ascending order according to the value of di, and get the positive sequence distance set D':

[0017] D'={ds1,ds2,…,dsn},d s1 ≤d s2 ≤…≤d sn ;

[0018] Sort the sample point set in ascending order according to the value of di, and get the positive order sample point set X':

[0019] X'={x'1,x'2,x'3,...,x' n};

[0020] Then define the sample point set B1, B2, ..., B in each block p for:

[0021]

[0022]

[0023] B1∪B2∪…∪B p =X, B1∩B2∩…∩B p =Φ;

[0024] Among them, the first p-1 blocks have the same number of sample points The remaining sample points are divided into the pth block.

[0025] In one embodiment, the same number of sample points are randomly selected from each divided block, and the anchor point set U is expressed as:

[0026] U=[u1,u2,u3,...,u p' ] T ∈R p'×d (U∈X)

[0027] Where p' represents the number of anchor points.

[0028] The present invention also provides an anchor-based adaptive multi-view clustering method, which comprises the following steps:

[0029] S1: Acquire multiple views of the same object as a multi-view dataset;

[0030] S2: Use the uniform random anchor point generation method to extract multiple anchor points in each view; take the union of the anchor points in each view to obtain the anchor point set;

[0031] S3: Establish an anchor-adaptive similarity graph model, and generate a corresponding anchor-adaptive similarity graph in the anchor-adaptive similarity graph model according to the anchor of each view;

[0032] Among them, the generation method of the anchor point adaptive similarity graph model is as follows:

[0033] S31: Build a classic anchor-based similarity graph model. The similarity graph model is used to generate an anchor similarity graph based on multiple anchors in each view.

[0034] S32: Introducing induction factors into similarity graph models to enable similarity graph models to overcome the occurrence of trivial solutions;

[0035] S33: Optimizing the constraints of the similarity graph model through the Lagrangian function to obtain an anchor-based adaptive similarity graph model;

[0036] S4: Assign a weight to each view according to the proportion of its anchor points in the anchor point set; fuse all anchor point similarity graphs according to the weight of each view to obtain the final anchor point similarity graph.

[0037] S5: Convert the final anchor point similarity map into a sample point similarity map through simple calculation.

[0038] In one embodiment, the anchor point adaptive similarity graph model is expressed as:

[0039]

[0040] Among them, z ij Represents the anchor point u j is the sample point x i The probability of a neighbor, 1 is a column vector whose elements are all 1.

[0041] The similarity graph model has a trivial solution, that is, there is only one anchor point and x i The distance is the closest, x i The similarity value with this anchor point is 1, and the similarity values with other anchor points are 0.

[0042] In one embodiment, an induction factor is introduced into the similarity graph model, and the anchor point adaptive similarity graph model is expressed as:

[0043]

[0044] Among them, β is the induction factor used to make each anchor point and x i The similarity is 1 / p'.

[0045] In one embodiment, the Lagrangian function for optimizing the constraints of the similarity graph model for:

[0046]

[0047]

[0048] Among them, η and ξ are Lagrange multipliers; d i is a vector; d ij is the jth element of the vector; z i is the i-th row element of the anchor point adaptive similarity matrix.

[0049] Then the anchor point adaptive similarity graph model is expressed as:

[0050]

[0051] In one embodiment, the anchor similarity graph Expressed as:

[0052]

[0053] Where k is the number of adaptive neighbors of the anchor point.

[0054] In one embodiment, the weight assigned to each view is w (v), the final anchor point similarity graph obtained by weighted fusion is:

[0055] Z=w (1) Z (1) +w (2) Z (2) +…+w (m) Z (m)

[0056] Among them, Z is the anchor point similarity matrix after fusion, w (v) is the proportion of the anchor point of the vth view in the anchor point set; and w (1) +w (2) +…+w m =1.

[0057] In one embodiment, the sample point similarity matrix S is:

[0058] S=ZΛ -1 Z T ,Λ∈R l×l ,Λ=diag(Z T 1).

