Feature selection method, computer device and computer program product

Abstract

The present application is applicable to the field of computer technology, and provides a feature selection method, computer equipment and computer program product, including: determining the symmetric uncertainty of each feature in the high-dimensional feature set of a specified business; constructing the first task and the second task of the feature selection task based on the symmetric uncertainty of each feature; performing multi-task learning based on the first particle population of the first task and the second particle population of the second task to obtain the specified particle of the feature selection task; if the number of updates for the specified particle does not reach the preset number, then adjusting the particle length of each particle in the first particle population and the second particle population respectively based on the adaptive length change mechanism, and performing multi-task learning again on the adjusted particle population until the target particle is obtained; taking the feature subset corresponding to the target particle as the target feature subset. The present application can improve the efficiency and accuracy of feature selection.

Description

Technical Field

[0001] The present application belongs to the field of computer technology, and in particular, relates to a feature selection method, a computer device, and a computer program product. Background Art

[0002] With the development of data technology, data sets in various fields are getting larger and larger, especially in the medical field, where the dimensions of data are very high. High-dimensional data can lead to the "curse of dimensionality" - the difficulty of algorithm calculation increases exponentially with the increase of dimensions, which prolongs the processing time and sometimes even makes it impossible to calculate. At the same time, high-dimensional data often contains many redundant or irrelevant features, which increases the computational burden, introduces noise, and leads to model performance degradation or overfitting. Overfitting refers to the model being too adaptable to the training data and cannot be effectively applied to new data. In addition, existing machine learning algorithms often encounter problems of low efficiency and inaccurate classification when processing high-dimensional data. To solve these problems, feature selection technology has emerged. It reduces the data dimension by screening out a small number of features that best represent the characteristics of the data, improves the accuracy and generalization ability of the model, and thus improves the accuracy of disease classification and prediction.

[0003] There are three existing feature selection methods: filtering, packaging, and embedded. The evolutionary algorithm in the packaging method is widely used in feature selection due to its excellent global optimization ability. The evolutionary algorithm can effectively avoid local optimal solutions, reduce computational costs, and efficiently find good feature combinations. However, it also has shortcomings: when selecting high-dimensional features, the evolutionary algorithm fails to distinguish the importance and redundancy of features well, which affects the search efficiency and results. Most methods use fixed-length encoding, which leads to a large search space under high-dimensional data, increasing the computational difficulty and cost. Although the variable-length evolutionary algorithm can adjust the length according to individual information, it ignores the global optimal information, resulting in slow convergence and affecting performance. These problems limit the performance of evolutionary algorithms in feature selection. Summary of the invention

[0004] In view of this, the embodiments of the present application provide a feature selection method, a computer device, and a computer program product, which can reduce the search space according to the symmetric uncertainty of the feature, and at the same time adopt flexible variable-length coding to improve the search efficiency, thereby reducing the search difficulty and improving the efficiency and accuracy of feature selection.

[0005] A first aspect of an embodiment of the present application provides a feature selection method, comprising:

[0006] Determine the symmetric uncertainty of each feature in a high-dimensional feature set for a specified business;

[0007] Based on the symmetric uncertainty of each of the features, construct a first task and a second task of the feature selection task;

[0008] Based on the first particle population of the first task and the second particle population of the second task, multi-task learning is performed to obtain designated particles of the feature selection task;

[0009] If the number of updates for the designated particle does not reach the preset number, based on the adaptive length change mechanism, the particle length of each particle in the first particle population and the second particle population is adjusted respectively, and the adjusted particle populations are subjected to multi-task learning again until the target particle is obtained;

[0010] The feature subset corresponding to the target particle is used as the target feature subset.

[0011] In an implementation of the first aspect, constructing the first task and the second task of the feature selection task based on the symmetric uncertainty of each feature includes:

[0012] Determine the adaptive threshold based on the symmetric uncertainty of the feature;

[0013] Selecting a feature subset having a symmetric uncertainty higher than the adaptive threshold from the high-dimensional feature set, and constructing a first task based on the feature subset;

[0014] Construct the second task based on the high-dimensional feature set.

[0015] In an implementation of the first aspect, determining the adaptive threshold based on the symmetric uncertainty of the feature includes:

[0016] Arrange all the features in the high-dimensional feature set in descending order of symmetric uncertainty to obtain an arrangement result;

[0017] Determining an adaptability threshold based on the symmetric uncertainty value of the specified position in the arrangement result and the first symmetric uncertainty value in the arrangement result;

[0018] In an implementation of the first aspect, the method further includes:

[0019] Based on the inflection point characteristics, the particle lengths of the particles with a specified proportion in each particle population are initialized through Gaussian distribution;

[0020] Based on the characteristic dimension, the particle lengths of the remaining particles in each particle population are initialized through uniform distribution.

