Non-equilibrium classification method based on multi-expert instance selection
Through multi-expert instance selection method, combining evolutionary computing and ensemble learning, expert objective function and environment selection strategy are designed, the classification performance instability of imbalanced data sets is solved and efficient classification performance improvement is achieved.
Patent Information
- Application Number
- CN202510554136.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-29
- Publication Date
- 2025-08-12
AI Technical Summary
The existing unbalanced data set processing methods have problems such as model overfitting, loss of most categories of information, redundancy of minority categories, and unstable classification performance caused by category imbalance. The existing methods have failed to effectively utilize the distribution information of the data set.
The multi-expert instance selection method is adopted, combining evolutionary computing and ensemble learning, and three expert objective functions are designed, individual searches are performed through the non-dominant sorting genetic algorithm NSGA-II, combined with environmental selection strategies, and example subsets are optimized to generate high-quality individuals to train the ensemble classifier.
Without affecting the performance of most classes, the recognition accuracy of a few classes is significantly improved, the overall classification performance is improved, and the convergence speed and computing efficiency of the algorithm are accelerated.
Smart Images
Figure CN120470397A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of evolutionary instance selection and ensemble learning for imbalanced data, and in particular to an imbalanced classification method based on multi-expert instance selection. Background Art
[0002] Existing common techniques for handling imbalanced datasets include resampling, data synthesis, cost-sensitive learning, ensemble learning, and instance selection based on evolutionary computation. Resampling and data synthesis reduce the bias caused by class imbalance at the data level. They reduce the number of instances in the majority class or increase the number of instances in the minority class to obtain balanced training data for model training. Cost-sensitive learning and ensemble learning improve the recognition ability of minority classes at the algorithm level by making the model pay more attention to the minority class. The instance selection method based on evolutionary computation integrates multiple technologies, including ensemble learning, undersampling, and multi-objective optimization algorithms, to extract high-quality instances from the training dataset. The selection sequence of these instances is regarded as an individual in the evolutionary algorithm. During the algorithm iteration process, the optimal combination of instances with high prediction accuracy and diversity is sought. These instances are used as training subsets to train the final set of classifiers.
[0003] Existing methods, while some of which incorporate multiple imbalanced learning techniques, still face numerous challenges. Resampling can result in loss of majority class information or redundancy of minority class information. If the class imbalance is high, generating too many synthetic samples can easily lead to model overfitting, resulting in poor performance on multi-classification problems. These methods fail to fully exploit the distribution of the majority and minority classes in the original data, resulting in a training process that favors minority class performance at the expense of majority class performance. Cost-sensitive learning uses a cost matrix to assign different penalties to instances with incorrect predictions, aiming to minimize the loss function. However, maintaining the cost matrix incurs additional computational costs, and setting it requires prior domain knowledge of the dataset, making it unsuitable for general applications, such as in medical image classification and fraud detection. Existing instance selection methods based on evolutionary computation and ensemble learning still use balanced resampling to initialize the population and lack effective constraints on individual instances. This makes it impossible to guarantee that the classification performance of the selected instances is superior to that of the original data, leading to unstable classification performance of the ensemble classifier. Summary of the Invention
[0004] Purpose of the invention: Taking into account the limitations of existing imbalanced learning technologies, the present invention combines evolutionary computing and ensemble learning to provide an unbalanced classification method based on multi-expert instance selection, giving full play to the advantages of multi-objective optimization algorithms in instance selection, constraining individuals, and evolving to search for high-quality instance subsets. The present invention designs three effective expert objectives for individuals, and each expert is designed to optimize an objective function that focuses on the distribution of a specific class. A random initialization strategy is adopted in the initial stage to alleviate the negative impact of balanced resampling on the original training set, thereby obtaining a set of high-quality and diverse initial individuals. The constraint processing mechanism is incorporated into the environmental selection process, giving priority to feasible individuals, that is, giving priority to instance subsets that produce better classification performance than using all training instances, so as to reduce the search range of feasible individuals and improve the search efficiency and convergence speed of the algorithm. The present invention comprehensively considers the distribution of majority and minority classes, effectively amplifying the performance of minority classes without affecting the performance of majority classes, thereby improving the overall classification performance.
