A multimodal data analysis method and system based on improved genetic algorithm

Abstract

The present invention belongs to the field of big data analysis, and specifically relates to a multimodal data analysis method, system, and device based on an improved genetic algorithm. The present invention optimizes the problem-solving layer in a partition order product space to obtain a target problem-solving layer. The method comprises the following steps: S1: constructing an adaptive genetic algorithm. The construction process is as follows: S11: designing a custom fitness function. S12: improving the selection operator to a hybrid selection operator. S13: improving the crossover operator to an adaptive crossover operator. S14: improving the mutation operator to an adaptive mutation operator. S2: using a partition order product space model to describe multimodal data: randomly generating an initialization population. S3: setting the maximum iteration rounds, and using an adaptive genetic algorithm to iteratively update the initialization population to obtain the desired target problem-solving layer. The present invention solves the problems of existing classical genetic algorithms having difficulty in convergence and being prone to falling into local optimality.

Description

Technical Field

[0001] The present invention belongs to the field of big data analysis, and specifically relates to a multimodal data analysis method, system, and device based on an improved genetic algorithm. Background Art

[0002] Multimodal data refers to data obtained from the same description object from different angles or perspectives, where each angle or perspective is called a modality. For example, human touch, hearing, vision, and smell; information media, such as voice, video, and text; and a variety of sensors, such as radar, infrared, and accelerometers. Each of the above can be called a modality. With the development of data collection methods and multimedia technology, multimodal data has been widely used in practical applications such as medical diagnosis.

[0003] In real life, each modality data often contains multiple levels, and different levels imply different knowledge. Problems can be solved more effectively and accurately at the appropriate level. For example, in medical diagnosis, doctors often need to perform multiple types of examinations on patients. Each examination contains multiple items at different levels of granularity. Each type of examination can be regarded as a modality. Doctors usually flexibly choose different modalities and different levels of modalities, and combine these examination results to obtain more accurate diagnosis results.

[0004] Multimodal data is a type of multivariate data with multiple levels and dimensions. The rich information contained in it can improve the accuracy of various conclusions based on multimodal data. However, the complexity of multivariate data also creates difficulties in the data analysis process. Partition order product space is a model based on granular computing that describes the same object from multiple perspectives and levels, where each perspective contains multiple levels. Therefore, we use partition order product space to describe multimodal data, where each perspective corresponds to a modality, and each perspective contains multiple levels of modal data. The partition order product space is essentially a lattice structure composed of multiple problem-solving layers, each of which is a combination of multiple single-level perspectives. It can describe the problem from different perspectives and different levels, and solve the problem on this basis.

[0005] The more perspectives and levels there are in the partition order product space, the more corresponding problem solving layers there are. Therefore, the selection of problem solving layers in the partition order product space is a typical NP-hard problem. Genetic algorithms, as an intelligent algorithm, can effectively solve this type of NP-hard problem; however, there are some limitations of classical genetic algorithms. For example, classical genetic algorithms will continuously reduce the diversity of the population during the optimization process, making it easy to fall into local optimality and cause immature convergence; secondly, once the population falls into a local optimal solution, it is difficult to get rid of the local optimal solution. Summary of the invention

[0006] In order to solve the problems of convergence difficulty and easy falling into local optimum in the problem-solving layer selection process of multimodal data models in existing classical genetic algorithms, the present invention provides a multimodal data analysis method, system and device based on an improved genetic algorithm.

[0007] The present invention is implemented by the following technical solutions:

[0008] A multimodal data analysis method based on an improved genetic algorithm is used to optimize the problem-solving layer in a partitioned order product space generated according to multimodal data to obtain the best target problem-solving layer; and ultimately achieve the purpose of determining the best solution based on the target problem-solving layer.

[0009] The multimodal data analysis method provided by the present invention specifically comprises the following steps:

[0010] S1: Construct an adaptive genetic algorithm based on the improved classical genetic algorithm. The construction process of the adaptive genetic algorithm is as follows:

[0011] S11: According to the classification accuracy and granularity requirements of the problem-solving layer, a custom fitness function is designed to balance the two indicators.

[0012] S12: Improve the selection operator of the classic genetic algorithm to a hybrid selection operator that changes with the iteration process. The hybrid selection operator uses both the elite retention operator and the "double-individual no-replacement" tournament selection operator in the early stage of evolution; in the late stage of evolution, only the classic tournament selection operator is used.

[0013] S13: Improve the crossover operator of the classic genetic algorithm into an adaptive crossover operator whose crossover probability is dynamically adjusted with the iteration process. This makes the crossover probability of individuals in the population in the early stage of evolution greater than that in the late stage of evolution, and the fitness value of the individual is negatively correlated with the crossover probability. Among them, the crossover method of the adaptive crossover operator is two-point crossover.

[0014] S14: Improve the mutation operator of the classic genetic algorithm into an adaptive mutation operator whose mutation probability is dynamically adjusted with the iteration process. The adaptive mutation operator mutates multiple times, and only when the similarity of the individual before and after mutation is less than a given threshold and the fitness value of the individual after mutation is greater than the fitness value of the individual before mutation, the mutated individual is retained; otherwise, the mutation continues until the maximum number of mutations is reached.

[0015] S2: Use the partitioned order product space model to describe multimodal data: convert the problem-solving layer into a chromosome of a specified length through real number encoding, and randomly generate an initialization population containing multiple chromosomes.

[0016] S3: Set the maximum iteration rounds, use the adaptive genetic algorithm to iteratively update the initial population, and obtain the required target problem solution layer after the iterative update. The detailed process of this step is as follows:

[0017] S31: Calculate the fitness value of each chromosome individual in the current population.

[0018] S32: Obtain the current iteration round and calculate the phase type of the iteration process. The phase type is divided into an early evolution phase and a late evolution phase.

[0019] S33: Execute different operation strategies according to the stage type of the current iteration round, and then update the population of the current round through selection operation, crossover operation and mutation operation.

[0020] S34: Determine whether the current iteration round has reached the maximum iteration round: if yes, calculate the fitness value of each chromosome individual in the current population, and take the individual with the maximum fitness value as the required target problem solving layer. Otherwise, loop the processing process of steps S31-S33.

[0021] As a further improvement of the present invention, in step S11, the formula of the fitness function f is as follows:

[0022]

[0023] In the above formula, l j represents a problem-solving layer, D is the decision attribute set of the decision system, γ β (l j , D) represents the problem solving layer l j Regarding the classification accuracy of D; d(l j ) represents the granularity of the problem solving layer lj; w 1 and w 2 are the weights of classification accuracy and granularity, and satisfy 0≤w 1 ≤1,0≤w 2 ≤1; γ is the given target classification accuracy; d(l n ) is the most detailed problem solving layer l n The granularity.