[0021] In an implementation of the first aspect, performing multi-task learning based on a first particle population of the first task and a second particle population of the second task to obtain designated particles for the feature selection task includes:

[0022] Based on the competitive group optimization method, the particle trajectory of each particle in the first particle group and the second particle group is iteratively updated to obtain the target particle group, and the particle trajectory includes the particle speed and the particle position

[0023] Based on the fitness of each particle in the target particle group, a designated particle for the feature selection task is obtained.

[0024] In an implementation of the first aspect,

[0025] Each update process includes:

[0026] Based on particle fitness, determine the winning particles and losing particles of each particle population;

[0027] Determining aggregate particles based on the winning particles of the first particle population and the winning particles of the second particle population;

[0028] Based on the current speeds of the aggregated particles, the winning particles, and the failed particles, updating the current speed of the failed particles;

[0029] Based on the updated speed of the failed particle and the current position of the failed particle, the current position of the failed particle is updated.

[0030] In an implementation of the first aspect, adjusting the particle length of each particle in the first particle population and the second particle population based on the adaptive length change mechanism respectively includes:

[0031] Based on the particle length of the designated particle, respectively determine the length change of each particle in the first particle population and the second particle population;

[0032] Based on the length change amount, a lengthening or shortening operation is performed on each particle.

[0033] In an implementation of the first aspect, performing a length lengthening or length shortening operation on each particle based on the length change amount includes:

[0034] If the length change is negative, delete at least one redundant feature gene bit at the end of the current particle;

[0035] If the length change is positive, the added characteristic gene bits are determined by combining the tournament selection method and the designated particles.

[0036] A second aspect of an embodiment of the present application provides a feature selection device, including:

[0037] A determination module, used to determine the symmetric uncertainty of each feature in the high-dimensional feature set of a specified business;

[0038] A construction module, configured to construct a first task and a second task of a feature selection task based on the symmetric uncertainty of each of the features;

[0039] An optimization module, performing multi-task learning based on a first particle population of the first task and a second particle population of the second task to obtain designated particles of the feature selection task;

[0040] The adjustment module adjusts the particle length of each particle in the first particle population and the second particle population respectively based on the adaptive length change mechanism if the number of updates for the specified particle does not reach the preset number, and performs multi-task learning on the adjusted particle population again until the target particle is obtained, and uses the feature subset corresponding to the target particle as the target feature subset.

[0041] A third aspect of an embodiment of the present application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the computer device implements the method described in the first aspect above.

[0042] A fourth aspect of an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method described in the first aspect above is implemented.

[0043] The first aspect of the embodiment of the present application can reduce the search space according to the symmetric uncertainty of the features, and at the same time adopt flexible variable-length coding to improve the search efficiency, thereby reducing the search difficulty and improving the efficiency and accuracy of feature selection.

[0044] It can be understood that the beneficial effects of the second to fourth aspects mentioned above can be found in the relevant description of the first aspect mentioned above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions in the embodiments of the present application, 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 some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1 is a flow chart of the feature selection method provided in the embodiment of the present application;

[0047] Figure 2 is a flowchart of a multi-task construction method provided in an embodiment of the present application;

[0048] Figure 3 It is a flow chart of the implementation method of the particle length initialization method provided in the embodiment of the present application;

[0049] Figure 4 is a flowchart for implementing the multi-task learning method provided in an embodiment of the present application;

[0050] Figure 5 is a flow chart of a method for updating the position and velocity of particles provided in an embodiment of the present application;

[0051] Figure 6 is a flow chart of a method for implementing an adaptive length change mechanism provided in an embodiment of the present application;

[0052] Figure 7 is a flow chart of the particle length adjustment method provided in the embodiment of the present application;

[0053] Figure 8 is another implementation flow chart of the feature selection method provided in an embodiment of the present application;

[0054] Fig. 9 is a schematic diagram of a feature selection device provided in an embodiment of the present application;

[0055] Fig.10 It is a schematic diagram of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0056] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0057] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or combinations thereof.

[0058] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0059] As used in the specification and appended claims of this application, the term "if" can be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [described condition or event] is detected" or "in response to detecting [described condition or event]", depending on the context.

