Text classification method based on space conversion multi-form multi-target feature selection

Abstract

The application discloses a text classification method based on space conversion multi-form multi-target feature selection, which converts original text into a machine processable form, selects appropriate features to represent in the form, adopts a space conversion multi-form multi-target algorithm to perform feature selection on the extracted features, inputs the selected features into a classifier, then uses a known K nearest neighbor classification technology to obtain a label of unknown text, and uses the label as a final result to output a result of the text to be identified and classified. The method has high text classification accuracy, strong text classification robustness and high speed, is simple to implement, has few selected features, saves computing resources, and has other advantages. The method can select truly relevant features, simplify a model, is easy to operate by a practitioner, eliminates irrelevant features, shortens model training time, and improves practical application value.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and specifically to a text classification method based on spatial transformation and multi-form, multi-target feature selection. Background Technology

[0002] With the rapid development of the internet and the explosive growth of text information online, relying on manual text classification has become extremely difficult and susceptible to subjective influences. However, advancements in computer science and artificial intelligence have enabled various technologies to automatically classify text. These emerging classification technologies offer advantages such as high stability, fast processing speed, and less susceptibility to subjective factors. Machine learning, in particular, is widely applied in various fields of text classification due to its high accuracy and efficiency. For example, in information retrieval, categorizing text into different topics provides more accurate and relevant search results; in sentiment analysis, classifying text as positive, negative, or neutral reveals users' attitudes and emotions towards products, services, or events; and in spam filtering, classifying emails as spam or non-spam effectively prevents an increase in spam. Furthermore, text classification plays a crucial role in areas such as public opinion analysis, automatic text summarization, and targeted advertising. In conclusion, text classification, as an important natural language processing technology, can extract valuable information from massive amounts of text data and support various application scenarios. Its background stems from the research and practical needs in the fields of information retrieval and natural language processing. With the advancement of machine learning, text classification technology is constantly developing and evolving, and has broad application prospects.

[0003] The goal of text classification is to assign one or more predefined class labels to a given text, extracting key information such as semantics and words as effective features to provide reliable input for subsequent classification tasks. However, the sheer volume and linguistic diversity of online text today result in high-dimensionality data, necessitating substantial computational resources for classification tasks. Furthermore, not all extracted features are valuable; many are irrelevant and redundant, wasting computational resources and potentially reducing classification accuracy. Therefore, dimensionality reduction is necessary.

[0004] Feature selection, as an effective data preprocessing technique, can reduce dimensionality by removing redundant and irrelevant features from the original feature set, thereby improving classification accuracy, reducing the risk of overfitting, and shortening the training time for building classification models. Currently, feature selection has been widely applied in fields such as biology, text classification, speech recognition, image classification, and computer vision, achieving excellent results. To date, numerous methods have been developed to address the feature selection problem, generally categorized into three types: filtering, wrapping, and embedded. Filtering selects features by evaluating each feature individually, ignoring interactions between features, but it boasts high computational efficiency. Wrapping methods utilize many excellent search algorithms to obtain feature subsets, which can be built and evaluated through learning algorithms. Therefore, compared to filtering, wrapping methods can achieve higher classification accuracy. Embedded methods, compared to filtering and wrapping methods, incorporate feature selection as part of the training model, considering both classification accuracy and computational efficiency to achieve a relative balance between the two. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a text classification method based on spatial transformation and multi-form, multi-target feature selection.

[0006] The technical solution adopted in this invention is: a text classification method based on spatial transformation and multi-form multi-target feature selection, comprising the following steps:

[0007] Step 1, Data Preprocessing: This includes text cleaning, word segmentation, and stop word removal, converting the raw text into a machine-processable format;

[0008] Step 2, Feature Extraction: Represent the text by selecting or designing appropriate features. Commonly used features include bag-of-words model, TF-IDF vectors, and word embeddings.

[0009] Step 3, Model Training and Evaluation: Feature selection is performed on the extracted features using a multi-form, multi-objective algorithm based on spatial transformation;

[0010] Step 4, Application Prediction: Input the selected feature combination into the classifier, then use the well-known K-nearest neighbor classification technique to obtain the label of the unknown text, and use this label as the final result to output the result of the text to be identified and classified.

[0011] Furthermore, step 3 includes the following steps:

[0012] Step 3-1: Divide the dataset into a training set and a test set, with the training set accounting for 70% and the test set accounting for 30%; input the training set for feature selection.