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] 1. This method measures the relative distances between sample points and divides the areas where the sample points are located based on the distances. It then randomly selects the same number of anchor points in each block, thereby generating more representative anchor points with lower runtime. This method better represents the potential structure of the sample point set for each view, achieving both high efficiency and high representativeness.

[0061] 2. In the process of generating anchor point similarity graphs through anchor points, an induction factor is introduced into the similarity graph model to overcome the problem of trivial solutions of the similarity graph model, and the similarity graph model is optimized through the Lagrangian function to generate an anchor point adaptive similarity graph model, thereby improving the robustness of the anchor point similarity graph generation method.

[0062] 3. The anchor-based adaptive multi-view clustering method of the present invention first adopts a uniform random anchor point generation strategy to generate more representative anchor points with a lower running time, better represent the potential structure of the sample point set of each view, and take the union of the anchor points generated for multiple views as the anchor point set to achieve anchor point consistency of each view. Subsequently, an anchor point similarity graph is constructed based on the anchor points, and the anchor point similarity graph of each view is weightedly fused, thereby better utilizing the difference complementarity between views, avoiding calculating the similarity graph of the restored sample points for each view, improving the efficiency of multi-view clustering, and improving the robustness of multi-view fusion. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 This is a step diagram of the uniform random anchor point generation method according to Example 1 of the present invention;

[0064] Figure 2 for Figure 1 Coordinate diagram of the sample point set of the anchor point generation method;

[0065] Figure 3 for Figure 1 Classification block diagram of the sample point set of the anchor point generation method;

[0066] Figure 4 for Figure 1 Anchor point generation graph of the sample point set of the anchor point generation method;

[0067] Figure 5 This is a step diagram of the anchor point adaptive multi-view clustering method according to Example 1 of the present invention;

[0068] Figure 6 for Figure 5 Flowchart of the anchor-adaptive multi-view clustering method;

[0069] Figure 7 This is a comparison chart of clustering performance of the anchor-adaptive multi-view clustering method according to Example 1 of the present invention;

[0070] Figure 8 This is a clustering time comparison diagram of the anchor-adaptive multi-view clustering method according to Example 1 of the present invention;

[0071] Figure 9 Performance comparison chart of different anchor generation strategies for the Caltech101-20 dataset;

[0072] Figure 10 Performance comparison chart of different anchor generation strategies for the Caltech101-all dataset;

[0073] Figure 11 Performance comparison chart of different anchor generation strategies for the Hidigit dataset

[0074] Figure 12 Performance comparison chart of different anchor generation strategies for the Reuters dataset;

[0075] Figure 13 Performance comparison chart of different anchor generation strategies for the NUS dataset;

[0076] Figure 14 Comparison of runtimes using different anchor generation strategies for five datasets. DETAILED DESCRIPTION

[0077] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0078] It should be noted that when a component is referred to as being "mounted on" another component, it may be directly on the other component or there may be a central component. When a component is considered to be "set on" another component, it may be directly set on the other component or there may be a central component. When a component is considered to be "fixed to" another component, it may be directly fixed to the other component or there may be a central component.

[0079] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this invention pertains. The terms used herein in the specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The term "or / and" as used herein includes any and all combinations of one or more of the associated listed items.

[0080] See also Figure 1 , Figure 1 A method for generating uniform random anchor points according to an embodiment of the present invention is shown in FIG. A method for generating uniform random anchor points according to an embodiment of the present invention comprises the following steps:

[0081] Step 1: Normalize the view to obtain multiple sample points. A view is a dataset formed by collecting and expressing information about any object, not just an image. The same object can be collected and expressed from multiple angles, resulting in multiple views. These views have connections and differences. Within each view's dataset, the points that represent the view's characteristics are called sample points.

[0082] Step 2: Map multiple sample points into a coordinate system and calculate the distance from each sample point to the origin. Then classify all sample points according to the distance to obtain multiple distance categories. Let the sample point set X be: X = [x1, x2, x3, ..., x n ] T ∈R n×d , the data origin x0 is: x0=[0,0,0,...,0]∈R 1×d ,

[0083] Among them, n represents the number of sample points and d represents the feature dimension.

