Class imbalance analysis method based on data life cycle evolution
By constructing a data lifecycle evolution model and using a Bayesian optimization adaptive fusion method, the problem of sample lifecycle anomalies and interactive entanglement in imbalanced data classification is solved, thereby improving the minority class recognition rate and overall classification performance.
Patent Information
- Application Number
- CN202610117809.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-28
- Publication Date
- 2026-05-08
AI Technical Summary
Existing technologies struggle to effectively distinguish between sample lifecycle anomalies and interactive entanglements when classifying imbalanced data, resulting in low recognition rates for minority class samples by classification models. Furthermore, traditional methods rely on complex parameter designs, limiting their generalization capabilities.
By constructing a data lifecycle evolution model, dynamic indicators of samples at different evolution stages are extracted, and Bayesian optimization is combined to achieve adaptive fusion of multidimensional evolution weights, which are then embedded into a randomly configured network for classification.
It achieves high-precision classification of imbalanced data, improves the recognition rate of minority class samples and the overall classification performance, and avoids the blindness of manual parameter tuning.
Smart Images

Figure CN121997026A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of imbalanced data analysis, and in particular relates to a quasi-imbalanced analysis method based on the evolution of data lifecycle. Background Technology
[0002] In real-world applications, data imbalance is a common problem. For example, in medical diagnosis, the number of diseased samples is extremely small compared to healthy samples; in financial risk control, fraudulent transactions account for only a tiny percentage of all transaction samples; and in cybersecurity, there is a significant difference between malicious intrusion traffic and normal traffic. This imbalance often causes traditional classifiers to favor the majority class during training, resulting in a significant decrease in the recognition rate of minority class samples, making them prone to false negatives and thus affecting the overall system performance and reliability.
[0003] To address the imbalanced data classification problem, researchers have proposed numerous data-level and algorithmic solutions. Among data-level methods, common approaches include undersampling, oversampling, and mixed sampling. Undersampling methods balance the data distribution by reducing majority class samples, but are prone to losing crucial information; oversampling methods expand the sample size by replicating or synthesizing minority class samples, but often introduce noise or exacerbate class overlap; mixed sampling combines the advantages of both, but may still disrupt the original data distribution structure. At the algorithmic level, ensemble learning, cost-sensitive learning, and clustering or feature selection-based strategies are widely used. These methods can improve minority class identification capabilities to some extent, but they are highly dependent on the data distribution when dealing with complex data, often rely on complex parameter designs, and have limited generalization ability.
[0004] However, in deep data analysis, simply maintaining quantitative balance often fails to resolve core contradictions because imbalanced data is usually accompanied by more complex evolutionary anomalies and entanglement issues. Specifically, outlier problems are not simply geometric deviations; they are essentially lifecycle anomalies exhibited by samples during the data structure generation process. Some samples may be transient noise that appears and disappears in an instant, or they may be lagging individuals outside the mainstream evolutionary path. If they are included in training indiscriminately, they will severely interfere with the model's convergence direction. Class overlap problems reflect atypical coexistence or excessive entanglement of samples from different classes in the early stages of evolution, manifested as strong interactive connections being established prematurely by samples from different classes. Traditional methods often lack the ability to discern this dynamic evolution, making it difficult to distinguish whether a minority class sample located at the boundary is a core skeleton with high information content or entangled noise causing classification confusion.
[0005] Therefore, how to break through the limitations of the static geometric perspective in imbalanced data classification, establish a life cycle evolution model that can simulate the entire process of data from isolation to symbiosis, and capture the dynamic response of samples in different evolutionary stages such as merging and assimilation, association closure, and configuration derivation, so as to achieve accurate identification of outliers, effective suppression of interactive overlap, and adaptive weighting of key evolutionary samples, has become an important technical challenge that urgently needs to be overcome in the field of imbalanced data classification. Summary of the Invention
[0006] Objective: The technical problem this invention aims to solve is to address the shortcomings of existing technologies by providing a class-imbalanced analysis method based on data lifecycle evolution. This method simulates the lifecycle process of samples evolving from independent individuals to complex structures, extracting indices representing the characteristics of each evolutionary stage, such as assimilation, association closure, and configuration derivation. Based on these indices, a sample-level comprehensive weight matrix that balances outlier and overlap suppression is constructed. Bayesian optimization is then used to adaptively fuse multidimensional evolutionary weights, which are finally embedded in an evolution-constrained randomized configuration network to achieve high-precision classification of imbalanced data.
[0007] This method includes the following steps:
[0008] Step 1: Construct a full life cycle evolution model of data samples to simulate the dynamic process of samples evolving from discrete independent states to complex structural forms. Extract dynamic indicators that characterize the structural response characteristics of samples at different evolution levels. The dynamic indicators include the life cycle merging and assimilation scale that describes the loss of individual independence of samples and integration into local manifolds, the life cycle association and closure scale that describes the formation of stable closed pathways in the neighborhood interaction network of samples, and the life cycle multidimensional configuration derivation scale that describes the participation of samples in constructing high-dimensional spatial topological skeletons and envelope structures.
[0009] Step 2: Perform a panoramic scan and activity statistical analysis of the dynamic indicators of structural response characteristics in different evolutionary levels. Based on the survival time and structural contribution of samples in the evolution process, identify abnormal behaviors and construct a life cycle outlier suppression mechanism to identify and weaken the weight contribution of transient noise samples that die out quickly in the early stage of evolution and free samples that participate in structural construction in the late stage of evolution during the classification model training process.
