A method for analyzing the improvement of self-learning ability of large data models

By constructing clustering space and dynamically adjusting clustering centers, the problem of K-mean algorithm falling into local optimal solutions during clustering is solved, and the clustering accuracy and model accuracy of large-model training data are improved.

CN118520323BActive Publication Date: 2024-11-08SHANDONG SHENGDE INTELLIGENT TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410968933.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-19
Publication Date
2024-11-08
Estimated Expiration
2044-07-19

AI Technical Summary

Technical Problem

The K-mean algorithm is prone to fall into local optimal solutions during the clustering process, resulting in inaccurate clustering results of large model training data, affecting the accuracy of the model.

Method used

By obtaining the training parameter set after dimensionality reduction, a cluster space is constructed, and the initial cluster center is determined using neighborhood density and initial cluster center distance characteristics, the training quality index is calculated by iterating the inter-cluster distance and density of the cluster center, and dynamically adjusting the cluster center Iterate the clustering center to avoid local optimal solutions.

Benefits of technology

It improves the accuracy of clustering results, reduces the local optimal solution, makes the training parameters of the large model more accurate, and improves the accuracy of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118520323B_ABST
    Figure CN118520323B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of data processing technology, and in particular to a method for analyzing the improvement of the self-learning ability of a large data model; obtaining an initial cluster center and performing iterative clustering according to the neighborhood density characteristics of training parameters and the distance characteristics from the determined initial cluster center; obtaining a training quality index according to the distance characteristics between the iterative cluster center and other iterative cluster centers, the area characteristics of the current cluster cluster, and the distribution characteristics of the training parameters within the cluster in the current iterative clustering; obtaining an iterative feature variation degree according to the difference characteristics of the area, training parameters, and training quality index between the historical cluster clusters where the iterative cluster center is located; and obtaining a correction coefficient according to the iterative feature variation degree and the training quality index. The present invention obtains the next round of iterative cluster centers and the final training cluster according to the correction coefficient and the moving distance value; and trains and learns the large model according to the training cluster to improve the accuracy of clustering and large model training.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method for analyzing the improvement of the self-learning ability of a large data model. Background Art

[0002] The data big model is a machine learning model that is trained and optimized based on massive data. It is mainly used to solve complex data analysis and prediction and is widely used in various fields, such as natural language processing, computer vision, speech recognition, etc. In order to improve the capabilities of the big model, a large amount of training data will be continuously added to allow the big model to perform self-training and learning; due to the large training data set, the quality of some training parameters is poor, and the role in the big model training process is small. Therefore, the training data set will be classified before training and learning, which is helpful for data enhancement and cleaning, and improves the training efficiency and accuracy of the big model.

[0003] K-means clustering algorithm is a commonly used clustering algorithm. It can be used to cluster training data sets to complete data classification and cleaning. Due to the wide range of sources of training parameters in the training data set, high degree of confusion, and strong randomness of features in the training data set, it is easy for K-means to fall into a local optimal solution during the clustering process and fail to reach a global optimal solution, resulting in inaccurate clustering results. This in turn makes the accuracy of large models trained and learned based on clustering results poor, affecting the accuracy of large models. Summary of the invention

[0004] In order to solve the technical problem that the clustering results obtained by the K-means algorithm are used for large model training and learning, which affects the accuracy of the large model, the purpose of the present invention is to provide an analysis method for improving the self-learning ability of a large data model. The technical solution adopted is as follows:

[0005] Obtain a set of training parameters after dimensionality reduction for large model learning, and construct a clustering space of the training parameters; obtain a preliminary degree in the clustering space according to the neighborhood density characteristics of the training parameters and the distance characteristics from the determined initial clustering center; and cyclically obtain a preset number of initial clustering centers according to the preliminary degree of the training parameters;

[0006] Iteratively clustering the training parameters using the K-means algorithm according to the initial cluster center; obtaining an inter-cluster distance characteristic value according to the distance characteristics between the iterative cluster center and other iterative cluster centers in the current iterative clustering; obtaining a density according to the interval characteristics of the training parameters in the current clustering cluster where the iterative cluster center is located, the quantity difference characteristics of the training parameters in different regions, and the area characteristics of the current clustering cluster; obtaining a training quality index of the current clustering cluster according to the inter-cluster distance characteristic value and the density;

[0007] Obtaining the iteration characteristic variation degree according to the area difference characteristics between the historical clustering clusters where the iteration clustering center is located in the historical iteration clustering, the quantity difference characteristics of the training parameters and the difference characteristics of the training quality index; obtaining the correction coefficient according to the iteration characteristic variation degree of the iteration clustering center and the training quality index of the current clustering cluster, and obtaining the moving distance value according to the distance characteristics between the iteration clustering center and the historical iteration clustering center; obtaining the iteration clustering center of the next round of iteration clustering according to the correction coefficient and the moving distance value;

[0008] Clustering is completed according to the iterative clustering centers obtained in a cycle to obtain different training clusters; and the large model is trained and learned according to the training clusters.

[0009] Furthermore, the step of obtaining a set of training parameters after dimension reduction for large model learning and constructing a clustering space of training parameters includes:

[0010] A two-dimensional set of training parameters is obtained through principal component analysis; the values ​​of two different dimensions of the training parameters are represented by the horizontal axis and the vertical axis in a plane rectangular coordinate system, the coordinates corresponding to the training parameters are obtained, and a clustering space is constructed according to the coordinates corresponding to all training parameters in the plane rectangular coordinate system.

