Relation cluster database optimization method based on multi-modal learning

By calculating the Laplace density gradient of data points and building an association matrix, embedding the data points into low-dimensional space, and designing a dynamic index structure, the problem of degradation of query performance in traditional databases in high-dimensional multi-modal data processing is solved, and the problem of degradation of query performance in traditional databases is realized. Efficient and accurate query and data processing.

CN119046316AActive Publication Date: 2024-11-29HENGHUI XINDA TECH CO LTD

Patent Information

Application Number
CN202411533349.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-10-31
Publication Date
2024-11-29
Estimated Expiration
2044-10-31

AI Technical Summary

Technical Problem

Traditional databases have deteriorated query performance when processing high-dimensional and multi-modal data. Traditional indexing methods such as B-tree and R-tree have reduced efficiency in high-dimensional space, and cannot effectively handle the complex relationship between multi-modal data points, resulting in query performance. and accuracy is limited.

Method used

By calculating the Laplace density gradient of data points, the correlation matrix and weighted undirected graph are constructed, the generalized Laplace matrix is ​​calculated, and the problem is optimized to embed data points into low-dimensional space, and dynamic index structures such as K-d trees and spherical trees are designed. Or LSH, optimize the query process.

Benefits of technology

It significantly reduces the dimensional complexity of the data, improves the efficiency and accuracy of database query, and the dynamic index structure can quickly respond to data changes, adapt to the complex relationships of multimodal data, and improves the flexibility and robustness of the system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119046316A_ABST
    Figure CN119046316A_ABST
Patent Text Reader

Abstract

The invention discloses a relation cluster database optimization method based on multi-modal learning, and relates to the technical field of data processing, and the method comprises the steps: 1, in a relation cluster database, taking each piece of stored data as a data point, calculating the Laplacian density gradient of each data point, and calculating the Laplacian density gradient of each data point according to the Laplacian density gradient of each data point; constructing an incidence matrix used for representing the correlation degree between the data points; step 2, constructing a weighted undirected graph based on the incidence matrix, and calculating a generalized Laplacian matrix of the weighted undirected graph to obtain embedded representation; 3, based on the low-dimensional embedding representation of the data points, defining an objective function of query optimization; and the query can be efficiently carried out in the embedding space. According to the method, the accuracy and depth of data analysis are improved, and the query efficiency of the database and the system performance are also remarkably improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a relational cluster database optimization method based on multimodal learning. Background Art

[0002] With the rapid development of information technology, database technology plays a vital role in modern information management systems. Especially driven by big data, artificial intelligence and multimodal learning technologies, the complexity and diversity of data have increased significantly, and traditional database optimization methods have gradually revealed their limitations in processing these complex data. Multimodal data usually comes from different sensors, platforms or information sources, and its characteristics are diversity, high dimensionality and heterogeneity. This makes the existing relational database face huge technical challenges when facing these complex data clusters.

[0003] At present, mainstream database query optimization technologies mainly rely on traditional indexing methods, such as B-trees and hash indexes. These methods work well when processing single-modality or low-dimensional data, but often perform poorly when processing high-dimensional and multimodal data. Especially in the context of multimodal learning, there are complex associations between data of different modalities, and traditional indexes cannot effectively handle these relationships, resulting in a significant decrease in query performance. In addition, as the amount of data increases, the query time grows exponentially, further limiting the application scenarios of the database. A common existing technology is a query optimization method based on index structures such as B-trees or R-trees. These methods perform relatively well in low-dimensional spaces and can quickly perform operations such as interval queries, point queries, and range queries. However, structures such as B-trees or R-trees will encounter the so-called "curse of dimensionality" problem in high-dimensional spaces, that is, as the dimension increases, the efficiency of the index structure decreases rapidly. This is because the distance difference between high-dimensional data points gradually becomes smaller, resulting in severe limitations on query performance and accuracy. In addition, in multimodal learning scenarios, the features of data points of different modalities are often distributed in different feature spaces, and it is difficult to fully capture the complex relationships between data points using a single index structure. Summary of the invention

[0004] The purpose of the present invention is to provide a relational cluster database optimization method based on multimodal learning, which improves the accuracy and depth of data analysis, and also significantly improves the query efficiency and system performance of the database. While ensuring the integrity of data structure and feature information, the method reduces the computational complexity and provides an effective solution for relational cluster database optimization in multimodal learning.

[0005] In order to solve the above technical problems, the present invention provides a relational cluster database optimization method based on multimodal learning, the method comprising: Step 1: In the relational cluster database, each stored data is regarded as a data point, and the Laplace density gradient of each data point is calculated to describe the density distribution and its change of the data point in its local area; according to the Laplace density gradient of each data point, an association matrix is ​​constructed to represent the degree of association between data points; Step 2: Based on the incidence matrix, construct a weighted undirected graph, calculate the generalized Laplace matrix of the weighted undirected graph, and construct an optimization problem based on the generalized Laplace matrix. By solving the optimization problem, find an embedding matrix, map the data points to a low-dimensional space, and obtain an embedded representation; Step 3: Based on the low-dimensional embedding representation of the data points, define the objective function of query optimization; obtain the update rule of the optimal embedding representation by taking the partial derivative of the objective function with respect to the embedding representation and setting it equal to zero; use the optimal embedding representation to design a dynamic index structure so that queries can be performed efficiently in the embedding space.

[0006] Furthermore, step 1 specifically includes: determining each data point Neighborhood range , thereby determining the data points Data points in the neighborhood of ; Calculate for each data point The local density gradient ; For each data point , through the Gaussian kernel function to the data points in the neighborhood The distances are weighted summed to obtain the data point pairs in the neighborhood. The influence of; combining the local density gradient and the neighborhood contribution, calculate the data point Laplace density gradient ; Local density gradient Reflects the data points Relative to the density change of its surrounding neighborhood, if ,illustrate The density of is higher than that of its neighborhood and is the cluster center. ,illustrate The density is low and is in a sparse area. According to the Laplace density gradient of each data point, an association matrix is ​​constructed to represent the degree of association between data points.

[0007] Furthermore, the Laplace density gradient The calculation formula is as follows: ; in, It is the data point The neighborhood density of is summed up, Represents data points The neighborhood of Determine, indicating that the radius is All data points in the region The neighborhood set ; Represents data points With neighboring data points The distance between With The relative position of right the extent of the impact; is a Gaussian kernel function used to transform the distance Converted to a weight, the smaller the distance, the greater the weight, indicating right The greater the impact, is the scale parameter, which controls the degree of smoothing within the neighborhood; is a density adjustment parameter used to control the data points in the neighborhood right The greater the impact The value represents the data point pair in the neighborhood The greater the impact, the smaller the Then it represents the data point pair of the neighborhood The smaller the impact.

[0008] Furthermore, the correlation matrix No. Row, No. Elements of a column Defined as: ; in, For data points and data points The association weight between them, the larger the value, the stronger the association; and The data points and data points Laplace density gradient of ; is the first-order L1 norm, is the second-order L1 norm; is a preset amplification parameter that controls the influence of density gradient direction similarity on the association weight; for and The Euclidean distance between is the preset gradient change rate smoothing parameter, which controls the influence of density gradient change rate difference on the association weight; and The data points and data points The gradient of the Laplace density gradient.

[0009] Furthermore, in step 2, a weighted undirected graph is constructed based on the association matrix , node data point set Contains data points corresponding to all data ; is the total number of data points; the edge set According to the correlation matrix Determine that the edge weight is , which represents the data point and The strength of the association between them; the degree matrix of a weighted undirected graph is a diagonal matrix whose elements are defined as: ; Represents node data points The degree of is the sum of the weights of the edges connected to it; define the generalized Laplace matrix for: ; in, is the standard graph Laplacian matrix; is the square of the Laplacian matrix, capturing the second-order neighborhood information; is the preset first regularization parameter; is the inverse square root of the degree matrix.

