High-dimensional data visualization method and device, high-dimensional data visualization model training method and device, equipment and medium

Abstract

The invention relates to the field of data processing, in particular to a high-dimensional data visualization method and device, a high-dimensional data model training method and device, equipment and a medium, and the model training method comprises the steps that a plurality of training samples are acquired, and each training sample comprises a high-dimensional data set and a corresponding historical successful visualization result; for each high-dimensional data set, calculating k different similarity matrixes of the high-dimensional data set; aiming at each similarity matrix of the high-dimensional data set, taking the similarity matrix as a fully-connected graph with a weight, and extracting graph position coding information of nodes in the fully-connected graph from the similarity matrix by adopting a spectral method; and performing model training by taking k different similarity matrixes of the high-dimensional data set and the corresponding graph position coding information as input data of the model and taking a historical successful visualization result corresponding to the high-dimensional data set as supervision information of the model to obtain a trained high-dimensional data visualization model. The technical scheme can be more efficient and higher in generalization performance.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of data processing, in particular to a high-dimensional data visualization method and model training method, device, equipment and medium. BACKGROUND

[0002] With the rapid development of information technology, the scale and complexity of data in various fields have increased significantly. In particular, in fields such as biological genomics, remote sensing monitoring, and financial analysis, researchers often need to process large-scale data sets with high-dimensional characteristics. High-dimensional data visualization (HDV) is a special dimensionality reduction (DR) technique that aims to project high-dimensional data sets into two or three-dimensional space, allowing researchers to intuitively understand the internal structure and potential patterns of the data, thereby supporting decision-making and innovation research.

[0003] Existing high-dimensional data visualization techniques can be roughly divided into linear and nonlinear methods. Linear methods such as Principal components analysis (PCA), Multi-Dimensional Scaling (MDS), and Linear Discriminant Analysis (LDA) have laid the foundation for dimensionality reduction to some extent, but due to their linear nature, they cannot effectively reveal the complex nonlinear structure in the data. To address these shortcomings, nonlinear methods such as Self-Organizing Map (SOM), Isometric Mapping (Isomap), Kernel PCA (Kernel Principal Components Analysis), autoencoder, stochastic neighbor embedding (SNE), and Laplacian Eigenmaps have been proposed. In recent years, methods such as t-SNE (t-distributed Stochastic Neighbor Embedding), UMAP (Uniform Manifold Approximation and Projection), and PaCMAP (Pairwise Controlled Manifold Approximation Projection) have gained widespread application by explicitly optimizing low-dimensional embeddings to effectively preserve local and global structural relationships in the data. However, these methods still have the following outstanding problems:

[0004] 1. Hyperparameter sensitivity problem: The effectiveness of existing methods is heavily dependent on the selection of hyperparameters, such as the perplexity parameter in t-SNE and the number of neighbors parameter (n_neighbors) in UMAP. Different parameter choices can lead to significant differences in results.

[0005] 2. Computational overhead problem: Existing methods require retraining the model every time new data is encountered, resulting in high computational costs, especially when dealing with large datasets.

[0006] 3. Insufficient generalization ability problem: Existing methods attempt to build parametric models to address the retraining problem, but these models often fail to effectively adapt to new datasets of different dimensions and are prone to overfitting, failing to effectively utilize historical data.

[0007] Therefore, there is an urgent need for a high-dimensional data visualization method to address the above problems. Summary of the Invention

[0008] In order to solve the problems in the related art, the embodiments of the present disclosure provide a high-dimensional data visualization and its model training method, device, equipment, and medium.

[0009] In a first aspect, the present disclosure provides a method for training a high-dimensional data visualization model, comprising:

[0010] Obtain multiple training samples, each of which includes a high-dimensional dataset and its corresponding historical successful visualization results;

[0011] For each high-dimensional data set, calculating k different similarity matrices of the high-dimensional data set;

[0012] For each similarity matrix of the high-dimensional data set, the similarity matrix is ​​used as a weighted fully connected graph, and a spectral method is used to extract graph position encoding information of nodes in the fully connected graph from the similarity matrix;

[0013] The k different similarity matrices of the high-dimensional data set and their corresponding graph position encoding information are used as the input data of the model, and the historical successful visualization results corresponding to the high-dimensional data set are used as the supervision information of the model for model training to obtain a trained high-dimensional data visualization model.

[0014] In a possible implementation, obtaining a plurality of training samples includes:

[0015] Obtain multiple high-dimensional datasets with supervised information;

[0016] Get the i-th high-dimensional dataset X according to the following formula i Corresponding historical successful visualization results

[0017] in, is the high-dimensional data visualization algorithm function, θ is The hyperparameters, y i is a high-dimensional dataset X i The corresponding supervision information, It is an evaluation index for visualization results.

[0018] In a possible implementation, calculating k different similarity matrices of the high-dimensional data set includes:

[0019] Calculate the j-th similarity matrix according to the following formula The element at row u and column v in

[0020]

[0021]

[0022] Among them, X i [u] is the high-dimensional dataset X i The u-th data point, X i [v] is the high-dimensional dataset X i The vth data point, γ (j) is the bandwidth parameter of the j-th similarity matrix.