[0060] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0061] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0062] In practical applications, the existing feature selection methods are mainly divided into three categories: filtering, packaging and embedded methods. Among them, the evolutionary algorithm in the packaging method is an excellent global optimization method and has been widely used in the field of feature selection. Its adaptive global search capability helps to overcome the two main difficulties commonly faced by feature selection methods: it is easy to fall into local optimal solutions and high computational costs, so that it can efficiently search for better feature subsets. Although the existing evolutionary algorithms can achieve good results in solving high-dimensional feature selection problems, there are still the following shortcomings: 1. The existing feature selection methods based on evolutionary algorithms do not fully consider the importance and redundancy of features, lack the ability to distinguish important features, resulting in poor search efficiency and search results; 2. Most of the existing methods use a fixed-length encoding mechanism, which means that the algorithm needs to explore the optimal feature subset in a fixed and large search space. Especially when facing higher-dimensional features, the fixed-length encoding mechanism makes the search process face huge challenges, which not only increases the difficulty of search, but also leads to reduced computational efficiency. In addition, the fixed-length encoding strategy may also lead to higher computational costs. 3. The existing variable-length evolutionary algorithm only adjusts the individual length according to the individual's own information, but does not fully consider the global effective information of the global optimal individual. This limitation leads to a slower convergence speed to a certain extent and affects the overall performance of the algorithm.

[0063] Based on this, an embodiment of the present application provides a feature selection method, which is a high-dimensional feature selection method based on adaptive length variable particle swarm multi-task learning. It is possible to obtain a smaller number of feature subsets with higher classification accuracy on a high-dimensional data set. The method is suitable for fusing evolutionary multi-task algorithms with variable length coding, reducing the search space according to feature importance, and using flexible variable length coding to improve search efficiency. Thereby reducing the difficulty of searching, reducing computational consumption, and achieving improved classification performance.

[0064] In one embodiment, Figure 1 As shown, a feature selection method is provided, which is described by taking the method applied to a computer device as an example, and includes the following steps S101 to S105:

[0065] Step S101, determining the symmetric uncertainty of each feature in a high-dimensional feature set of a specified business.

[0066] In applications, Symmetrical Uncertainty (SU) is an indicator that measures the correlation between two variables by calculating the mutual information between the two variables and their respective information entropy, and then normalizing them. The larger the value, the stronger the correlation between the two variables. The computer device determines the symmetric uncertainty of each feature in the high-dimensional feature set according to the preset symmetric uncertainty determination method. The numerical value of symmetric uncertainty is used to measure the degree of association between the feature and the label, and is also used to characterize the importance of the feature. For example, in the medical disease classification scenario, the correlation between each feature in the high-dimensional lesion feature set and the disease classification result determined by the lesion feature set.

[0067] Step S102: constructing a first task and a second task of a feature selection task based on the symmetric uncertainty of each feature.

[0068] In the application, the feature selection task refers to selecting features with high feature importance from a high-dimensional feature set to form a feature subset. In this way, the specified business function originally implemented by the high-dimensional feature set can be converted to be performed by a low-dimensional feature subset. The computer device converts the feature selection task of the high-dimensional feature set for the specified business into a multi-task learning problem for solution (i.e., multi-objective optimization), and constructs two tasks, the first task and the second task, based on the symmetric uncertainty of each feature in the high-dimensional feature set. The first task refers to the features whose symmetric uncertainty of the features is higher than the adaptive threshold, and the second task is composed of the original high-dimensional feature set.

[0069] Step S103 , performing multi-task learning based on the first particle population of the first task and the second particle population of the second task to obtain designated particles for the feature selection task.

[0070] In the application, the computer device evolves the information sharing between tasks based on the preset multi-task learning optimization method to complete the multi-task learning and obtain the designated particle of the feature selection task. Such as the competitive group optimization method. The designated particle is the designated particle whose particle fitness is higher than the preset fitness threshold after at least one multi-task learning. Usually, the particle with the highest particle fitness is taken as the designated particle.

[0071] Step S104: If the number of updates for the specified particle does not reach the preset number, based on the adaptive length change mechanism, the particle length of each particle in the first particle population and the second particle population is adjusted respectively, and the adjusted particle populations are subjected to multi-task learning processing again until the target particle is obtained.

[0072] In the application, during the iterative search process, if the number of updates of the specified particle obtained does not reach the preset number, that is, during the entire iterative search process, the specified particle has not been updated for more than β times, the adaptive length change mechanism is executed. The essence of the adaptive length change mechanism is to calculate the particle length change amount based on the difference between the length of the specified particle and the length of each particle; then, the particle shortening or lengthening operation is performed. When executing the shortening operation operator, the ΔL redundant feature gene bits at the end of the particle are deleted. When executing the lengthening operation operator, the added feature gene bits are determined by combining the tournament selection method and the specified particle.

[0073] In the application, after the adaptive length change mechanism is executed, multi-task learning is iteratively performed based on the updated first particle population of the first task and the updated second particle population of the second task until the stop condition is met and the specified particle is output. When the stop condition is met, the specified particle of the current population is output. At this point, after the algorithm has been continuously iteratively evolved, the performance of the specified particles in the population has been greatly improved.