[0010] Step 3: Monitor the interaction behavior between different categories of samples in real time during the dynamic advancement of the full life cycle evolution model, identify early interaction structures including direct cross-class connections and high-order mixed entanglement, and construct a life cycle class overlap suppression mechanism by calculating the evolution stage position and entanglement depth of the first cross-class contact of a sample. This mechanism is used to suppress the negative impact of boundary fuzzy samples that exhibit atypical cross-class affinity in the early stages of evolution on the classification decision surface.
[0011] Step 4: Construct a quantitative balance mechanism based on the abundance of life cycle evolution, and fuse the quantitative balance mechanism with the life cycle outlier suppression mechanism and the life cycle class overlap suppression mechanism in a convex combination form; transform the search for the optimal fusion coefficient into a hyperparameter optimization problem, and use a Bayesian optimization algorithm with Gaussian process regression as the surrogate model and the expectation improvement criterion to automatically find the optimal solution, thereby determining the life cycle comprehensive weight vector that can maximize the system evolution stability.
[0012] Step 5: Construct a sample-level diagonal weighted matrix from the lifetime integrated weight vector and embed it into the loss function and hidden node generation constraints of the randomly configured network. Train the classification model using the weighted objective function and apply the classification model to the test set to achieve accurate classification of imbalanced data. Output performance metrics including geometric mean (G-mean), recall (Sensitivity), F1 score (F1-score), and precision.
[0013] Step 1 includes:
[0014] Step 1-1: Define the imbalanced dataset ,in Indicates the first one sample 3D feature vectors Indicates the first The category label of each sample, The total number of samples, Represent the real number space; calculate the heterogeneous interaction strength matrix between samples. , where matrix elements Indicates sample With sample Structural affinity between them; based on the heterogeneous interaction strength matrix, construct the evolutionary time series... Dynamically driven full life cycle evolution model If and only if the heterogeneous interaction strength between any two samples in the imbalanced dataset is less than or equal to At that time, it was determined that active connection pathways had been established between the samples and formed a local symbiotic structural unit;
[0015] Steps 1-2: Extract the lifecycle merging and assimilation scale, establish a community evolution simulation mechanism, and simulate the process of independent samples merging into a local manifold; extract the lifecycle merging and assimilation scale of the samples. All potential connections in the heterogeneous interaction strength matrix According to structural affinity Sort the connections from smallest to largest, then iterate through the sorted connections. If the sample and If they belong to different independent evolutionary communities, then perform community assimilation; define samples. Lifecycle assimilation scale as follows:
[0016] ,
[0017] in, and The sample index represents the two sample numbers corresponding to any potential connection in the heterogeneous interaction strength matrix, i.e., the sample numbers corresponding to the first two samples in the sample set. Sample With the Sample , The critical interaction strength that leads to structural fusion of independent communities;
[0018] Steps 1-3: Extract lifecycle-related closure scales and define samples. In evolutionary timeline The following interactive neighborhood set When the correlation matrix third-order trace This indicates the emergence of interconnected cyclical pathways in the full life cycle evolution model, which is considered the formation of stable closed associations; a sample is defined. Life cycle related closed-scale approximation For the sample The average minimum interaction strength that forms a closed loop within the interaction neighborhood:
[0019] ,
[0020] in, The cardinality of the set of interactive neighborhoods;
[0021] Step 14: Extract the multidimensional configuration-derived scale of the life cycle. The multidimensional configuration-derived scale of the life cycle of the sample is extracted using the neighborhood Voronoi approximation method. When the sample The number of neighborhood samples satisfies the supporting condition. At that time, calculate the high-dimensional configuration-derived scale:
[0022] ,
[0023] in, This is the preset configuration magnification factor; if Then define .
[0024] Step 2 includes:
[0025] Step 2-1: Determine the full lifecycle evolution scanning interval. Based on the average heterogeneous interaction intensity of minority class samples and the mean of the global lifecycle merging and assimilation scale, determine the observation interval of the evolutionary time series. and in the observation interval Internal generation includes Multi-stage scanning sequence of evolutionary time nodes ,in This represents the threshold for the initial heterogeneous interaction intensity in the lifecycle evolution scan. This represents the threshold for the termination of heterogeneous interaction intensity in the lifecycle evolution scan. Indicates the first The heterogeneous interaction intensity thresholds corresponding to each life cycle evolution stage; the minority class samples correspond to the minority class category, and the number of minority class samples is denoted as... The majority class samples correspond to the majority class, and the number of samples in the majority class is denoted as . Among them, the minority class is the class with fewer samples in the current training data, and the majority class is the class with more samples, i.e. ;
[0026] Step 2-2: Calculate the average activity weight across stages, based on the observed samples in the evolutionary time series. The life cycle dynamics changes are calculated using an exponential mapping function. In the Average weight across stages in the evolutionary dimension:
[0027] ,
[0028] in, These correspond to the three lifecycle dimensions of merging and assimilation, association closure, and configuration derivation, respectively. For the sample A set of time indices that demonstrate an effective evolutionary response; For the first The mean of global dynamic indicators in a given dimension, which is the average of the dynamic indicators extracted from all samples in that evolutionary dimension. It is a smoothing factor;
[0029] Step 2-3: Introduce category correction coefficients The average activity weight across different stages is adjusted to obtain the corrected lifecycle weight. The correction coefficient for minority class samples is set to be greater than that for majority class samples;
[0030] Step 2-4: Construct evolutionary hysteresis suppression logic. Determine hysteresis behavior based on the lifecycle response time sequence of samples in the multi-stage evolutionary scan sequence. When the lifecycle assimilation scale of a sample is significantly lagging behind the overall level, or its effective evolutionary response is mainly concentrated in the latter part of the evolutionary scan interval, the sample is determined to be a lifecycle-hysteresis individual. For samples determined to be lifecycle-hysteresis individuals, apply exponential compression to the original lifecycle weights of the samples, so that the sample weights decrease monotonically as the degree of hysteresis increases, thereby weakening the influence of the samples in the training of the classification model. For samples that do not meet the hysteresis determination conditions, keep the weights obtained from Step 2-2 to Step 2-3 unchanged.