[0011] Furthermore, the step of obtaining the preliminary degree in the cluster space according to the neighborhood density characteristics of the training parameters and the distance characteristics from the determined initial cluster center includes:

[0012] In the clustering space, the average value of the Euclidean distances between all training parameters and other training parameters closest to them is calculated to obtain a neighborhood radius; the number of other training parameters in a circular area with the training parameter as the center and the neighborhood radius as the radius is calculated and normalized to obtain a neighborhood concentration of the training parameter; the average value of the Euclidean distances between the training parameter and the determined initial clustering center is calculated and normalized to obtain a clustering deviation of the training parameter; the sum of the neighborhood concentration and the clustering deviation is calculated to obtain a preliminary degree of the training parameter.

[0013] Furthermore, the step of cyclically acquiring a preset number of initial cluster centers according to the initial degree of the training parameters includes:

[0014] The training parameter corresponding to the maximum value of the preliminary selection degree is selected as the initial cluster center; after obtaining an initial cluster center, the preliminary selection degree corresponding to the training parameter of the non-initial cluster center is recalculated until a preset number of initial cluster centers are obtained.

[0015] Furthermore, the step of obtaining the inter-cluster distance characteristic value according to the distance characteristic between the iterative cluster center and other iterative cluster centers in the current iterative clustering includes:

[0016] The sum of the Euclidean distances between the iterative cluster center and other iterative cluster centers is calculated and normalized to obtain the inter-cluster distance characteristic value of the iterative cluster center.

[0017] Furthermore, the step of obtaining the density according to the interval characteristics of the training parameters in the current cluster where the iterative cluster center is located, the quantity difference characteristics of the training parameters in different regions and the area characteristics of the current cluster includes:

[0018] Calculate the average value of the Euclidean distance between all training parameters in the current cluster and other training parameters that are closest to them and perform negative correlation mapping to obtain the nearest neighbor eigenvalue of the current cluster; calculate the area of ​​the minimum circumscribed rectangle of the current cluster in the cluster space and perform negative correlation mapping to obtain the first clustering eigenvalue of the current cluster; obtain the minimum circumscribed circle of the current cluster in the cluster space, obtain a preset second number of concentric circles with the same center as the minimum circumscribed circle within the minimum circumscribed circle, and the radius difference between adjacent concentric circles is the same; obtain the circular rings between adjacent concentric circles, calculate the sum of the quantity difference of the training parameters in two adjacent circular rings from the center of the concentric circle to the outside direction and normalize them to obtain the second clustering eigenvalue of the current cluster;

[0019] The product of the nearest neighbor eigenvalue, the first cluster eigenvalue and the second cluster eigenvalue is calculated to obtain the density of the current cluster.

[0020] Furthermore, the step of obtaining the training quality index of the current clustering cluster according to the inter-cluster distance characteristic value and density includes:

[0021] The average value of the inter-cluster distance characteristic value and the density is calculated to obtain the training quality index of the current clustering cluster.

[0022] Furthermore, the step of obtaining the iteration feature variation degree according to the area difference characteristics between the historical clustering clusters where the iteration cluster center is located in the historical iteration clustering, the quantity difference characteristics of the training parameters and the difference characteristics of the training quality index comprises:

[0023] Any historical iteration cluster of the iteration cluster center and the adjacent next historical iteration cluster are taken as adjacent historical iteration clusters; the absolute value of the difference in the area of ​​the minimum circumscribed rectangle of the two historical cluster clusters where the iteration cluster center is located in the adjacent historical iteration cluster is calculated and normalized to obtain the first iteration characteristic value of the arbitrary historical iteration cluster; the absolute value of the difference in the number of training parameters of the two historical cluster clusters where the iteration cluster center is located in the adjacent historical iteration cluster is calculated and normalized to obtain the second iteration characteristic value of the arbitrary historical iteration cluster; the absolute value of the difference in the training quality index of the two historical cluster clusters where the iteration cluster center is located in the adjacent historical iteration cluster is calculated to obtain the third iteration characteristic value of the arbitrary historical iteration cluster;

[0024] Calculate the sum of the iteration orders of all historical iteration clusters to obtain an iteration order benchmark; calculate the ratio of the iteration order of any historical iteration cluster to the iteration order benchmark to obtain the iteration weight of the arbitrary historical iteration cluster; calculate the product of the iteration weight, the first iteration eigenvalue, the second iteration eigenvalue and the third iteration eigenvalue to obtain the iteration change characteristic factor of the arbitrary historical iteration cluster; calculate the sum of the iteration change characteristic factors of all historical iteration clusters of the iteration cluster center to obtain the iteration characteristic change degree of the iteration cluster center.

[0025] Furthermore, the step of obtaining a correction coefficient according to the iteration characteristic change degree of the iteration cluster center and the training quality index of the current cluster cluster, and obtaining a moving distance value according to the distance characteristic between the iteration cluster center and the historical iteration cluster center includes:

[0026] The training quality index is negatively correlated and mapped to obtain the clustering effect characteristic value of the current clustering cluster, and the average value of the clustering effect characteristic value and the iterative characteristic change degree is calculated to obtain the correction factor; the sum of the correction factor and the constant 1 is calculated to obtain the correction coefficient; the Euclidean distance between the iterative cluster center and the previous historical iterative cluster center in the current cluster cluster is calculated to obtain the moving distance value.

[0027] Furthermore, the step of obtaining the iterative clustering center of the next round of iterative clustering according to the correction coefficient and the moving distance value includes:

[0028] Calculate the product of the correction coefficient and the moving distance value to obtain an adaptive moving distance value; obtain a line segment with a length of the adaptive moving distance value from the previous historical iteration cluster center as the starting point to the iteration cluster center, and find the training parameter with the closest distance at the end point of the line segment as the iteration cluster center of the next round of iterative clustering.