[0024] As a further improvement of the present invention, the construction process of the fitness function f is as follows:

[0025] (1) Given a decision system DS = (U, AT = C ∪ D, V, f), where U = {x 1 , x 2 , ..., x n} is a non-empty finite set of objects, called the domain; AT is a non-empty finite attribute set, C is the conditional attribute set, and D is the decision attribute set; V is the value range of the attribute set AT, f: U×AT→V is the information function, f(x, a)∈V a Represents the value of object x∈U on attribute a∈AT; the partition order product space corresponding to the decision system DS From m perspectives v i In the problem solving layer j middle, Represents the viewing angle v i The jth i Levels, n i is the viewing angle v i The number of levels, 1≤i≤m, 1≤j i ≤n i .

[0026] (2) Assume that U / D = {X 1 , X 2 , ..., X r}, then the problem solving layer l j Classification accuracy γ about D β (l j , D) is defined as:

[0027]

[0028] in,

[0029]

[0030] In the above formula, X j represents the decision class of the domain U with respect to the decision attribute set D, j = {1, 2, ..., r}; β is the multi-granularity fusion threshold, and 0≤β≤1.

[0031] (3) For a given problem, find the solution layer Then the particle size g(l j ) is defined as:

[0032]

[0033] From the definition, we know that g(l j ) The larger the problem solving level l j The smaller the particle size, the coarser it is.

[0034] (4) Based on classification accuracy γ β (l j , D) and particle size g(l j), in order to make the granularity of the problem-solving layer as coarse as possible while meeting the classification accuracy requirements, the following fitness function f is constructed:

[0035]

[0036] As a further improvement of the present invention, in the adaptive genetic algorithm, the type division rule of the iterative process is as follows:

[0037] Assume that the maximum number of iterations is set to I, the current number of iterations is i, 1≤i≤I; a threshold α is given for classifying the types of iterative processes, 0<α<1; then when When , the i-th iteration is judged to be in the early stage of evolution; when , it is determined that the i-th iteration is in the late stage of evolution.

[0038] As a further improvement of the present invention, in step S12, the process of the hybrid selection operator simultaneously executing two selection strategies in the early stage of evolution is as follows:

[0039] First, use the elite retention operator to copy some individuals with larger fitness values ​​in the current population to the next generation population. Then use the "double-individual without replacement" tournament selection operator to select individuals from the current population until the number of individuals in the next generation population is consistent with the original size of the current population.

[0040] Among them, the "two individuals without replacement" tournament selection operator only selects two individuals to participate in the tournament in each round of iteration, and copies the winning individual to the next generation population. At the same time, the winning individual will be deleted from the current population.

[0041] As a further improvement of the present invention, in the adaptive crossover operator of step S13, the two parent individuals p to be crossovered 1 、p 2 Adaptive crossover probability P c The adjustment formula is as follows:

[0042]

[0043] In the above formula, f(p 1 ) and f(p 2 ) represent p 1 and p 2 The fitness value of max represents f(p 1 ) and f(p 2 ) maximum value; P cpre Indicates the maximum crossover probability in the early stage of population evolution; P cbeh Represents the maximum crossover probability in the late stage of population evolution; and satisfies: P cpre >P cbeh; σ is the standard deviation of the fitness values ​​of all individuals in the population.

[0044] As a further improvement of the present invention, in the adaptive mutation operator of step S14, the adaptive mutation probability P of the individual p currently undergoing mutation is m The adjustment formula is as follows:

[0045]

[0046] In the above formula, f(p) is the fitness value of chromosome individual p; P mpre is the maximum mutation probability in the early stage of population evolution; P mbeh is the maximum mutation probability in the late stage of population evolution; and P mpre <P mbeh ;σ is the standard deviation of all individual fitness values ​​in the population;

[0047] For any two chromosomes p i and p j , the similarity SD(p i , p j ) is calculated as follows:

[0048]

[0049] Among them, p sub Yes i and p j The largest common substring of sub =sub(p i , p j );len(p sub ) represents the length of the largest common substring; m represents the length of the chromosome, and the value of m also corresponds to the number of perspectives in the problem-solving layer.

[0050] As a further improvement of the present invention, in step S14, in each chromosome generated by real number coding in step S2, each gene position corresponds to a perspective, and the gene value on each gene position corresponds to the hierarchical number under the perspective.

[0051] The present invention also provides a multimodal data analysis system based on an improved genetic algorithm. The analysis system adopts the multimodal data analysis method based on the improved genetic algorithm as mentioned above to optimize a problem solving layer in a partitioned order product space generated based on multimodal metadata; and obtain the optimal target problem solving layer.

[0052] The multimodal data analysis system provided by the present invention comprises: a real number encoding module, a population initialization module, a genetic algorithm processing module, and an output module.

[0053] The real number coding module is used to describe the data structure of multimodal metadata using a partition order product space model, and then convert the problem solving layer in the partition order product space into a chromosome of a specified length through real number coding.

[0054] The population initialization module is used to randomly generate an initial population including multiple chromosomes.

[0055] The genetic algorithm processing module is used to iteratively update the initial population through selection, crossover and mutation operations. Among them, the genetic algorithm processing module includes a progress classification unit, a loop control unit, a fitness calculation unit, a hybrid selection unit, an adaptive crossover unit, and an adaptive mutation unit. The progress classification unit is used to determine the stage type corresponding to the current iteration round according to a preset rule, and the stage type is divided into the early evolutionary stage and the late evolutionary stage. The loop control unit is used to determine whether the current iteration round has reached the maximum number of iterations. If so, it will jump out of the iteration loop, otherwise it will continue to iteratively update the current population through the fitness calculation unit, the hybrid selection unit, the adaptive crossover unit, and the adaptive mutation unit. The fitness calculation unit is used to calculate the fitness value of each chromosome in the current population according to a custom fitness function before each round of iteration. The hybrid selection unit is used to select different selection operators according to the stage type corresponding to the current iteration round, and then select samples from the current population to enter the next iteration round. The adaptive crossover unit uses an adaptive crossover operator whose crossover probability is dynamically adjusted with the iteration process to match individuals in the current population in pairs, and perform double-point crossover on a pair of chromosomes that meet the requirements. The adaptive mutation unit uses an adaptive mutation operator whose mutation probability is dynamically adjusted along with the iteration process to mutate the selected chromosome multiple times until the mutated chromosome meets the requirements.

[0056] The output module is used to select the individual with the largest fitness in the population after iterative updating as the required target problem solving layer after the iterative updating of the population by the genetic algorithm processing module reaches the maximum number of iterations.