[0013] Step 3-2, set the parameters of the feature selection evolutionary algorithm: population size P is 100, number of iterations N is 100, number of features is D, select K nearest neighbor classifier, and the distance value K of K nearest neighbor classifier is 5;

[0014] Step 3-3: Set the encoding method of each individual in the population to binary encoding method, where 1 represents selecting this set of features and 0 represents not selecting it;

[0015] Steps 3-4 define the first grouping format, including the following steps:

[0016] Step 3-4-1: Evaluate the features using SU and ReliefF, and project the rankings of these two evaluation metrics onto a two-dimensional coordinate axis.

[0017] Step 3-4-2: Treat the feature as a solution, and the SU and ReliefF rankings of the feature as two objectives;

[0018] Step 3-4-3: Use the non-dominated ranking method to divide these features into levels, and treat the features at each level as a feature group of an individual.

[0019] Features at the non-dominated front are considered the most important, features at the second level are considered the second most important, and so on. Since the non-dominated ranking automatically divides features into different levels, the number of feature groups is adaptively determined as Gnum.

[0020] Steps 3-5 define the second grouping method: use the K-means algorithm to group the features projected onto the two-dimensional coordinate axes; where the size of K determines the size of the number of groups, so K is determined according to the number of groups Gnum obtained from the first grouping method;

[0021] Steps 3-6: Construct a multi-form framework: Combine the two grouping forms into a multi-form framework, and treat each feature group as a dimension; then assign a population P1 and a population P2 to the two grouping forms respectively for search;

[0022] Steps 3-7, Population Initialization: Since each digit of an individual represents a set of features, an N1-row D-column matrix and an N2-row D-column matrix are randomly generated. Each value in the matrix is ​​randomly generated between 0 and 1, where each row represents an individual and each column of an individual represents a feature. The average value of each set of features after evaluation by SU and ReliefF is calculated, and each is assigned a weight of 0.5. The weighted value is used as the threshold for that set of features. If the generated value is greater than the set threshold, it is set to 1; otherwise, it is set to 0.

[0023] Steps 3-8: Evaluate the initial population: The 5-fold cross-validation evaluation method is used to evaluate the classification error rate of each individual in the initial population and the ratio of the selected feature to the total number of features using the fitness function of formula (1) to obtain the target value of the initial population. The classification error rate is the first column of the target value, and the ratio of the selected indicator features is the second column of the target value.

[0024]

[0025] In the formula: f error f represents the classification error rate; ratio Indicates the proportion of selected features; c represents the number of categories; FP i S represents the number of prediction errors in category i; i The number of samples representing category i;

[0026] Steps 3-9: Identify the optimal solution set in the population: Sort the target values ​​of the multi-objective population according to the Pareto non-dominated sorting method, and take the solution set of the first frontier as the optimal solution set of the multi-objective population. If it exceeds 100, select the top 100 individuals according to the Euclidean distance of the distance solutions.

[0027] Step 3-10, Crossover: Perform a crossover operation on each individual in the two populations from Step 3-7. First, the selection of parents is done using a binary tournament method. Then, considering the preservation of important features, if both parents select a certain set of features, then that set of features is considered important and will be retained in the offspring; conversely, if neither parent selects, then the offspring will not select. If one parent selects and the other does not, the probability P of whether the offspring selects or not is determined. c The calculation is as follows:

[0028]

[0029] In the formula: and These represent the average rankings of the j-th feature group after SU and ReliefF evaluations, respectively; D represents the total number of features.

[0030] Step 3-11, Mutation: The offspring obtained in Step 3-10 are subjected to mutation with a certain probability, where the mutation probability P is... m Set to 1 / Gnum;

[0031] Step 3-12, Evaluation: Using the 5-fold cross-validation evaluation method, taking Step 3-8 as an example, evaluate the classification error rate and the proportion of selected features for each individual in the new offspring population generated after mutation in Step 3-11, and obtain the target value of the new offspring population, where the classification error rate is the first column of the target value and the proportion of selected features is the second column of the target value.

[0032] Step 3-13, Selection: The Pareto non-dominated sorting method is used to sort the target values ​​of all individuals in the generated offspring and parent sets of the two populations to obtain the index of all individuals. The first 50 individuals by index are selected as the next generation of multi-objective population.