[0023] In a possible implementation, extracting graph position encoding information of nodes in the fully connected graph from the similarity matrix using a spectral method includes:

[0024] According to the following formula, from the i-th high-dimensional dataset X i The j-th similarity matrix Extract the graph position encoding information P of the nodes in the fully connected graph i (j) :

[0025] P i (j) =US 1 / 2 ;

[0026] Among them, U represents the The maximum d obtained after singular value decomposition e The matrix composed of the left singular vectors corresponding to the singular values, S is the maximum d e A diagonal matrix consisting of singular values.

[0027] In one possible implementation, the high-dimensional data visualization model includes a graph neural network sub-model, a graph transformer conversion sub-model, and a multi-layer perceptron sub-model;

[0028] Among them, the input of the graph neural network sub-model is k different similarity matrices of the high-dimensional data set and their corresponding graph position encoding information, and the output is the output information of k graph neural network sub-models; the input of the graph Transformer sub-model is the splicing information of the output information of k graph neural network sub-models.

[0029] In one possible implementation, the loss function of the high-dimensional data visualization model is for:

[0030]

[0031] Among them, L is the number of training samples, N i is the number of data points in the high-dimensional dataset, is the Bregman divergence, σ is the similarity conversion function, is the u-th data point of the visualization result output by the high-dimensional data visualization model, is the vth data point of the visualization result output by the high-dimensional data visualization model, is the u-th data point of the historical successful visualization result, The vth data point of the historical successful visualization results.

[0032] In one possible implementation, when the historical successful visualization result is obtained based on the t-distributed stochastic neighbor embedding t-SNE algorithm, the loss function of the high-dimensional data visualization model is for:

[0033]

[0034] in, The uth visualization result output by the high-dimensional data visualization model ′ data points, The vth visualization result output by the high-dimensional data visualization model ′ data points, The uth historical successful visualization result ′ data points, The vth historical success visualization result ′ data points;

[0035] When the historical successful visualization results are obtained based on the unified manifold approximation and projection UMAP algorithm, the loss function of the high-dimensional data visualization model is for:

[0036]

[0037] Among them, a and b are two similarity conversion parameters, M i is the index set of k neighboring points of each point in the high-dimensional dataset, and λ(t) is a weighted term that decays with the number of training iterations.

[0038] In a second aspect, the present disclosure provides a high-dimensional data visualization method, including:

[0039] Obtain the target high-dimensional dataset to be processed;

[0040] Calculating k different target similarity matrices of the target high-dimensional dataset;

[0041] Taking the target similarity matrix as a weighted fully connected graph, and extracting graph position encoding information of nodes in the fully connected graph from the target similarity matrix using a spectral method;

[0042] inputting the k different similarity matrices of the target high-dimensional data set and the corresponding graph position encoding information thereof into the pre-trained high-dimensional data visualization model, executing the pre-trained high-dimensional data visualization model, and obtaining a visualization result output by the pre-trained high-dimensional data visualization model;

[0043] The pre-trained high-dimensional data visualization model is trained based on the method of any one of the first aspect.

[0044] In a third aspect, the embodiments of the present disclosure provide a high-dimensional data visualization model training apparatus, comprising:

[0045] A sample obtaining module is configured to obtain a plurality of training samples, each training sample comprising a high-dimensional data set and corresponding historical successful visualization result thereof;

[0046] A first calculating module is configured to calculate, for each high-dimensional data set, k different similarity matrices of the high-dimensional data set.

[0047] A first graph information extracting module is configured to, for each similarity matrix of the high-dimensional data set, take the similarity matrix as a weighted fully connected graph, and extract graph position encoding information of nodes in the fully connected graph from the similarity matrix by using a spectral method.

[0048] A training module is configured to take the k different similarity matrices of the high-dimensional data set and the corresponding graph position encoding information thereof as input data of a model, take the historical successful visualization result corresponding to the high-dimensional data set as supervision information of the model, and perform model training to obtain a trained high-dimensional data visualization model.

[0049] In a fourth aspect, the embodiments of the present disclosure provide a high-dimensional data visualization apparatus, comprising:

[0050] A data obtaining module is configured to obtain a target high-dimensional data set to be processed.

[0051] A second calculating module is configured to calculate k different target similarity matrices of the target high-dimensional data set.

[0052] A second graph information extracting module is configured to take the target similarity matrix as a weighted fully connected graph, and extract graph position encoding information of nodes in the fully connected graph from the target similarity matrix by using a spectral method.

[0053] a visualization module configured to input the k different similarity matrices of the target high-dimensional data set and the corresponding graph position encoding information thereof into a pre-trained high-dimensional data visualization model, execute the pre-trained high-dimensional data visualization model, and obtain a visualization result output by the pre-trained high-dimensional data visualization model;

[0054] The pre-trained high-dimensional data visualization model is trained based on the device of the third aspect.