[0010] Furthermore, the process of calculating the low-dimensional embedding representation of the data point in step 2 specifically includes: constructing an optimization problem whose objective function is for: ; The objective function aims to find an embedding matrix , which can preserve the local structure and keep consistent with the original features; is the low-dimensional embedding matrix of the data points, is the embedding dimension; the relational cluster database is regarded as a matrix, and the original feature matrix of the relational cluster database is obtained through feature extraction , is the dimension of the original feature; is the trace operation of the matrix, i.e. the sum of the diagonal elements; is the first-order F-norm; is the preset second regularization parameter, which controls the degree of deviation between the embedding result and the original feature; is the identity matrix; the constraints are: ; Ensure that the embedded data is orthogonal in the new space to prevent degenerate embedding results.

[0011] Furthermore, in step 2, in order to solve the optimization problem, a Lagrangian function is constructed for: ; in, is the Lagrange multiplier matrix; for Take the derivative and set it to zero: ; The solution is obtained by simplifying the following formula: ; Among them, due to ,get ; Using the above equation, solve the following generalized eigenvalue problem: ; Solving this equation, we get the embedding matrix , where each element It is a data point A low-dimensional embedding representation of .

[0012] Furthermore, in step 3, based on the low-dimensional embedding representation of the data points, the objective function of query optimization is defined for: ; in, Represents the set of data point pairs involved in the query workload; For data point pairs The query weights of , which reflect the frequency with which they appear in the query; is the preset third regularization parameter.

[0013] Furthermore, in step 3, the objective function about Taking the partial derivative and setting it equal to zero, we get ; In this way, we can get the objective function of query optimization The first item The partial derivative of is: ; in, To query the relevant Laplacian matrix, it is defined as: ; in, and and Same, both are subscript indexes; the objective function of query optimization The second item The partial derivative of ; Item 3 The partial derivative of ; The objective function of query optimization Adding the first, second, and third partial derivatives of gives the following equation: ; After simplifying the equation, we get: ; Finding the optimal embedding representation for: .

[0014] Furthermore, in step 3, the optimal embedding representation is used The process of designing a dynamic index structure includes: since the embedding space is low-dimensional, use Kd tree, ball tree or LSH as the index structure; represent the embedding Each data point in Insert into the index structure; when the data in the database changes, update the embedded representation accordingly and index structures.

[0015] The multimodal learning-based relational cluster database optimization method of the present invention has the following beneficial effects: The present invention significantly reduces the dimensional complexity of data by embedding high-dimensional data into low-dimensional space. Multimodal data often comes from different modes or information sources, and the data points of each mode may have different characteristic dimensions. Directly processing these high-dimensional data will lead to the "curse of dimensionality" problem, that is, the increase in dimensions leads to a significant decrease in the efficiency of indexing and querying. Through the optimization step, the data points are mapped from the high-dimensional space to the low-dimensional embedding space, which effectively reduces the dimension while retaining the structure and characteristic information of the data points, and significantly improves the speed and efficiency of database queries. Low-dimensional embedding not only reduces the computational complexity of data processing, but also lays the foundation for subsequent query optimization, enabling the system to retrieve data in a more efficient manner.

[0016] Secondly, the present invention designs a dynamic index structure in the embedding space, which further improves the response speed and flexibility of the query. Traditional index structures such as B-tree, R-tree and LSH show poor performance when processing high-dimensional data, especially when facing multimodal data, it is difficult to effectively handle the complex relationship between data points. By adopting index structures suitable for low-dimensional space such as Kd tree, ball tree or local sensitive hashing (LSH), the present invention can efficiently organize embedded data points and support fast nearest neighbor queries or similarity queries. A significant advantage of the dynamic index structure is that it can be quickly updated as the data in the database changes. As the data in the database is inserted, deleted or updated, the embedded representation and index structure can be automatically adjusted to ensure the accuracy and timeliness of the query results. This dynamic indexing mechanism enables the database to adapt to real-time changes in data and improves the flexibility and robustness of the system.

[0017] Third, the present invention improves the accuracy and relevance of queries by incorporating the data point pair relationships in the query workload into the optimization objective function. In actual database applications, some data point pairs frequently appear in queries or have high relevance. Traditional index structures are usually unable to dynamically identify and process these high-frequency data point pairs, thereby ignoring important relationships between data points during the query process. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0019] Figure 1 A schematic diagram of a method flow of a relational cluster database optimization method based on multimodal learning provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] Example 1: Reference Figure 1 , a relational cluster database optimization method based on multimodal learning, the method comprising: Step 1: In the relational cluster database, each stored data is regarded as a data point, and the Laplace density gradient of each data point is calculated to describe the density distribution and its change of the data point in its local area; according to the Laplace density gradient of each data point, an association matrix is ​​constructed to represent the degree of association between data points; In the context of multimodal learning, data usually comes from different modalities, such as text, images, or time series, and the associations between these modal data are complex and multidimensional. Therefore, when processing these data, simple Euclidean distance or other traditional similarity metrics cannot effectively capture the complex relationships between data. Therefore, a density gradient method based on the Laplace operator is introduced, which not only considers the local density distribution of data points, but also captures their complex relationships in high-dimensional space. In this way, the density gradient can effectively distinguish points in high-density areas from points located at the edge or abnormal positions, thereby more accurately representing the mutual influence between data points. The process of calculating the Laplace density gradient can be regarded as a density evaluation of the local neighborhood of each data point. The local neighborhood here refers to other points that are close to the data point in high-dimensional space, which together constitute the local environment of the data point. In traditional density estimation methods, it is often assumed that the distribution of data points in their neighborhood is uniform, but this assumption does not hold in multimodal data. Therefore, the Laplace density gradient can more accurately reflect the actual distribution of data points by considering the density change rate of data points in the neighborhood. In this way, it can be determined whether a data point is in a dense data cluster, at the edge of the data cluster, or even near an outlier. Each element of the association matrix represents the similarity or correlation between two data points in the database. Compared with the traditional similarity measurement method, the uniqueness of the association matrix in the present invention is that it is not only based on the distance information of the data points, but also combines the local density information of the data points. Specifically, if the density gradients of two data points are similar and their distance in the high-dimensional space is close, then their similarity value in the association matrix will be higher. On the contrary, if the two data points are close, but the density gradient difference is large, then their correlation will be low. This density gradient-based association matrix can more accurately reflect the complex potential relationships in multimodal data because it not only considers the spatial position, but also the local data structure. In this way, the Laplace density gradient not only helps to understand the distribution of data points in high-dimensional space, but also provides a basis for building a more accurate association matrix. This process effectively solves the limitations of traditional methods in processing multimodal data, so that the relationship between data points can be better captured and quantified. The construction of the association matrix is ​​the basis for subsequent weighted undirected graphs and optimization problems. Through this matrix, complex multidimensional data relationships can be expressed in the form of a graph, laying a solid foundation for further low-dimensional embedding and database optimization.

[0022] Step 2: Based on the incidence matrix, construct a weighted undirected graph, calculate the generalized Laplace matrix of the weighted undirected graph, and construct an optimization problem based on the generalized Laplace matrix. By solving the optimization problem, find an embedding matrix, map the data points to a low-dimensional space, and obtain an embedded representation; The association matrix is ​​an important tool for reflecting the relationship between data points in a database. Based on this matrix, a weighted undirected graph can be constructed. In this weighted undirected graph, the nodes of the graph correspond to the data points in the database, and the edges between the nodes represent the relationship or similarity between the data points. The weight of the edge is determined by the similarity information in the association matrix. The higher the weight, the closer the relationship between the two nodes. In this way, the weighted undirected graph can intuitively show the mutual connection between data points, which is particularly suitable for processing multimodal data, because multimodal data comes from different sources or channels, and its inherent association has high complexity. The weighted undirected graph not only reveals the relationship between data points in their respective modes, but also can uniformly process multidimensional data from multiple modes, so that it can be integrated into a graph structure, which is convenient for subsequent optimization analysis. After constructing the weighted undirected graph, the generalized Laplace matrix of the graph needs to be calculated. The generalized Laplace matrix is ​​an important concept in graph theory, which can reflect the structural properties of the graph, especially the connectivity between nodes. Through the generalized Laplace matrix, the global relationship between data points and the local similarity can be effectively analyzed. The generalized Laplace matrix is ​​not only suitable for processing traditional homogeneous data, but also can process multimodal data, which is one of the innovations of the present invention. Unlike ordinary Laplace matrices, generalized Laplace matrices can process heterogeneity and multidimensionality in multimodal data, retain the characteristics of each modal data, and uniformly analyze the graph structure. Therefore, when dealing with multimodal learning problems, the generalized Laplace matrix is ​​a key tool for analyzing and optimizing data structures.