[0031] Steps 2-5: Construct an adaptive cooperative coupling model for multidimensional lifecycle signals, treating the dynamic weights of the three dimensions of assimilation, correlation closure, and configuration derivation as independent evolutionary signal channels, and defining a single lifecycle outlier suppression weight. The result of coordinated modulation of multi-channel signals:
[0032] ,
[0033] Among them, the weighting coefficient The search is determined by particle swarm optimization under simplex constraints. The objective function is the verification index including geometric mean (G-mean), recall (Sensitivity), F1 score (F1-score), and precision. The three-dimensional update is adopted to make the contribution of the three dimensions consistent with the actual evolution signal strength of the data.
[0034] Steps 2-6: Perform global normalization on the lifetime outlier suppression weights of all samples and construct a sample-level diagonal lifetime outlier suppression matrix. Where diag represents a diagonal matrix, For the first Outlier suppression weights corresponding to each sample.
[0035] Step 3 includes the following steps:
[0036] Step 3-1: Real-time monitoring of cross-class interaction structures during the evolutionary process; during the advancement of the full life-cycle evolutionary model, when samples from different categories... In evolutionary timeline The following satisfies the heterogeneous interaction strength When a direct cross-class connection path is established, it is denoted as a set. :
[0037] ,
[0038] If three sample points In evolutionary timeline Active connectivity pathways were established in each pair of samples, and three sample points were used. If the category labels are not completely consistent, it is determined that a higher-order hybrid entangled structure has been formed, denoted as set. :
[0039] ;
[0040] Step 3-2: Evaluate the lifecycle entanglement initiation index and basic purity confidence weight of the sample; track the sample. The critical evolution timing for the first time involved in constructing a cross-class interaction structure is defined as the entanglement initiation point. The basic purity confidence weights are constructed using a sigmoid mapping function with shift and scaling factors. :
[0041] ,
[0042] in, For entanglement boundary threshold; The coefficient of discrimination is exp; exp represents the natural exponential function.
[0043] Step 3-3: Calculate the inter-class conflict intensity and density suppression factor;
[0044] Steps 3-4: Generate a lifecycle class overlap suppression mechanism.
[0045] Step 3-3 includes: calculating the sample Cross-class connectivity at a specified evolutionary time sequence :
[0046] ,
[0047] in, For the elements of the adjacency matrix of the evolutionary model, As an indicator function, a density suppression factor is constructed based on cross-class connectivity. .
[0048] Steps 3-4 include: assigning confidence weights to the basic purity. With density inhibition factor Couple the components to generate the final lifecycle class overlap suppression weights. and to Global normalization is performed to construct a sample-level diagonal lifetime class overlap suppression matrix. ,in Indicates the first The lifetime class overlap suppression weights corresponding to each sample.
[0049] Step 4 includes the following steps:
[0050] Step 4-1: Construct a sample quantity balancing mechanism based on lifecycle evolution abundance. First, calculate the imbalance ratio between the majority and minority classes, take the square root of the imbalance ratio, and set an upper limit threshold. In the majority dataset scenario, since the imbalance ratio is usually not too large, the upper limit threshold is set to 10, resulting in a smoothed basic quantity factor. Then, calculate the ratio of the effective evolutionary response quantity to the total number of samples in the three lifecycle evolution dimensions of merging and assimilation, association closure, and configurational derivation for both classes (majority and minority). Take the arithmetic mean of the ratios in the three lifecycle dimensions to obtain the comprehensive evolutionary abundance ratios for the minority and majority classes. Finally, calculate the ratio of the comprehensive evolutionary abundance ratio of the majority class to the comprehensive evolutionary abundance ratio of the minority class. Define this ratio as the evolutionary adjustment coefficient and multiply it by the basic quantity factor as the final weighting multiple for the minority class samples, constructing a diagonal quantity weight matrix. ;
[0051] Step 4-2: Construct a convex combination fusion model with multidimensional constraint weights; combine the diagonal weight matrix Sample-level diagonal lifetime outlier suppression matrix and sample-level diagonal lifetime class overlap suppression matrix The fusion is performed using a convex combination method to obtain the comprehensive weight vector W:
[0052] ,
[0053] Among them, the fusion coefficient , , satisfy and ;
[0054] Step 4-3: Establish a surrogate model based on Gaussian process regression with fusion parameters; use the fusion coefficients to adjust the weights for sample size, lifecycle outlier suppression, and lifecycle class overlap suppression. The search problem is transformed into a hyperparameter optimization problem for imbalanced classification performance; a Bayesian optimization method is introduced, and Gaussian process regression is used as a surrogate model to approximate the fusion coefficients. The nonlinear mapping relationship between the weights and the performance evaluation index of imbalanced classification enables adaptive fusion of multi-source lifecycle weights.