[0057] The present invention also includes a multimodal data analysis device based on an improved genetic algorithm, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The functional modules in the multimodal data analysis device are deployed in the manner of the multimodal data analysis system based on the improved genetic algorithm as described above. When the processor executes the computer program, the steps of the multimodal data analysis method based on the improved genetic algorithm as described above are implemented, and then the best target problem solution layer is selected in the partitioned order product space.

[0058] The technical solution provided by the present invention has the following beneficial effects:

[0059] When dealing with multimodal data analysis problems, the present invention first adopts a novel multimodal multi-granularity fusion model to describe the multimodal data, divides the data under each modality into multiple levels, and can mine useful information at multiple levels of the multimodal data.

[0060] Secondly, this embodiment also designs a new type of adaptive genetic algorithm to complete the task of finding the best problem-solving layer from the partition order product space. The selection operator, crossover operator and mutation operator of the improved adaptive genetic algorithm of the present invention can be automatically adjusted as the state changes in the population iteration stage, thereby achieving the purpose of ensuring classification accuracy and improving population diversity in the full cycle of population iteration.

[0061] The improvement of the present invention not only improves the convergence speed of the genetic algorithm, but also improves the ability of the population in the late stage of evolution to get rid of the local optimal solution. The problem solving layer selected by the adaptive genetic algorithm provided by the present invention has a coarser granularity level while ensuring the classification accuracy, thereby improving the efficiency of problem solving and reducing the cost of data acquisition. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 A flowchart of the steps of a multimodal data analysis method based on an improved genetic algorithm provided in Example 1 of the present invention.

[0063] Figure 2 A flowchart of a multimodal data analysis method based on an improved genetic algorithm provided in Example 1 of the present invention.

[0064] Figure 3 It is a schematic diagram of the structure of a typical chromosome according to the problem solving layer conversion in Example 1 of the present invention.

[0065] Figure 4 It is a structural diagram of the partitioned order product space drawn according to the data structure of the student score table in Example 1 of the present invention.

[0066] Figure 5 This is a curve showing the change of the difference between the optimal individual classification accuracy of the population and the target classification accuracy and the ratio thereof to the granularity of the finest problem solving layer in each iteration of IGA in Example 1 of the present invention.

[0067] Figure 6 It is a variation curve of the difference between the optimal individual classification accuracy of the population and the target classification accuracy and the ratio to the finest problem solving layer granularity in each iteration of IGA and IGA1 in Example 1 of the present invention.

[0068] Figure 7 This is a curve showing the change of the difference between the optimal individual classification accuracy of the population and the target classification accuracy and the ratio to the finest problem solving layer granularity in each iteration of IGA and IGA2 in Example 1 of the present invention.

[0069] Figure 8 This is a curve showing the change of the difference between the optimal individual classification accuracy of the population and the target classification accuracy and the ratio to the finest problem solving layer granularity in each iteration of IGA and IGA3 in Example 1 of the present invention.

[0070] Fig. 9 This is a curve showing the change of the difference between the optimal individual classification accuracy of the population and the target classification accuracy and the ratio to the finest problem solving layer granularity in each iteration of IGA and IGA4 in Example 1 of the present invention.

[0071] Fig.10 A schematic diagram of the system framework of a multimodal data analysis system based on an improved genetic algorithm provided in Example 2 of the present invention.

[0072] Fig.11 A module schematic diagram of a multimodal data analysis system based on an improved genetic algorithm provided in Example 2 of the present invention. DETAILED DESCRIPTION

[0073] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0074] Example 1

[0075] This embodiment provides a multimodal data analysis method based on an improved genetic algorithm, which is used to optimize the problem-solving layer in a partitioned order product space generated according to multimodal data to obtain the best target problem-solving layer; and ultimately achieve the purpose of determining the best solution based on the target problem-solving layer.

[0076] like Figure 1 and Figure 2 As shown, the multimodal data analysis method provided in this embodiment specifically includes the following steps:

[0077] S1: Construct an adaptive genetic algorithm based on the improved classical genetic algorithm. The construction process of the adaptive genetic algorithm is as follows:

[0078] S11: According to the classification accuracy and granularity requirements of the problem-solving layer, a custom fitness function is designed to balance the two indicators.

[0079] The formula of the fitness function f is as follows:

[0080]

[0081] In the above formula, l jrepresents a problem-solving layer, D is the decision attribute set of the decision system, γ β (l j , D) represents the problem solving layer l j Regarding the classification accuracy of D; d(l j ) represents the problem solving layer l j The particle size of 1 and w 2 are the weights of classification accuracy and granularity, and satisfy 0≤w 1 ≤1,0≤w 2 ≤1; γ is the given target classification accuracy; d(l n ) is the most detailed problem solving layer l n The granularity.

[0082] S12: Improve the selection operator of the classic genetic algorithm to a hybrid selection operator that changes with the iteration process. The hybrid selection operator uses both the elite retention operator and the "double-individual no-replacement" tournament selection operator in the early stage of evolution; in the late stage of evolution, only the classic tournament selection operator is used.

[0083] The process of the hybrid selection operator executing two selection strategies simultaneously in the early stage of evolution is as follows:

[0084] First, use the elite retention operator to copy some individuals with larger fitness values ​​in the current population to the next generation population. Then use the "double-individual without replacement" tournament selection operator to select individuals from the current population until the number of individuals in the next generation population is consistent with the original size of the current population.

[0085] Among them, the "two individuals without replacement" tournament selection operator only selects two individuals to participate in the tournament in each round of iteration, and copies the winning individual to the next generation population. At the same time, the winning individual will be deleted from the current population.

[0086] The rules for classifying iterative processes are as follows:

[0087] Assume that the maximum number of iterations is set to I, the current number of iterations is i, 1≤i≤I; a threshold α is given for classifying the types of iterative processes, 0<α<1; then when When , the i-th iteration is judged to be in the early stage of evolution; when , it is determined that the i-th iteration is in the late stage of evolution.

[0088] S13: Improve the crossover operator of the classic genetic algorithm into an adaptive crossover operator whose crossover probability is dynamically adjusted with the iteration process. This makes the crossover probability of individuals in the population in the early stage of evolution greater than that in the late stage of evolution, and the fitness value of the individual is negatively correlated with the crossover probability. Among them, the crossover method of the adaptive crossover operator is two-point crossover.