[0023] Through the calculated generalized Laplace matrix, an optimization problem can be constructed. The goal of this optimization problem is to find a low-dimensional embedding space so that the relationship between data points in this space can be preserved as accurately as possible. In other words, the core of the optimization problem lies in minimizing the difference between the embedded representation of data points and the original high-dimensional data representation. This optimization problem not only involves how to map data points from high-dimensional space to low-dimensional space, but also requires that the important relationship characteristics between data points be preserved in low-dimensional space. By solving this optimization problem, an embedding matrix can be obtained, which converts the original data points into low-dimensional representations. These low-dimensional representations can effectively reduce the computational complexity while retaining the key information of the original data. This process is particularly important in the context of multimodal learning. The dimensionality of multimodal data is usually very high, and it is complex and time-consuming to directly optimize or query calculations in high-dimensional space. Therefore, by embedding data into a low-dimensional space, not only can the computational overhead be greatly reduced, but also the query performance of the database can be improved. In addition, since the data points in the low-dimensional embedding space retain the structural information of the original data, the accuracy of the query results is also guaranteed. This low-dimensional embedding space actually provides an effective way for the database to process complex data from different modalities more efficiently. Especially on large-scale data sets, the calculation of embedded representation can significantly improve the performance of the system. At the same time, the introduction of the generalized Laplace matrix also makes this method essentially different from traditional database optimization methods. Traditional database optimization methods usually rely on simple index structures or distance-based query optimization, which often perform poorly when processing high-dimensional or multimodal data. The optimization method based on the generalized Laplace matrix can handle more complex data structures and find a more suitable data representation through the analysis and optimization of the graph structure. Especially in the face of a database environment with huge data volumes and diverse modalities, the combination of the generalized Laplace matrix and the low-dimensional embedding representation provides an effective solution that can greatly improve query efficiency while ensuring query accuracy.

[0024] Step 3: Based on the low-dimensional embedding representation of the data points, define the objective function of query optimization; obtain the update rule of the optimal embedding representation by taking the partial derivative of the objective function with respect to the embedding representation and setting it equal to zero; use the optimal embedding representation to design a dynamic index structure so that queries can be performed efficiently in the embedding space.

[0025] The low-dimensional embedding representation is obtained by optimizing the generalized Laplacian matrix in step 2. This embedding representation is essentially a compression of high-dimensional data. It not only retains the important associations between data points, but also expresses these relationships in a simplified way in the low-dimensional space. Therefore, querying in the low-dimensional embedding space can greatly improve the efficiency of the query while retaining sufficient accuracy. Based on this low-dimensional embedding representation, the objective function of query optimization is defined. The design of this objective function takes into account two key factors: on the one hand, the accuracy of the query, that is, the data points in the low-dimensional space need to reflect the data relationship in the original high-dimensional space as accurately as possible; on the other hand, the efficiency of the query, that is, the computational cost of query operations in the low-dimensional space is significantly lower than that in the high-dimensional space. The core idea of ​​the objective function is to ensure the accuracy of the query by minimizing the difference between the low-dimensional embedding representation and the query result. This difference can be regarded as an error measure of the query in the low-dimensional space, and the objective function is designed to reduce this error by optimizing the embedding representation. In the context of multimodal learning, the data points come from different modalities and their relationships are very complex, so query optimization must accurately capture these complex relationships in the low-dimensional space. By taking partial derivatives of the objective function, we obtain the update rule for the optimal embedding representation. This update rule ensures that the embedding representation can be dynamically adjusted as the query is continuously optimized, thereby gradually approaching the global optimal solution.

[0026] In the process of solving the objective function, the update rule of the embedding representation plays a key role. The update rule makes the distribution of data points in the low-dimensional space gradually tend to the optimal state by continuously adjusting the parameters of the embedding representation. As the update proceeds, the low-dimensional embedding representation can not only better express the relationship between the original data points, but also dynamically adjust according to the needs of query optimization, so that each query operation can be performed in the optimal state. In this way, the data representation in the low-dimensional space not only simplifies the complexity of the data, but also provides higher flexibility for query operations. Based on the update of the embedding representation, a dynamic index structure is further designed. The index structure is a key tool for accelerating queries in databases. Traditional index structures are usually designed for static data and are difficult to adapt to the dynamic changes of multimodal data. In the relational cluster database optimization method based on multimodal learning, the dynamic index structure can be adjusted in real time according to the update of the embedding representation to ensure that the query can be performed efficiently in the low-dimensional embedding space. The design idea of ​​the dynamic index structure is that the index not only needs to be able to quickly locate data points, but also needs to be able to respond quickly when the embedding representation changes to ensure that the query efficiency is not affected. This dynamic index structure is closely integrated with the embedded representation, so that the query process can be continuously optimized as the data changes, greatly improving the query response speed and accuracy. An important feature of the dynamic index structure is its adaptability. Traditional index structures usually rely on predefined rules or structures, while dynamic indexes can be adjusted in real time according to changes in data points. For example, when the relationship between certain data points in the database changes, the dynamic index structure can quickly update the corresponding index information so that the query operation can always remain in the optimal state. In addition, because the index structure is closely integrated with the low-dimensional embedded representation, when the embedded representation is updated, the index structure will also be adjusted synchronously, thereby ensuring that the index can accurately reflect the relationship between the current data points.

[0027] Embodiment 2: Step 1 specifically includes: determining each data point Neighborhood range , thereby determining the data points Data points in the neighborhood of ; Calculate for each data point The local density gradient ; For each data point , through the Gaussian kernel function to the data points in the neighborhood The distances are weighted summed to obtain the data point pairs in the neighborhood. The influence of; combining the local density gradient and the neighborhood contribution, calculate the data point Laplace density gradient ; Local density gradient Reflects the data points Relative to the density change of its surrounding neighborhood, if ,illustrate The density of is higher than that of its neighborhood and is the cluster center. ,illustrate The density is low and is in a sparse area. According to the Laplace density gradient of each data point, an association matrix is ​​constructed to represent the degree of association between data points.

[0028] Specifically, the neighborhood range of the data point Determine which data points will have an impact on the target data point. For multimodal data, data of different modes have different distribution characteristics, so the selection of the neighborhood range is particularly critical. Too large a neighborhood range may introduce irrelevant data points, causing the calculation results to deviate from reality, while too small a neighborhood range may miss important local information. Therefore, The selection of needs to balance the features under different modes. Based on this neighborhood, each data point Data points in the neighborhood of Will The introduction of Gaussian kernel function is an effective method to weight these effects according to the distance between data points. The closer the point, the greater the impact, and the farther the point, the smaller the impact. The form of Gaussian kernel function Indicates the distance and The square distance between Here, Controls the influence of distance on weight. The value makes the farther points still have some influence, while the smaller The value of limits the impact to only very close data points. With this weighting method, data points in the neighborhood The contribution of is effectively quantified and used to calculate the data points The local density gradient The local density gradient is actually A quantitative expression of the density change relative to its surrounding neighborhood. , indicating the data point The area is denser than its neighbors, indicating is a cluster center. When The sparse areas with lower density may be the boundary points or outliers of the cluster. Through the calculation of the local density gradient, the present invention can effectively identify the important clustering structures and clustering properties of the data points in the database.