[0029] The present invention has the following beneficial effects:

[0030] In the present invention, obtaining the preliminary selection degree can determine the suitability of the training parameters as the initial cluster center. Determining the initial cluster center according to the preliminary selection degree can avoid falling into the local optimal solution during initial clustering, and improve the clustering speed. Obtaining the characteristic value of the inter-cluster distance can characterize the distance difference between different cluster clusters, thereby analyzing the training effect of different cluster clusters on the large model; obtaining the density can determine the training effect of the cluster cluster on the large model according to the distribution characteristics of the training parameters in the cluster cluster; obtaining the training quality index can characterize the training effect of the training parameters in the current cluster cluster, thereby determining the cluster clusters that need to be optimized in the future, and improving the accuracy of the clustering results. Obtaining the iterative feature change degree can determine the training effect of the cluster cluster where the iterative cluster center is located on the large model according to the change characteristics of the cluster cluster in the dynamic process of iterative clustering; thereby determining the cluster cluster that needs to be optimized in the future. Obtaining the correction coefficient can determine the mobile expansion range of the iterative cluster center according to the training effect of the cluster cluster on the large model, and obtaining the moving distance value can determine the distance basis for adjusting the iterative cluster center. Finally, the training clusters obtained after improving the iterative clustering centers can reduce the occurrence of local optimal solutions in the clustering process, thereby improving the accuracy of the clustering results, making the training parameters in the large model training process more accurate and improving the accuracy of the large model. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0032] Figure 1 A flow chart of an analysis method for improving the self-learning capability of a large data model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0033] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following is a detailed description of the method for analyzing the improvement of the self-learning ability of a large data model proposed by the present invention, its specific implementation method, structure, characteristics and effects, in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0034] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0035] The following is a detailed description of a specific scheme of an analysis method for improving the self-learning ability of a large data model provided by the present invention in conjunction with the accompanying drawings.

[0036] See also Figure 1 , which shows a flow chart of an analysis method for improving the self-learning ability of a large data model provided by an embodiment of the present invention, the method comprising the following steps:

[0037] Step S1, obtain a set of reduced-dimensional training parameters for large model learning, and construct a clustering space for the training parameters; obtain a preliminary degree in the clustering space based on the neighborhood density characteristics of the training parameters and the distance characteristics from the determined initial clustering centers; and cyclically obtain a preset number of initial clustering centers based on the preliminary degree of the training parameters.

[0038] In an embodiment of the present invention, the implementation scenario is to cluster the training parameter set for large model training, and the large model selects a suitable clustering cluster for training and learning to improve the accuracy of the large model. First, the reduced-dimensional training parameter set for large model learning is obtained, and a clustering space of the training parameters is constructed. Preferably, in an embodiment of the present invention, the steps of obtaining the reduced-dimensional training parameter set and constructing the clustering space of the training parameters include: obtaining a two-dimensional training parameter set by principal component analysis; because the dimension of the training data used for large model training is relatively high, in order to facilitate the use of the K-means algorithm for clustering and improve the analysis speed of the clustering process, the training data is reduced in dimension by principal component analysis to obtain two-dimensional training parameters. The two-dimensional training parameters can reflect the main features of the original training data, and a training parameter set is obtained based on all the reduced-dimensional training parameters; it should be noted that the principal component analysis method belongs to the prior art, and the specific dimensionality reduction steps will not be repeated. In the plane rectangular coordinate system, the horizontal axis and the vertical axis are used to represent the values ​​of two different dimensions of the training parameters, respectively, and the coordinates corresponding to each training parameter are obtained. A clustering space is constructed according to the coordinates corresponding to all training parameters in the plane rectangular coordinate system. Each training parameter in the clustering space corresponds to a coordinate position. Subsequently, the clustering results can be optimized according to the distribution characteristics of the training parameters.

[0039] Due to the large amount of data in the training parameter set, the wide source of training parameters, and the strong randomness of the characteristics of training parameters, the existing K-means clustering algorithm is prone to local optimal solutions, resulting in inaccurate clustering results, unsatisfactory cleaning and classification of training parameter sets, and ultimately causing the large model to lose accuracy after training based on the clustering results; in order to achieve better self-training learning effects for large models, it is necessary to improve the clustering process and reduce the occurrence of local optimal solutions. First, it is necessary to determine the selection of the initial clustering center. The existing K-means algorithm randomly specifies the initial clustering center, but it is easy to fall into the local optimal solution after the initial clustering, and the number of iterative clustering is more, which affects the clustering speed; therefore, in order to avoid the local optimal solution in the initial stage of clustering and to improve the clustering speed, it is necessary to select the initial clustering center. In order to avoid falling into the local optimal solution, the initial clustering centers are as far apart as possible, and the density of the training parameters in the neighborhood of each initial clustering center is as high as possible. Therefore, the preliminary degree is obtained in the clustering space according to the neighborhood density characteristics of the training parameters and the distance characteristics from the determined initial clustering center, and the initial clustering center is determined according to the preliminary degree.

[0040] Preferably, in an embodiment of the present invention, the step of obtaining the preliminary selection degree includes: calculating the average value of the Euclidean distance between all training parameters and the other training parameters closest to them in the clustering space to obtain the neighborhood radius; the neighborhood radius represents the average level of the distance between the training parameters and the other training parameters closest to them in the clustering space. Calculate and normalize the number of other training parameters in a circular area with the training parameter as the center and the neighborhood radius as the radius to obtain the neighborhood concentration of the training parameter; the more the number of other training parameters in the circular area of ​​the training parameter, the greater the neighborhood concentration, which means that the denser the distribution of other training parameters in the neighborhood of the training parameter, the more suitable it is as the initial clustering center. Calculate and normalize the average value of the Euclidean distance between the training parameter and the determined initial clustering center to obtain the clustering deviation of the training parameter; the farther the distance between the training parameter and the determined initial clustering center, the greater the clustering deviation, the greater the difference in the characteristics represented by the training parameter and the determined initial clustering center, the more likely it is to be divided into different clusters, and the more likely it is that the training parameter will be used as the next initial clustering center. The sum of the neighborhood concentration and the cluster deviation is calculated to obtain the preliminary degree of the training parameter. The larger the values ​​of the neighborhood concentration and the cluster deviation of the training parameter, the larger the preliminary degree. The larger the neighborhood density is and the farther the training parameter is from the determined initial cluster center, the more likely it is that the training parameter will be the next initial cluster center.