[0089] In the adaptive crossover operator, the two parent individuals p that undergo crossover 1 、p 2 Adaptive crossover probability P c The adjustment formula is as follows:

[0090]

[0091] In the above formula, f(p 1 ) and f(p 2 ) represent p 1 and p 2 The fitness value of max represents f(p 1 ) and f(p 2 ) maximum value; P cpre Indicates the maximum crossover probability in the early stage of population evolution; P cbeh Represents the maximum crossover probability in the late stage of population evolution; and satisfies: P cpre >P cbeh ; σ is the standard deviation of the fitness values ​​of all individuals in the population.

[0092] S14: Improve the mutation operator of the classic genetic algorithm into an adaptive mutation operator whose mutation probability is dynamically adjusted with the iteration process. The adaptive mutation operator mutates multiple times, and only when the similarity of the individual before and after mutation is less than a given threshold and the fitness value of the individual after mutation is greater than the fitness value of the individual before mutation, the mutated individual is retained; otherwise, the mutation continues until the maximum number of mutations is reached.

[0093] In the adaptive mutation operator, the adaptive mutation probability P of the individual p currently undergoing mutation is m The adjustment formula is as follows:

[0094]

[0095] In the above formula, f(p) is the fitness value of chromosome individual p; P mpre is the maximum mutation probability in the early stage of population evolution; P mbeh is the maximum mutation probability in the late stage of population evolution; and P mpre <P mbeh ;σ is the standard deviation of all individual fitness values ​​in the population;

[0096] S2: Use the partitioned order product space model to describe multimodal data: convert the problem-solving layer into a chromosome of a specified length through real number encoding, and randomly generate an initialization population containing multiple chromosomes.

[0097] In each chromosome generated by real number coding in this embodiment, each gene position corresponds to a perspective, and the gene value on each gene position corresponds to the hierarchical number under the perspective. Chromosomes such as Figure 3 As shown:

[0098] S3: Set the maximum iteration rounds, use the adaptive genetic algorithm to iteratively update the initial population, and obtain the required target problem solution layer after the iterative update. The detailed process of this step is as follows:

[0099] S31: Calculate the fitness value of each chromosome individual in the current population.

[0100] S32: Obtain the current iteration round and calculate the phase type of the iteration process. The phase type is divided into an early evolution phase and a late evolution phase.

[0101] S33: Execute different operation strategies according to the stage type of the current iteration round, and then update the population of the current round through selection operation, crossover operation and mutation operation.

[0102] S34: Determine whether the current iteration round has reached the maximum iteration round: if yes, calculate the fitness value of each chromosome individual in the current population, and take the individual with the maximum fitness value as the required target problem solving layer. Otherwise, loop the processing process of steps S31-S33.

[0103] The technical solution provided by the present invention mainly converts the solution decision problem based on multimodal data analysis into the problem of finding the best problem-solving layer in the partition order product space, and then uses a genetic algorithm to process the optimization task of the problem-solving layer.

[0104] The following example uses a scholarship evaluation to illustrate how to convert a student's multi-course score table into a multimodal data, and then generate the corresponding partition order product space. Table 1 below gives a typical student score statistics table:

[0105] Table 1: Comprehensive performance list of students participating in the scholarship evaluation

[0106]

[0107] The data in Table 1 describe students' performance from three perspectives: Chinese performance perspective, 1 、Mathematics performance perspective 2 and English performance perspective 3 , 8 students form a domain U = {x 1 , x 2 , x 3 , x 4 , x 5, x 6 , x 7 , x 8}. To simplify the discussion, each perspective v i Contains only one attribute.

[0108] From the perspective of language performance 1 In, a 1 、a 2 The attribute value sequences corresponding to two scales are {O, G} and {90-100, 50-60, 40-50}. 1 、a 2 We can get the partition order on U That is v 1 =(P 1 ,≤), and Corresponding to v 1 Similarly, from the perspective of mathematics performance 2 Middle, b 1 , b 2 Corresponding to the attribute value sequence of 2 scales, based on b 1 , b 2 We can get the partition order on U Right now and Corresponding to v 2 From the perspective of English performance 3 In, c 1 、c 2 Corresponding to the attribute value sequence of 2 scales, based on c 1 、c 2 We can get the partition order on U Right now and Corresponding to v 3 There are two levels of O, where O stands for excellent and G stands for good.

[0109] For the viewing angle v 1 , and its two levels are described as follows:

[0110] For the viewing angle v 2 , and its two levels are described as follows:

[0111] For the viewing angle v 3 , and its two levels are described as follows:

[0112] It can be seen that the viewing angle v 1 Level of satisfaction The corresponding equivalence class satisfies Viewing angle 2 Level of satisfaction The corresponding equivalence class satisfies Viewing angle 3 Level of satisfaction The corresponding equivalence class satisfies

[0113] These three perspectives can form a partition order product space The structure of the partition order product is as follows Figure 4 As shown. Combined Figure 4 It can be found that the partitioned order product space combines a multi-level granular structure and a multi-perspective granular structure, with a total of 8 problem-solving layers. Therefore, when conducting scholarship reviews, you can flexibly select and switch between different perspectives and different levels as needed to obtain more accurate review results. The multimodal data analysis method based on an improved genetic algorithm provided in this embodiment is used to select the best solution for evaluating student performance from multiple strategies at different perspectives and different levels.

[0114] This embodiment uses a genetic algorithm to select the best problem-solving layer that meets the preset restrictive conditions in the partition order product space. The processing method of the genetic algorithm is to iteratively update the chromosomes representing each problem-solving layer and select the best problem-solving layer. In order to overcome the defects of the slow convergence speed and easy to fall into the local optimum of the classical genetic algorithm, this embodiment improves the genetic algorithm from four aspects: fitness function, selection operator, crossover operator and mutation operator. Then, an algorithm is obtained that can accurately evaluate the performance of the problem-solving layer, and can dynamically adjust the various operators used for population update in the iteration with the number of iterations. The processing strategy of the adjusted operator can be adapted to the characteristics of the population at different stages, and maintain the diversity of the population on the basis of accelerating the convergence rate of the algorithm; thereby significantly improving the disadvantage that the algorithm is prone to fall into the local optimum.

[0115] The following describes in detail the four improvements and advantages of the improved adaptive genetic algorithm provided in this embodiment:

[0116] 1. Fitness Function

[0117] When dealing with the problem of optimizing the problem-solving layer, the genetic algorithm can describe the problem more abstractly and briefly by omitting detailed information at a coarse level in order to improve the processing efficiency of the algorithm. However, the coarse-level analysis ignores too much detailed information and may also lead to low accuracy of the solution results. On the contrary, at a finer level, a more specific and detailed problem description will reduce the efficiency of problem solving, but the accuracy of the solution result is higher. Therefore, this embodiment sets corresponding restrictive conditions to ensure that the problem-solving layer selected in the partition order product space can meet the requirements of classification accuracy; at the same time, it is also necessary to ensure that the granularity of the problem-solving layer is as coarse as possible, thereby improving the efficiency of problem solving.