[0005] Technical solution: The unbalanced classification method based on multi-expert instance selection described in the present invention includes the following steps:
[0006] (1) Load an unbalanced dataset, divide it into training and test sets according to stratified sampling, train the model with all instances in the training set, and calculate the individual constraint threshold based on the objective function;
[0007] (2) Encode the individuals 0-1 and use 90% of the number of minority class instances as the baseline number. For each class, generate a random number between the baseline and the number of all instances of the class as the initial number of instances selected for the class to randomly initialize all individuals in the population;
[0008] (3) The non-dominated sorting genetic algorithm NSGA-II is used in the instance subset search process. The tournament selection algorithm is used in the initial population to select individuals for crossover and mutation operations to obtain offspring, and the offspring and parent populations are merged;
[0009] (4) Remove individuals with the same instance selection combination in the offspring and parent populations. During environment selection, the instance selection subset is obtained from the individuals in the current population to train the model and calculate the three expert objectives. The degree of violation is calculated based on the constraints, and the new population is selected to enter the next generation based on the number of feasible individuals.
[0010] (5) Integrate feasible individuals from the final population to obtain an integrated classifier.
[0011] Furthermore, in step (1), the objective function is:
[0012] max F(x)=(acc ocd (x),acc bcd(x),acc icd (x)) T
[0013] stg1(x)=acc ocd (x)-acc ocd (x full )≥0
[0014] g2(x)=acc bcd (x)-acc bcd (x full )≥0
[0015] g3(x)=acc icd (x)-acc icd (x full )≥0
[0016] where x=(x1,x2...,x n ) T ∈{0,1} n
[0017] x full =(1,1...,1) T
[0018] Where n is the total number of instances in the training dataset. x represents an individual, x full is the individual that selects all training instances. The above formula is to satisfy g i (x) under the condition that F(x) is maximized.
[0019] Furthermore, in step (1), the three expert objective functions are:
[0020] acc ocd (x), based on the original non-equilibrium distribution:
[0021]
[0022] acc bcd (x), based on the equilibrium distribution:
[0023]
[0024] acc icd (x), based on the inverse non-equilibrium distribution:
[0025]
[0026] Where TP (True Positives) is the number of correctly predicted instances represents the total number of instances in the training subset, Represents the original training set. K is the number of categories in the training dataset, TP i and denote the number of correctly predicted instances of class i and the total number of selected instances, respectively. for Considering these three objectives simultaneously can effectively handle different class distributions and balance the performance of minority and majority classes.
[0027] Furthermore, in step (1), the combined constraints are: use all training instances to train the classifier and calculate acc ocd (x full ),acc bcd (x full ),acc icd (x full ) as the constraint threshold. Determine whether the three expert goals of each individual meet the conditions:
[0028]
[0029] When CV(x)=0, the individual is a feasible individual, otherwise it is an infeasible individual.
[0030] Furthermore, the specific method of population initialization in step (2) is to encode the individuals from 0 to 1, where 1 indicates that the instance is selected and 0 indicates that it is discarded. Taking 90% of the number of minority class instances in the training set as a benchmark, a random number between the benchmark and the total number of minority class instances is generated to alleviate the imbalance of the initial instance subset of the individuals.
[0031] Furthermore, in step (3), the non-dominated sorting genetic algorithm NSGA-II is used in the instance subset search process, and the size of the tournament selection algorithm is set to 3.
[0032] Furthermore, in step (4), the strategy of environmental selection includes: if the number of feasible individuals exceeds the population size, screening is performed according to the non-dominated level and crowding degree; if the number of feasible individuals does not reach the population size, the missing individuals are supplemented from the infeasible individuals according to the degree of breach of contract.
[0033] Furthermore, in step (5), the integration strategy includes: if the number of feasible individuals exceeds the population size, the integrated individuals are selected according to the non-dominated sorting of the feasible individuals; if the feasible individuals do not reach the population size or there are no feasible individuals, the feasible individuals or the infeasible individuals with the least degree of default are integrated first.
[0034] Furthermore, in step (5), the final classification results are integrated by voting among the base classifiers.
[0035] An electronic device described in the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, it implements any one of the unbalanced classification methods based on multi-expert instance selection.
[0036] The storage medium of the present invention stores a computer program, wherein the computer program, when executed by a processor, implements any one of the unbalanced classification methods based on multi-expert instance selection.