[0084] Let f be the sample point x in X iThe distance metric function to the data origin x0, d i (i=1,2,…,m) represents x i The distance to x0 about f, d i =f(x0,x i ) ; Combine the distances from all sample points in the sample point set to the data origin to obtain the distance set D:

[0085] D={d1,d2,…,dn}.

[0086] Sort in ascending order according to the value of di, and get the positive sequence distance set D':

[0087] D'={ds1,ds2,…,dsn},d s1 ≤d s2 ≤…≤d sn .

[0088] Sort the sample point set in ascending order according to the value of di, and get the positive order sample point set X':

[0089] X'={x'1,x'2,x'3,...,x' n}.

[0090] Among them, the first p-1 blocks have the same number of sample points The remaining sample points are divided into the pth block.

[0091] Step 3: Divide the area where multiple sample points are located into several blocks according to the distance category, and make each block have the same number of sample points. Then define the sample point set B1, B2, ..., B in each block p for:

[0092]

[0093]

[0094] B1∪B2∪…∪B p =X, B1∩B2∩…∩B p =Φ.

[0095] Step 4: Randomly extract the same number of sample points in each block as anchor points. Multiple anchor points are merged into an anchor point set. The anchor point set U is expressed as:

[0096] U=[u1,u2,u3,...,u p' ] T ∈R p'×d (U∈X)

[0097] Where p' represents the number of anchor points.

[0098] Please combine Figure 2 、 Figure 3 and Figure 4 , Figure 2 for Figure 1 Coordinate diagram of the sample point set of the anchor point generation method; Figure 3 for Figure 1 Classification block diagram of the sample point set of the anchor point generation method; Figure 4 for Figure 1 Anchor point generation diagram of the sample point set of the anchor point generation method. This embodiment uses a sample point set with 100 sample points. First, a plane coordinate system is established so that all 100 sample points are located in the first quadrant. Then, the Euclidean distance of each data point to the origin is calculated, and the 100 sample points are evenly divided into five blocks according to the Euclidean distance value, namely B1, B2, ..., B5, with 20 sample points in each block. Five sample points are randomly selected as anchor points in each block. It can be seen that the selected anchor points are relatively evenly distributed in the area where the entire sample point set is located, and can better represent the global structure of the sample point set.

[0099] In other embodiments, other algorithms may be used to calculate the relative distance between sample points and the origin, thereby classifying multiple sample points. For example, Hamming distance, Manhattan distance, Chebyshev distance, or Minn distance may be used. Of course, when calculating the distance between sample points using other methods, other coordinate systems may also be established, such as a spatial coordinate system, a spherical coordinate system, etc.

[0100] Please combine Figure 5 and Figure 6 , Figure 5 1 is a step diagram of the anchor point adaptive multi-view clustering method of this embodiment; Figure 6 for Figure 5 This embodiment also provides an anchor-based adaptive multi-view clustering method, which includes the following steps:

[0101] S1: Acquire multiple views of the same object as a multi-view dataset. Collecting information from different angles of the same object can obtain multiple views of the same object. The union of these multiple views is the multi-view dataset.

[0102] S2: Use the uniform random anchor point generation method to extract multiple anchor points in each view. Take the union of the anchor points in each view to get the anchor point set.

[0103] When processing multi-view data, it is often necessary to fuse multiple views to achieve better clustering results. To preserve the original local geometric structure of multi-view data, anchor points can be used to represent the underlying structure of the data while reducing computational cost. Because existing anchor point generation methods generate inconsistent anchor points for each view in multi-view data, the similarity graph constructed based on the sample points and anchor points of each view cannot be directly used for fusion. Restoring a sample-based similarity graph for each view significantly increases the computational effort as the number of views and sample points in the multi-view data increases.