[0118] Based on the above reasons, this embodiment customizes a new fitness function for the genetic algorithm according to the classification accuracy and granularity of the problem solving layer. The construction process of the customized fitness function f in this embodiment is as follows:

[0119] (1) Given a decision system DS = (U, AT = C ∪ D, V, f), where U = {x 1 , x 2 , ..., x n} is a non-empty finite set of objects, called the domain; AT is a non-empty finite attribute set, C is the conditional attribute set, and D is the decision attribute set; V is the value range of the attribute set AT, f: U×AT→V is the information function, f(x, a)∈V a Represents the value of object x∈U on attribute a∈AT; the partition order product space POPS corresponding to the decision system DS m =(×P i , ≤P) from m perspectives v i In the problem solving layer j middle, Represents the viewing angle v i The jth i Levels, n i is the viewing angle v i The number of levels, 1≤i≤m, 1≤j i ≤n i .

[0120] (2) Assume that U / D = {X 1 , X 2 , ..., X r}, then the problem solving layer l j Classification accuracy γ about D β (l j , D) is defined as:

[0121]

[0122] in,

[0123]

[0124] In the above formula, X j represents the decision class of the domain U with respect to the decision attribute set D, j = {1, 2, ..., r}; β is the multi-granularity fusion threshold, and 0≤β≤1.

[0125] (3) For a given problem, find the solution layer Then j The particle size g(l j ) is defined as:

[0126]

[0127] From the definition, we know that g(l j ) The larger the problem solving level l j The smaller the particle size, the coarser it is.

[0128] (4) Based on classification accuracy γ β (l j , D) and particle size g(l j ), in order to make the granularity of the problem-solving layer as coarse as possible while meeting the classification accuracy requirements, the following fitness function f is constructed:

[0129]

[0130] 2. Hybrid Selection Operator

[0131] Elite retention operator and tournament selection operator are commonly used population selection strategies in genetic algorithms. Among them, although the elite retention operator can ensure that the best individuals of the current population are not eliminated, individuals with small fitness values ​​but excellent gene fragments will be eliminated. The tournament selection operator is a selection operator with replacement sampling. During the selection process, when the scale of individuals participating in the tournament is large, individuals with small fitness values ​​but excellent gene fragments still find it difficult to enter the next generation population. And this replacement strategy will make the repetition of individuals in the population higher, thereby reducing the diversity of the population. This embodiment organically combines these two conventional sample selection strategies, and designs a special hybrid selection operator in combination with the characteristics of the population at different stages.

[0132] Considering that the sample diversity of the population is high in the early stage of iteration, the hybrid selection operator provided in this embodiment first uses the elite retention operator to copy some individuals with larger fitness values ​​in the current population to the next generation population in the early stage of evolution, and then uses the "two individuals without replacement" tournament selection operator to select individuals until the number of individuals in the next generation population reaches the original scale. The "two individuals without replacement" tournament selection operator only selects two individuals to participate in the tournament in each iteration, and then copies the winning individuals to the next generation population. At the same time, the operator will also delete the obtained individuals from the current population to avoid repeated selection and improve the diversity of the next generation population. In the late stage of evolution, the individuals in the population gather near a local optimal solution, so only the classic tournament selection operator is used at this time.

[0133] 3. Adaptive Crossover Operator

[0134] This embodiment takes into account that in the early stage of evolution, the overall fitness value of the population is small; therefore, in order to expand the search range and produce better individuals, the crossover probability of chromosomes is increased. In the late stage of evolution, individuals in the population gather near a local optimal solution, and it is difficult to produce new excellent individuals by crossover; therefore, in order to retain excellent individuals, the crossover probability is reduced. At the same time, in order to ensure that the overall genotype of the population is more optimized during the entire evolutionary process, this embodiment intends to make individuals with smaller fitness values ​​obtain a larger crossover probability to optimize their genotypes, while individuals with larger fitness values ​​should obtain a smaller crossover probability to retain their excellent genotypes.

[0135] In order to achieve the above objectives, this embodiment defines an adaptive crossover operator that takes into account the evolution of local individuals and the overall population. The characteristics of the adaptive crossover operator are as follows:

[0136] Specifically, the crossover operation in this embodiment is completed by a two-point crossover method. 1 and p 2 When f(p 1 ) and f(p 2 ) is p 1 and p 2 The fitness value, f max =max(f(P 1 ), f(P 2 )) is f(p 1 ) and f(p 2 ). cpre is the maximum crossover probability in the early stage of population evolution, P cbeh is the maximum crossover probability in the late stage of population evolution, and P cpre >P cbeh σ is the standard deviation of the fitness values ​​of all individuals in the population.

[0137] Then p1 、p 2 Adaptive crossover probability P c Defined as:

[0138]

[0139] 4. Adaptive Mutation Operator

[0140] The mutation operation enriches the genotype of the population by mutating the genes on a single chromosome, thereby achieving the purpose of optimizing the population. In the early stage of evolution, there are fewer excellent individuals in the population, and the mutation probability needs to be reduced to retain the excellent individuals; in the late stage of evolution, the individuals in the population gather near a local optimal solution, and the mutation probability needs to be increased to get rid of the local optimal solution. At the same time, during the entire evolutionary process, individuals with smaller fitness values ​​should obtain a larger mutation probability to better get rid of the local optimal solution; individuals with larger fitness values ​​should obtain a smaller mutation probability to retain their excellent genotypes.

[0141] In order to achieve the above objectives, this embodiment defines an adaptive mutation operator that considers the similarity of individuals after mutation and can be dynamically adjusted with the iteration cycle. The characteristics of the adaptive mutation operator are as follows:

[0142] Given the individual p that is currently undergoing mutation, f(p) is the fitness value of p. mpre is the maximum mutation probability in the early stage of population evolution, P mbeh is the maximum mutation probability in the late stage of population evolution, and P mpre <P mbeh σ is the standard deviation of the fitness values ​​of all individuals in the population. Then the adaptive mutation probability P of p m Defined as:

[0143]

[0144] In addition, in the classic genetic algorithm, when the mutation probability of an individual meets the conditions, the individual is only mutated once, that is, the value of a gene position on the chromosome is converted to another value. Therefore, the similarity between the individual after mutation in the classic genetic algorithm and the original individual is often high, and it is difficult to get rid of the local optimal solution in the late stage of evolution.