[0037] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: The present invention provides an unbalanced classification method based on multi-expert instance selection, which not only takes into account the original distribution of classes in the data, but also comprehensively considers the uniform distribution and inverse unbalanced distribution of the majority class and the minority class, and determines three expert objective functions based on the three distributions. In addition, in the search of instance subsets, the instance selection subsets of individuals are different, and the category weights involved in the inverse unbalanced distribution in the objective function remain unchanged and do not decrease with the increase in the number of minority class instances, so as to alleviate the deviation caused by category imbalance. In the initial stage, the population is randomly initialized to avoid the large-scale loss of majority class information in the data set and reduce the impact of class imbalance on the classification effect. Each individual is effectively constrained and the decision-making mechanism is optimized, so that the feasible individuals in the integrated individuals are better than the original training set, thereby improving the overall classification performance of the integrated classifier. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 It is a schematic flow diagram of the present invention;
[0039] Figure 2 is the comparison result of Gmean of the present invention;
[0040] Figure 3 This is the comparison result of mAUC of the present invention. DETAILED DESCRIPTION
[0041] To help those skilled in the art better understand the present invention, the following will use a real-world imbalanced dataset to illustrate the specific implementation and experimental analysis of the present invention, and provide a clear and complete description of the method of the present invention.
[0042] Note: The following experiments and descriptions are conducted using the multilayer perceptron as the basic classifier. However, the method proposed in this invention is not limited to the multilayer perceptron. In theory, any classifier can be used as the base classifier, such as support vector machine, decision tree, neural network, etc.
[0043] Table 1 Basic characteristics of the dataset
[0044]
[0045] Table 2 Parameter description of multilayer perceptron
[0046]
[0047] To validate the effectiveness of the present invention, we used the aforementioned method steps to analyze an unbalanced dataset from the UCI database. The UCI Machine Learning Repository is a collection of databases, domain theories, and data generators for empirical analysis of machine learning algorithms. The unbalanced dataset and its basic parameters are shown in Table 1. A multilayer perceptron (MLP) was used as the base classifier, with its parameters shown in Table 2. The learning rate was set to 0.1, and all other parameters were default.
[0048] Some parameter settings involved in the present invention are as follows:
[0049] 1) Population size: 30;
[0050] 2) Iterations: 30;
[0051] 3) Tournament selection: size is set to 3;
[0052] 4) Crossover rate, mutation rate: 100%, random numbers between 5% and 20%.
[0053] In order to evaluate the classification performance of the classifier, two evaluation indicators, Gmean and mAUC, are used to measure the classification performance on the test set. Gmean is the geometric mean of the recall rate of each category, and mAUC is the extension of AUC from the binary classification problem to the multi-classification problem. The classification performance of the present invention is compared with other imbalanced learning techniques. Figure 2 、 Figure 3 The specific experimental results of Gmean and mAUC are shown respectively, comparing the present invention with other methods, and the average value is obtained by running 30 times on different data sets.
[0054] like Figure 1 As shown, the embodiment of the present invention provides an unbalanced classification method based on multi-expert instance selection, comprising the following steps:
[0055] Step 1: Load the unbalanced dataset, including feature data and labels, divide the training set and test set according to the original distribution, train the model and calculate the constraint threshold.
[0056] Step 2: Based on the training set, encode the individuals in the population from 0 to 1, keeping the length of the individual encoding consistent with the length of the training set. Using the 90% of the class with the smallest number of instances in the training set as the benchmark, generate a random number between the benchmark and the total number of instances in the class, and randomly initialize the population.
[0057] Step 3: First, perform a non-dominated sort on the initial population and use tournament selection to select individuals from the parent generation. If there is a tie among the selected individuals, consider the crowding distance. Perform single-point crossover and mutation on the selected individuals to obtain the offspring. The parent and offspring generations are then merged to form a new population.
[0058] Step 4: Remove duplicate individuals, divide the population into feasible individuals and infeasible individuals by CV(x), and select individuals to enter the next generation based on the number of feasible individuals.
[0059] Step 5: Repeat steps 1-4 until the iteration ends, integrate the feasible individuals in the final population, and integrate the classification results of the classifier by voting.
[0060] Figure 2 、 Figure 3 The present invention is compared with the E-MOSAIC, SMOTE, RUS, and NOS methods. A bar graph is provided for each dataset to illustrate the performance comparison of these methods. The first one is the method proposed by the present invention to highlight the performance difference with other methods. Figure 2 、 Figure 3 The results analysis shows that in most cases, the present invention outperforms other methods and has higher Gmean and mAUC. In the PageBlocks and Dermatology datasets, the Gmean is lower than some experimental methods. Although the present invention did not achieve the best results on these datasets, it did not show the worst results in any dataset. The present invention showed the best performance overall on all datasets. This is because random initialization enhances the diversity of instance subsets, which reduces the class imbalance ratio to a certain extent and is conducive to generating high-quality initial individuals. In addition, the constraints applied to the three expert objectives guide the search process in the desired direction, thereby accelerating convergence and improving the classification performance of the ensemble classifier.