[0041] Furthermore, a preset number of initial cluster centers can be obtained cyclically according to the preliminary degree of the training parameters, specifically including: selecting the training parameter corresponding to the maximum value of the preliminary degree as the initial cluster center; after obtaining an initial cluster center, recalculating the preliminary degree corresponding to the training parameter of the non-initial cluster center until a preset number of initial cluster centers are obtained. It should be noted that when the first initial cluster center is obtained, the value of the cluster deviation is 0, so the preliminary degree is only related to the neighborhood concentration, and the training parameter with the largest neighborhood density feature is selected as the first initial cluster center. Then recalculate the preliminary degrees of the training parameters of all non-initial cluster centers, and select the training parameter corresponding to the maximum value as the next initial cluster center; repeat the step of obtaining the preliminary degree until a preset number of initial cluster centers are obtained. This preset number is related to the number of cluster centers and cluster clusters. A larger or smaller value will affect the accuracy of the clustering results. Therefore, it is necessary to select the most appropriate value. In the embodiment of the present invention, the number of cluster centers is obtained by the elbow rule. The elbow rule is a method for determining the optimal number of clusters. By changing the number of clusters and calculating the loss function, an inflection point is found, that is, the point where the increase in the number of clusters no longer significantly reduces the loss function. It should be noted that the elbow rule for obtaining the number of cluster centers belongs to the prior art, and the specific calculation steps will not be repeated. The result obtained by the elbow rule is used as the preset number, and the implementer can determine the preset number according to the implementation scenario.

[0042] Step S2, iteratively clustering the training parameters by using the K-means algorithm according to the initial cluster center; obtaining the inter-cluster distance characteristic value according to the distance characteristics between the iterative cluster center and other iterative cluster centers in the current iterative clustering; obtaining the density according to the interval characteristics of the training parameters in the current clustering cluster where the iterative cluster center is located, the quantity difference characteristics of the training parameters in different regions and the area characteristics of the current clustering cluster; obtaining the training quality index of the current clustering cluster according to the inter-cluster distance characteristic value and the density.

[0043] After determining a preset number of initial cluster centers in the clustering space, it is possible to initially avoid falling into a local optimal solution and to improve the clustering speed; then, the training parameters can be iteratively clustered according to the initial cluster centers through the K-means algorithm. It should be noted that the K-means algorithm belongs to the prior art and the specific clustering process will not be described in detail; after obtaining a preset number of initial cluster centers, the algorithm calculates the distance between each object and each cluster center, assigns each object to the cluster center closest to it, recalculates the position of the cluster center according to the existing objects in the cluster, and repeats this step for iterative clustering until termination conditions such as the maximum number of iterations, the minimum cluster center change, or the local minimum sum of squared errors are met.

[0044] During the training and learning process of a large model, the features between the training parameters that can improve the accuracy of the large model are usually highly similar, and it is easier to form clusters with good clustering effects during the clustering process; while the features between the training parameters with poor training effects are usually dissimilar and random, and it is not easy to form clusters with good clustering effects during the clustering process. It is more likely to have a local optimal solution, and even affect the clustering results of clusters with good training effects. Therefore, it is necessary to analyze the features of different clusters during the clustering process to determine the clusters with good training effects and the clusters with poor training effects. Since the characteristics of the training parameters with better training effects and the training parameters with worse training effects are significantly different, there is a significant distance interval between the clustering clusters with better training effects and the clustering clusters with worse training effects in the clustering space; therefore, in the current iterative clustering, the inter-cluster distance characteristic value is obtained according to the distance characteristics between the iterative clustering center and other iterative clustering centers, and the current iterative clustering refers to the latest iterative clustering; preferably, in the embodiment of the present invention, the step of obtaining the inter-cluster distance characteristic value includes: calculating the sum of the Euclidean distances between the iterative clustering center and other iterative clustering centers in the current iterative clustering and normalizing them to obtain the inter-cluster distance characteristic value of the iterative clustering center. The iterative clustering center refers to the new clustering center obtained in each iterative clustering. Since the characteristics between the training parameters with worse training effects are not similar and random, the distances between the corresponding clustering clusters are relatively close; while the characteristics between the training parameters with better training effects are highly similar, and there are significant differences with the characteristics of the training parameters with worse training effects, so the training parameters with better training effects are easy to become a concentrated cluster cluster in the clustering process, and there is a significant distance interval between the clustering cluster and other clustering clusters. Therefore, when the inter-cluster distance characteristic value of the iterative cluster center is larger, the cluster of the iterative cluster center is more likely to be a cluster with better training effect.

[0045] Furthermore, since the features between the training parameters with better training effects are highly similar, the distribution of the training parameters in the corresponding clustering clusters is highly concentrated, while the features between the training parameters with poor training effects are dissimilar and random, and the distribution of the training parameters in the corresponding clustering clusters is relatively discrete; therefore, the density is obtained according to the interval characteristics of the training parameters in the current clustering cluster where the iterative clustering center is located, the quantity difference characteristics of the training parameters in different regions and the area characteristics of the current clustering cluster, and the current clustering cluster is the clustering cluster where the iterative clustering center is located in the current iterative clustering.