[0055] In a fifth aspect, an electronic device is provided, including a memory and a processor, where the memory is configured to store one or more computer instructions, and the one or more computer instructions are executed by the processor to implement the method of any one of the first or second aspect.

[0056] In a sixth aspect, a computer readable storage medium is provided, which stores computer instructions, and the computer instructions are executed by a processor to implement the method of any one of the first or second aspect.

[0057] According to the technical scheme provided in the embodiments of the present disclosure, a plurality of training samples can be obtained, each of which includes a high-dimensional data set and a corresponding historical successful visualization result of the high-dimensional data set. For each high-dimensional data set, k different similarity matrices of the high-dimensional data set are calculated. For each similarity matrix of the high-dimensional data set, the similarity matrix is taken as a weighted full connection graph, and a spectral method is used to extract graph position encoding information of nodes in the full connection graph from the similarity matrix. In this way, the high-dimensional data set can be converted into a multi-view graph structure and the graph position encoding information can be extracted, and the data set of any dimension can be successfully converted into graph structure information with fixed dimension. Then, the k different similarity matrices of the high-dimensional data set and the corresponding graph position encoding information thereof are taken as input data of a model, and the historical successful visualization result corresponding to the high-dimensional data set is taken as supervision information of the model for model training, to obtain a trained high-dimensional data visualization model. In this way, the historical successful visualization result of the high-dimensional data set of different dimensions and different fields can be fully utilized for model training, which can be effectively applied to new data of different dimensions and different fields, improve the generalization ability of the model, avoid the repeated training process of the new data in the prior art, greatly reduce the computational overhead, and improve the processing efficiency. Moreover, the training method significantly reduces the sensitivity of the model to hyperparameters and enhances the robustness of the model.

[0058] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0059] Other features, objects, and advantages of the present disclosure will become more apparent from the following detailed description when read in conjunction with the accompanying drawings. In the drawings:

[0060] Figure 1 A flow chart of a high-dimensional data visualization model training method provided by an embodiment of the present disclosure is shown.

[0061] Figure 2 A flow chart of a high-dimensional data visualization method provided by an embodiment of the present disclosure is shown.

[0062] Figure 3 A structural block diagram of a high-dimensional data visualization model training device provided by an embodiment of the present disclosure is shown.

[0063] Figure 4 A structural block diagram of a high-dimensional data visualization device provided by an embodiment of the present disclosure is shown.

[0064] Figure 5 A structural block diagram of an electronic device according to an embodiment of the present disclosure is shown.

[0065] Figure 6 A structural diagram of a computer system suitable for implementing the method of an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0066] Hereinafter, exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings so as to be easily implemented by those skilled in the art. Also, portions irrelevant to the description of the exemplary embodiments are omitted in the drawings for the sake of clarity.

[0067] In the present disclosure, it should be understood that terms such as "include" or "have" are intended to indicate that there are features, numbers, steps, actions, components, parts or combinations thereof disclosed in the specification, and do not exclude the possibility of adding one or more other features, numbers, steps, actions, components, parts or combinations thereof.

[0068] It should also be further noted that the embodiments in the present disclosure and the features in the embodiments can be combined with each other without conflict. The present disclosure will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0069] Figure 1 A flow chart of a high-dimensional data visualization model training method provided by an embodiment of the present disclosure is shown. As shown in Figure 1 The high-dimensional data visualization model training method includes the following steps S101-S104:

[0070] In step S101, a plurality of training samples are obtained, each of which includes a high-dimensional data set and its corresponding historical successful visualization result;

[0071] In step S102, for each high-dimensional data set, k different similarity matrices of the high-dimensional data set are calculated;

[0072] In step S103, for each similarity matrix of the high-dimensional data set, the similarity matrix is taken as a weighted full-connected graph, and a spectral method is used to extract graph position encoding information of nodes in the full-connected graph from the similarity matrix;

[0073] In step S104, the k different similarity matrices of the high-dimensional data set and the corresponding graph position encoding information are taken as input data of a model, and historical successful visualization results corresponding to the high-dimensional data set are taken as supervision information of the model for model training, to obtain a trained high-dimensional data visualization model.

[0074] In a possible implementation, the high-dimensional data visualization model training method can be applied to an electronic device such as a computer, a server, a computer cluster, or the like, which can perform training of a high-dimensional data visualization model.