[0104] Because multiple views of the same object share certain connections or similarities, the anchor points selected from each view may overlap. Therefore, taking the union of the anchor points generated for each view provides the most concise and comprehensive anchor point set. Furthermore, due to the use of uniform random anchor point generation, the anchor point similarity graph generated by taking the union of the anchor points from each view also exhibits a certain degree of similarity and can be directly used for fusion.

[0105] S3: Establish an anchor point adaptive similarity graph model, and generate a corresponding anchor point similarity graph in the adaptive similarity graph model according to the anchor point of each view.

[0106] Among them, the method for establishing the anchor point adaptive similarity graph model is as follows:

[0107] S31: Build a classic similarity graph model based on anchor points. The similarity graph model is used to generate an anchor similarity graph based on multiple anchor points in each view. Anchor Adaptive Similarity Graph Model Expressed as:

[0108]

[0109] Among them, z ij Represents the anchor point u j is the sample point x i The probability of the neighbor, its value represents the size of the similarity value, 1 is a column vector whose elements are all 1.

[0110] It is easy to observe that the similarity graph model has a trivial solution, that is, there is only one anchor point with x i The distance is the closest, x i The similarity value with this anchor point is 1, and the similarity values with other anchor points are 0.

[0111] S32: Introducing an induction factor into the similarity graph model to overcome the appearance of trivial solutions. Introducing the induction factor can overcome the appearance of trivial solutions and transform the similarity graph model into:

[0112]

[0113] Among them, β is the induction factor used to make each anchor point and x i The similarity is 1 / p'.

[0114] make d i is a vector and d ij represents the jth element of the vector, then the similarity graph model can be rewritten in vector form as:

[0115]

[0116] Among them, z i is the i-th row element of the anchor point adaptive similarity matrix, and T is the transposed matrix.

[0117] S33: Optimize the constraints of the similarity graph model through the Lagrangian function to obtain the anchor point adaptive similarity graph model. Lagrangian function for optimizing the constraints of the similarity graph model for:

[0118]

[0119] Among them, η and ξ are Lagrange multipliers.

[0120] Find the Lagrangian function with respect to z i Taking the partial derivative of 0, we can get z i The jth element of can be expressed as:

[0121]

[0122] According to KKT conditions, z ij ξ j =0.

[0123] KKT conditions are a method used to solve optimization problems. The optimization problem mentioned here usually refers to finding the global minimum of a given function in a specified scope.

[0124] represents z ij The optimal solution of It can be expressed as:

[0125]

[0126] Assume d i1 ,d i2 ,…,d ip' Arrange in ascending order and combine the k-nearest neighbor idea, i.e. z i There are at most k non-zero values in . and constraint Z i 1=1 can be deduced:

[0127]

[0128]

[0129] In order to make The number of neighbors of the anchor point can be adapted to k, so the parameter value is set The anchor point similarity graph can be obtained as:

[0130]

[0131] S4: Assign a weight to each view according to the proportion of its anchor points in the anchor point set. According to the weight of each view, all anchor point similarity graphs are fused to obtain the final anchor point similarity graph. Assume that w (v) is the weight of the v-th view, then:

[0132] Z=w (1) Z (1) +w (2) Z (2) +…+w (m) Z (m) .

[0133] make You can get:

[0134]

[0135] Let the similarity graph matrix be S, which combines the similarities within and between views of multiple views, then

[0136] S=ZΛ -1 Z T ,Λ∈R l×l ,Λ=diag(Z T 1)

[0137] Among them, Z is the final anchor point similarity graph matrix; ZT is the transposed matrix of Z.

[0138] Accordingly, the objective function is set as follows:

[0139]

[0140] Where tr(,) is the trace of the matrix, F∈R n×c Is a low-dimensional embedding matrix used to output the clustering results, c is the number of clusters in the final clustering. L = DS is the Laplace matrix, and L is a semi-positive matrix, D is the degree matrix,

[0141]

[0142] D=diag(S1)=diag(ZΛ -1 Z T 1)

[0143] =diag(ZΛ -1 Λ)

[0144] =diag(Z1)

[0145] = Ι

[0146] The time complexity of eigendecomposition on the Laplacian matrix L of the graph is high and is not suitable for large-scale data.