[0145] In this embodiment, in order to improve the ability of the genetic algorithm to get rid of the local optimal solution, the adaptive genetic algorithm first sets a maximum number of mutations N when performing the mutation operation, allowing individuals to mutate multiple times within the maximum number of mutations. At the same time, the similarity between individuals is defined. Only when the similarity before and after the individual mutation is less than a given threshold λ and the fitness value of the individual after the mutation is greater than the fitness value of the individual before the mutation, the mutated individual is retained, otherwise the mutation continues until the maximum number of mutations is reached. In this way, it can be ensured that each round of mutation operation can achieve the best gene mutation effect, thereby greatly improving the population optimization effect in the iterative process. Specifically, the characteristics of the similarity between chromosomes defined in this embodiment are as follows:

[0146] For any two chromosomes p i and p j , the similarity SD(p i , p j ) is calculated as follows:

[0147]

[0148] Among them, P sub Yes i and p j The largest common substring of sub =sub(p i , p j );len(p sub ) represents the length of the largest common substring; m represents the length of the chromosome, and the value of m also corresponds to the number of perspectives in the problem-solving layer.

[0149] Performance Testing

[0150] In order to verify the performance of the improved adaptive genetic algorithm provided by this embodiment, this embodiment also formulates corresponding verification experiments.

[0151] 1. Dataset selection

[0152] This implementation selected 9 data sets from UCI and Kaggle to verify the effectiveness of the improved genetic algorithm in this paper. The detailed information of the data sets is shown in Table 2.

[0153] Table 2: Data characteristics of each test data set used in the validation test

[0154]

[0155]

[0156] For each selected data set, the corresponding multi-view and multi-level granular structure is constructed for each data set. First, the conditional attribute set C is divided into multiple subsets Each attribute subset C i are defined as a perspective v i , 1≤i≤m. Then, the viewing angle v i The corresponding attribute subset C i Divide into multiple subsets They satisfy and Constructing perspectives based on nested attribute set sequences i Finally, the partition order product space is constructed based on the multi-level granular structure at multiple perspectives. Table 3 gives the description of the multi-level granular structure of all data sets.

[0157] Table 3: Multi-view and multi-level granular structures corresponding to each data set

[0158]

[0159]

[0160] 2. Performance indicators

[0161] Based on the aforementioned multimodal data, this embodiment evaluates the performance of the provided improved genetic algorithm (IGA) through the following performance indicators and items:

[0162] (1) Convergence of IGA

[0163] In order to verify the convergence of IGA, the following is the definition of the convergence condition: Assuming the target classification accuracy is γ, for any given ε ≥ 0, the optimal individual in the population in the i-th iteration is p i , the corresponding problem is to find the hierarchy l i , l i The classification accuracy is denoted as γ β (l i , D), when it satisfies the formula: γ β (l i , D)≥γ-ε: and l i The particle size g(l i ) no longer changes, the IGA converges and the convergence is independent of the initial population.

[0164] Among all the data sets in Table 2, the spatial complexity of the particle structure corresponding to the data set SCADI is the largest, with a total of 5.8 25In the granular structure corresponding to SCADI, assuming that the multi-granularity fusion threshold β = 0.1, γ is the classification accuracy of the finest problem-solving layer, the difference between the optimal individual classification accuracy of the population and γ in each iteration of IGA and its ratio to the granularity of the finest problem-solving layer is as follows: Figure 5 Analysis Figure 5 From the data in , we can see that after 10 population iterations, the difference between the classification accuracy of the best individual in the population and γ is 0 and remains unchanged. At the same time, the granularity of the optimal problem solving layer remains unchanged after 35 population iterations, which indicates that IGA converges at this time. In addition, in the granular structure and multi-granularity fusion threshold of other data sets, the number of population iterations when IGA converges is less than 50. Therefore, this embodiment sets the maximum number of iterations of IGA in each evolution to 50.

[0165] (2) Ablation experiment

[0166] This embodiment also verifies the effectiveness of all improvements in IGA through a series of ablation experiments. In the ablation experiments, multiple adaptive genetic algorithms with different improvements are designed for control experiments. The differences between the different improved algorithms of the control group and the IGA of the test group are shown in Table 4:

[0167] Table 4: Characteristics of different improved algorithms used as control groups in ablation experiments

[0168]

[0169] In this embodiment, an ablation experiment is conducted in the granular structure corresponding to the dataset SCADI, and the multi-granularity fusion threshold β is set to 0.1. During the experiment, the difference between the optimal individual classification accuracy of the population and the target classification accuracy and the ratio to the finest problem solving layer granularity of IGA, IGA1, IGA2, IGA3 and IGA4 in each iteration are as follows: Figure 6-9 shown.

[0170] Combination Figure 6 It can be seen that the convergence speed of IGA1 is greater than that of IGA, but the classification accuracy of the optimal individual in the population does not reach the target classification accuracy when it converges. At this time, IGA1 converges to the local optimal solution. Figure 7 It can be seen that IGA2 only uses the classic tournament operator to select individuals, and the individuals entering the next generation population have a high degree of repetition, which causes the algorithm to converge to the local optimal solution in advance. Figure 8 It can be seen that IGA3 uses a fixed crossover probability for crossover operations. In the later stages of evolution, a larger crossover probability will destroy the genotypes of excellent individuals, resulting in a larger granularity of the selected problem solution layer. Fig. 9 It can be seen that IGA4 uses the classic mutation operator to mutate the individuals that meet the mutation conditions only once, and it is difficult to get rid of the local optimal solution in the later stage.

[0171] In particular, the control experiments of other ablation experiments conducted in this embodiment on other experimental data sets have similar results to the above example. It can be seen that the performance of each control group in the ablation experiment is inferior to the IGA algorithm provided in this embodiment, that is, the method provided in this embodiment is the best solution to achieve convergence rate improvement and overcome the defects of local optimal solution.

[0172] Example 2

[0173] This embodiment provides a multimodal data analysis system based on an improved genetic algorithm. Fig.10 As shown, the analysis system adopts the multimodal data analysis method based on the improved genetic algorithm as in Example 1 to optimize a problem solving layer in a partitioned order product space generated based on multimodal metadata; obtain the optimal target problem solving layer, and then solve the corresponding multimodal data analysis problem.

[0174] like Fig.11 As shown, the multimodal data analysis system provided in this embodiment includes: a real number encoding module, a population initialization module, a genetic algorithm processing module, and an output module.

[0175] The real number coding module is used to describe the data structure of multimodal metadata using a partition order product space model, and then convert the problem solving layer in the partition order product space into a chromosome of a specified length through real number coding.

[0176] The population initialization module is used to randomly generate an initial population including multiple chromosomes.