[0075] In a possible implementation, a plurality of training samples for model training can be obtained, and each training sample includes a high-dimensional data set and a historical successful visualization result that has been successfully dimensionally reduced and visualized in history. Here, the high-dimensional data set in the i th training sample can be denoted as X i , and X i includes N i d i dimensional data points, and can be represented as an N i xd i matrix. Since the high-dimensional data sets in different training samples can be high-dimensional data of different fields and different types, the data dimensions and the data amount of each dimension of different high-dimensional data sets can be the same or different, which is not limited here. That is, when i takes different values, N i may take different or same values, and d i may take different or same values. The corresponding historical successful visualization result of the high-dimensional data set X i may be denoted as Y , which is an N i d' dimensional low-dimensional vector representation, where a d' dimensional low-dimensional vector is a data point of X , and d ′ = 2 or 3. In this way, L training samples can be obtained:

[0076] In a possible implementation, for the i th high-dimensional data set X i, k different similarity matrices of the high-dimensional data set can be calculated, and the j-th similarity matrix among the k different similarity matrices can be recorded as j takes the value of 1, 2, ..., k. Here, the similarity matrix is ​​a symmetric matrix or an asymmetric matrix (depending on the definition of similarity) used to represent the pairwise similarity between a set of data (such as vectors, samples, or nodes). For a high-dimensional dataset X i , (N i ×d i The data matrix of

[0077]

[0078] Among them, X i [u] is the high-dimensional dataset X i The u-th data point, X i [v] is the high-dimensional dataset X i The vth data point, sim is a similarity function (such as cosine similarity, Gaussian kernel similarity, etc.). For the same high-dimensional data set, different similarity matrices can be obtained by calculating different parameters of the sim function.

[0079] In a possible implementation, for the i-th high-dimensional dataset X i The j-th similarity matrix The high-dimensional dataset X i Each data point is a node, and the similarity matrix Elements in is the similarity weight between node u (the u-th data point) and node v (the v-th data point).

[0080] In one possible implementation, Graph Positional Encoding is a technique for generating structured low-dimensional vectors for nodes in a graph, which is used to represent the topological position or structural role of the node in the graph. Unlike node features, positional encoding focuses on the structural information of the graph. Spectral methods can use the eigendecomposition of matrices to extract the structural information of the graph, i.e., the graph position encoding information. For example, the following formula can be used to express the similarity matrix: The graph position encoding information P extracted from i (j) :

[0081] Among them, h is the graph position encoding extraction function. So, for the i-th high-dimensional dataset X ifor each similarity matrix of the i-th high-dimensional data set X

[0082] In a possible implementation, the i-th high-dimensional data set X i corresponding historical successful visualization result As input data of the model, the i-th high-dimensional data set X i corresponding historical successful visualization result As supervision information of the model, L is used as follows and Model training can be performed by updating parameters in the high-dimensional data visualization model using the gradient descent method until the loss function of the model converges or the number of iterations reaches the maximum number, and the training is ended, and a trained high-dimensional data visualization model is obtained.

[0083] In the embodiment, a plurality of training samples can be obtained, each of which includes a high-dimensional data set and a corresponding historical successful visualization result; for each high-dimensional data set, k different similarity matrices of the high-dimensional data set are calculated; for each similarity matrix of the high-dimensional data set, the similarity matrix is taken as a weighted full connection graph, and the graph position encoding information of the nodes in the full connection graph is extracted from the similarity matrix using a spectral method; in this way, the high-dimensional data set can be converted into a multi-view graph structure and the graph position encoding information is extracted, and the data set of any dimension is successfully converted into a graph structure information with fixed dimension; then the k different similarity matrices of the high-dimensional data set and the corresponding graph position encoding information are taken as input data of the model, and the historical successful visualization result corresponding to the high-dimensional data set is taken as supervision information of the model for model training, and a trained high-dimensional data visualization model is obtained. In this way, the historical successful visualization results of high-dimensional data sets of different dimensions and different fields can be fully utilized for model training, which can effectively be applied to new data of different dimensions and different fields, improve the generalization ability of the model, avoid the repeated training process of the new data in the prior art, greatly reduce the computational overhead, and improve the processing efficiency; moreover, the training method significantly reduces the sensitivity of the model to hyperparameters and enhances the robustness of the model.

[0084] In a possible implementation, the obtaining of the plurality of training samples comprises:

[0085] Obtaining a plurality of high-dimensional data sets with supervision information

[0086] The i-th high-dimensional data set X i corresponding historical successful visualization result

[0087] wherein, is a high-dimensional data visualization algorithm function, and θ is a hyper-parameter, y i is a high-dimensional data set X i corresponding supervision information, is a visualization result evaluation index.

[0088] In this embodiment, for each high-dimensional data set, the visualization result evaluation index the maximum visualization result as its corresponding historical successful visualization result.

[0089] In this embodiment, the i-th high-dimensional data set X i corresponding supervision information is y i , y i is represented as an N i dimensional vector.

[0090] In this embodiment, the high-dimensional data visualization algorithm function may be any algorithm function such as t-SNE or UMAP, refers to using the to perform dimension reduction visualization on the high-dimensional data set X i The transformation of different hyper-parameters θ will lead to different visualization results, and the visualization result evaluation index and y i can be calculated based on the Example, the visualization result evaluation index may be normalized mutual information (Normalized Mutual Information, NMI), which is an index for evaluating the consistency with the real supervision information (i.e. y i ), and is especially suitable for the evaluation of the visualization result after the dimension reduction of the high-dimensional data.