[0055] Step 4-4: In the Bayesian optimization process, the expected improvement criterion is used as the sampling strategy. The classification performance indicators, including G-mean, Sensitivity, F1-score, and Precision, are used as the objective function. The surrogate model is iteratively evaluated and updated to finally obtain the optimal combination of fusion coefficients. Thus, the comprehensive weight vector X1 is determined; These represent the optimal sample size weight, the optimal lifetime outlier suppression weight, and the optimal lifetime class overlap suppression weight, respectively.
[0056] Step 5 includes the following steps:
[0057] Step 5-1: Initialize the parameter system of the evolution-constrained stochastic network; set the maximum hidden layer node capacity of the network. Maximum number of candidate searches generated by the node The system expects to tolerate residuals. Random weight generation interval and contractile factor ,in This represents the upper bound (radius of the value range) of the magnitude of the random weights of hidden layer nodes in a randomly configured network.
[0058] Step 5-2: During the iterative addition of nodes, generate the interval from the random weights. A set of input weights and biases for candidate hidden layer nodes are randomly generated, and the output vectors of the candidate hidden layer nodes are calculated.
[0059] Step 5-3: Perform supervised constraint-based node selection; calculate the effectiveness index of candidate nodes according to the supervised inequality constraint criterion, and select effective evolving hidden layer nodes that can reduce the current residual (if a candidate node satisfies the supervised inequality constraint and further reduces the network residual after its introduction, then the candidate node is determined to reduce the current residual) and satisfy the linear independence constraint, add them to the network structure, and synchronously update the hidden layer output matrix. ;
[0060] Step 5-4: Embed the lifecycle comprehensive weight matrix to solve the output layer; construct the comprehensive weight vector X1 into a sample-level diagonal weighted matrix. The output layer weights are analytically solved using the weighted least squares method. ,in, For the training set label matrix, Represents the generalized inverse of a matrix. Indicates transpose;
[0061] Step 5-5: Evaluate the model's evolution convergence state; calculate the current model's training residuals and determine whether the training residuals are less than the expected tolerance error. Or has the current number of hidden layer nodes reached the capacity limit? If any stopping condition is met, the model construction is considered complete; otherwise, the current residual is updated and the process returns to step 5-2 to continue incrementally adding nodes.
[0062] Steps 5-6: Use the trained model parameters to build the final classifier, perform classification prediction on the test set, and calculate the classification performance index.
[0063] The present invention also provides an electronic device, including a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method.
[0064] The present invention also provides a storage medium storing a computer program or instructions that, when the computer program or instructions are run on a computer, execute the steps of the method described.
[0065] The beneficial effects of this invention are as follows: By constructing a full lifecycle evolution model, this invention can dynamically reconstruct the growth and connection process of data from isolated individuals to complex groups at the sample level. Through the identification of lifecycle characteristics such as sample merging, closure, and derivation, as well as real-time monitoring of cross-class interaction behavior, this invention effectively distinguishes between valid information and interference noise in the data, accurately suppressing the negative impacts of outlier samples with delayed evolution and overlapping samples that have prematurely contacted each other. Thus, it achieves accurate weighting and correction of complex imbalanced distributions during the data preprocessing stage.
[0066] Building upon this foundation, this invention employs a multidimensional adaptive fusion of quantitative weights based on evolutionary abundance, outlier suppression weights, and overlap suppression weights. The optimal fusion coefficients are automatically determined using a Bayesian optimization method with Gaussian process regression as the surrogate model, avoiding the blindness of manual parameter tuning and enabling the weight generation process to flexibly adapt to the distribution characteristics of different datasets. The fused comprehensive weights are embedded in the training process of a randomly configured network, retaining the network's advantages of fast modeling speed and lightweight structure while forcing the model to focus on key samples with high lifetime confidence, thereby significantly improving the recognition rate of minority class samples and overall classification performance. Attached Figure Description
[0067] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.
[0068] Figure 1 This is a flowchart of the class imbalance analysis method based on data lifecycle evolution of the present invention;
[0069] Figure 2 This is a schematic diagram of the evolution-constrained random allocation network in this invention. Detailed Implementation
[0070] like Figure 1 As shown, this invention provides a method for analyzing multi-scale persistent homology-type imbalanced data, comprising the following steps:
[0071] Step 1: Construct a full lifecycle evolution model for data samples to simulate the dynamic process of samples evolving from discrete, independent states to complex structural forms. In this process, dynamic indicators characterizing the structural response characteristics of samples at different evolutionary levels are extracted. These indicators specifically include the lifecycle merging and assimilation scale, which describes the loss of individual independence and integration into the local manifold; the lifecycle association and closure scale, which describes the formation of stable closed pathways in the sample's neighborhood interaction network; and the lifecycle multidimensional configuration derivation scale, which describes the sample's participation in constructing a high-dimensional spatial topological skeleton and envelope structure. Step 1 specifically includes the following steps:
[0072] Step 1-1: Define the imbalanced dataset ,in Indicates the first one sample 3D feature vectors Indicates the category label of the sample. Given the total number of samples. Calculate the heterogeneous interaction strength matrix among the samples. , where matrix elements Indicates sample With sample The structural affinity between them. Based on the distance matrix, construct the evolution time series... Dynamically driven full life cycle evolution model If and only if the heterogeneous interaction strength between any two samples in the sample set is less than or equal to At that time, it was determined that active connection pathways had been established between these samples, forming a local symbiotic structural unit;
[0073] Steps 1-2: Extract the lifecycle merging and assimilation scale to establish a community evolution simulation mechanism, simulating the process of independent samples merging into a local manifold. Extract the lifecycle merging and assimilation scale of the samples. All potential connections in the heterogeneous interaction strength matrix. According to strength Sort the connections from smallest to largest. Iterate through the sorted connections, processing each connection sequentially. If the sample and If the population belongs to different independent evolutionary communities, then community assimilation is performed. Define the sample. Lifecycle assimilation scale as follows:
[0074] ,
[0075] in, and The sample index represents the two sample numbers corresponding to any potential connection in the heterogeneous interaction strength matrix, i.e., the sample numbers corresponding to the first two samples in the sample set. Sample With the Sample , The critical interaction strength that leads to structural fusion of independent communities, Characterized the sample The timing of merging when individuals lose their independence and integrate into the local manifold can be used to identify outliers that are slow to integrate during the evolutionary process.