[0074] Step S105: taking the feature subset corresponding to the target particle as the target feature subset.

[0075] In the application, all the features corresponding to the target particle are the target feature subset. That is to say, the features selected in the specified particle are the optimal feature subset output at last.

[0076] In this embodiment, based on the symmetric uncertainty of each feature, the first task and the second task of the feature selection task are constructed; based on the first particle population of the first task and the second particle population of the second task, multi-task learning is performed to obtain the designated particle of the feature selection task; if the number of updates for the designated particle does not reach the preset number, based on the adaptive length change mechanism, the particle length of each particle in the first particle population and the second particle population is adjusted respectively, and the adjusted particle population is multi-task learning is performed again until the target particle is obtained; the feature subset corresponding to the target particle is used as the target feature subset. In this way, the search space can be reduced according to the symmetric uncertainty of the feature, and flexible variable-length encoding can be used to improve the search efficiency, thereby reducing the search difficulty, reducing the computational consumption, and improving the classification performance.

[0077] In one embodiment, Figure 2 As shown, the implementation process of step S102 includes the following steps S201 to S203:

[0078] Step S201, determining an adaptive threshold based on the symmetric uncertainty of the feature.

[0079] In the application, the computer device determines the symmetric uncertainty of each feature in the high-dimensional feature set and obtains the maximum symmetric uncertainty value SU max , and the numerical value of the symmetric uncertainty of the specified index Indicates that the features in the high-dimensional feature set are sorted in descending order according to SU. The adaptive threshold is determined based on the following formula:

[0080]

[0081] Among them, ρ0 represents the adaptive threshold, SU max represents the maximum value of SU, Represents the first The value of the bit. And the larger value of the two is used as the adaptive threshold through the max function. Two optimization subtasks are constructed under the guidance of the adaptive threshold. Among them, the first task refers to the feature subset containing symmetric uncertainty values ​​higher than the adaptive threshold. These features are considered to have higher feature importance and are more likely to be selected. The second task contains the complete original feature set, that is, the high-dimensional feature set.

[0082] The maximum value of the symmetric uncertainty SU max and all features are arranged in descending order of symmetric uncertainty. Symmetric uncertainty value of the characteristic of the bit The larger of the two is multiplied by a coefficient of 0.2, and D represents the total number of features.

[0083] Guided by the adaptive threshold ρ0, two optimization subtasks can be constructed. One of the subtasks includes those features whose symmetric uncertainty is greater than or equal to ρ0, which are considered to have higher importance or contribution. The purpose of this design is to more effectively identify and utilize the most valuable features, thereby improving the overall optimization effect and model performance.

[0084] Step S202: Select a feature subset whose symmetric uncertainty value is higher than an adaptive threshold from the high-dimensional feature set, and construct a first task based on the feature subset.

[0085] In the application, the first task and the second task to be optimized are constructed under the guidance of the adaptive threshold, wherein the first task contains a feature subset whose symmetric uncertainty value is higher than the threshold, and these features are considered to have higher feature importance and are more likely to be selected.

[0086] Step S203: construct a second task based on the high-dimensional feature set.

[0087] In applications, the second task contains a complete set of high-dimensional features.

[0088] In this embodiment, two tasks are constructed based on adaptive thresholds to distinguish features according to their importance, thereby increasing the importance of subsequent feature selection.

[0089] In one embodiment, the implementation process of step S201 is as follows: all features in the high-dimensional feature set are arranged in descending order of symmetric uncertainty to obtain an arrangement result; based on the symmetric uncertainty value at a specified position in the arrangement result and the first symmetric uncertainty value in the arrangement result, an adaptability threshold is determined.

[0090] In one embodiment, Figure 3 As shown, the implementation process of initializing the particle population includes the following steps S301 to S302:

[0091] Step S301 : based on the inflection point feature, the particle lengths of particles of a specified proportion in each particle population are initialized through Gaussian distribution.

[0092] In applications, feature importance is based on the symmetric uncertainty characterization of features. Inflection point features are used to distinguish important features. The numerical value of the symmetric uncertainty of a feature is used to measure the degree of association between a feature and a label. The larger the symmetric uncertainty value, the higher the degree of association between the feature and the label. Based on feature importance, all features in the high-dimensional feature set are sorted in descending order to obtain a sorted high-dimensional feature set. The features whose symmetric uncertainty values ​​in the sorted high-dimensional feature set are equal to the adaptive threshold are called inflection point features.

[0093] The inflection point index refers to the index of the inflection point feature in the descending sorted high-dimensional feature set. Based on the inflection point index and the inflection point ratio, a Gaussian distribution is constructed, and the particle length (also called gene length) of the particles with a specified ratio (90%) in the first particle group is randomly initialized, and the particle length of the particles with a specified ratio (90%) in the second particle group is randomly initialized.