[0091] In this embodiment, the Bayesian optimization method can be used to find the hyper-parameter θ when the visualization result evaluation index is maximum, and then the visualization result when the visualization result evaluation index is maximum is obtained as the historical successful visualization result.

[0092] In one possible embodiment, the calculation of the k different similarity matrices corresponding to the high-dimensional data set comprises:

[0093] The element in the u-th row and the v-th column in the j-th similarity matrix is calculated according to the following formula

[0094]

[0095]

[0096] where X i [u] is the u-th data point of the high-dimensional data set X i , X i [v] is the v-th data point of the high-dimensional data set X i , and γ (j) is the bandwidth parameter of the j-th similarity matrix.

[0097] In this embodiment, the Gaussian kernel function can be used to calculate the similarity matrix, and ‖X i [u]-X i [v]‖2 represents the Euclidean distance between the data points X i [u] and X i [v], exp(·) is the exponential function, γ (j) is the bandwidth parameter, and different similarity matrices have different γ (j) . For example, k = 5, and γ (j) can be set to {0.1, 0.5, 1, 2, 5}, so that 5 different similarity matrices can be obtained for each high-dimensional data set.

[0098] In one possible embodiment, the graph position encoding information of the nodes in the complete graph is extracted from the similarity matrix by using the spectral method, comprising:

[0099] The graph position encoding information P i of the nodes in the complete graph is extracted from the j-th similarity matrix of the i-th high-dimensional data set X i (j) according to the following formula:

[0100] P i (j) = US 1 / 2 ;

[0101] where U represents a matrix composed of left singular vectors corresponding to the largest d e singular values obtained by singular value decomposition of X e , and S is a diagonal matrix composed of the largest d i singular values.

[0102] In this embodiment, the spectral method can be used to perform singular value decomposition (SVD) on the similarity matrix , which is a method of decomposing any matrix into the product of three specific matrices. For the similarity matrix , its SVD is expressed as:

[0103]

[0104] where W is N i × N i is an orthogonal matrix (left singular vector), the column vector is an orthonormal basis, Σ is an N i × N i diagonal matrix, the diagonal elements are singular values (non-negative and arranged in descending order), V T is the transpose of an N i × N i orthogonal matrix (right singular vector), and the data points are orthonormal bases.

[0105] In this embodiment, the largest d e singular values in Σ can be selected to obtain U composed of the left singular vectors corresponding to the largest d e singular values in W, the diagonal matrix S composed of the largest d e singular values, and P i (j) = US 1 / 2 , P i (j) is an N i × d e matrix, where d e is the dimension of the graph position encoding, and in an example, d e may be set to 64.

[0106] In one possible embodiment, the high-dimensional data visualization model includes a graph neural network (GNN) sub-model, a graph transformer sub-model, and a multilayer perceptron (MLP) sub-model.

[0107] In this embodiment, k pairs of and P i (j) may be input into the graph neural network sub-model to obtain k graph neural network sub-model output information, then the k graph neural network sub-model output information is spliced to obtain spliced information, and the spliced information is input into a graph transformer sub-model. Finally, the output is reduced to d' dimensions by a multilayer perceptron sub-model. Specifically, the expression of the high-dimensional data visualization model is as follows:

[0108]

[0109] where, represents splicing along the feature dimension, and GT is a graph transformer sub-model.

[0110] In this embodiment, the graph neural network sub-model can use GIN (Graph Isomorphism Network) as the backbone network, and the parameter settings of the graph Transformer conversion sub-model can be 8 layers, 4 attention heads, and 2048 implicit dimensions.

[0111] In one possible implementation, the loss function of the high-dimensional data visualization model is for:

[0112]

[0113] Among them, L is the number of training samples, N i is the number of data points in the high-dimensional dataset, is the Bregman divergence, σ is the similarity conversion function, is the u-th data point of the visualization result output by the high-dimensional data visualization model, is the vth data point of the visualization result output by the high-dimensional data visualization model, is the u-th data point of the historical successful visualization result, is the vth data point of the historical successful visualization result. It should be noted here that and Same, also N i A d′-dimensional low-dimensional vector representation, where a d′(d ′ =2 or 3)-dimensional low-dimensional vector is A data point of The vth data point of the visualization result output by the high-dimensional data visualization model.

[0114] In this embodiment, an affine invariant loss function can be used as the loss function for model training, and the difference of the similarity matrix can be used as the training target. and The calculation formula is and The elements in the similarity matrix of for The element in the uth row and vth column of the similarity matrix of for The element in the uth row and vth column of the similarity matrix.

[0115] In this embodiment, the historical success visualization results are used Different, the above σ and The parameter ψ in is different.

[0116] In a possible implementation, when the historical successful visualization result is obtained based on the t-SNE algorithm, the loss function of the high-dimensional data visualization model is for:

[0117]

[0118]

[0119] in, The uth visualization result output by the high-dimensional data visualization model ′ data points, The vth visualization result output by the high-dimensional data visualization model ′ data points, The uth historical successful visualization result ′ data points, The vth historical success visualization result ′ data points.