[0033] Steps 3-14: Knowledge transfer;

[0034] Step 3-15: When one of the populations gets trapped in a local optimum three times, project the two grouped features back to the original feature space, cancel the grouping, and then perform the search using one bit representing one feature. If the non-dominated front distance between two generations of the population is greater than log... 10 When D, the cumulative count is recalculated.

[0035] Step 3-16, Iteration: After each execution of step 3-14, the iteration count is incremented by one, and then the operation is returned to step 3-10.

[0036] Step 3-17, Termination: When the number of iterations reaches the number of iterations T in step 3-2, i.e., 100, the feature selection evolutionary algorithm terminates and outputs the last generation of multi-objective population and its target value;

[0037] Step 3-18: Sort the target values ​​of all individuals output in Step 3-17 using Pareto non-dominated sorting to obtain the individuals on the first Pareto front. Record the individual with the highest classification accuracy in the first front, which is the multiple important feature combinations obtained by the feature selection evolutionary algorithm after feature extraction in Step 3.

[0038] Furthermore, step 3-14 includes the following steps:

[0039] Step 3-14-1: First, obtain the non-dominated fronts of the two newly generated populations, and calculate the Euclidean distance between the centroids of the fronts of the two generations and the fronts of the previous generation populations. If the distance between the fronts of one of the populations is 0, it is considered that the population is trapped in a local optimum.

[0040] Step 3-14-2 triggers the knowledge transfer mechanism. Assuming that P1 is trapped in a local optimum, the number of times individuals in the P2 population select feature groups is ranked from high to low. The feature group selected most often is considered an important group, and the feature group selected least often is considered a worthless group. The feature groups ranked in the top F are selected as the best feature groups (BestG).

[0041] Step 3-14-3: Find the same features as BestG in the grouping pattern of population P1, and select these features to form the same group as BestG; for the newly formed groups in population P1, except for the individuals at the forefront of population P1, all other individuals are selected from these newly formed groups.

[0042] Step 3-14-4: Re-evaluate all individuals in the P1 population and archive the non-dominant frontier of the P1 individuals before migration.

[0043] The beneficial effects of this invention are:

[0044] (1) The text classification has high accuracy, strong robustness and fast speed;

[0045] (2) It has the advantages of simple implementation, fewer selected features, and saving computing resources;

[0046] (3) Selecting truly relevant features can further simplify the model and make it easier for practitioners to operate;

[0047] (4) Removing irrelevant features can further shorten the model training time and enhance its practical application value; Attached Figure Description

[0048] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0049] The invention will now be further described with reference to the accompanying drawings.

[0050] like Figure 1 As shown, this invention is a text classification method based on spatial transformation and multi-form multi-target feature selection, comprising the following steps:

[0051] Step 1, Data Preprocessing: This includes text cleaning, word segmentation, and stop word removal, converting the raw text into a machine-processable format.

[0052] Step 2, Feature Extraction: Represent the text by selecting or designing appropriate features. Commonly used features include bag-of-words model, TF-IDF vectors, and word embeddings.

[0053] Step 3, Model Training and Evaluation: A multi-form, multi-objective algorithm based on spatial transformation is used to select features from the extracted features, including the following steps:

[0054] Step 3-1: Divide the dataset into a training set and a test set, with the training set accounting for 70% and the test set accounting for 30%; input the training set for feature selection.

[0055] Step 3-2: Set the parameters of the feature selection evolutionary algorithm: population size P is 100, number of iterations N is 100, number of features is D, select K-nearest neighbor classifier, and the distance value K of K-nearest neighbor classifier is 5.

[0056] Step 3-3: Set the encoding method of each individual in the population to binary encoding method, where 1 represents selecting this set of features and 0 represents not selecting.

[0057] Steps 3-4 define the first grouping format, including the following steps:

[0058] Step 3-4-1: Evaluate the features using SU and ReliefF, and project the rankings of these two evaluation metrics onto a two-dimensional coordinate axis.

[0059] Step 3-4-2: Treat the feature as a solution, and the SU and ReliefF rankings of the feature as two objectives;

[0060] Step 3-4-3: Use the non-dominated ranking method to divide these features into levels, and treat the features at each level as a feature group of an individual.

[0061] Features at the non-dominated front are considered the most important features, features at the second level are considered the second most important, and so on. Since the non-dominated ranking automatically divides features into different levels, the number of feature groups is adaptively determined as Gnum.