[0094] Step S302 : based on the characteristic dimension, the particle lengths of the remaining particles in each particle population are initialized by uniform distribution.

[0095] In the application, in order to consider the weakly correlated features with low partial symmetric uncertainty SU values, the computer device initializes the remaining particles in the first particle group based on a specified uniform distribution, and the length of the initialized particles conforms to the specified uniform distribution ranging from 1 to D, where D is the feature dimension in the high-dimensional feature set.

[0096] Specifically, the formula for initializing particle length based on Gaussian distribution and uniform distribution is as follows:

[0097]

[0098] Among them, L(i) represents the length of particle i, G represents Gaussian distribution, KPoint is the inflection point index, vp is the inflection point ratio in the Gaussian distribution variance, and Popsize is the number of particles in the population. At the same time, in Gaussian distribution, KPoint is the mean, is the variance. D represents the feature dimension, and unifrnd represents uniform distribution.

[0099] In the application, Gaussian distribution is initialized, and KPoint is the inflection point index after the feature is sorted by symmetric uncertainty (SU), which indicates the importance and discrimination of the feature. It is used to control the variance of Gaussian distribution and ensure the diversity of particle length. For the first 90% of particles, the particle length L(i) is distributed by Gaussian distribution. initialization.

[0100] For the remaining 10% particles, the particle length L(i) is initialized by the uniform distribution unifrnd(1,D). The uniform distribution method ensures that some diversity is retained even when the feature importance is low, preventing potential useful information from being ignored.

[0101] In this embodiment, by comprehensively considering the importance and diversity of features, Gaussian distribution and uniform distribution are used to initialize the gene length of particles, aiming to improve the performance and robustness of the particle swarm optimization algorithm.

[0102] In one embodiment, Figure 4 As shown, the implementation process of step S103 includes the following steps S401 to S402:

[0103] Step S401, based on the competitive swarm optimization method, iteratively update the particle trajectory of each particle in the first particle population and the second particle population to obtain a target particle population, where the particle trajectory includes the particle speed and the particle position.

[0104] In the application, at the beginning, the computer device randomly divides each particle group into failed particles and winning particles by competitive group optimization. Based on the competitive group optimization method, the particle trajectory of each particle in the first particle population and the second particle population is iteratively updated to obtain the target particle population, and the particle trajectory includes the particle speed and the particle position.

[0105] Step S402, based on the fitness of each particle in the target particle group, obtain the designated particle for the feature selection task.

[0106] In the application, after each round of update processing, a new target particle swarm is obtained, the fitness of each particle in the target particle swarm is determined, and the current designated particle is obtained.

[0107] In this embodiment, based on the competitive particle swarm optimization method, the convergence speed of the feature selection task is accelerated through effective knowledge transfer between tasks, and the global search capability of the feature selection task is improved.

[0108] In one embodiment, Figure 5 As shown, the implementation process of each update process includes the following steps S501 to S503:

[0109] Step S501, based on the particle fitness, determine the winning particles and failed particles of each particle population.

[0110] In the application, the computer device determines the fitness of each particle in the particle population, and determines the winning particles and the losing particles of each particle population by comparing the fitness sizes of the particles.

[0111] Step S502 : determining aggregate particles based on the winning particles of the first particle population and the winning particles of the second particle population.

[0112] In the application, the current state of the particle is described by the particle's speed and position. In order to share information between tasks, the computer device determines the aggregated particle based on the winning particle in the first particle group of the first task and the winning particle in the second particle group. The calculation formula for the aggregated particle is as follows:

[0113]

[0114] Among them, X W (t) represents the position of the winning particle in the current task, X′ W (t) represents the position of the winning particle in another task, and λ is the weight of the weighted sum. is the aggregate particle generated by weighted sum of the winning particles in the two tasks. That is, the process of determining the aggregate particle is to calculate the position X′ of the winning particles in the two tasks. W (t) and X W (t) Perform weighted summation to obtain the position of the aggregated particles.

[0115] Step S503, based on the current speeds of the aggregated particles, the winning particles and the failed particles, the speed of the failed particles is updated.

[0116] In applications, the location of aggregate particles The current position X of the failed particle L (t), the current position X of the winning particle W (t) and the current velocity V of the failed particle L (t), update the velocity of the failed particle to obtain the updated velocity v of the failed particle L (t+1). The speed update formula for the failed particles in the population to learn from the winning particles is as follows:

[0117]

[0118] Where t and t+1 represent the current and next generation respectively. r1, r2 and r3 represent random numbers ranging from 0 to 1. w (t)-X L (t) represents the position difference between the winning particle and the losing particle, Characterize the second position difference between failed particles and aggregated particles.