[0120] In a possible implementation, when the historical successful visualization results are obtained based on the UMAP algorithm, the loss function of the high-dimensional data visualization model is for:

[0121]

[0122] Among them, a and b are two similarity conversion parameters, for example, a=1.93, b=0.79; M i is the index set of k neighboring points of each point in the high-dimensional dataset, λ(t) is a weighted term that decays with the number of training iterations. For example, λ(t) = 1-t / T, where T is the maximum number of iterations.

[0123] In this embodiment, a point in a high-dimensional dataset refers to a data point in the high-dimensional dataset. For the i-th high-dimensional dataset X i For each point in , find its nearest k points. The set of subscript indices of each point and its nearest k points is the index set M of k neighboring points. i , represents l belonging to k neighboring points (u,v) (u,v) the sum of represents l that does not belong to the k neighboring points (u,v) (u,v) sum.

[0124] The present disclosure also provides a high-dimensional data visualization method, Figure 2 A flow chart of a high-dimensional data visualization method provided by an embodiment of the present disclosure is shown as follows: Figure 2As shown, the high-dimensional data visualization method comprises the following steps S201-S204:

[0125] In step S201, a target high-dimensional data set to be processed is acquired;

[0126] In step S202, k different target similarity matrices of the target high-dimensional data set are calculated;

[0127] In step S203, the target similarity matrices are taken as weighted complete graphs, and the graph position encoding information of nodes in the complete graphs is extracted from the target similarity matrices by using a spectral method;

[0128] In step S204, the k different similarity matrices of the target high-dimensional data set and the corresponding graph position encoding information are input into a pre-trained high-dimensional data visualization model, the pre-trained high-dimensional data visualization model is executed, and a visualization result output by the pre-trained high-dimensional data visualization model is obtained.

[0129] In a possible implementation, the high-dimensional data visualization method can be applied to an electronic device such as a computer, a server, a computer cluster, or the like that can perform high-dimensional data visualization.

[0130] In a possible implementation, a target high-dimensional data set X new to be processed is acquired. new The target high-dimensional data set X new may be represented as an N new ×d new matrix.

[0131] In a possible implementation, k different similarity matrices of the target high-dimensional data set X new are calculated, denoted as is a similarity matrix with a size of N new ×N new .

[0132] In a possible implementation, a spectral method can be used to extract the corresponding graph position encoding of denoted as is a matrix with a size of N new ×d e . The operation is performed on each of , thereby extracting the corresponding graph position encoding information

[0133] In a possible implementation, the above-obtained and the corresponding are input into a pre-trained high-dimensional data visualization model.The pre-trained high-dimensional data visualization model is inputted, executed, and a visualization result output by the pre-trained high-dimensional data visualization model is obtained. The visualization result output by the high-dimensional data visualization model can be represented by the following formula

[0134]

[0135] The disclosure also provides a high-dimensional data visualization model training device, Figure 3 A structural block diagram of a high-dimensional data visualization model training device provided by an embodiment of the disclosure is shown, which can be realized as part of or the whole of an electronic device by software, hardware, or a combination of both. As shown in Figure 3 The high-dimensional data visualization model training device comprises:

[0136] The sample acquisition module 301 is configured to acquire a plurality of training samples, each of which comprises a high-dimensional data set and a corresponding historical successful visualization result;

[0137] The first calculation module 302 is configured to calculate, for each high-dimensional data set, k different similarity matrices of the high-dimensional data set;

[0138] The first graph information extraction module 303 is configured to, for each similarity matrix of the high-dimensional data set, take the similarity matrix as a weighted full-connected graph, and extract graph position encoding information of nodes in the full-connected graph from the similarity matrix by using a spectral method;

[0139] The training module 304 is configured to take the k different similarity matrices of the high-dimensional data set and the corresponding graph position encoding information as input data of a model, take the historical successful visualization result corresponding to the high-dimensional data set as supervision information of the model, and perform model training to obtain a trained high-dimensional data visualization model.

[0140] The disclosure also provides a high-dimensional data visualization device, Figure 4 A structural block diagram of a high-dimensional data visualization device provided by an embodiment of the disclosure is shown, which can be realized as part of or the whole of an electronic device by software, hardware, or a combination of both. As shown in Figure 4 The high-dimensional data visualization device comprises:

[0141] The data acquisition module 401 is configured to acquire a target high-dimensional data set to be processed;

[0142] The second calculation module 402 is configured to calculate k different target similarity matrices of the target high-dimensional data set;

[0143] The second graph information extraction module 403 is configured to extract graph position encoding information of nodes in a full connection graph from the target similarity matrix as a weighted full connection graph by using a spectral method.

[0144] The visualization module 404 is configured to input the k different similarity matrices of the target high-dimensional data set and the corresponding graph position encoding information into a pre-trained high-dimensional data visualization model, execute the pre-trained high-dimensional data visualization model, and obtain a visualization result output by the pre-trained high-dimensional data visualization model.