[0147] Since L = DS = 1-S, the objective function can be written as:

[0148]

[0149] make So: S=HH T ,

[0150] Using the singular value decomposition technique H=P∑Q T have to:

[0151] S=HH T =( P ∑Q T )( P ∑Q T ) T =P(∑∑ T )P T

[0152] Representing the graph in the form of an adjacency matrix, denoted as S, the sum of the weights of all edges between subgraph A and subgraph B can be expressed as follows:

[0153]

[0154] Among them, s ij Represents the weight between vertex i and vertex j, if s ij =0, it means that vertex i and vertex j are not connected.

[0155] S5: Convert the final anchor point similarity map into a sample point similarity map through a simple calculation matrix.

[0156] The sample point similarity matrix S is:

[0157] S=ZΛ -1 Z T ,Λ∈R l×l ,Λ=diag(Z T 1).

[0158] In this embodiment, the uniform random anchor point selection is only affected by the number of samples, and the time complexity is O(n). The time complexity of the anchor point-based similarity matrix Z is O(np'∑ v d v ). Calculate the eigendecomposition of the fusion similarity matrix S by Z∈R n×p' Perform SVD with a time complexity of O(np' 2 ), if we only need c (c<<n) largest singular values, the time complexity can be reduced to O(nc 2 ), which greatly reduces the computation and storage burden of large-scale data, and is comparable to the existing time complexity of O(n 3 ) has significant advantages over the eigenvalue decomposition of the Laplace matrix.

[0159] Experimental verification

[0160] 1. Clustering Algorithm Performance Experiment

[0161] To validate the performance of our proposed anchor-adaptive multi-view clustering method (SWAP), we compared our proposed SWAP method with currently available large-scale multi-view clustering algorithms on five benchmark datasets. To ensure a fair comparison, all the code for the compared methods is source code provided by the authors. To mitigate the effects of random steps in the algorithm and ensure more representative experimental results, each set of experiments was run 10 times and the average result was used as the final result.

[0162] Among them, the Co-reg algorithm is based on the co-regularization technology to explore the true potential space of multiple views, which can find cross-view Figure 1 The consistent clustering can obtain more accurate clustering results than single-view clustering.

[0163] The MVSC algorithm uses local manipulative fusion to integrate heterogeneous features, which can solve the problem of data being represented by different features in practical applications.

[0164] The MSC_IAS algorithm builds similarity based on complete space learning technology, which can construct an information-completely perceived similarity to improve clustering performance.

[0165] The LMVSC algorithm first learns a smaller graph for each view and then performs spectral clustering on the smaller graph.

[0166] Please combine Figure 7 , which is a comparison chart of the clustering performance of the anchor-adaptive multi-view clustering method of this embodiment. Tables 1, 2, and 3 respectively compare the ACC (adaptivity), NMI (normalized mutual information), and Purity results of the five clustering methods on five datasets.

[0167] In terms of the NMI metric, the SWAP method outperformed all comparison algorithms, achieving the best results on all datasets. In terms of the ACC metric, the SWAP method achieved the best results on four of the five datasets. Furthermore, the SWAP method significantly improved performance compared to the four comparison methods, particularly on the Caltech101-20 dataset, where it outperformed the MVSC method by nearly 9 percentage points. In terms of the Purity metric, the SWAP method achieved the best results on all four datasets, outperforming the LMVSC method by nearly 5 percentage points on the Caltech101-all dataset. These results demonstrate the effectiveness of the SWAP method.

[0168] Please combine Figure 8 , which is a clustering time comparison diagram of the anchor-adaptive multi-view clustering method of this embodiment.