[0177] The genetic algorithm processing module is used to iteratively update the initial population through selection, crossover and mutation operations. Among them, the genetic algorithm processing module includes a progress classification unit, a loop control unit, a fitness calculation unit, a hybrid selection unit, an adaptive crossover unit, and an adaptive mutation unit. The progress classification unit is used to determine the stage type corresponding to the current iteration round according to a preset rule, and the stage type is divided into the early evolutionary stage and the late evolutionary stage. The loop control unit is used to determine whether the current iteration round has reached the maximum number of iterations. If so, it will jump out of the iteration loop, otherwise it will continue to iteratively update the current population through the fitness calculation unit, the hybrid selection unit, the adaptive crossover unit, and the adaptive mutation unit. The fitness calculation unit is used to calculate the fitness value of each chromosome in the current population according to a custom fitness function before each round of iteration. The hybrid selection unit is used to select different selection operators according to the stage type corresponding to the current iteration round, and then select samples from the current population to enter the next iteration round. The adaptive crossover unit uses an adaptive crossover operator whose crossover probability is dynamically adjusted with the iteration process to match individuals in the current population in pairs, and perform double-point crossover on a pair of chromosomes that meet the requirements. The adaptive mutation unit uses an adaptive mutation operator whose mutation probability is dynamically adjusted along with the iteration process to mutate the selected chromosome multiple times until the mutated chromosome meets the requirements.

[0178] The output module is used to select the individual with the largest fitness in the population after iterative updating as the required target problem solving layer after the iterative updating of the population by the genetic algorithm processing module reaches the maximum number of iterations.

[0179] Example 3

[0180] This embodiment provides a multimodal data analysis device based on an improved genetic algorithm, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The functional modules in the multimodal data analysis device are deployed in the manner of a multimodal data analysis system based on an improved genetic algorithm as in Example 2. When the processor executes the computer program, the steps of the multimodal data analysis method based on an improved genetic algorithm as in Example 1 are implemented, and then the best target problem solution layer is selected in the partitioned order product space.

[0181] The computer device may be a smart phone, tablet computer, laptop computer, desktop computer, rack server, blade server, tower server or cabinet server (including an independent server or a server cluster composed of multiple servers) capable of executing a program. The computer device of this embodiment includes at least but is not limited to: a memory and a processor that can be interconnected through a system bus.

[0182] In this embodiment, the memory (i.e., readable storage medium) includes flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory, etc.), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory may be an internal storage unit of a computer device, such as a hard disk or memory of the computer device. In other embodiments, the memory may also be an external storage device of a computer device, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the computer device. Of course, the memory may also include both the internal storage unit of the computer device and its external storage device. In this embodiment, the memory is generally used to store an operating system and various application software installed on the computer device. In addition, the memory may also be used to temporarily store various types of data that have been output or are to be output.

[0183] In some embodiments, the processor may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor is generally used to control the overall operation of a computer device. In this embodiment, the processor is used to run the program code stored in the memory or process data to implement the processing of the multimodal data analysis method based on the improved genetic algorithm in the aforementioned embodiment 1, and then according to the metadata in the given multimodal data model, according to the requirements of the customized fitness function, the best problem solving layer is obtained to solve the actual big data analysis and processing problems.

[0184] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A multimodal data analysis method based on an improved genetic algorithm, wherein the multimodal data is voice, video, or text data, and is used to optimize the problem-solving layer in a partitioned order product space generated according to the multimodal data to obtain the best target problem-solving layer; and ultimately achieve the purpose of determining the best solution according to the target problem-solving layer; characterized in that, The multimodal data analysis method comprises the following steps: S1: Construct an adaptive genetic algorithm based on the improvement of the classical genetic algorithm; the construction process of the adaptive genetic algorithm is as follows: S11: According to the classification accuracy and granularity requirements of the problem solving layer, a custom fitness function is designed to balance the two indicators; S12: improving the selection operator of the classical genetic algorithm into a hybrid selection operator that changes with the iteration process; the hybrid selection operator simultaneously adopts the elite retention operator and the "double-individual no-replacement" tournament selection operator in the early stage of evolution, and only adopts the classical tournament selection operator in the late stage of evolution; S13: improving the crossover operator of the classical genetic algorithm into an adaptive crossover operator whose crossover probability is dynamically adjusted with the iteration process; so that the crossover probability of individuals in the population in the early stage of evolution is greater than that in the late stage of evolution, and the fitness value of the individual is negatively correlated with the crossover probability; the crossover mode of the adaptive crossover operator is two-point crossover; S14: The mutation operator of the classical genetic algorithm is improved into an adaptive mutation operator whose mutation probability is dynamically adjusted with the iteration process; the mutation mode of the adaptive mutation operator is multiple mutations, and only when the similarity of the individual before and after the mutation is less than a given threshold and the fitness value of the individual after the mutation is greater than the fitness value of the individual before the mutation, the mutated individual is retained, otherwise the mutation continues until the maximum number of mutations is reached; S2: Use the partitioned order product space model to describe multimodal data: convert the problem-solving layer into a chromosome of a specified length through real number encoding, and randomly generate an initialization population containing multiple chromosomes; S3: Setting the maximum iteration rounds, using the adaptive genetic algorithm to iteratively update the initialization population, and obtaining the required target problem solution layer after the iterative update. The detailed process is as follows: S31: Calculate the fitness value of each chromosome individual in the current population; S32: Obtain the current iteration round and calculate the stage type of the iteration process; the stage type is divided into early evolution stage and late evolution stage; S33: executing different operation strategies according to the stage type of the current iteration round, and then updating the population of the current round through selection operation, crossover operation and mutation operation; S34: Determine whether the current iteration round reaches the maximum iteration round: if yes, calculate the fitness value of each chromosome individual in the current population, and take the individual with the maximum fitness value as the required target problem solving layer; Otherwise, the processing of steps S31-S33 is looped.

2. The multimodal data analysis method based on improved genetic algorithm according to claim 1, characterized in that: In step S11, the formula of the fitness function f is as follows: In the above formula, l j represents a problem-solving layer, m represents the number of perspectives in the problem-solving layer; D is the decision attribute set of the decision system, γ β (l j ,D) represents the problem solving layer l j Regarding the classification accuracy of D; d(l j ) represents the problem solving layer l j granularity; w1 and w2 are the weights of classification accuracy and granularity, respectively, and satisfy 0≤w1≤1, 0≤w2≤1; γ is the given target classification accuracy; d(l n ) is the most detailed problem solving layer l n The granularity.