[0145] The pre-trained high-dimensional data visualization model is trained based on the high-dimensional data visualization model training device.

[0146] The technical terms and technical features mentioned in the device embodiments are the same as or similar to those mentioned in the above method embodiments, and the explanation and description of the technical terms and technical features involved in the device can refer to the explanation and description of the above method embodiments, which will not be repeated here.

[0147] The present disclosure also discloses an electronic device, Figure 5 A structural block diagram of an electronic device according to an embodiment of the present disclosure is shown.

[0148] As Figure 5 shown, the electronic device 500 includes a memory 501 and a processor 502, wherein the memory 501 is configured to store one or more computer instructions, and the one or more computer instructions are executed by the processor 502 to implement the method according to the embodiment of the present disclosure.

[0149] Figure 6 A structural schematic diagram of a computer system suitable for implementing the method of the embodiment of the present disclosure is shown.

[0150] As Figure 6 shown, the computer system 600 includes a processing unit 601, which can perform various processes in the above embodiments according to programs stored in a read-only memory (ROM) 602 or loaded into a random access memory (RAM) 603 from a storage portion 608. In the RAM 603, various programs and data required for the operation of the computer system 600 are also stored. The processing unit 601, the ROM 602, and the RAM 603 are connected to each other through a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0151] The following components are connected to the I / O interface 605: an input part 606 including a keyboard, a mouse, and the like; an output part 607 including a cathode ray tube (CRT), a liquid crystal display (LCD), and the like, and a speaker, and the like; a storage part 608 including a hard disk, and the like; and a communication part 609 including a network interface card such as a LAN card, a modem, and the like. The communication part 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the I / O interface 605 as necessary. A removable medium 611 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, and the like is attached to the drive 610 as necessary, so that a computer program read out therefrom is installed in the storage part 608 as necessary. Among them, the processing unit 601 can be implemented as a CPU, a GPU, a TPU, an FPGA, an NPU, and the like.

[0152] In particular, according to embodiments of the present disclosure, the method described above can be implemented as a computer software program. For example, embodiments of the present disclosure include a computer program product including computer instructions that, when executed by a processor, implement the method steps described above. In such embodiments, the computer program product can be downloaded and installed from a network by the communication part 609, and / or installed from the removable medium 611.

[0153] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowcharts and block diagrams can represent a module, a segment, or a portion of code, which comprises one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures. For example, two blocks noted in succession can in fact be executed substantially concurrently or in the reverse order, depending on the functionality involved. It will also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented by dedicated hardware-based systems that perform the specified functions or operations, or can be implemented by a combination of dedicated hardware and computer instructions.

[0154] The units or modules described in the embodiments of the present disclosure can be implemented by software, or can be implemented by programmable hardware. The described units or modules can also be arranged in a processor, and the names of these units or modules do not constitute a limitation on the units or modules themselves in some cases.

[0155] As another aspect, the disclosure also provides a computer readable storage medium, which can be the computer readable storage medium contained in the electronic device or the computer system in the above embodiments; or can be a computer readable storage medium existing separately and not assembled into a device. The computer readable storage medium stores one or more programs used by one or more processors to execute the method described in the disclosure.

[0156] The above description is merely the preferred embodiments of the disclosure and the explanation of the principles of the applied technology. It should be understood by those skilled in the art that the inventive scope of the disclosure is not limited to the technical solutions formed by the specific combinations of the above technical features, and should also cover other technical solutions formed by the combinations of the above technical features or equivalent features without departing from the inventive concept. For example, the technical solutions formed by the mutual replacement of the above features and the technical features disclosed in the disclosure (but not limited to) having similar functions.

Claims

1. A high-dimensional data visualization model training method, characterized in that: include: Obtain multiple training samples, each of which includes a high-dimensional dataset and its corresponding historical successful visualization results; For each high-dimensional data set, calculate k different similarity matrices of the high-dimensional data set, where k is a positive integer; For each similarity matrix of the high-dimensional data set, the similarity matrix is ​​used as a weighted fully connected graph, and a spectral method is used to extract graph position encoding information of nodes in the fully connected graph from the similarity matrix; The k different similarity matrices of the high-dimensional data set and their corresponding graph position encoding information are used as the input data of the model, and the historical successful visualization results corresponding to the high-dimensional data set are used as the supervision information of the model for model training to obtain a trained high-dimensional data visualization model.

2. The method according to claim 1, characterized in that The obtaining of multiple training samples includes: Obtain multiple high-dimensional datasets with supervised information; Get the i-th high-dimensional dataset X according to the following formula i Corresponding historical successful visualization results in, is the high-dimensional data visualization algorithm function, θ is The hyperparameters, y i is a high-dimensional dataset X i The corresponding supervision information, It is the evaluation index of visualization results, and i is a positive integer.