[0169] As can be seen, the SWAP method is 23 times faster than the Co-reg method on the smallest dataset. The advantage becomes more pronounced as the dataset size increases. On the largest NUSWIDEOBJ dataset, the SWAP method runs nearly 300 times faster than the Co-reg method. The SWAP method is on average about seven times faster than the MVSC method, which uses k-means to determine anchor points by combining multiple views. Because the MSC_IAS method constructs similarity based on full-space learning technology, it consumes a large amount of memory and cannot be executed on the largest NUSWIDEOBJ dataset due to insufficient memory. The most efficient method compared is the LMVSC method, which runs in linear time, but the SWAP method is still faster than LMVSC on all datasets. Compared with existing large-scale multi-view clustering methods, it can be said that the SWAP method has an absolute advantage in runtime and is more suitable for processing large-scale multi-view datasets.

[0170] 2. Performance Experiment of Anchor Point Generation Method

[0171] To verify that the uniform random anchor point generation method proposed in this paper can generate representative anchor points at a low computational cost, we conducted an experimental comparison of our anchor point generation method with a random anchor point selection strategy and a K-Means anchor point selection strategy. Using the same experimental environment and the same multi-view clustering algorithm, we compared the clustering performance and runtime of the three anchor point selection strategies on five benchmark datasets.

[0172] Please combine Figures 9 to 13 , Figure 9 Performance comparison chart of different anchor generation strategies for the Caltech101-20 dataset; Figure 10 Performance comparison chart of different anchor generation strategies for the Caltech101-all dataset; Figure 11Performance comparison chart of different anchor generation strategies for the Hidigit dataset; Figure 12 Performance comparison chart of different anchor generation strategies for the Reuters dataset; Figure 13 Performance comparison of different anchor generation strategies for the NUS dataset.

[0173] Because the Reuters dataset has a high number of features, the k-Means strategy cannot handle it due to its high computational complexity. As can be seen from the multiple comparison figures above, the random anchor point selection strategy has the worst clustering performance. The anchor point generation method of the present invention and the k-Means strategy have roughly equivalent clustering performance, and sometimes even better performance.

[0174] Please combine Figure 14 , which is a comparison chart of the running time of five datasets using different anchor generation strategies.

[0175] In terms of runtime, the K-Means-based anchor selection strategy runs much faster than the random anchor selection strategy and the anchor generation method of the present invention. The random anchor selection strategy has the shortest runtime, while the anchor generation method of the present invention runs slightly faster than the random anchor selection strategy.

[0176] It can be seen that compared with the random anchor point selection strategy and the K-Means-based anchor point selection strategy, the anchor point generation method of the present invention has low computational complexity, and the selected anchor points are more representative and can better capture the potential structure of the data.

[0177] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0178] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be determined by the appended claims.