[0029] Next, the Laplace density gradient of the data point is calculated based on the combination of the local density gradient and the neighborhood contribution The core of Laplace density gradient is that it describes the data points by further processing local density information. The rate of change of density in its local area. The Laplace density gradient is closely related to the topological structure of the data points and can reveal the changing trend of the data points in their local environment. Unlike simple density estimation, the Laplace density gradient can capture the complex relationship between data points in a larger range, especially when dealing with multimodal data, this method can better reflect the interactive relationship and global structure of different modal data. This Laplace density gradient calculation based on Gaussian kernel weighted summation lays the foundation for the construction of the association matrix. By calculating the density gradient of each data point, the constructed association matrix can not only reflect the distance relationship between the data points, but also capture their local density changes. This is especially important for multimodal data, because data of different modalities are often distributed in different high-dimensional spaces, and the relationship between them is not only determined by the distance, but also closely related to their respective distribution characteristics in their respective modalities. By introducing the Laplace density gradient, these complex inter-modal relationships can be more accurately characterized, thus providing a reliable basis for subsequent optimization. In addition, this Laplace density gradient calculation also has strong robustness. When faced with noise or outliers, traditional density estimation methods may produce large deviations due to the presence of these abnormal points, while the Laplace density gradient effectively reduces the impact of noise points on the overall density estimation by combining the weighted summation of data points in the neighborhood with the Gaussian kernel function. This means that even if there are certain abnormal points in the data, their impact on the calculation results will be limited to a certain range and will not significantly change the identification results of cluster centers or sparse areas. This feature is particularly important when processing complex multimodal data, because the distribution between different modes may introduce various unpredictable noises, and the calculation method of the Laplace density gradient provides a more stable processing mechanism for this.

[0030] Example 3: Laplace density gradient The calculation formula is as follows: ; in, It is the data point The neighborhood density of is summed up, Represents data points The neighborhood of Determine, indicating that the radius is All data points in the region The neighborhood set ; Represents data points With neighboring data points The distance between With The relative position of right the extent of the impact; is a Gaussian kernel function used to transform the distance Converted to a weight, the smaller the distance, the greater the weight, indicating right The greater the impact, is the scale parameter, which controls the degree of smoothing within the neighborhood; is a density adjustment parameter used to control the data points in the neighborhood right The greater the impact The value represents the data point pair in the neighborhood The greater the impact, the smaller the Then it represents the data point pair of the neighborhood The smaller the impact.

[0031] Specifically, the formula It contains two main parts: one is the local density gradient , and the second is the neighborhood density summation based on the weighted Gaussian kernel function. Local density gradient Reflects the data points The density change compared to its neighbors. If , which means The density of is higher than that of its surrounding neighborhood, it may be a cluster center; on the contrary, if ,but may be located in sparse areas with lower density. Therefore, It is used to preliminarily determine the density state of the area where the data point is located. The second part of the formula For data points The neighborhood density is weighted summed to correct the local density gradient Neighborhood Is caused by Distance in Data points within The set of these points will be In the context of multimodal learning, data points may come from different modalities, and their spatial distribution is complex. Traditional Euclidean distance or simple neighborhood definition is often not enough to accurately capture the relationship between data. Therefore, the weighted summation of data points in the neighborhood can better reflect the mutual influence between these multimodal data. In the summation term, is the Gaussian kernel function, Represents data points The data points in its neighborhood The Gaussian kernel function is introduced to assign weights according to distance. The greater the influence of the neighboring data points on the target data point, the greater the influence of the neighboring data points on the target data point. This weighting mechanism makes the formula not only consider the distance, but also more accurately considers the influence of each data point in the neighborhood on The smaller the distance, the greater the weight, which means that the data points with closer distance in the neighborhood have more influence on The density of the data points has a greater impact; conversely, the farther away the data points, the smaller the impact. is a key parameter in the Gaussian kernel function, which controls the degree of density smoothing within the neighborhood. A larger value indicates a wider range of influence, allowing points that are farther apart to has a greater effect, making the density calculation smoother. Conversely, a smaller The value will limit the range of influence, and only data points that are close to each other will significantly affect the density estimate. The choice of needs to be reasonably set according to the modal distribution characteristics of the data to balance the smoothness and calculation accuracy in the local neighborhood. Used to control the data point pairs within the neighborhood The overall impact strength of The value represents the data point pair in the neighborhood The Laplace density gradient will depend more on the density distribution of other data points in the neighborhood. On the contrary, a smaller The value means that the data point in the neighborhood The impact is small. The density of is mainly determined by its own local environment. The adjustment of plays an important role in the final calculation result of the Laplace density gradient, which balances the relationship between local density and neighborhood density. In multimodal learning scenarios, data points of different modalities may have different neighborhood density distribution characteristics, so it is necessary to select the appropriate one according to the modal characteristics of the specific data. Value. In the calculation process of the whole formula, the Laplace density gradient The final result not only describes the data points The density change of the data point is also combined with the influence of other data points in its neighborhood. This calculation method makes the density estimation of data points more accurate, especially when processing multimodal data, it can effectively identify cluster centers and sparse areas, thus providing reliable support for subsequent database query optimization. Compared with traditional density estimation methods, Laplace density gradient captures the complex relationships between multimodal data more flexibly through the weighted neighborhood summation of Gaussian kernel functions, especially in high-dimensional space. This method can better handle data interactions and mutual influences between different modalities.

[0032] Example 4: Correlation Matrix No. Row, No. Elements of a column Defined as: ; in, For data points and data points The association weight between them, the larger the value, the stronger the association; and The data points and data points Laplace density gradient of ; is the first-order L1 norm, is the second-order L1 norm; is a preset amplification parameter that controls the influence of density gradient direction similarity on the association weight; for and The Euclidean distance between is the preset gradient change rate smoothing parameter, which controls the influence of density gradient change rate difference on the association weight; and The data points and data points The gradient of the Laplace density gradient.

[0033] Specifically, the first part of the formula Reflects the data points and The similarity between the Laplace density gradient of . Laplace density gradient and Describes the data points and The density changes in their respective neighborhoods, so by calculating their dot product, we can measure whether the density gradient directions of the two data points are consistent. The more consistent the directions, the larger the dot product value, indicating that the density changes of the data points are similar in direction and the stronger the correlation between the data points. By normalizing this value, that is, dividing it by their first-order Norm , which ensures that gradients of different scales can be compared fairly. The introduction of is to control the influence of density gradient direction similarity on the association weight. A larger value will amplify the influence of Laplace density gradient similarity on the association weight, while a smaller value will reduce this influence, making the association weight more dependent on other factors. Secondly, the second part of the formula is a Gaussian kernel function based on Euclidean distance, which is based on the data points and The space distance between The closer the distance, the greater the weight, which is consistent with the fact that adjacent data points in physical space are more likely to have similar characteristics. Controls the decay rate of the weight. A value of 1 means that data points farther away will also have a greater impact on the associated weight, while smaller The value means that only very close data points will significantly affect the association weight. This part reflects the geometric similarity of data points in high-dimensional space. For multimodal data sets, geometric similarity is an important factor affecting the association of data points, because data points of different modes may show similar geometric distributions in different dimensions. The last part of the formula The rate of change of the Laplace density gradient is further introduced as a factor affecting the association weight. and The data points are and The gradient of the Laplace density gradient reflects the rate of change of the density of the data point. This part is introduced to capture more detailed density change patterns between data points, especially in multimodal data, where data of different modes may show significant differences in the rate of change of density. By calculating the difference in the rate of change of the Laplace density gradient of two data points, the association between the data points can be further refined. Smoothing parameter The influence of the difference in density gradient change rate on the association weight is controlled. The value makes the difference in density change rate have a greater impact on the association weight, while smaller The value weakens this effect.

[0034] Example 5: In step 2, a weighted undirected graph is constructed based on the association matrix , node data point set Contains data points corresponding to all data ; is the total number of data points; the edge set According to the correlation matrix Determine that the edge weight is , which represents the data point and The strength of the association between them; the degree matrix of a weighted undirected graph is a diagonal matrix whose elements are defined as: ; Represents node data points The degree of is the sum of the weights of the edges connected to it; define the generalized Laplace matrix for: ; in, is the standard graph Laplacian matrix; is the square of the Laplacian matrix, capturing the second-order neighborhood information; is the preset first regularization parameter; is the inverse square root of the degree matrix.