3. The method according to claim 1, characterized in that The calculating k different similarity matrices of the high-dimensional data set includes: Calculate the i-th high-dimensional dataset X according to the following formula i The j-th similarity matrix The element at row u and column v in Among them, X i [u] is the high-dimensional dataset X i The u-th data point, X i [v] is the high-dimensional dataset X i The vth data point, γ (j) is the bandwidth parameter of the j-th similarity matrix.

4. The method according to claim 1, wherein The extracting graph position encoding information of nodes in the fully connected graph from the similarity matrix using a spectral method includes: According to the following formula, from the i-th high-dimensional dataset X i The j-th similarity matrix Extract the graph position encoding information P of the nodes in the fully connected graph i (j) : P i (j) =US 1 / 2 ; Among them, U represents the The maximum d obtained after singular value decomposition e The matrix composed of the left singular vectors corresponding to the singular values, S is the maximum d e A diagonal matrix consisting of singular values.

5. The method according to claim 1, wherein The high-dimensional data visualization model includes a graph neural network sub-model, a graph conversion Transformer sub-model and a multi-layer perceptron sub-model; Among them, the input of the graph neural network sub-model is k different similarity matrices of the high-dimensional data set and their corresponding graph position encoding information, and the output is the output information of k graph neural network sub-models; the input of the graph Transformer sub-model is the splicing information of the output information of k graph neural network sub-models.

6. The method according to claim 1, characterized in that Loss function of the high-dimensional data visualization model for: Among them, L is the number of training samples, N i is the number of data points in the high-dimensional dataset, is the Bregman divergence, σ is the similarity conversion function, is the u-th data point of the visualization result output by the high-dimensional data visualization model, is the vth data point of the visualization result output by the high-dimensional data visualization model, is the u-th data point of the historical successful visualization result, The vth data point of the historical successful visualization results.

7. The method according to claim 6, characterized in that When the historical successful visualization results are obtained based on the t-distributed random neighbor embedding t-SNE algorithm, the loss function of the high-dimensional data visualization model is for: in, The uth visualization result output by the high-dimensional data visualization model ′ data points, The vth visualization result output by the high-dimensional data visualization model ′ data points, The uth historical successful visualization result ′ data points, The vth historical success visualization result ′ data points; When the historical successful visualization results are obtained based on the unified manifold approximation and projection UMAP algorithm, the loss function of the high-dimensional data visualization model is for: Among them, a and b are two similarity conversion parameters, M i is the index set of k neighboring points of each point in the high-dimensional dataset, and λ(t) is a weighted term that decays with the number of training iterations.

8. A high-dimensional data visualization method, characterized in that: include: Obtain the target high-dimensional dataset to be processed; Calculating k different target similarity matrices of the target high-dimensional data set, where k is a positive integer; Taking the target similarity matrix as a weighted fully connected graph, and extracting graph position encoding information of nodes in the fully connected graph from the target similarity matrix using a spectral method; Inputting k different similarity matrices of the target high-dimensional data set and their corresponding graph position encoding information into a pre-trained high-dimensional data visualization model, executing the pre-trained high-dimensional data visualization model, and obtaining a visualization result output by the pre-trained high-dimensional data visualization model; The pre-trained high-dimensional data visualization model is trained based on the method described in any one of claims 1 to 7.

9. A high-dimensional data visualization model training device, characterized in that: include: a sample acquisition module configured to acquire a plurality of training samples, each training sample including a high-dimensional data set and its corresponding historical successful visualization results; A first calculation module is configured to calculate k different similarity matrices of each high-dimensional data set, where k is a positive integer; a first graph information extraction module configured to extract graph position encoding information of nodes in each similarity matrix of the high-dimensional data set, using the similarity matrix as a weighted fully connected graph, and employing a spectral method to extract graph position encoding information of nodes in the fully connected graph from the similarity matrix; The training module is configured to use k different similarity matrices of the high-dimensional data set and their corresponding graph position encoding information as input data of the model, and use the historical successful visualization results corresponding to the high-dimensional data set as the supervision information of the model to perform model training, thereby obtaining a trained high-dimensional data visualization model.

10. A high-dimensional data visualization device, characterized in that: include: A data acquisition module is configured to acquire a target high-dimensional data set to be processed; A second calculation module is configured to calculate k different target similarity matrices of the target high-dimensional data set, where k is a positive integer; a second graph information extraction module configured to use the target similarity matrix as a weighted fully connected graph and extract graph position encoding information of nodes in the fully connected graph from the target similarity matrix using a spectral method; a visualization module configured to input k different similarity matrices of the target high-dimensional data set and their corresponding graph position encoding information into a pre-trained high-dimensional data visualization model, execute the pre-trained high-dimensional data visualization model, and obtain a visualization result output by the pre-trained high-dimensional data visualization model; Wherein, the pre-trained high-dimensional data visualization model is trained based on the device described in claim 9.

11. An electronic device, characterized in that: The method comprises a memory and a processor, wherein the memory is used to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the method according to any one of claims 1 to 8.

12. A readable storage medium, characterized in that: Computer instructions are stored thereon, and when the computer instructions are executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images