[0119] Updating the speed of failed particles is essentially a weighted summation of the current speed of the failed particles, the position difference between the winning particles and the failed particles, and the position difference between the failed particles and the aggregated particles to obtain the updated speed of the failed particles.

[0120] Step S504: updating the position of the failed particle based on the updated speed of the failed particle and the current position of the failed particle.

[0121] In the application, for failed particles, the next generation position of the failed particles is determined by the sum of the current position of the failed particles and the updated speed of the failed particles. That is, the position update formula for failed particles in the population to learn from the winning particles is as follows:

[0122] X L (t+1)=X L (t)+V L (t+1)

[0123] In this embodiment, the update rate can be increased by updating the position and velocity of the failed particles based on the position and velocity of the aggregated particles and the winning particles.

[0124] In one embodiment, Figure 6 As shown, the process of respectively adjusting the particle length of each particle in the first particle population and the second particle population in step S104 includes the following steps S601 to S602:

[0125] Step S601 : determining the length change of each particle in the first particle population and the second particle population respectively based on the particle length of the designated particle.

[0126] In the application, for each particle in the first particle population and the second particle population, the length difference between the designated particle and the targeted particle (i.e., the length of the designated particle minus the particle length of the targeted particle) is determined as the length change of the targeted particle.

[0127] Step S602: Based on the length change amount, a lengthening or shortening operation is performed on each particle.

[0128] In the application, if the length change is positive, it means that the particle length of the targeted particle is shorter than the specified particle, and the length lengthening operation is performed on the targeted particle. If the length change is negative, it means that the particle length of the targeted particle is longer than the specified particle, and the length shortening operation is performed on the targeted particle.

[0129] In this embodiment, based on the adaptive length change mechanism, by adaptively and dynamically adjusting the particle length during the evolution process, the particles can explore the best feature subset in the dynamically adjusted promising subspace, which significantly reduces the search space and computational cost and improves the overall performance of the algorithm.

[0130] In one embodiment, Figure 7 The implementation process of the length lengthening or length shortening operation shown in the steps includes the following steps S701 to S702:

[0131] Step S701: If the length change is negative, at least one redundant feature gene bit at the end of the current particle is deleted.

[0132] In the application, if the length change is negative, it indicates that the particle length of the targeted particle is longer than the specified particle, and the length shortening operation is performed on the targeted particle. Specifically, at least one redundant feature gene bit at the end of the current particle is directly deleted.

[0133] Step S702: If the length change is positive, the added feature gene bits are determined by combining the tournament selection method and the designated particles.

[0134] In the application, if the length change is positive, it indicates that the particle length of the targeted particle is shorter than the designated particle, and the length change operation is performed on the targeted particle. The specific method is to combine the tournament selection method and the designated particle to determine the added characteristic gene bit.

[0135] In this embodiment, the optimization process can be accelerated by the adaptive mechanism of dynamically adjusting the particle length, and a high-quality feature subset can be obtained in a relatively short time.

[0136] In one embodiment, Figure 8 As shown in Figure 2, the specific implementation process of feature selection includes the following steps:

[0137] Step S801: input a high-dimensional feature data set.

[0138] Step S802, calculating feature importance based on feature uncertainty.

[0139] Step S803: construct a multi-task model according to feature importance.

[0140] Step S804: perform multi-task learning processing.

[0141] In the application, the multi-task learning process includes the following steps S805 to S810.

[0142] Step S805: Initialize the particle length based on the variable length initialization strategy.

[0143] In the application, the aforementioned Gaussian distribution and uniform distribution are combined to initialize the particle length of each particle in the particle population.

[0144] Step S806: Share information and update individual speed and position through competitive group optimization algorithm.

[0145] In the application, the competitive swarm optimization algorithm is combined to determine the winning particles and the failed particles, and the speed and position of each failed particle are updated according to the position update formula and speed update formula of the failed particle.

[0146] Step S807: lengthen or shorten the particle length based on the adaptive length change mechanism.

[0147] Step S808: recalculate the fitness value of the particle after the length is changed, and update the optimal particle.

[0148] Step S809, determine whether the iteration termination condition is met, then go to step S10, otherwise return to step S6.

[0149] Step S810: output the particle with the best fitness.

[0150] The particle with the best fitness is the target particle mentioned above.

[0151] Step S811: The feature subset selected by the particle with the best fitness is the feature subset output as a result.

[0152] In the application, after continuous iteration and evolution of the algorithm, the performance of the specified particles in the population has been greatly improved. The features selected in the specified particles are the optimal feature subsets output at the end.