[0076] Steps 1-3: Extract lifecycle-related closure scales and define samples. In evolutionary timeline The following interactive neighborhood set When the correlation matrix third-order trace This indicates the emergence of interconnected cyclical pathways in the evolutionary model, which is considered the formation of stable closed associations. (Definition of samples) Life cycle related closed-scale approximation The average minimum interaction strength that forms a closed loop within its interaction neighborhood:
[0077] ,
[0078] in, This metric represents the cardinality of the set of interactive neighborhoods. It reflects the tightness of the formation of stable closed pathways in the local neighborhood interaction network of the sample, and is used to identify complex entangled structures at the boundary of data distribution;
[0079] Steps 1-4: Extract the multidimensional configuration-derived scale of the life cycle. The multidimensional configuration-derived scale of the life cycle of the sample is extracted using the neighborhood Voronoi approximation method. When the sample The number of neighborhood samples satisfies the supporting condition. At that time, calculate the high-dimensional configuration-derived scale:
[0080] ,
[0081] in, The preset configuration magnification factor ( ), used to quantify the density threshold required for the development of high-dimensional envelope structures; if Then define This metric is used to characterize the hierarchical position of a sample in constructing the high-dimensional spatial topological skeleton and envelope structure.
[0082] Step 2: Perform a comprehensive scan and activity statistical analysis of the dynamic indicators of structural response characteristics at different evolutionary levels. Based on the survival time and structural contribution of samples during the evolutionary process, identify abnormal behaviors and construct a lifecycle outlier suppression mechanism to identify and weaken the weight contribution of transient noise samples that rapidly disappear in the early stages of evolution and free samples that only participate in structural construction in the late stages of evolution during the classification model training process. Step S2 specifically includes the following sub-steps:
[0083] Step 2-1: Determine the full lifecycle evolution scan interval, based on minority class samples (the number of samples in the corresponding class is the minimum of the two classes, denoted as ). The number of samples in the majority class is ,and The average heterogeneous interaction strength and the mean of the global lifetime assimilation scale determine the observation interval of the evolutionary time series. and in the observation interval Internal generation includes Multi-stage scanning sequence of evolutionary time nodes ,in The threshold representing the initial heterogeneous interaction intensity for lifecycle evolution scanning is the 5th percentile of the distance sets between two different samples. The threshold for the termination of heterogeneous interaction intensity in a lifecycle evolution scan is the 95th percentile of the distance between two different sample sets. Indicates the first The threshold for heterogeneous interaction intensity corresponding to each life cycle evolution stage;
[0084] Step 2-2: Calculate the average activity weight across stages, based on the observed samples in the evolution time series. The life cycle dynamics of the changes. The sample is calculated using an exponential mapping function. In the Cross-stage average weights in the evolutionary dimension
[0085] ,
[0086] in, These correspond to the three lifecycle dimensions of merging and assimilation, association closure, and configuration derivation, respectively. For the sample A set of time indices that demonstrate an effective evolutionary response; For the first The mean of global dynamic indicators in a given dimension, i.e., the average of the dynamic indicators extracted from all samples in that evolutionary dimension, such as when hour It represents the average response level of the samples in the assimilation and evolution dimension, and is used to normalize differences in evolution scale; It is a smoothing factor, typically ranging from 0.1 to 1.
[0087] Step 2-3: Introduce category correction coefficients The average activity weight across different stages is adjusted to obtain the corrected lifecycle weight. Define a minority class sample (the number of samples in its corresponding class is the minimum of the two classes, denoted as ). The number of samples in the majority class is ,and The correction coefficient for the ) is greater than the correction coefficient for the majority class samples;
[0088] Step 2-4: Construct evolutionary hysteresis suppression logic. Determine hysteresis behavior based on the lifecycle response time sequence of samples in the multi-stage evolutionary scan sequence. When the lifecycle assimilation scale of a sample lags significantly behind the overall level, or its effective evolutionary response is mainly concentrated in the latter part of the evolutionary scan interval, the sample is determined to be a lifecycle-hysteresis individual. For samples determined to be lifecycle-hysteresis individuals, apply exponential compression to their original lifecycle weights, so that the sample weights monotonically decrease as the degree of hysteresis increases, thereby weakening their impact on classification model training. Samples that do not meet the hysteresis determination criteria retain the weights obtained from steps 2-2 to 2-3 unchanged.
[0089] Steps 2-5: Construct an adaptive cooperative coupling model for multidimensional lifecycle signals. Treat the dynamic weights of the three dimensions—merging and assimilation, correlation closure, and configuration derivation—obtained in the preceding steps as independent evolutionary signal channels, and define a single lifecycle outlier suppression weight. The result of coordinated modulation of multi-channel signals:
[0090] ,
[0091] Among them, the weighting coefficient The particle swarm optimization is used to search under simplex constraints. The objective function is to use validation metrics such as G-mean and F1-scores. The three-term update is adopted, which includes inertia, individual and population, so that the contributions of the three dimensions are consistent with the actual evolution signal strength of the data.