3. The multimodal data analysis method based on improved genetic algorithm according to claim 2, characterized in that: The construction process of the fitness function f is as follows: (1) Given a decision system DS = (U, AT = C ∪ D, V, f), where U = {x1, x2, ..., x n } is a non-empty finite set of objects, called the domain; AT is a non-empty finite attribute set, C is the conditional attribute set, and D is the decision attribute set; V is the value range of the attribute set AT, f:U×AT→V is the information function, f(x,a)∈V a Represents the value of object x∈U on attribute a∈AT; the partition order product space corresponding to the decision system DS From m perspectives v i Composition; at the problem solving layer j middle, Represents the viewing angle v i The jth i Levels, n i is the viewing angle v i The number of levels, 1≤i≤m, 1≤j i ≤n i ; (2) Assume that U / D = {X1,X2,...,X r }, then the problem solving layer l j Classification accuracy γ about D β (l j ,D) is defined as: in, In the above formula, X j represents the decision class of the domain U with respect to the decision attribute set D, j = {1, 2, ..., r}; β is the multi-granularity fusion threshold, And 0≤β≤1; (3) For a given problem, find the solution layer l j The particle size g(l j ) is defined as: From the definition, we know that g(l j ) The larger the problem solving level l j The finer the particle size, the coarser the particle size; (4) Based on classification accuracy γ β (l j ,D) and particle size g(l j ), in order to make the granularity of the problem-solving layer as coarse as possible while meeting the classification accuracy requirement, the fitness function f is constructed.

4. The multimodal data analysis method based on improved genetic algorithm according to claim 1, characterized in that: In the improved adaptive genetic algorithm, the type division rules of the iterative process are as follows: Assume that the maximum number of iterations is set to I, the current number of iterations is i, 1≤i≤I; a threshold α is given for classifying the types of iterative processes, 0<α<1; then when When , the i-th iteration is judged to be in the early stage of evolution; when , it is determined that the i-th iteration is in the late stage of evolution.

5. The multimodal data analysis method based on improved genetic algorithm according to claim 4, characterized in that: In step S12, the process of the hybrid selection operator executing two selection strategies simultaneously in the early stage of evolution is as follows: First, the elite retention operator is used to copy some individuals with large fitness values ​​in the current population to the next generation population; then the "two individuals without replacement" tournament selection operator is used cyclically to select individuals from the current population until the number of individuals in the next generation population is consistent with the original size of the current population; among them, the "two individuals without replacement" tournament selection operator only selects 2 individuals to participate in the tournament in each round of iteration, and copies the winning individual to the next generation population. At the same time, the winning individual will also be deleted from the current population.

6. The multimodal data analysis method based on improved genetic algorithm according to claim 4, characterized in that: In the adaptive crossover operator of step S13, the adaptive crossover probability P of the two parent individuals p1 and p2 undergoing crossover is c The adjustment formula is as follows: In the above formula, f(p1) and f(p2) represent the fitness values ​​of p1 and p2 respectively; f max represents the maximum value of f(p1) and f(p2); P cpre Indicates the maximum crossover probability in the early stage of population evolution; P cbeh Represents the maximum crossover probability in the late stage of population evolution; and satisfies: P cpre >P cbeh ; σ is the standard deviation of the fitness values ​​of all individuals in the population.

7. The multimodal data analysis method based on improved genetic algorithm according to claim 4, characterized in that: In the adaptive mutation operator of step S14, the adaptive mutation probability P of the individual p currently undergoing mutation is m The adjustment formula is as follows: In the above formula, f(p) is the fitness value of chromosome individual p; P mpre is the maximum mutation probability in the early stage of population evolution; P mbe h is the maximum mutation probability in the late stage of population evolution; and P mpre <P mbe h; σ is the standard deviation of all individual fitness values ​​in the population; For any two chromosomes p i and p j , the similarity SD(p i ,p j ) is calculated as follows: Among them, p sub Yes i and p j The largest common substring of sub =sub(p i ,p j );len(p sub ) represents the length of the largest common substring; m′ represents the length of the chromosome, and the value of m′ corresponds to the number of perspectives in the problem-solving layer.

8. The multimodal data analysis method based on improved genetic algorithm according to claim 1, characterized in that: In step S14, in each chromosome generated by real number coding in step S2, each gene position corresponds to a perspective, and the gene value on each gene position corresponds to the hierarchical number under the perspective.

9. A multimodal data analysis system based on an improved genetic algorithm, characterized in that: It adopts the multimodal data analysis method based on the improved genetic algorithm as described in any one of claims 1 to 8 to optimize the problem solving layer in a partitioned order product space generated based on multimodal metadata; Get the best target problem solving layer; The multimodal data analysis system comprises: A real number encoding module, which is used to describe the data structure of the multimodal metadata by adopting a partition order product space model, and then convert the problem solving layer in the partition order product space into a chromosome of a specified length by means of real number encoding; A population initialization module, which is used to randomly generate an initial population including a plurality of chromosomes; A genetic algorithm processing module is used to iteratively update the initial population through selection, crossover and mutation operations; the genetic algorithm processing module includes a progress classification unit, a loop control unit, a fitness calculation unit, a hybrid selection unit, an adaptive crossover unit, and an adaptive mutation unit; the progress classification unit is used to determine the stage type corresponding to the current iteration round according to a preset rule, and the stage type is divided into an early evolutionary period and a late evolutionary period; the loop control unit is used to determine whether the current iteration round has reached the maximum number of iterations, if so, it will jump out of the iteration loop, otherwise it will continue to iteratively update the current population through the fitness calculation unit, the hybrid selection unit, the adaptive crossover unit, and the adaptive mutation unit; the fitness The calculation unit is used to calculate the fitness value of each chromosome in the current population according to a user-defined fitness function before each round of iteration; the hybrid selection unit is used to select different selection operators according to the stage type corresponding to the current iteration round, and then select samples from the current population to enter the next iteration round; the adaptive crossover unit uses an adaptive crossover operator with a crossover probability dynamically adjusted with the iteration process to match individuals in the current population in pairs, and performs a two-point crossover on a pair of chromosomes that meet the requirements; the adaptive mutation unit uses an adaptive mutation operator with a mutation probability dynamically adjusted with the iteration process to perform multiple mutations on the selected chromosomes until the mutated chromosomes meet the requirements; and The output module is used to select the individual with the largest fitness in the population after iterative update as the required target problem solving layer after the iterative update of the population by the genetic algorithm processing module reaches the maximum number of iterations.

10. A multimodal data analysis device based on an improved genetic algorithm, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The functional modules in the multimodal data analysis device are deployed in the manner of a multimodal data analysis system based on an improved genetic algorithm as described in claim 9. When the processor executes the computer program, the steps of the multimodal data analysis method based on an improved genetic algorithm as described in any one of claims 1-8 are implemented, and the optimal target problem solution layer is selected in the partitioned order product space.

Images