[0153] In this embodiment, a feature importance evaluation mechanism based on symmetric uncertainty and a multi-task construction method are first introduced, which can effectively screen out important features and reduce redundant features, significantly improving the accuracy of feature selection and the search efficiency of the algorithm. Secondly, by introducing the idea of ​​multi-task learning, the high-dimensional feature selection problem is modeled as a multi-task learning problem, and an adaptive length-variable competitive particle swarm optimization algorithm is proposed, which effectively improves the convergence speed of the algorithm and the feature selection effect. Finally, by dynamically adjusting the adaptive mechanism of particle length, the optimization process can be accelerated, and a high-quality feature subset can be obtained in a relatively short time.

[0154] The present application also provides a feature selection device for executing the steps in the above feature selection method embodiment. Fig. 9 As shown, the feature selection device 900 provided in the embodiment of the present application includes:

[0155] The determination module 910 is used to determine the symmetric uncertainty of each feature in the high-dimensional feature set of the specified service.

[0156] The construction module 920 is used to construct the first task and the second task of the feature selection task based on the symmetric uncertainty of each of the features.

[0157] The optimization module 930 is used to perform multi-task learning based on the first particle population of the first task and the second particle population of the second task to obtain the designated particles of the feature selection task.

[0158] The adjustment module 940 adjusts the particle length of each particle in the first particle population and the second particle population respectively based on the adaptive length change mechanism if the number of updates for the specified particle does not reach the preset number, and performs multi-task learning on the adjusted particle population again until the target particle is obtained, and uses the feature subset corresponding to the target particle as the target feature subset.

[0159] In one embodiment, the determination module is also used to arrange all features in the high-dimensional feature set in descending order of symmetric uncertainty to obtain an arrangement result; and determine the adaptability threshold based on the symmetric uncertainty value at a specified position in the arrangement result and the first symmetric uncertainty value in the sorting result.

[0160] In one embodiment, the construction module is also used to determine an adaptive threshold based on the symmetric uncertainty of the feature; select a feature subset whose symmetric uncertainty is higher than the adaptive threshold from the high-dimensional feature set, and construct a first task based on the feature subset; and construct a second task based on the high-dimensional feature set.

[0161] In one embodiment, the construction module is also used to initialize the particle length of a specified proportion of particles in each particle population through Gaussian distribution based on the inflection point feature; and initialize the initial particle length of the remaining particles in each particle population through uniform distribution based on the feature dimension.

[0162] In one embodiment, the optimization module is further used to iteratively update the particle trajectory of each particle in the first particle group and the second particle group based on a competitive group optimization method to obtain a target particle group, where the particle trajectory includes the particle speed and the particle position; based on the fitness of each particle in the target particle group, a designated particle for the feature selection task is obtained.

[0163] In one embodiment, the optimization module is further used to determine the winning particles and failed particles of each particle group based on the particle fitness; determine the aggregated particles based on the winning particles of the first particle group and the winning particles of the second particle group; update the speed of the failed particles based on the current speeds of the aggregated particles, the winning particles and the failed particles; and update the position of the failed particles based on the updated speed of the failed particles and the current position of the failed particles.

[0164] In one embodiment, the optimization module is further used to determine the length change of each particle in the first particle population and the second particle population based on the particle length of the specified particle; based on the length change, perform a length lengthening or shortening operation on each particle.

[0165] In one embodiment, the optimization module is further used to delete at least one redundant characteristic gene bit at the end of the current particle if the length change is negative; if the length change is positive, determine the added characteristic gene bit by combining the tournament selection method and the designated particle.

[0166] In applications, each module in the image processing device may be a software program module, or may be implemented by different logic circuits integrated in a processor, or may be implemented by multiple distributed processors.

[0167] Fig.10 This is a schematic diagram of the structure of a computer device for implementing a method for clearing freight charges for rail transit provided in one embodiment of the present application. Fig.10 As shown, the computer device 1000 of this embodiment includes: at least one processor 1010 ( Fig.10 Only one is shown in the figure) a processor, a memory 1020, and a computer program 1030 stored in the memory 1020 and executable on the at least one processor 1010, and when the processor 1010 executes the computer program 1030, the steps in any of the above-mentioned exception handling method embodiments are implemented.

[0168] The computer device 1000 may be a computing device such as a desktop computer, a notebook, a PDA, or a cloud server. The computer device 1000 may include, but is not limited to, a processor 1010 and a memory 1020. Those skilled in the art will appreciate that Fig.10 The computer device 1000 is merely an example and does not constitute a limitation on the computer device 1000 . The computer device 1000 may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the computer device 1000 may also include input and output devices, network access devices, etc.