Claims

1. An anchor-based adaptive multi-view clustering method, which is used to extract sample points from multiple view data of the same object and fuse each view according to the sample points to form a representative sample point similarity graph; characterized in that: The adaptive multi-view clustering method includes the following steps: S1: Acquire multiple views of the same object as a multi-view dataset; S2: A uniform random anchor point generation method is used to extract multiple anchor points in each view; a union of the anchor points of each view is taken to obtain an anchor point set; wherein the uniform random anchor point generation method includes the following steps: normalizing the view to obtain multiple sample points; mapping the multiple sample points to a coordinate system, and calculating the distance of each sample point to the data origin; then classifying all sample points according to the distance to obtain multiple distance categories; dividing the area where the multiple sample points are located into a number of blocks according to the distance categories, and ensuring that each block has the same number of sample points; randomly extracting the same number of multiple sample points in each block as multiple anchor points, wherein any view is multiple views of the same object, and each view is fused according to the sample points to form a representative sample point similarity graph; The block division method is as follows: Assume that the sample point set X is: X=[x1,x2,x3,...,x n ] T ∈R n×d , the data origin x0 is: x0=[0,0,0,...,0]∈R 1 ×d ; Among them, n represents the number of sample points, and d represents the feature dimension; Let f be the sample point x in X i The distance metric function to the data origin x0, d i (i=1,2,...,m) represents x i The distance to x0 about f, d i =f(x0,x i ) ; Combine the distances from all sample points in the sample point set to the data origin to obtain the distance set D: D = {d1, d2, ..., dn}; Sort in ascending order according to the value of di, and get the positive sequence distance set D': D’={ds1,ds2,…,dsn},d s1 ≤d s2 ≤…≤d sn ; Sort the sample point set in ascending order according to the value of di, and get the positive order sample point set X': X'={x1',x'2,x'3,...,x' n }; Then define the sample point set B1, B2, ..., B in each block p for: B1∪B2∪…∪B p =X,B1∩B2∩…∩B p =Φ; Among them, the first p-1 blocks have the same number of sample points The remaining sample points are divided into the pth block; S3: establishing an anchor point adaptive similarity graph model, and generating a corresponding anchor point similarity graph in the anchor point adaptive similarity graph model according to the anchor point of each view; Among them, the generation method of the anchor point adaptive similarity graph model is as follows: S31: constructing a classic similarity graph model based on anchor points, wherein the similarity graph model is used to generate an anchor similarity graph according to multiple anchor points of each view; S32: introducing an induction factor into the similarity graph model to enable the similarity graph model to overcome the occurrence of trivial solutions; S33: Optimizing the constraint conditions of the similarity graph model by using a Lagrangian function to obtain an anchor-based adaptive similarity graph model; S4: Assign a weight to each view according to the proportion of its anchor points in the anchor point set; fuse all anchor point similarity graphs according to the weight of each view to obtain the final anchor point similarity graph; S5: Convert the final anchor point similarity map into a sample point similarity map through simple calculation.

2. The anchor-based adaptive multi-view clustering method according to claim 1, characterized in that: In S31, the anchor point adaptive similarity graph model is expressed as: Among them, z ij Represents the anchor point u j is the sample point x i The probability of a neighbor, 1 is a column vector whose elements are all 1; The similarity graph model has a trivial solution, that is, there is only one anchor point with x i The distance is the closest, x i The similarity value with this anchor point is 1, and the similarity values with other anchor points are 0.

3. The anchor-based adaptive multi-view clustering method according to claim 1, characterized in that: By introducing the induction factor into the similarity graph model, the anchor point adaptive similarity graph model is expressed as: Among them, β is the induction factor used to make each anchor point and x i The similarity is 1p'.

4. The anchor-based adaptive multi-view clustering method according to claim 1, characterized in that: Lagrangian function for optimizing the constraints of the anchor point adaptive similarity graph model for: Among them, η and ξ are Lagrange multipliers; d i is a vector; d ij is the jth element of the vector; z i is the i-th row element of the anchor point adaptive similarity matrix; Then the adaptive similarity graph model is expressed as:

5. The anchor-based adaptive multi-view clustering method according to claim 1, characterized in that: In S3, the anchor similarity graph Expressed as: Where k is the number of adaptive neighbors of the anchor point.

6. The anchor-based adaptive multi-view clustering method according to claim 1, characterized in that: In S4, the weight given to each view is w (v) , the final anchor point similarity graph obtained by weighted fusion is: Z=w (1) WITH (1) +in (2) WITH (2) +…+in (m) WITH (m) Among them, Z is the anchor point similarity matrix after fusion, w (v) is the proportion of the anchor point of the vth view in the anchor point set; and w (1) +w (2) +…+w m =1.

7. The anchor-based adaptive multi-view clustering method according to claim 1, characterized in that: In S5, the sample point similarity graph matrix S is: S=ZΛ- 1 WITH T ,Λ∈R l ×l,Λ=diag(Z T 1)。 8. The anchor-based adaptive multi-view clustering method according to claim 1, characterized in that: In the fourth step, the anchor point set U is expressed as: U=[u1,u2,u3,...,u p' ] T ∈R p'×d (U∈X) Where p' represents the number of anchor points.

Images