[0046] Preferably, in one embodiment of the present invention, the step of obtaining the density includes: calculating the average value of the Euclidean distance between all training parameters in the current cluster and other training parameters that are closest to each other and performing negative correlation mapping to obtain the nearest neighbor feature value of the current cluster; when the distance of the training parameters in the cluster is closer, the smaller the average value of the Euclidean distance, the larger the nearest neighbor feature value, and the more likely the cluster is to be a cluster with a better training effect. Calculate the area of ​​the minimum bounding rectangle of the current cluster in the cluster space and perform negative correlation mapping to obtain the first cluster feature value of the current cluster; when the area of ​​the minimum bounding rectangle is smaller, the larger the first cluster feature value is, which means that the training parameters in the cluster are denser and the features of the training parameters in the cluster are more highly similar. The minimum circumscribed circle of the current cluster cluster is obtained in the cluster space, and a preset second number of concentric circles with the same center as the minimum circumscribed circle are obtained within the minimum circumscribed circle, and the radius difference between adjacent concentric circles is the same; in the embodiment of the present invention, the preset second number is 5, and there are a total of 5 circles with the same center including the minimum circumscribed circle. If the radius of the minimum circumscribed circle is 10, the radii of the other 4 concentric circles are 8, 6, 4, and 2 respectively, so that the radius difference between adjacent concentric circles is the same. The implementer can determine the preset second number and radius difference according to the implementation scenario. Get the rings between adjacent concentric circles, for example, there is a ring with a width of 2 between concentric circles with radii of 6 and 4, and there is a ring with a width of 2 between concentric circles with radii of 8 and 6; calculate the sum of the number differences of the training parameters in two adjacent rings from the center of the concentric circle to the outside and normalize them to obtain the second clustering eigenvalue of the current clustering cluster; because the training parameters in the clustering cluster with better training effect are highly concentrated, the number of training parameters closer to the center area is more, so the number of training parameters in the ring closer to the center is more, and then when the sum of the number differences of the training parameters in the two adjacent rings is larger, the second clustering eigenvalue is larger, and the clustering cluster is more likely to be a clustering cluster with better training effect. Calculate the product of the nearest neighbor eigenvalue, the first clustering eigenvalue and the second clustering eigenvalue to obtain the density of the current clustering cluster; when the density is larger, the current clustering cluster is more likely to be a clustering cluster with better training effect, and when the density is smaller, the current clustering cluster is more likely to be a clustering cluster with poor training effect.

[0047] Furthermore, the training quality index of the current cluster cluster can be obtained according to the inter-cluster distance characteristic value and the density. Preferably, in the implementation of the present invention, the step of obtaining the training quality index includes: calculating the average value of the inter-cluster distance characteristic value and the density, and obtaining the training quality index of the current cluster cluster. The larger the training quality index, the better the training effect of the training parameters in the current cluster cluster on the large model. The smaller the training quality index, the worse the training effect of the training parameters in the current cluster cluster on the large model. The formula for obtaining the training quality index includes:

[0048]

[0049] In the formula, B represents the training quality index of the current cluster, represents the normalization function, represents an exponential function with a natural constant as the base, through which the function value is negatively correlated; N represents the number of other iterative clustering centers, Indicates the Euclidean distance between the iterative cluster center corresponding to the current cluster and the nth other iterative cluster center. represents the inter-cluster distance feature value; I represents the number of training parameters in the current clustering cluster, represents the Euclidean distance between the ith training parameter and the other nearest training parameters, S represents the area of ​​the minimum circumscribed rectangle of the current cluster, and H represents the sum of the differences in the number of training parameters in two adjacent circles calculated from the center of the concentric circles outward; represents the nearest neighbor eigenvalue, represents the first clustered eigenvalue, represents the second aggregated eigenvalue; Indicates the degree of density.

[0050] Step S3, obtaining the iteration characteristic variation degree according to the area difference characteristics between the historical clustering clusters where the iteration clustering center is located in the historical iteration clustering, the quantity difference characteristics of the training parameters and the difference characteristics of the training quality index; obtaining the correction coefficient according to the iteration characteristic variation degree of the iteration clustering center and the training quality index of the current clustering cluster, obtaining the moving distance value according to the distance characteristics between the iteration clustering center and the historical iteration clustering center; obtaining the iteration clustering center of the next round of iteration clustering according to the correction coefficient and the moving distance value.

[0051] Since the iterative clustering of K-means is a dynamic process, we cannot only analyze the characteristics of the fixed clusters of a certain iterative clustering, but also need to analyze according to the change characteristics of the clusters; the characteristics of the training parameters with better training effects are highly similar, and the characteristics of the training parameters with poor training effects are dissimilar and have strong randomness; therefore, in the dynamic iterative clustering process, the clusters composed of training parameters with better training effects are relatively stable, and there will not be a large change in the training parameters within the cluster, while the clusters composed of training parameters with poor training effects have stronger change characteristics. With the iteration of clustering, the training parameters within the cluster are prone to large changes. Therefore, the iterative feature change degree is obtained based on the area difference characteristics between the historical clusters where the iterative cluster center is located in the historical iterative clustering, the number difference characteristics of the training parameters, and the difference characteristics of the training quality index.