[0092] Steps 2-6: Perform global normalization on the lifetime outlier suppression weights of all samples and construct a sample-level diagonal lifetime outlier suppression matrix. Where diag represents a diagonal matrix, For the first Outlier suppression weights corresponding to each sample.
[0093] Step 3: During the dynamic advancement of the full lifecycle evolution model, monitor the interaction behavior between samples of different categories in real time, identify early interaction structures including direct cross-class connections and higher-order mixed entanglement, and construct a lifecycle class overlap suppression mechanism by calculating the evolutionary stage position and entanglement depth of the first cross-class contact triggered by a sample. This mechanism is used to suppress the negative impact of boundary-ambiguous samples that exhibit atypical cross-class affinity in the early stages of evolution on the classification decision surface. Step 3 specifically includes the following sub-steps:
[0094] Step 3-1: Real-time monitoring of cross-class interaction structures during the evolutionary process. During the advancement of the full lifecycle evolutionary model, when samples from different classes... In evolutionary timeline The following satisfies the heterogeneous interaction strength When a direct cross-class connection is established between the two, it is denoted as a set. :
[0095] ,
[0096] If three sample points In evolutionary timeline If active connectivity pathways are established in every pair of samples, and the class labels of these three samples are not completely consistent (i.e., they include samples from different classes), then a higher-order hybrid entangled structure is determined to have formed, denoted as set.
[0097] ;
[0098] Step 3-2: Evaluate the lifetime entanglement initiation index and basic purity confidence weights of the sample. Track the sample. The critical evolutionary sequence of the first participation in constructing the above cross-class interaction structure is defined as the entanglement initiation index. The smaller the index, the more likely the sample came into contact with outliers very early in its evolution, indicating a high-risk boundary or noisy sample. A basic purity confidence weight is constructed using a sigmoid mapping function with shift and scaling factors. :
[0099] ,
[0100] in, The entanglement boundary threshold is the median of the set of entanglement initiation indices for all samples. The discrimination coefficient is set to be the inverse multiple of the standard deviation of the set. This weight assigns higher confidence to samples where entanglement occurs later in the process.
[0101] Step 3-3: Calculate the inter-class conflict intensity and density suppression factor. Calculate the sample... Cross-class connectivity at a specified evolutionary time sequence This is used to quantify the intensity of the competitive environment of different species surrounding a sample.
[0102] ,
[0103] in, For the elements of the adjacency matrix of the evolutionary model, As an indicator function, a density suppression factor is constructed based on the cross-class connectivity. Specifically, this involves first calculating the sample... exist Connectivity below The connectivity of all samples is normalized by the maximum value of the connectivity, and then the normalized connectivity is exponentially mapped to penalize samples in high-conflict regions.
[0104] Steps 3-4: Generate a lifecycle-class overlap suppression mechanism. The basic purity confidence weights are then... With density inhibition factor Couple the components to generate the final lifecycle class overlap suppression weights. The sample-level diagonal lifetime class overlap suppression matrix is then globally normalized. .
[0105] Step 4: Construct a quantitative balancing mechanism based on lifecycle evolution abundance, and fuse this mechanism with the lifecycle outlier suppression mechanism and the lifecycle class overlap suppression mechanism in a convex combination form; transform the search for the optimal fusion coefficient into a hyperparameter optimization problem, and use a Bayesian optimization algorithm with Gaussian process regression as the surrogate model and the expected improvement criterion to automatically find the optimal solution, thereby determining the lifecycle comprehensive weight vector that maximizes the system's evolutionary stability; Step 4 specifically includes the following sub-steps:
[0106] Step 4-1: Construct a sample quantity balancing mechanism based on lifecycle evolutionary abundance. First, calculate the imbalance ratio between the majority and minority classes, take its square root, and set an upper limit threshold. In the majority dataset scenario, since the imbalance ratio is usually not too large, an upper limit threshold of 10 is sufficient to obtain the smoothed basic quantity factor. Then, calculate the ratio of the number of effective evolutionary responses to the total number of samples in the three lifecycle evolutionary dimensions of merging and assimilation, association closure, and configurational derivation for both classes (majority and minority). Take the arithmetic mean of these three ratios to obtain the comprehensive evolutionary abundance ratios for both the minority and majority classes. Finally, calculate the ratio of the comprehensive evolutionary abundance ratio of the majority class to that of the minority class, define it as the evolutionary adjustment coefficient, and multiply it by the basic quantity factor to obtain the final weighting multiple for the minority class samples, constructing a diagonal quantity weight matrix. ;
[0107] Step 4-2: Construct a convex combination fusion model of multidimensional constraint weights. This involves combining the quantitative weights obtained in Step 4-1. Step B yields the lifetime outlier suppression weights. The lifetime class overlap suppression weights obtained in step C The fusion is performed using a convex combination method to obtain the comprehensive weight vector:
[0108] ,
[0109] Wherein, the fusion coefficient satisfies and These represent the relative importance of quantity balance, outlier suppression, and overlap suppression mechanisms in the final decision-making process, respectively.
[0110] Step 4-3: Establish a surrogate model for fusion parameters based on Gaussian process regression. Optimal fusion coefficients... The search problem is transformed into a hyperparameter optimization problem for the classification system performance. A Bayesian optimization method is introduced, using Gaussian process regression as a surrogate model to approximate the nonlinear functional relationship between the classification performance index and the three fusion coefficients, thereby capturing the synergistic effect between different mechanisms.