[0061] The present invention proposes a new classification method for multi-class unbalanced data. Based on the multi-objective genetic algorithm, three effective expert objective functions are set for each individual, and the distribution characteristics of the majority class and the minority class in the data set are comprehensively considered. Without interfering with the classification performance of the majority class, the classification ability of the minority class is effectively enhanced, thereby significantly improving the recognition accuracy of the minority class. At the same time, by imposing constraints on the individuals, the multi-layer perceptron model trained with the original training set is used as the base classifier to constrain all individuals. The constraints require that each individual has better classification performance than all training instances in the three expert objectives, thereby optimizing the search space of feasible individuals, avoiding unnecessary calculations, and further improving the convergence speed and computational efficiency of the algorithm. Through this method, the class imbalance problem can be fully considered within the framework of multi-objective optimization, the performance of the classification model on the minority class can be optimized, and at the same time, the improvement of the overall classification effect can be ensured.
Claims
1. An unbalanced classification method based on multi-expert instance selection, characterized in that: The following steps are involved: (1) Load the unbalanced dataset, divide it into training set and test set according to stratified sampling, train the model with all instances of the training set, and calculate the constraint threshold based on the objective function; (2) Individuals are coded 0-1, where 0 and 1 represent whether an instance is selected for the training dataset, respectively. 90% of the number of minority class instances is used as a benchmark. For each class, a random number is generated between the benchmark and the number of all instances of the class, which is used as the initial number of instances selected for that class. This is used to randomly initialize all individuals in the population. (3) Use the non-dominated sorting genetic algorithm NSGA-II to search for a subset of instances. Use the tournament selection algorithm to select parents from the initial population, perform single-point crossover and mutation operations to obtain offspring, and merge the offspring and parent populations; (4) Remove individuals with the same instance selection combination in the offspring and parent populations. When selecting the environment, constrain the current population, then calculate the breach degree CV(x), and select a new population to enter the next generation based on the number of feasible individuals and the size of CV(x); (5) Integrate feasible individuals from the final population to obtain an integrated classifier.
2. The unbalanced classification method based on multi-expert instance selection according to claim 1, characterized in that: In step (1), the three expert objectives in the objective function include: acc ocd (x), based on the original non-equilibrium distribution: acc bcd (x), based on the equilibrium distribution: acc icd (x), based on the inverse non-equilibrium distribution: Where TP (True Positives) is the number of correctly predicted instances, represents the total number of instances in the training subset, represents the original training set. K is the number of categories in the training dataset, TP i and denote the number of correctly predicted instances of class i and the total number of selected instances, respectively. for The number of instances of category i in .
3. The unbalanced classification method based on multi-expert instance selection according to claim 1, characterized in that: In step (1), the constraint is: use all training instances to train the classifier and calculate acc ocd (x full ),acc bcd (x full ),acc icd (x full ) as the constraint threshold. A subset of instances of each individual is selected, and after training the model, its three expert objectives are calculated and the conditions are met: When CV(x)=0, the individual is a feasible individual, otherwise it is an infeasible individual.
4. The unbalanced classification method based on multi-expert instance selection according to claim 1, characterized in that: In step (2), the specific method of random initialization of the population is: taking 90% of the number of minority class instances in the training set as a benchmark, generating a random number between the benchmark and the total number of instances of the class as the number of instances selected for the class, so as to alleviate the imbalance ratio of the initial instance subset of the individual.
5. The unbalanced classification method based on multi-expert instance selection according to claim 1, characterized in that: In step (3), the non-dominated sorting genetic algorithm NSGA-II is used in the instance subset search process.
6. The unbalanced classification method based on multi-expert instance selection according to claim 1, characterized in that: In step (4), the strategy of environmental selection includes: if the number of feasible individuals exceeds the population size, screening is performed according to the non-dominated level and crowding degree; if the number of feasible individuals does not reach the population size, the missing individuals are supplemented from the infeasible individuals according to the degree of breach of contract.
7. The unbalanced classification method based on multi-expert instance selection according to claim 1, characterized in that: In step (5), the integration strategy includes: if the number of feasible individuals exceeds the population size, the integrated individuals are selected according to the non-dominated sorting of the feasible individuals; if the number of feasible individuals does not reach the population size or there are no feasible individuals, the feasible individuals or the infeasible individuals with the smallest degree of default are integrated first.
8. The unbalanced classification method based on multi-expert instance selection according to claim 1, characterized in that: In step (5), the classification results of the classifiers are finally integrated by voting.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the computer program is loaded into a processor, the computer program implements the unbalanced classification method based on multi-expert instance selection according to any one of claims 1 to 8.
10. A storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the computer program implements the unbalanced classification method based on multi-expert instance selection according to any one of claims 1 to 8.