[0062] Steps 3-5 define the second grouping method: use the K-means algorithm to group the features projected onto the two-dimensional coordinate axes; where the size of K determines the size of the number of groups, so K is determined according to the number of groups Gnum obtained from the first grouping method.

[0063] Steps 3-6: Construct a multi-form framework: Combine the two grouping forms into a multi-form framework, and treat each feature group as a dimension; then assign a population P1 and a population P2 to the two grouping forms respectively for search.

[0064] Steps 3-7, Population Initialization: Since each digit of an individual represents a set of features, an N1-row D-column matrix and an N2-row D-column matrix are randomly generated. Each value in the matrix is ​​randomly generated between 0 and 1, where each row represents an individual and each column of an individual represents a feature. The average value of each set of features after evaluation by SU and ReliefF is calculated and assigned a weight of 0.5. The weighted value is used as the threshold for that set of features. If the generated value is greater than the set threshold, it is set to 1; otherwise, it is set to 0.

[0065] Steps 3-8: Evaluate the initial population: The 5-fold cross-validation evaluation method is used to evaluate the classification error rate of each individual in the initial population and the ratio of the selected feature to the total number of features using the fitness function of formula (1) to obtain the target value of the initial population. The classification error rate is the first column of the target value, and the ratio of the selected indicator features is the second column of the target value.

[0066]

[0067] In the formula: f error f represents the classification error rate; ratio Indicates the proportion of selected features; c represents the number of categories; FP i S represents the number of prediction errors in category i; i This represents the number of samples in category i.

[0068] Steps 3-9: Identify the optimal solution set in the population: Sort the target values ​​of the multi-objective population according to the non-dominated sorting method, and take the solution set of the first frontier as the optimal solution set of the multi-objective population. If it exceeds 100, select the top 100 individuals according to the Euclidean distance of the distance solutions.

[0069] Step 3-10, Crossover: Perform a crossover operation on each individual in the two populations from Step 3-7. First, the selection of parents is done using a binary tournament method. Then, considering the preservation of important features, if both parents select a certain set of features, then that set of features is considered important and will be retained in the offspring; conversely, if neither parent selects, then the offspring will not select. If one parent selects and the other does not, the probability P of whether the offspring selects or not is determined. c The calculation is as follows:

[0070]

[0071] In the formula: and represents the average ranking of the j-th feature group after SU and ReliefF evaluations, respectively; D represents the total number of features.

[0072] Step 3-11, Mutation: The offspring obtained in Step 3-10 are subjected to mutation with a certain probability, where the mutation probability P is... m Set to 1 / Gnum.

[0073] Step 3-12, Evaluation: Using the 5-fold cross-validation evaluation method, taking Step 3-8 as an example, evaluate the classification error rate and the proportion of selected features for each individual in the new offspring population generated after mutation in Step 3-11, and obtain the target value of the new offspring population, where the classification error rate is the first column of the target value and the proportion of selected features is the second column of the target value.

[0074] Step 3-13, Selection: The two populations are sorted using the non-dominated sorting method to obtain the index of all individuals in the generated offspring and parent sets. The first 50 individuals by index are selected as the next generation of multi-objective population.

[0075] Steps 3-14, knowledge transfer, include the following steps:

[0076] Step 3-14-1: First, obtain the non-dominated fronts of the two newly generated populations, and calculate the Euclidean distance between the centroids of these fronts and the fronts of the previous generation population. That is, obtain the centroids of these fronts and the fronts of the previous generation population in the target space, and then calculate the distance using the Euclidean distance formula between the two centroids. Where x and y are another objective value f. error (X) and f ratio (X); If the front distance between two generations of a population is 0, then the population is considered to be trapped in a local optimum.

[0077] Step 3-14-2 triggers the knowledge transfer mechanism. Assuming that P1 is trapped in a local optimum, the number of times individuals in the P2 population select feature groups is ranked from high to low. The feature group selected most often is considered an important group, and the feature group selected least often is considered a worthless group. The feature groups ranked in the top F are selected as the best feature groups (BestG).

[0078] Step 3-14-3: Find the same features as BestG in the grouping pattern of population P1, and select these features to form the same group as BestG; for the newly formed groups in population P1, except for the individuals at the forefront of population P1, all other individuals are selected from these newly formed groups.

[0079] Step 3-14-4: Re-evaluate all individuals in the P1 population and archive the non-dominant frontier of the P1 individuals before migration.