[0111] Step 4-4: In the Bayesian optimization process, the Expected Improvement Criterion (EI+) is used as the sampling strategy. The classification performance indicators, including G-mean, Sensitivity, F1-score, and Precision, are used as the objective functions to iteratively evaluate and update the surrogate model, ultimately obtaining the optimal combination of fusion coefficients. , These represent the optimal sample size weight, the optimal lifetime outlier suppression weight, and the optimal lifetime class overlap suppression weight, respectively, thus determining the comprehensive weight vector.
[0112] Step 5: Construct the comprehensive weight vector into a sample-level diagonal weighted matrix and embed it into the loss function and hidden layer node generation constraints of the randomly configured network. Train the classification model using the weighted objective function and apply it to the test set to achieve accurate classification of imbalanced data and output performance metrics including G-means, Sensitivity, F1-score, and Precision. Step 5 specifically includes the following sub-steps:
[0113] Step 5-1: As Figure 2 As shown, the parameter system of the evolution-constrained stochastic network is initialized. The maximum hidden layer node capacity of the network is set. (50) Maximum number of candidate searches generated by the node (20) The system expects to tolerate residuals (0.001), random weight generation interval [-50,50] and contractility factor [0.9, 0.99], where This represents the upper bound (radius of the value range) of the magnitude of the random weights of hidden layer nodes in a randomly configured network.
[0114] Step 5-2: During the iterative addition of nodes, a set of input weights and biases for candidate hidden layer nodes are randomly generated from the generation interval, and the output vectors of the candidate hidden layer nodes are calculated.
[0115] Step 5-3: Perform node selection based on supervised constraints. Calculate the effectiveness index of candidate nodes according to the supervised inequality constraint criterion, and select effective evolving hidden layer nodes that can reduce the current residual (a candidate node is considered to reduce the current residual if it satisfies the supervised inequality constraint and further reduces the network residual after its introduction) and satisfy the linear independence constraint. Add these nodes to the network structure and simultaneously update the hidden layer output matrix. ;
[0116] Step 5-4: Solve the output layer by embedding the lifecycle comprehensive weight matrix. Introduce the comprehensive weight vector obtained in Step 4, which contains multi-dimensional evolutionary information such as merging and assimilation, association closure, configuration derivation, and inter-class interactions, and construct it into a sample-level diagonal weighted matrix. The weighted least squares method is used to analytically solve for the output layer weights. ,in, This is the hidden layer output matrix. For the training set label matrix, Represents the generalized inverse of a matrix. This step represents transposition; it forces the classifier to prioritize fitting high-confidence samples with significant lifecycle features during the error minimization process.
[0117] Step 5-5: Evaluate the model's evolution convergence state. Calculate the training residuals of the current model and determine whether the residuals are less than the expected tolerance error. Or has the current number of hidden layer nodes reached the capacity limit? If any stopping condition is met, the model construction is considered complete; otherwise, the current residual is updated and the process returns to step 5-2 to continue incrementally adding nodes.
[0118] Steps 5-6: Use the trained model parameters to build the final classifier, perform classification prediction on the test set, and calculate the classification performance index.
[0119] To verify the performance of this invention in the low-quality imbalanced data classification task, a portion of the imbalanced benchmark datasets from the KEEL (Knowledge Extraction based on Evolutionary Learning) database were selected for experimental evaluation. These datasets include ecoli1, glass1, yeast1, and vehicle0; the above English names are the original names / numbers used by the KEEL database for each benchmark dataset, and this naming convention is retained for ease of reproduction and comparison. The quantitative characteristics of each dataset are shown in Table 1.
[0120] Table 1. Dataset Description Diagram
[0121]
[0122] The experiment selected five methods for comparison in class imbalance classification: Incremental Random Vector Function Connection Network (IRVFL), Randomized Network (SCN), Randomized Network based on sample size and density (SD-SCN), Class Dependency Randomized Network based on Mixed Density (HD-CSCN), and Randomized Network based on Density and Probability (JP-SCN). IRVFL and SCN do not include any mechanisms for addressing class imbalance and are therefore used as baseline models without imbalance handling for comparison. SD-SCN, HD-CSCN, and JP-SCN incorporate features such as sample density, mixed structure metric, or joint probability into the incremental construction mechanism of traditional SCN, respectively, enhancing the model's adaptability to class imbalance and sample distribution bias from different perspectives. These are improved models with class sensitivity in the comparative experiment. The method of this invention was compared with the above five methods in terms of Sensitivity, G-Mean, F1-Score, and Precision score. The calculation methods for these indicators are as follows:
[0123] ,
[0124] ,
[0125] ,
[0126] ,
[0127] ,
[0128] The TP value represents the number of samples that are actually in the minority class but are predicted as minority; the FN value represents the number of samples that are actually in the minority class but are predicted as majority; the FP value represents the number of samples that are actually in the majority class but are predicted as minority; and the TN value represents the number of samples that are actually in the majority class but are predicted as majority. The comparative experimental results are shown in Tables 2, 3, 4, and 5. As can be seen from the tables, the method of this invention achieves optimal performance on all four imbalance evaluation metrics, showing a significant improvement over IRVFL and SCN, and maintaining its advantage over improved algorithms such as SD-SCN, HD-CSCN, and JP-SCN, proving the effectiveness of this method.
[0129] Table 2. Comparison of different methods on the ecoli1 dataset
[0130]
[0131] Table 3. Comparison of different methods on the glass1 dataset
[0132]
[0133] Table 4 Comparison of different methods on the yeast1 dataset
[0134]
[0135] Table 5. Comparison of different methods on the vehicle0 dataset.