[0169] The processor 1010 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0170] In some embodiments, the memory 1020 may be an internal storage unit of the computer device 1000, such as a hard disk or memory of the computer device 1000. In other embodiments, the memory 1020 may also be an external storage device of the computer device 1000, 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 1000. Further, the memory 1020 may also include both an internal storage unit of the computer device 1000 and an external storage device. The memory 1020 is used to store an operating system, an application program, a boot loader (BootLoader), data, and other programs, such as the program code of the computer program, etc. The memory 1020 may also be used to temporarily store data that has been output or is to be output.

[0171] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of the present application. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0172] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0173] An embodiment of the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in the above-mentioned method embodiments can be implemented.

[0174] An embodiment of the present application provides a computer program product. When the computer program product runs on a mobile terminal, the mobile terminal can implement the steps in the above-mentioned method embodiments when executing the computer program product.

[0175] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, which can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may at least include: any entity or device that can carry the computer program code to the device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electric carrier signal, a telecommunication signal, and a software distribution medium. For example, a USB flash drive, a mobile hard disk, a magnetic disk or an optical disk. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electric carrier signals and telecommunication signals.

[0176] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0177] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0178] In the embodiments provided in the present application, it should be understood that the disclosed devices / network equipment and methods can be implemented in other ways. For example, the device / network equipment embodiments described above are merely schematic. For example, the division of the modules or units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0179] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0180] The embodiments described above are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A feature selection method, characterized in that the method include: Determine the symmetric uncertainty of each feature in a high-dimensional feature set for a specified business; Based on the symmetric uncertainty of each of the features, construct a first task and a second task of the feature selection task; Based on the first particle population of the first task and the second particle population of the second task, multi-task learning is performed to obtain designated particles of the feature selection task; If the number of updates for the designated particle does not reach the preset number, based on the adaptive length change mechanism, the particle length of each particle in the first particle population and the second particle population is adjusted respectively, and the adjusted particle populations are subjected to multi-task learning again until the target particle is obtained; The feature subset corresponding to the target particle is used as the target feature subset.

2. The feature selection method according to claim 1, characterized in that: The constructing the first task and the second task of the feature selection task based on the symmetric uncertainty of each feature includes: Determine the adaptive threshold based on the symmetric uncertainty of the feature; Selecting a feature subset having a symmetric uncertainty higher than the adaptive threshold from the high-dimensional feature set, and constructing a first task based on the feature subset; Construct the second task based on the high-dimensional feature set.

3. The feature selection method according to claim 2, characterized in that: The step of determining an adaptive threshold based on the symmetric uncertainty of the feature comprises: Arrange all the features in the high-dimensional feature set in descending order of symmetric uncertainty to obtain an arrangement result; An adaptability threshold is determined based on the symmetric uncertainty value of a specified position in the arrangement result and the first symmetric uncertainty value in the arrangement result.

4. The feature selection method according to claim 1, characterized in that: The method further comprises: Based on the inflection point characteristics, the particle lengths of the particles with a specified proportion in each particle population are initialized through Gaussian distribution; Based on the characteristic dimension, the particle lengths of the remaining particles in each particle population are initialized through uniform distribution.

5. The feature selection method according to claim 1, characterized in that: The performing multi-task learning based on the first particle population of the first task and the second particle population of the second task to obtain the designated particles of the feature selection task includes: Based on a competitive group optimization method, iteratively updating the particle trajectory of each particle in the first particle population and the second particle population to obtain a target particle population, wherein the particle trajectory includes a particle velocity and a particle position; Based on the fitness of each particle in the target particle group, a designated particle for the feature selection task is obtained.

6. The feature selection method according to claim 5, characterized in that: Each update process includes: Based on particle fitness, determine the winning particles and losing particles of each particle population; Determining aggregate particles based on the winning particles of the first particle population and the winning particles of the second particle population; Based on the current speeds of the aggregated particles, the winning particles, and the failed particles, updating the speed of the failed particles; The position of the failed particle is updated based on the updated speed of the failed particle and the current position of the failed particle.

7. The feature selection method according to claim 1, characterized in that: The step of adjusting the particle length of each particle in the first particle population and the second particle population based on the adaptive length change mechanism comprises: Based on the particle length of the designated particle, respectively determine the length change of each particle in the first particle population and the second particle population; Based on the length change amount, a lengthening or shortening operation is performed on each particle.

8. The feature selection method according to claim 7, characterized in that: The step of performing a length lengthening or shortening operation on each particle based on the length change amount includes: If the length change is negative, delete at least one redundant feature gene bit at the end of the current particle; If the length change is positive, the added characteristic gene bits are determined by combining the tournament selection method and the designated particles.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the computer device is caused to implement the method according to any one of claims 1 to 8.

10. A computer program product, characterized in that The invention comprises a computer program, which, when being executed, enables the method according to any one of claims 1 to 8 to be performed.

Images