[0052] Preferably, in an embodiment of the present invention, the step of obtaining the degree of change of the iteration feature includes: taking any historical iteration cluster of the iteration cluster center and the adjacent next historical iteration cluster as adjacent historical iteration clusters; if 10 iterations have been performed up to the current iteration cluster, the current iteration cluster is the 10th iteration, for example, the 7th iteration cluster and the 8th iteration cluster can be called adjacent historical iteration clusters. Calculate the absolute value of the difference in the area of ​​the minimum circumscribed rectangle of the two historical cluster clusters where the iteration cluster center is located in the adjacent historical iteration clusters and normalize it to obtain the first iteration characteristic value of the arbitrary historical iteration cluster; for example, the greater the difference in the area of ​​the historical cluster clusters where the iteration cluster center is located in the 7th and 8th iteration clusters, the greater the difference in the distribution of the training parameters in the two historical cluster clusters, the greater the change in the cluster cluster, and the greater the first iteration characteristic value; conversely, when the first iteration characteristic value is smaller, it means that the area change of the two historical cluster clusters where the iteration cluster center is located is smaller, and the training parameters in the two historical cluster clusters are more beneficial to the training of the large model. The absolute value of the difference in the number of training parameters of the two historical clustering clusters where the iterative cluster center is located in adjacent historical iterative clusters is calculated and normalized to obtain the second iteration characteristic value of any historical iterative cluster; the greater the difference in the number of training parameters in the two historical clustering clusters, the larger the second iteration characteristic value, which means that the historical clustering cluster where the iterative cluster center is located has undergone a more obvious change in the iterative clustering, and the training effect of the training parameters in the historical clustering cluster is worse. The absolute value of the difference in the training quality index of the two historical clustering clusters where the iterative cluster center is located in adjacent historical iterative clusters is calculated to obtain the third iteration characteristic value of any historical iterative cluster; the greater the change in the training quality index, the larger the third iteration characteristic value, which means that the training parameters in the two historical clustering clusters have undergone a significant change, and the training effect of the training parameters in the historical clustering cluster is worse.

[0053] Furthermore, the sum of the iteration orders of all historical iteration clusters is calculated to obtain the iteration order benchmark; the ratio of the iteration order of any historical iteration cluster to the iteration order benchmark is calculated to obtain the iteration weight of the arbitrary historical iteration cluster; since the changes of each cluster cluster of the initial iteration cluster in the iterative clustering process are relatively large, the obtained result should be given a smaller weight, so the larger the iteration order and the closer it is to the current iterative clustering process, the larger the iteration weight of the historical iteration cluster. The product of the iteration weight, the first iteration eigenvalue, the second iteration eigenvalue and the third iteration eigenvalue is calculated to obtain the iteration change characteristic factor of the arbitrary historical iteration cluster; the sum of the iteration change characteristic factors of all historical iteration clusters of the iteration cluster center is calculated to obtain the iteration characteristic change degree of the iteration cluster center. When the iterative change characteristic factor corresponding to each historical iterative clustering of the iterative clustering center is smaller, the iterative characteristic change degree is smaller, the change characteristic of the historical clustering cluster where the iterative clustering center is located is smaller, and the training effect of the training parameters in the cluster is better; conversely, when the iterative change characteristic factor corresponding to each historical iterative clustering of the iterative clustering center is larger, the iterative characteristic change degree is larger, the change characteristic of the historical clustering cluster where the iterative clustering center is located is larger, and the training effect of the training parameters in the cluster is worse. The formula for obtaining the iterative characteristic change degree includes:

[0054]

[0055] In the formula, R represents the iterative characteristic change degree of the iterative cluster center. Indicates the number of historical iterative clustering up to the current iterative clustering. represents the iteration weight of the xth historical iteration clustering, represents the first iteration eigenvalue of the xth historical iteration clustering, represents the second iteration eigenvalue of the xth historical iteration clustering, represents the third iteration eigenvalue of the xth historical iteration clustering; Represents the iterative change characteristic factor of the x-th historical iteration clustering.

[0056] After obtaining the training quality index of the current clustering cluster where the iterative clustering center is located, and the iterative feature change degree of the iterative clustering center in the historical iterative clustering; when the training quality index is larger and the iterative feature change degree is smaller, it means that the training parameters in the current clustering cluster where the iterative clustering center is located and the historical clustering cluster have a better training effect on the large model. Since the characteristics of the training parameters within the cluster are highly similar, the clustering clusters are relatively stable during the iterative clustering process and the clustering results are better. When the training quality index is smaller and the iteration feature variation is larger, it means that the training parameters in the current clustering cluster where the iteration clustering center is located can have a worse training effect on the large model. Since the characteristics of such training parameters are relatively different and highly random, the clustering clusters change greatly during the iteration clustering process, and are prone to fall into the situation of local optimal solutions during the iteration clustering, resulting in low final clustering accuracy, and even causing some training parameters with poor training effects to be included in the clustering clusters with better training effects. Therefore, in order to reduce the occurrence of local optimal solutions, it is necessary to make the clustering clusters with poor training effects expand the selection range of the iteration clustering center as much as possible during the iteration clustering process, so that more training parameters with poor training effects are included in the corresponding clustering clusters. Therefore, the correction coefficient is obtained according to the iteration feature variation of the iteration clustering center and the training quality index of the current clustering cluster, and the moving distance value is obtained according to the distance feature between the iteration clustering center and the historical iteration clustering center.

[0057] Preferably, in an embodiment of the present invention, the step of obtaining the correction coefficient and the moving distance value includes: mapping the training quality index in a negative correlation to obtain the clustering effect characteristic value of the current cluster cluster, when the training quality index is smaller, the clustering effect characteristic value is larger, and the training parameters in the current cluster cluster belong to the type with poor training effect. Calculate the average value of the clustering effect characteristic value and the iterative characteristic variation degree to obtain the correction factor; when the clustering effect characteristic value and the iterative characteristic variation degree are larger, the correction factor is larger, which means that the training parameters in the current cluster cluster and the historical cluster cluster where the iterative cluster center is located belong to the type with poor training effect; and then it is more necessary to expand the selection range of the iterative cluster center so that more training parameters with poor training effect are included therein, thereby avoiding confusion of training parameters in cluster clusters with poor training effect and good training effect. Calculate the sum of the correction factor and the constant 1 to obtain the correction coefficient, so that the value range of the correction coefficient is 1 to 2; calculate the Euclidean distance between the iterative cluster center and the previous historical iterative cluster center in the current cluster cluster to obtain the moving distance value. The moving distance value represents the distance difference after the iterative clustering center changes as the new training parameter is added to the previous historical iterative clustering cluster during the iterative clustering process; the larger the moving distance value, the more obvious the change of the iterative clustering center, and the closer the change direction is to the new training parameter, which means that the new training parameter is similar to the training parameter in the current clustering cluster. In order to avoid the local optimal solution, the moving distance needs to be expanded, and the iterative clustering center point needs to be closer to the new training parameter, so that other similar training parameters near the new training parameter are further classified into the clustering cluster; therefore, the iterative clustering center of the next round of iterative clustering is obtained according to the correction coefficient and the moving distance value.