[0080] Steps 3-15: When one of the populations gets trapped in a local optimum three times, project the two grouped features onto the original feature space, cancel the grouping, and then perform the search with one bit representing one feature. If the non-dominated front distance between two generations of the population is greater than log... 10 When D, the cumulative count is recalculated.

[0081] Step 3-16, Iteration: After each execution of step 3-14, the iteration count is incremented by one, and then the operation is returned to step 3-10.

[0082] Step 3-17, Termination: When the number of iterations reaches the number of iterations T in step 3-2, i.e., 100, the feature selection evolutionary algorithm terminates and outputs the last generation of multi-objective population and its target value.

[0083] Step 3-18: Sort the target values ​​of all individuals output in Step 3-17 in a non-dominated manner to obtain the individuals on the first Pareto front. Record the individual with the highest classification accuracy in the first front, which is the multiple important feature combinations obtained by the feature selection evolutionary algorithm after feature extraction in Step 3.

[0084] Step 4, Application Prediction: Input the selected feature combination into the classifier, then use the well-known K-nearest neighbor classification technique to obtain the label of the unknown text, and use this label as the final result to output the result of the text to be identified and classified.

[0085] Regarding the evaluation of the test results, this invention used three text classification datasets: DBworld, PCMAC, and REATHE. The performance of this invention in recognizing and classifying text using all features of the used datasets was compared. The number of features in the subset with the highest classification accuracy obtained in each run, averaged after 30 runs, is the number of output features of this invention. The classification accuracy of this invention is also the average after 30 runs. As shown in Table 1, this invention can achieve higher classification accuracy using a much smaller combination of features than all features combined. This experiment demonstrates that this invention can effectively reduce the dimensionality of the original data, removing irrelevant and redundant features, and using a small number of useful features for more accurate text recognition and classification.

[0086] Table 1 compares the classification accuracy (perfection) using all text features versus using the combination of output features from this invention.

[0087]

Claims

1. A text classification method based on spatial transformation and multi-form, multi-target feature selection, characterized in that, Includes the following steps: Step 1, Data Preprocessing: This includes text cleaning, word segmentation, and stop word removal, converting the raw text into a machine-processable format; Step 2, Feature Extraction: Represent the text by selecting or designing appropriate features. Commonly used features include bag-of-words model, TF-IDF vectors, and word embeddings. Step 3, Model Training and Evaluation: A multi-form, multi-objective algorithm based on spatial transformation is used to select features from the extracted features, including the following steps: Step 3-1: Divide the dataset into a training set and a test set, with the training set accounting for 70% and the test set accounting for 30%; input the training set for feature selection. Step 3-2, set the parameters of the feature selection evolutionary algorithm: population size P is 100, number of iterations N is 100, number of features is D, select K nearest neighbor classifier, and the distance value K of K nearest neighbor classifier is 5; Step 3-3: Set the encoding method of each individual in the population to binary encoding method, where 1 represents selecting this set of features and 0 represents not selecting it; Steps 3-4 define the first grouping format, including the following steps: Step 3-4-1: Evaluate the features using SU and ReliefF, and project the rankings of these two evaluation metrics onto a two-dimensional coordinate axis. Step 3-4-2: Treat the feature as a solution, and the SU and ReliefF rankings of the feature as two objectives; Step 3-4-3: Use the non-dominated ranking method to divide these features into levels, and treat the features at each level as a feature group of an individual. Features at the non-dominated front are considered the most important, features at the second level are considered the second most important, and so on. Since the non-dominated ranking automatically divides features into different levels, the number of feature groups is adaptively determined as Gnum. Steps 3-5 define the second grouping method: use the K-means algorithm to group the features projected onto the two-dimensional coordinate axes; where the size of K determines the size of the number of groups, so K is determined according to the number of groups Gnum obtained from the first grouping method; Steps 3-6: Construct a multi-form framework: Combine the two grouping forms into a multi-form framework, and treat each feature group as a dimension; then assign a population P1 and a population P2 to the two grouping forms respectively for search; Steps 3-7, Population Initialization: Since each digit of an individual represents a set of features, an N1-row D-column matrix and an N2-row D-column matrix are randomly generated. Each value in the matrix is ​​randomly generated between 0 and 1, where each row represents an individual and each column of an individual represents a feature. The average value of each set of features after evaluation by SU and ReliefF is calculated, and each is assigned a weight of 0.