[0035] Specifically, in a weighted undirected graph In the vertex set contains all data points, and the edge set From the incidence matrix The edge weight is determined by the elements in the incidence matrix Given, represents the data point and The correlation matrix not only considers the geometric distance between data points, but also combines factors such as their Laplace density gradient and density gradient change rate. In this way, the graph structure can effectively express the relationship between data points, not only capturing directly adjacent data points, but also reflecting the deeper connection between individual data points through the weight of the edge. is an important element in a weighted undirected graph. Its diagonal elements Represents the degree of each data point, which is the sum of the weights of all edges associated with the data point. The degree matrix reflects the importance of the data point in the graph by encoding the degree of each node as a diagonal element of the matrix. A larger degree value indicates that the data point has a stronger correlation with more other data points and may be at the center of a data cluster; a smaller degree value means that the data point may be in a sparse area. , the importance of the nodes can be normalized to ensure that the contributions of different data points are reasonably distributed in subsequent calculations. Generalized Laplace Matrix It is the key part of the entire graph structure analysis. The standard graph Laplacian matrix It reflects the difference between each data point and its neighbors and captures the local topological structure in the graph. By combining the degree matrix and the association matrix, the Laplacian matrix can quantify the position and association of each data point in the graph. In the scenario of multimodal learning, data points of different modalities may show different relationships in different feature spaces, so a single geometric distance often cannot fully describe these complex relationships. The Laplacian matrix can more comprehensively capture the interdependencies in multimodal data by encoding the local relationships of data points into a matrix form.

[0036] The definition of the generalized Laplacian matrix further extends the functionality of the standard Laplacian matrix. The generalized Laplace matrix not only considers the direct neighbors of the data points, but also captures the information of the second-order neighbors, that is, the relationship between the data points and the indirect neighbors. The second-order neighborhood information is particularly important in multimodal data, because multimodal data usually has a complex interactive structure, and the relationship between data points is not limited to directly connected points, but also has a large number of indirect connections through other points. By introducing , these indirect relationships can be taken into account to enhance the understanding of the global structure of the graph and make the correlation analysis of data points more accurate. At the same time, the generalized Laplace matrix also introduces a regularization term , which is used to adjust the weight of the standard Laplace matrix in the overall optimization process. The role of the regularization term is to balance the influence of local and global information and avoid over-reliance on certain local structures. For multimodal learning, regularization can help deal with data differences between different modalities and prevent optimization deviations caused by over-prominence of certain modal features. By adjusting the parameters , which can achieve a dynamic balance between local and global structures, ensuring the stability and robustness of correlation calculations. is the inverse square root of the degree matrix, which plays a normalization role and ensures that the contribution of each data point in the calculation is balanced. Since the degrees of different data points may vary greatly, directly using the degree will cause the data points with larger degrees to occupy too much weight in the calculation, thus affecting the fairness of the optimization results. , the degrees of different nodes can be normalized, thereby eliminating the imbalance caused by degree differences. This normalization process ensures that the contributions of different data points can be reasonably considered in the complex environment of multimodal data, avoiding some data points from excessively dominating the entire optimization process. The overall design idea of ​​the generalized Laplace matrix is ​​to comprehensively capture the complex relationship between data points through in-depth analysis of the graph structure, combined with second-order neighborhood information and regularization terms. In the context of relational cluster database optimization, the relationship between data points not only affects the efficiency of the query, but also directly affects the accuracy of the query results. By using the generalized Laplace matrix, the correlation information between data points can be retained to the maximum extent while ensuring query efficiency. This method is particularly suitable for processing multimodal data, because the complexity of multimodal data is often manifested in multi-level correlation structures, and the generalized Laplace matrix can effectively capture and express these structures.

[0037] Example 6: The process of calculating the low-dimensional embedding representation of the data point in step 2 specifically includes: constructing an optimization problem, whose objective function objective function for: ; The objective function aims to find an embedding matrix , which can preserve the local structure and keep consistent with the original features; is the low-dimensional embedding matrix of the data points, is the embedding dimension; the relational cluster database is regarded as a matrix, and the original feature matrix of the relational cluster database is obtained through feature extraction , is the dimension of the original feature; is the trace operation of the matrix, i.e. the sum of the diagonal elements; is the first-order F-norm; is the preset second regularization parameter, which controls the degree of deviation between the embedding result and the original feature; is the identity matrix; the constraints are: ; Ensure that the embedded data is orthogonal in the new space to prevent degenerate embedding results.

[0038] Specifically, the first term of the objective function Represents the embedding matrix Relationship with graph structure. Matrix is the aforementioned generalized Laplace matrix, which captures the complex relationship between data points in the graph, especially the local and global structural features. , the objective function measures the embedding matrix The degree to which these structural features are preserved in the new low-dimensional space. The trace operation is essentially the sum of the diagonal elements of the matrix. In this formula, it reflects the graph structural relationship between each data point and other data points in the low-dimensional space. Specifically, The smaller the value of, the more consistent the structural features of the data point in the low-dimensional space are with its structural features in the high-dimensional space. Therefore, this part of the objective function ensures that the embedding matrix The local and global structural characteristics of data points can still be preserved in low-dimensional space, which is especially important for the processing of multimodal data, because multimodal data often has a complex structure, and this structure preservation helps to maintain the original relationship between data points. To ensure that the embedding matrix With the original feature matrix Here the similarity between is the original feature matrix of the relational cluster database, which contains the feature information of the data points in the original space. and The first-order Frobenius norm is calculated for the difference between the two, and the objective function can measure the degree of deviation between the embedded representation and the original features. The Frobenius norm is a matrix norm that measures the difference between two matrices by taking the square root of the sum of all elements of the matrix. By introducing this term, the objective function can retain the original feature information as much as possible in the low-dimensional embedding space, ensuring that the embedding result is consistent with the characteristics of the original data point. For multimodal data, the original features may come from different data sources of multiple modalities. The retention of these features in the low-dimensional embedding space is crucial for subsequent query optimization and data analysis.

[0039] is the second regularization parameter, which controls the degree of deviation between the embedding result and the original feature. The value means the embedding matrix More inclined to retain the original feature matrix consistency, while the smaller A value of indicates that the embedding matrix is ​​more inclined to optimize the preservation of graph structure. , a reasonable balance can be found between structure preservation and feature consistency. In the scenario of multimodal learning, data of different modalities may have different importance. The adjustment can help the optimizer adjust the embedding focus according to specific needs. In addition, the constraints of the objective function It ensures that the embedded data points remain orthogonal in the new low-dimensional space. This constraint is introduced to prevent the embedding matrix Degeneration, that is, avoiding the overlap or loss of resolution of different data points in the embedded space. Equal to the identity matrix , which ensures that the representation of the embedded data points in the low-dimensional space is independent and will not be over-compressed or confused with each other. Maintaining the independence between data points is crucial for the optimization of relational cluster databases, because the representation after embedding will directly affect the accuracy and efficiency of queries. If the embedding result is degraded, it may lead to inaccurate or even invalid query results, so this orthogonality constraint plays a key role in the embedding process. By constructing such an objective function, the embedding matrix It is able to maintain both the local structure of data points and their original features in low-dimensional space, thus providing effective support for database query optimization. This optimization process is particularly suitable for processing multimodal data, because the complexity of multimodal data is reflected in multiple dimensions and modes, and it is necessary to maintain the global relationship of data points and ensure the inherent consistency between features. In the optimization process of relational cluster databases, the calculation of low-dimensional embedding representation directly affects the subsequent query speed and accuracy. Through the optimization of this objective function, complex data in the original high-dimensional space can be mapped to a low-dimensional space, thereby significantly reducing the computational overhead while maintaining the core features and structural information of the data.