[0058] Preferably, in an embodiment of the present invention, the step of obtaining the iterative clustering center of the next round of iterative clustering includes: calculating the product of the correction coefficient and the moving distance value to obtain the adaptive moving distance value; obtaining a line segment with a length of the adaptive moving distance value from the previous historical iterative clustering center as the starting point to the iterative clustering center, and finding the training parameter with the closest distance at the end point of the line segment as the iterative clustering center of the next round of iterative clustering. When the correction coefficient of the iterative clustering center in the current iterative clustering is larger, the adaptive moving distance value is larger, and the iterative clustering center of the next round of iterative clustering is closer to the new training parameter added in the current iterative clustering, thereby avoiding falling into a local optimal solution and improving clustering accuracy. When the correction coefficient of the iterative clustering center in the current iterative clustering is smaller, it means that the training effect of the training parameters in the corresponding current clustering cluster is better, the clustering cluster is relatively stable, and there is no need to expand the clustering range, so as to avoid classifying training parameters with dissimilar features into it, thereby improving clustering accuracy.

[0059] Step S4, completing clustering according to the iterative clustering centers obtained in the loop to obtain different training clusters; and training and learning the large model according to the training clusters.

[0060] Clustering is completed according to the iterative clustering center obtained in the cycle to obtain different training clusters. After each round of iterative clustering is completed, the iterative clustering center of the next round of iterative clustering is obtained according to the moving distance of the iterative clustering center and the characteristics of the historical clustering cluster and the current clustering cluster where the iterative clustering center is located. The iterative clustering steps are repeated until the clustering results converge and different training clusters are obtained. Finally, the training clusters with better training effects are put into the self-training learning of the large model. By selecting the iterative clustering center in the clustering process, it is avoided to fall into the situation of local optimal solution, and the accuracy of the clustering results is improved, so that the accuracy of the large model after training is higher.

[0061] In summary, the embodiment of the present invention provides an analysis method for improving the self-learning ability of a large data model; the initial cluster center is obtained and iterative clustering is performed according to the neighborhood density characteristics of the training parameters and the distance characteristics from the determined initial cluster center; in the current iterative clustering, the training quality index is obtained according to the distance characteristics between the iterative cluster center and other iterative cluster centers, the area characteristics of the current cluster cluster, and the distribution characteristics of the training parameters within the cluster; the iterative feature variation degree is obtained according to the difference characteristics of the area, training parameters, and training quality index between the historical cluster clusters where the iterative cluster center is located; the correction coefficient is obtained according to the iterative feature variation degree and the training quality index. The present invention obtains the next round of iterative cluster centers and the final training cluster according to the correction coefficient and the moving distance value; the large model is trained and learned according to the training cluster to improve the accuracy of clustering and large model training.

[0062] It should be noted that the sequence of the above embodiments of the present invention is only for description and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0063] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