5. The weighted value is used as the threshold for that set of features. If the generated value is greater than the set threshold, it is set to 1; otherwise, it is set to 0. Steps 3-8: Evaluate the initial population: The 5-fold cross-validation evaluation method is used to evaluate the classification error rate of each individual in the initial population and the ratio of the selected feature to the total number of features using the fitness function of formula (1) to obtain the target value of the initial population. The classification error rate is the first column of the target value, and the ratio of the selected indicator features is the second column of the target value. (1) In the formula: Indicates the classification error rate; c represents the proportion of selected features; c represents the number of categories; This represents the number of prediction errors in category i; The number of samples representing category i; Steps 3-9: Identify the optimal solution set in the population: Sort the target values ​​of the multi-objective population according to the Pareto non-dominated sorting method, and take the solution set of the first frontier as the optimal solution set of the multi-objective population. If it exceeds 100, select the top 100 individuals according to the Euclidean distance of the distance solutions. Step 3-10, Crossover: Perform a crossover operation on each individual in the two populations from Step 3-7. First, the selection of parents is done using a binary tournament method. Then, considering the preservation of important features, if both parents select a certain set of features, then that set of features is considered important and will be retained in the offspring; conversely, if neither parent selects, then the offspring will not select. If one parent selects and the other does not, the probability P of whether the offspring selects or not is determined. c The calculation is as follows: （2） In the formula: and These represent the average rankings of the j-th feature group after SU and ReliefF evaluations, respectively; D represents the total number of features. Step 3-11, Mutation: The offspring obtained in Step 3-10 are subjected to mutation with a certain probability. Set to 1 / Gnum; Step 3-12, Evaluation: Using the 5-fold cross-validation evaluation method, taking Step 3-8 as an example, evaluate the classification error rate and the proportion of selected features for each individual in the new offspring population generated after mutation in Step 3-11, and obtain the target value of the new offspring population, where the classification error rate is the first column of the target value and the proportion of selected features is the second column of the target value. Step 3-13, Selection: The Pareto non-dominated sorting method is used to sort the target values ​​of all individuals in the generated offspring and parent sets of the two populations to obtain the index of all individuals. The first 50 individuals by index are selected as the next generation of multi-objective population. Steps 3-14: Knowledge transfer; Step 3-15: When one of the populations gets trapped in a local optimum three times, project the two grouped features onto the original feature space, cancel the grouping, and then search using one feature per bit. If the non-dominated front distance between two generations of the population is greater than... When this happens, the cumulative count is recalculated; Step 3-16, Iteration: After each execution of step 3-14, the iteration count is incremented by one, and then the operation is returned to step 3-10. Step 3-17, Termination: When the number of iterations reaches the number of iterations T in step 3-2, i.e., 100, the feature selection evolutionary algorithm terminates and outputs the last generation of multi-objective population and its target value; Step 3-18: Sort the target values ​​of all individuals output in Step 3-17 using Pareto non-dominated sorting to obtain the individuals in the first Pareto front after sorting; record the individual with the highest classification accuracy in the first front, which is the multiple important feature combinations obtained by the feature selection evolutionary algorithm after feature extraction in Step 3. Step 4, Application Prediction: Input the selected feature combination into the classifier, then use the well-known K-nearest neighbor classification technique to obtain the label of the unknown text, and use this label as the final result to output the result of the text to be identified and classified.

2. The text classification method based on spatial transformation and multi-form multi-target feature selection according to claim 1, characterized in that, Step 3-14 includes the following steps: Step 3-14-1: First, obtain the non-dominated fronts of the two newly generated populations, and calculate the Euclidean distance between the centroids of the fronts of the two generations and the fronts of the previous generation populations. If the distance between the fronts of one of the populations is 0, it is considered that the population is trapped in a local optimum. Step 3-14-2 triggers the knowledge transfer mechanism. Assuming that P1 is trapped in a local optimum, the number of times individuals in the P2 population select feature groups is ranked from high to low. The feature group selected most often is considered an important group, and the feature group selected least often is considered a worthless group. The feature groups ranked in the top F are selected as the best feature groups (BestG). Step 3-14-3: Find the same features as BestG in the grouping pattern of population P1, and select these features to form the same group as BestG; for the newly formed groups in population P1, except for the individuals at the forefront of population P1, all other individuals are selected from these newly formed groups. Step 3-14-4: Re-evaluate all individuals in the P1 population and archive the non-dominant frontier of the P1 individuals before migration.

Images