[0040] Example 7: In step 2, in order to solve the optimization problem, construct the Lagrangian function for: ; in, is the Lagrange multiplier matrix; for Take the derivative and set it to zero: ; The solution is obtained by simplifying the following formula: ; Among them, due to ,get ; Using the above equation, solve the following generalized eigenvalue problem: ; Solving this equation, we get the embedding matrix , where each element It is a data point A low-dimensional embedding representation of .

[0041] Specifically, the key to the optimization problem is to construct a Lagrangian function ,in is the low-dimensional embedding matrix of the data points, is the Lagrange multiplier matrix used to handle constraints , ensuring that the embedded data points maintain orthogonality in the low-dimensional space. Orthogonality constraints are very important for low-dimensional embedding, especially when dealing with relational cluster databases for multimodal learning, ensuring that the embedding matrix The orthogonality of can prevent the embedding results from degenerating, thereby ensuring that the data points maintain good distribution and identification in the low-dimensional space. If the orthogonality constraint is not imposed, the embedded data points may overlap, lose the effective expression of the original data structure, and ultimately affect the query efficiency and accuracy of the database. Lagrangian function It consists of three main parts: Part I is the graph Laplacian matrix With the embedding matrix It reflects whether the embedded data points can retain their original graph structure. Graph Laplacian matrix Through the adjacency matrix Sum degree matrix It describes the relationship between data points, especially the relative position of data points in their local and global neighborhoods. , the optimization problem ensures that the embedding matrix in the low-dimensional space can still reflect the local and global structural features of the original graph. In other words, the distribution of data points in the low-dimensional space still maintains their topological relationship in the high-dimensional space, thus providing effective support for subsequent database queries. As a specific function of a matrix, the sum of the diagonal elements of the matrix is ​​calculated, which in this case provides a simplified way to measure how well the embedding matrix preserves the graph structure.

[0042] Part 2 It is about the embedding matrix and the original feature matrix The original feature matrix contains the original high-dimensional feature information of each data point in the relational cluster database, while the embedding matrix The goal is to approximate these feature information in a low-dimensional space. Through the Frobenius norm calculation in this part, the objective function can measure the difference between the embedding matrix and the original feature matrix, ensuring that the embedding result can not only retain the structural information of the graph, but also reflect the characteristics of the original data points as much as possible. The Frobenius norm is a matrix norm that measures the distance between two matrices by taking the square root of the sum of the squares of all elements in the matrix. Therefore, it can intuitively reflect the degree of deviation between the embedding result and the original feature. Regularization parameter After controlling for the effect size of this deviation, the larger The value means that the embedding matrix tends to be closer to the original feature matrix , while the smaller The value allows the embedding matrix to have more freedom in preserving the graph structure. is the Lagrange multiplier term, which is used to process the embedding matrix This orthogonality constraint ensures that the embedded data points are independent of each other in the low-dimensional space, preventing them from overlapping or degenerating. Without this constraint, the embedded data points may be clustered together in the low-dimensional space, resulting in reduced recognition between data points, thus affecting the accuracy of database queries. Therefore, the Lagrange multiplier This orthogonality constraint is incorporated into the optimization problem through trace operation, ensuring that the embedded data points can maintain a proper distribution in the low-dimensional space and avoid degradation. After constructing the Lagrangian function, the optimization problem is solved by embedding the matrix This is done by taking the derivative and setting the derivative to zero. Through this derivation process, the relationship between the embedded matrix and the Laplace matrix, the original characteristic matrix, and the Lagrange multiplier is obtained. Further simplifying the equation: , It shows that the generalized Laplace matrix With the embedding matrix By introducing the regularization term , which not only smoothes the optimization process, but also prevents overfitting, making the embedding matrix In low-dimensional space, the structural characteristics of the data can be preserved while maintaining good generalization ability. In addition, due to the orthogonality constraint of the embedding matrix , we can further solve the Lagrange multiplier matrix by substituting , and finally the optimization problem is transformed into a generalized eigenvalue problem. The form of this generalized eigenvalue problem is: , It describes how to obtain the embedding matrix by solving the generalized eigenvalues . The solution to the generalized eigenvalue problem has a wide range of applications in graph embedding and dimensionality reduction. It can effectively extract the main structural features of the data by calculating the eigenvalues ​​and eigenvectors of the matrix. In the relational cluster database optimization method of multimodal learning, the solution to the generalized eigenvalue problem can find a low-dimensional embedding matrix so that the representation of data points in the low-dimensional space can both retain its global structure and approximate its original features, thereby achieving the goal of optimizing queries. The result of solving this generalized eigenvalue problem is a low-dimensional embedding matrix , where each element Corresponding to data points in the relational cluster database Embedded representation in low-dimensional space. This embedded representation can not only significantly reduce the complexity of high-dimensional data, but also retain the main structural features and attribute information of the data. For query optimization of relational cluster databases, low-dimensional embedded representation provides an effective way to reduce the computational overhead of queries while ensuring the accuracy of query results. Through this embedded representation, the database can significantly improve the query speed and maintain high query accuracy when processing large-scale multimodal data.

[0043] Example 8: Defining the objective function of query optimization based on the low-dimensional embedding representation of data points in step 3 for: ; in, Represents the set of data point pairs involved in the query workload; For data point pairs The query weights of , which reflect the frequency with which they appear in the query; is the preset third regularization parameter.

[0044] Specifically, the first part of the objective function is for the data points involved in the query workload The calculations performed here are represents the set of data point pairs contained in the query workload, and is the frequency weight of these data point pairs in the query. The core idea of ​​this part is to increase the weight of the data point pairs that appear frequently in the query process, so that the relationship between these data point pairs can be better maintained in the low-dimensional embedding representation. In other words, point pairs that appear frequently in the query workload should maintain a small distance in the low-dimensional space, reflecting their close relationship in the high-dimensional space. Therefore, by taking the weighted sum of the distances of these data point pairs, the optimization problem ensures that the similarity of the data point pairs in the query process can be reflected in the low-dimensional space, thereby improving the accuracy and efficiency of the query. Part II is a regularization term, which is used to ensure that the embedding representation It can maintain the relationship between data points in the original graph structure. is the graph Laplacian matrix, which reflects the topological characteristics of data points in the global graph structure by capturing the adjacency relationship and degree information between data points. , which measures the embedding representation Whether it is possible to effectively preserve the local and global relationships in the graph structure in a low-dimensional space. The introduction of the graph Laplacian matrix ensures that the structural features between data points are not lost during the embedding process. Especially in the context of query optimization, structural information is crucial for understanding the similarities and relationships between data points. Controls the weight of this regularization term, the larger A value of 0 means that the preservation of the graph structure plays a more important role in the optimization process, while a smaller The value allows a greater degree of embedding freedom, allowing the embedding matrix to be adjusted according to the needs of the query workload. is another regularization term used to control the difference between the embedded representation and the original features. is the original feature matrix, which contains the high-dimensional feature information of data points in the relational cluster database, and the embedding matrix It is an approximate representation of these features in a low-dimensional space. , the objective function measures the deviation between the embedded representation and the original features, ensuring that the embedded result can maintain the integrity of the original feature information as much as possible. The Frobenius norm is a matrix norm that measures the difference between two matrices by calculating the square root of the square of the elements of the matrix. Here, it ensures that the embedded representation does not deviate too much from the characteristics of the original data point, so that the query optimization result can take into account both structural information and feature information. Parameters Controls the weight of this regularization term, and the larger The value means that the optimization process pays more attention to the consistency between the embedded representation and the original features, while a smaller A larger value allows more embedding freedom and thus better adapts to query requirements.