Claims

1. A method for improving the self-learning ability of a large data model, applicable to the fields of natural language processing, computer vision or speech recognition, characterized in that: The method comprises the following steps: Obtain a set of training parameters after dimensionality reduction for large model learning, and construct a clustering space of the training parameters; obtain a preliminary degree in the clustering space according to the neighborhood density characteristics of the training parameters and the distance characteristics from the determined initial clustering center; and cyclically obtain a preset number of initial clustering centers according to the preliminary degree of the training parameters; Iteratively clustering the training parameters using the K-means algorithm according to the initial cluster center; obtaining an inter-cluster distance characteristic value according to the distance characteristics between the iterative cluster center and other iterative cluster centers in the current iterative clustering; obtaining a density according to the interval characteristics of the training parameters in the current clustering cluster where the iterative cluster center is located, the quantity difference characteristics of the training parameters in different regions, and the area characteristics of the current clustering cluster; obtaining a training quality index of the current clustering cluster according to the inter-cluster distance characteristic value and the density; Obtaining the iteration characteristic variation degree according to the area difference characteristics between the historical clustering clusters where the iteration clustering center is located in the historical iteration clustering, the quantity difference characteristics of the training parameters and the difference characteristics of the training quality index; obtaining the correction coefficient according to the iteration characteristic variation degree of the iteration clustering center and the training quality index of the current clustering cluster, and obtaining the moving distance value according to the distance characteristics between the iteration clustering center and the historical iteration clustering center; obtaining the iteration clustering center of the next round of iteration clustering according to the correction coefficient and the moving distance value; Clustering is completed according to the iterative clustering centers obtained in a loop to obtain different training clusters; and the large model is trained and learned according to the training clusters; The step of obtaining the density according to the interval characteristics of the training parameters in the current cluster where the iterative cluster center is located, the quantity difference characteristics of the training parameters in different regions and the area characteristics of the current cluster includes: Calculate the average value of the Euclidean distance between all training parameters in the current cluster and other training parameters that are closest to them and perform negative correlation mapping to obtain the nearest neighbor eigenvalue of the current cluster; calculate the area of ​​the minimum circumscribed rectangle of the current cluster in the cluster space and perform negative correlation mapping to obtain the first clustering eigenvalue of the current cluster; obtain the minimum circumscribed circle of the current cluster in the cluster space, obtain a preset second number of concentric circles with the same center as the minimum circumscribed circle within the minimum circumscribed circle, and the radius difference between adjacent concentric circles is the same; obtain the circular rings between adjacent concentric circles, calculate the sum of the quantity difference of the training parameters in two adjacent circular rings from the center of the concentric circle to the outside direction and normalize them to obtain the second clustering eigenvalue of the current cluster; Calculate the product of the nearest neighbor eigenvalue, the first cluster eigenvalue and the second cluster eigenvalue to obtain the density of the current cluster; The step of obtaining the iteration feature variation degree according to the area difference characteristics between the historical clustering clusters where the iteration clustering center is located in the historical iteration clustering, the quantity difference characteristics of the training parameters and the difference characteristics of the training quality index comprises: Any historical iteration cluster of the iteration cluster center and the adjacent next historical iteration cluster are taken as adjacent historical iteration clusters; the absolute value of the difference in the area of ​​the minimum circumscribed rectangle of the two historical cluster clusters where the iteration cluster center is located in the adjacent historical iteration cluster is calculated and normalized to obtain the first iteration characteristic value of the arbitrary historical iteration cluster; the absolute value of the difference in the number of training parameters of the two historical cluster clusters where the iteration cluster center is located in the adjacent historical iteration cluster is calculated and normalized to obtain the second iteration characteristic value of the arbitrary historical iteration cluster; the absolute value of the difference in the training quality index of the two historical cluster clusters where the iteration cluster center is located in the adjacent historical iteration cluster is calculated to obtain the third iteration characteristic value of the arbitrary historical iteration cluster; Calculate the sum of the iteration orders of all historical iteration clusters to obtain an iteration order benchmark; calculate the ratio of the iteration order of any historical iteration cluster to the iteration order benchmark to obtain the iteration weight of the arbitrary historical iteration cluster; calculate the product of the iteration weight, the first iteration eigenvalue, the second iteration eigenvalue and the third iteration eigenvalue to obtain the iteration change characteristic factor of the arbitrary historical iteration cluster; calculate the sum of the iteration change characteristic factors of all historical iteration clusters of the iteration cluster center to obtain the iteration characteristic change degree of the iteration cluster center; The step of obtaining a correction coefficient according to the iterative characteristic change degree of the iterative cluster center and the training quality index of the current cluster cluster, and obtaining a moving distance value according to the distance characteristic between the iterative cluster center and the historical iterative cluster center comprises: Mapping the training quality index negatively correlated to obtain the clustering effect characteristic value of the current cluster cluster, calculating the average value of the clustering effect characteristic value and the iterative characteristic change degree to obtain the correction factor; calculating the sum of the correction factor and a constant 1 to obtain the correction coefficient; calculating the Euclidean distance between the iterative cluster center and the previous historical iterative cluster center in the current cluster cluster to obtain the moving distance value; The step of obtaining the iterative clustering center of the next round of iterative clustering according to the correction coefficient and the moving distance value comprises: Calculate the product of the correction coefficient and the moving distance value to obtain an adaptive moving distance value; obtain a line segment with a length of the adaptive moving distance value from the previous historical iteration cluster center as the starting point to the iteration cluster center, and find the training parameter with the closest distance at the end point of the line segment as the iteration cluster center of the next round of iterative clustering.

2. According to claim 1, a method for analyzing the improvement of the self-learning ability of a large data model is characterized in that: The step of obtaining a set of training parameters after dimension reduction for large model learning and constructing a clustering space of training parameters comprises: A two-dimensional set of training parameters is obtained through principal component analysis; the values ​​of two different dimensions of the training parameters are represented by the horizontal axis and the vertical axis in a plane rectangular coordinate system, the coordinates corresponding to the training parameters are obtained, and a clustering space is constructed according to the coordinates corresponding to all training parameters in the plane rectangular coordinate system.

3. The method for improving the self-learning ability of a large data model according to claim 1 is characterized in that: The step of obtaining the preliminary degree in the cluster space according to the neighborhood density characteristics of the training parameters and the distance characteristics from the determined initial cluster center includes: In the clustering space, the average value of the Euclidean distances between all training parameters and other training parameters closest to them is calculated to obtain a neighborhood radius; the number of other training parameters in a circular area with the training parameter as the center and the neighborhood radius as the radius is calculated and normalized to obtain a neighborhood concentration of the training parameter; the average value of the Euclidean distances between the training parameter and the determined initial clustering center is calculated and normalized to obtain a clustering deviation of the training parameter; the sum of the neighborhood concentration and the clustering deviation is calculated to obtain a preliminary degree of the training parameter.

4. The method for improving the self-learning ability of a large data model according to claim 1 is characterized in that: The step of cyclically acquiring a preset number of initial cluster centers according to the initial selection degree of the training parameters comprises: The training parameter corresponding to the maximum value of the preliminary selection degree is selected as the initial cluster center; after obtaining an initial cluster center, the preliminary selection degree corresponding to the training parameter of the non-initial cluster center is recalculated until a preset number of initial cluster centers are obtained.

5. The method for improving the self-learning ability of a large data model according to claim 1 is characterized in that: The step of obtaining the inter-cluster distance characteristic value according to the distance characteristics between the iterative cluster center and other iterative cluster centers in the current iterative clustering comprises: The sum of the Euclidean distances between the iterative cluster center and other iterative cluster centers is calculated and normalized to obtain the inter-cluster distance characteristic value of the iterative cluster center.

6. The method for analyzing the improvement of the self-learning ability of a large data model according to claim 1 is characterized in that: The step of obtaining the training quality index of the current clustering cluster according to the inter-cluster distance characteristic value and the density comprises: The average value of the inter-cluster distance characteristic value and the density is calculated to obtain the training quality index of the current clustering cluster.