[0045] In the entire objective function In the design, the data point pairs of the query workload, the preservation of the original graph structure, and the consistency between the embedded representation and the original features are organically combined to achieve the purpose of optimizing queries. By weighting the data point pairs, the high-frequency query patterns in the query workload can be reflected in the low-dimensional embedding space, thereby improving query efficiency. and the original feature matrix , the optimization problem ensures that the embedded representation can be flexibly represented in a low dimension while retaining the original graph structure and feature information, so as to adapt to the changes in different query requirements. This optimization method has important application value in the optimization of relational clustering databases for multimodal learning. Multimodal data usually contains complex information from different sources, and the relationship between data points is not only reflected in the geometric distance, but also in its original features and structural information. Therefore, in the query optimization process, these multi-level information must be considered simultaneously. Through the objective function The definition of can take into account the complexity of multimodal data in the process of query optimization, ensuring that the query results can not only maintain the similarity between data points, but also reflect the original characteristics and global structure of the data points.

[0046] Example 9: In step 3, the objective function about Taking the partial derivative and setting it equal to zero, we get ; In this way, we can get the objective function of query optimization The first item The partial derivative of is: ; in, To query the relevant Laplacian matrix, it is defined as: ; in, and and Same, both are subscript indexes; the objective function of query optimization The second item The partial derivative of ; Item 3 The partial derivative of ; The objective function of query optimization Adding the first, second, and third partial derivatives of gives the following equation: ; After simplifying the equation, we get: ; Finding the optimal embedding representation for: .

[0047] Specifically, the objective function It aims to preserve the complex relationships between data points by optimizing the low-dimensional embedding representation. The distance between pairs of data points that appear frequently in the query workload in the embedding space should be minimized, so the first term of the objective function introduces a weighted distance sum, which partially reflects the importance of the data point pairs in the query. By calculating the square of the distance between these data point pairs in the low-dimensional space and assigning them weights related to the query frequency, the optimization process is able to prioritize retaining these important relationships. This helps to better represent the similarities between high-frequency data point pairs in the low-dimensional space, thereby improving the accuracy of the query results. Secondly, the optimization process needs to consider not only the query-related data point pairs, but also the global structural information of the data. This is achieved through the graph Laplacian matrix The graph Laplacian matrix is ​​based on the adjacency matrix and the degree matrix, which can reflect the topological structure between data points. The second item in the optimization objective It is through this matrix that constraints are imposed on the low-dimensional embedding representation, ensuring that the data points still maintain their original global structural characteristics after embedding. This is crucial for the processing of multimodal data, because in multimodal learning, data points may have complex structural relationships, including both local neighboring point relationships and global topological structures. If the preservation of the global structure is ignored during the optimization process, the embedded data representation may lose its key contextual information, thereby affecting the query results of the database. Therefore, as a representative of the global structure, the graph Laplacian matrix ensures that the embedding representation not only focuses on the local but also maintains global consistency.

[0048] In order to ensure that the embedded representation in the low-dimensional space can reflect the characteristics of the original data as much as possible, the third term of the objective function measures the embedded representation by the Frobenius norm With the original feature matrix The difference between the original feature matrix Contains the characteristics of each data point in high-dimensional space, while the embedding representation is the approximation of these features in a low-dimensional space. By introducing this term, the optimization process can control the deviation between the embedded representation and the original features, ensuring that the embedding result does not deviate too much from the original data. This retention of feature consistency is particularly important for multimodal data, because different features of multimodal data may come from different data sources and represent the multidimensional properties of data points. By retaining these properties during the optimization process, the database can accurately capture the multidimensional information of each data point in a low-dimensional embedding space. Throughout the optimization process, three key factors—data point pairs in the query workload, global structural information, and feature consistency—are considered simultaneously and balanced by corresponding parameters. For example, the regularization parameter Controls the weight of global structural information, The consistency between the embedded representation and the original features is adjusted. By properly selecting these parameters, a balance can be achieved between different optimization objectives to meet the complex needs of multimodal learning. When solving this optimization problem, we first need to find the partial derivative of the objective function and set it equal to zero. The weighted distance and partial derivative of the query workload give the Laplace matrix associated with the query. This matrix reflects the importance of data point pairs in the query process by weighting the data point pairs with high query frequency. Therefore, in the optimization process, the data point pairs with higher query frequency will be retained more in the embedding representation. By taking the derivative, we obtain an optimization condition related to the global graph structure, which ensures that the embedded representation can continue to retain the global topological information in the low-dimensional space. Finally, we take the derivative of the feature consistency term to ensure that the embedded representation does not deviate too much from the original features. By adding the partial derivatives of these three parts, we obtain a comprehensive optimization condition, which is expressed as: .

[0049] This equation shows how different factors work together to influence the generation of embedding representations during query optimization. The matrix on the left takes into account query frequency, global structure, and regularization terms, while the matrix on the right takes into account the original feature matrix Keep the embedded representation consistent with the original data. Finally, by solving this equation, the optimal embedded representation can be obtained , which is the best representation of data points in low-dimensional space. This result is of great significance for the optimization of relational clustering databases in multimodal learning. By optimizing the embedding representation, the complexity of high-dimensional data can be effectively reduced, the computational overhead of queries can be reduced, and the accuracy and consistency of query results can be maintained. Especially when processing multimodal data, this optimization method can simultaneously process data features of multiple modalities and retain their global and local relationships in low-dimensional space, which significantly improves the query performance of the database.

[0050] Example 10: In step 3, using the optimal embedding representation The process of designing a dynamic index structure includes: since the embedding space is low-dimensional, use Kd tree, ball tree or LSH as the index structure; represent the embedding Each data point in Insert into the index structure; when the data in the database changes, update the embedded representation accordingly Neutralize the index structure.

[0051] Specifically, the embedding representation is a low-dimensional representation obtained through the previous optimization process, containing the embedding vector of each data point . These vectors not only retain the complex structure between the original data points and the relationship in the query workload, but also compress the information in the high-dimensional space into the low-dimensional space. Therefore, the design goal of the index structure is to effectively organize these embedded representations in the low-dimensional space so that query requests can be quickly responded to. Since the embedding space is low-dimensional, the traditional high-dimensional indexing method may no longer be applicable, and the index structure that is more suitable for low-dimensional data becomes the best choice. In this process, index structures such as Kd tree, ball tree and locality sensitive hashing are recommended for processing data points in low-dimensional embedding space. The common characteristics of these index structures are that they can quickly and efficiently organize data in low-dimensional space and support efficient nearest neighbor queries. Kd tree is a common tree data structure for multidimensional space, which is suitable for performing nearest neighbor search in low-dimensional space. It can quickly narrow the scope of the query by recursively dividing the space into hyper-rectangular regions. For the embedding representation of relational cluster databases, Kd tree can handle data point queries in the embedding space well. Ball tree is an index structure suitable for processing non-uniform data distribution. It groups data points into spherical regions, which effectively narrows the search scope when performing nearest neighbor queries. Compared with Kd-tree, spherical tree has higher query efficiency when data points are unevenly distributed, which makes it very suitable for relational clustering databases in multimodal learning, especially when the embedding representations of different modal data have different distribution characteristics.

[0052] Locality Sensitive Hashing (LSH) is an indexing method based on hash functions that can map similar points to the same hash bucket. LSH is particularly suitable for handling approximate nearest neighbor query problems in high-dimensional and low-dimensional spaces. It divides the space into multiple hash buckets so that similar vectors are more likely to be mapped to the same bucket. Therefore, when faced with large-scale embedded data, LSH provides an efficient approximate query method that can significantly reduce the computational complexity of the query while maintaining a high query accuracy. Each data point After being calculated, these data points are inserted into the selected index structure. In the index structure, the position and relationship of the data points are preserved, so that similar points can be quickly found when performing queries. With the dynamic index structure, the query system can quickly respond to various query requests in the low-dimensional embedding space, whether it is similarity-based queries or other complex query types. The core feature of the dynamic index structure is its adaptability to changes in database data. When the data in the database changes, such as inserting new data points or deleting existing data points, the embedded representation The embedding representation directly affects the index structure, so when the data is updated, the index structure must also be updated synchronously to ensure the accuracy and timeliness of the query results. This update process involves recalculating the embedding representation of the relevant data points and inserting or deleting them from the index structure. For the addition of new data points, the feature vector of the new data point needs to be mapped to the low-dimensional space through the previous embedding method to obtain the corresponding embedding representation. , and then insert it into the existing index structure. This process requires the index structure to have good scalability and be able to efficiently handle the insertion of new data points. For the deletion of data points, the index structure needs to support fast deletion operations to ensure that the query results after deletion do not contain invalid data points. The update of data points requires the synchronous adjustment of the position of their embedded representation in the index structure so that it can reflect the latest data relationship. The advantage of the dynamic index structure is that it can be adjusted in time according to the changes in the database, thereby ensuring the flexibility and efficiency of the query system when processing large-scale data. Especially in multimodal learning applications, data updates may be very frequent, and the representation of data points in the embedding space will also be continuously adjusted as the data changes. Therefore, the dynamic index structure must not only be able to handle addition, deletion and update operations, but also must ensure that these operations do not affect the overall performance of the system and the real-time nature of the query.

[0053] The present invention is described in detail above. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. It should be pointed out that for ordinary technicians in this technical field, without departing from the principle of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. Relational cluster database optimization method based on multimodal learning, It is characterized in that The method comprises: Step 1: In the relational cluster database, each stored data is regarded as a data point, and the Laplace density gradient of each data point is calculated to describe the density distribution and its change of the data point in its local area; according to the Laplace density gradient of each data point, an association matrix is ​​constructed to represent the degree of association between data points; Step 2: Based on the incidence matrix, construct a weighted undirected graph, calculate the generalized Laplace matrix of the weighted undirected graph, and construct an optimization problem based on the generalized Laplace matrix. By solving the optimization problem, find an embedding matrix, map the data points to a low-dimensional space, and obtain an embedded representation; Step 3: Based on the low-dimensional embedding representation of the data points, define the objective function of query optimization; obtain the update rule of the optimal embedding representation by taking the partial derivative of the objective function with respect to the embedding representation and setting it equal to zero; use the optimal embedding representation to design a dynamic index structure so that queries can be performed efficiently in the embedding space.

2. The method for optimizing a relational cluster database based on multimodal learning according to claim 1, It is characterized in that Step 1 specifically includes: determining each data point Neighborhood range , thereby determining the data points Data points in the neighborhood of ; Calculate for each data point The local density gradient ; For each data point , through the Gaussian kernel function to the data points in the neighborhood The distances are weighted summed to obtain the data point pairs in the neighborhood. The influence of; combining the local density gradient and the neighborhood contribution, calculate the data point Laplace density gradient ; Local density gradient Reflects the data points Relative to the density change of its surrounding neighborhood, if ,illustrate The density of is higher than that of its neighborhood and is the cluster center. ,illustrate The density is low and is in a sparse area. According to the Laplace density gradient of each data point, an association matrix is ​​constructed to represent the degree of association between data points.

3. The method for optimizing relational cluster database based on multimodal learning as claimed in claim 2, It is characterized in that Laplace density gradient The calculation formula is as follows: ; in, It is the data point The neighborhood density of is summed up, Represents data points The neighborhood of Determine, indicating that the radius is All data points in the region The neighborhood set ; Represents data points With neighboring data points The distance between With The relative position of right the extent of the impact; is a Gaussian kernel function used to transform the distance Converted to a weight, the smaller the distance, the greater the weight, indicating right The greater the impact, is the scale parameter, which controls the degree of smoothing within the neighborhood; is a density adjustment parameter used to control the data points in the neighborhood right The greater the impact strength, The value represents the data point pair in the neighborhood The greater the impact, the smaller the Then it represents the data point pair of the neighborhood The smaller the impact.

4. The method for optimizing a relational cluster database based on multimodal learning according to claim 3, It is characterized in that Incidence Matrix No. Row, No. Elements of a column Defined as: ; in, For data points and data points The association weight between them, the larger the value, the stronger the association; and The data points and data points Laplace density gradient of ; is the first-order L1 norm, is the second-order L1 norm; is a preset amplification parameter that controls the influence of density gradient direction similarity on the association weight; for and The Euclidean distance between is the preset gradient change rate smoothing parameter, which controls the influence of density gradient change rate difference on the association weight; and The data points and data points The gradient of the Laplace density gradient.

5. The method for optimizing relational cluster database based on multimodal learning as claimed in claim 4, It is characterized in that In step 2, a weighted undirected graph is constructed based on the association matrix , node data point set Contains data points corresponding to all data ; is the total number of data points; the edge set According to the correlation matrix Determine that the edge weight is , which represents the data point and The strength of the association between them; the degree matrix of a weighted undirected graph is a diagonal matrix whose elements are defined as: ; Represents node data points The degree of is the sum of the weights of the edges connected to it; define the generalized Laplace matrix for: ; in, is the standard graph Laplacian matrix; is the square of the Laplacian matrix, capturing the second-order neighborhood information; is the preset first regularization parameter; is the inverse square root of the degree matrix.

6. The method for optimizing relational cluster database based on multimodal learning as claimed in claim 5, It is characterized in that The process of calculating the low-dimensional embedding representation of the data point in step 2 specifically includes: constructing an optimization problem whose objective function is for: ; The objective function aims to find an embedding matrix , which can preserve the local structure and keep consistent with the original features; is the low-dimensional embedding matrix of the data points, is the embedding dimension; the relational cluster database is regarded as a matrix, and the original feature matrix of the relational cluster database is obtained through feature extraction , is the dimension of the original feature; is the trace operation of the matrix, i.e. the sum of the diagonal elements; is the first-order F-norm; is the preset second regularization parameter, which controls the degree of deviation between the embedding result and the original feature; is the identity matrix; the constraints are: ; Ensure that the embedded data is orthogonal in the new space to prevent degenerate embedding results.

7. The method for optimizing relational cluster database based on multimodal learning according to claim 6, It is characterized in that In step 2, in order to solve the optimization problem, the Lagrangian function is constructed for: ; in, is the Lagrange multiplier matrix; for Take the derivative and set it to zero: ; The solution is obtained by simplifying the following formula: ; Among them, due to ,get ; Using the above equation, solve the following generalized eigenvalue problem: ; Solving this equation, we get the embedding matrix , where each element It is a data point A low-dimensional embedding representation of .

8. The method for optimizing relational cluster database based on multimodal learning as claimed in claim 7, It is characterized in that In step 3, we define the objective function of query optimization based on the low-dimensional embedding representation of the data points. for: ; in, Represents the set of data point pairs involved in the query workload; For data point pairs The query weights of , which reflect the frequency with which they appear in the query; is the preset third regularization parameter.

9. The method for optimizing relational cluster database based on multimodal learning as claimed in claim 8, It is characterized in that In step 3, the objective function about Taking the partial derivative and setting it equal to zero, we get ; In this way, we can get the objective function of query optimization The first item The partial derivative of is: ; in, To query the relevant Laplacian matrix, it is defined as: ; in, and and Same, both are subscript indexes; the objective function of query optimization The second item The partial derivative of ; Item 3 The partial derivative of ; The objective function of query optimization Adding the first, second, and third partial derivatives of gives the following equation: ; After simplifying the equation, we get: ; Finding the optimal embedding representation for: 。 10. The method for optimizing relational cluster database based on multimodal learning according to claim 9, It is characterized in that In step 3, the optimal embedding representation is used The process of designing a dynamic index structure includes: since the embedding space is low-dimensional, use Kd tree, ball tree or LSH as the index structure; represent the embedding Each data point in Insert into the index structure; when the data in the database changes, update the embedded representation accordingly and index structures.

Citation Information

Patent Citations

  • Node representation learning and community division method based on multi-order neighbor information fusion

    CN118568534A

  • Supervised Nonnegative Matrix Factorization

    US20120041906A1

  • Unsupervised data dimensionality reduction method based on adaptive nearest neighbor graph embedding

    WO2022178977A1

Cited By

  • Multi-modal test data uplink method, system and equipment

    CN119938792A

  • Data aggregation method based on multi-modal features

    CN120372555A

  • A data aggregation method based on multi-modal features

    CN120372555B