[0136]
[0137] This invention provides a class-imbalanced analysis method based on data lifecycle evolution. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.
Claims
1. A class-imbalanced analysis method based on data lifecycle evolution, characterized in that, Includes the following steps: Step 1: Construct a full life cycle evolution model of data samples to simulate the dynamic process of samples evolving from discrete independent states to complex structural forms. Extract dynamic indicators that characterize the structural response characteristics of samples at different evolution levels. The dynamic indicators include the life cycle merging and assimilation scale that describes the loss of individual independence of samples and integration into local manifolds, the life cycle association and closure scale that describes the formation of stable closed pathways in the neighborhood interaction network of samples, and the life cycle multidimensional configuration derivation scale that describes the participation of samples in constructing high-dimensional spatial topological skeletons and envelope structures. Step 2: Perform a panoramic scan and activity statistical analysis of the dynamic indicators of structural response characteristics in different evolutionary levels. Based on the survival time and structural contribution of samples in the evolution process, identify abnormal behaviors and construct a life cycle outlier suppression mechanism to identify and weaken the weight contribution of transient noise samples that die out quickly in the early stage of evolution and free samples that participate in structural construction in the late stage of evolution during the classification model training process. Step 3: Monitor the interaction behavior between different categories of samples in real time during the dynamic advancement of the full life cycle evolution model, identify early interaction structures including direct cross-class connections and high-order mixed entanglement, and construct a life cycle class overlap suppression mechanism by calculating the evolution stage position and entanglement depth of the first cross-class contact of a sample. This mechanism is used to suppress the negative impact of boundary fuzzy samples that exhibit atypical cross-class affinity in the early stages of evolution on the classification decision surface. Step 4: Construct a quantitative balance mechanism based on the abundance of life cycle evolution, and fuse the quantitative balance mechanism with the life cycle outlier suppression mechanism and the life cycle class overlap suppression mechanism in a convex combination form; transform the search for the optimal fusion coefficient into a hyperparameter optimization problem, and use a Bayesian optimization algorithm with Gaussian process regression as the surrogate model and the expectation improvement criterion to automatically find the optimal solution, thereby determining the life cycle comprehensive weight vector that can maximize the system evolution stability. Step 5: Construct a sample-level diagonal weighted matrix from the lifetime integrated weight vector and embed it into the loss function and hidden node generation constraints of the randomly configured network. Train the classification model using the weighted objective function and apply the classification model to the test set to achieve accurate classification of imbalanced data. Output performance metrics including geometric mean (G-mean), recall (Sensitivity), F1 score (F1-score), and precision.
2. The method according to claim 1, characterized in that, Step 1 includes: Step 1-1: Define the imbalanced dataset ,in Indicates the first one sample 3D feature vectors Indicates the first The category label of each sample, The total number of samples, Represent the real number space; calculate the heterogeneous interaction strength matrix between samples. , where matrix elements Indicates sample With sample Structural affinity between them; based on the heterogeneous interaction strength matrix, construct the evolutionary time series... Dynamically driven full life cycle evolution model If and only if the heterogeneous interaction strength between any two samples in the imbalanced dataset is less than or equal to At that time, it was determined that active connection pathways had been established between the samples and formed a local symbiotic structural unit; Steps 1-2: Extract the lifecycle merging and assimilation scale, establish a community evolution simulation mechanism, and simulate the process of independent samples merging into a local manifold; extract the lifecycle merging and assimilation scale of the samples. All potential connections in the heterogeneous interaction strength matrix According to structural affinity Sort the connections from smallest to largest, then iterate through the sorted connections. If the sample and If they belong to different independent evolutionary communities, then perform community assimilation; define samples. Lifecycle assimilation scale as follows: , in, and The sample index represents the two sample numbers corresponding to any potential connection in the heterogeneous interaction strength matrix, i.e., the sample numbers corresponding to the first two samples in the sample set. Sample With the Sample , The critical interaction strength that leads to structural fusion of independent communities; Steps 1-3: Extract lifecycle-related closure scales and define samples. In evolutionary timeline The following interactive neighborhood set When the correlation matrix third-order trace This indicates the emergence of interconnected cyclical pathways in the full life cycle evolution model, which is considered the formation of stable closed associations; a sample is defined. Life cycle related closed-scale approximation For the sample The average minimum interaction strength that forms a closed loop within the interaction neighborhood: , in, The cardinality of the set of interactive neighborhoods; Step 14: Extract the multidimensional configuration-derived scale of the life cycle. The multidimensional configuration-derived scale of the life cycle of the sample is extracted using the neighborhood Voronoi approximation method. When the sample The number of neighborhood samples satisfies the supporting condition. At that time, calculate the high-dimensional configuration-derived scale: , in, This is the preset configuration magnification factor; if Then define .
3. The method according to claim 2, characterized in that, Step 2 includes: Step 2-1: Determine the full lifecycle evolution scanning interval. Based on the average heterogeneous interaction intensity of minority class samples and the mean of the global lifecycle merging and assimilation scale, determine the observation interval of the evolutionary time series. and in the observation interval Internal generation includes Multi-stage scanning sequence of evolutionary time nodes ,in This represents the threshold for the initial heterogeneous interaction intensity in the lifecycle evolution scan. This represents the threshold for the termination of heterogeneous interaction intensity in the lifecycle evolution scan. Indicates the first The heterogeneous interaction intensity thresholds corresponding to each life cycle evolution stage; the minority class samples correspond to the minority class category, and the number of minority class samples is denoted as... The majority class samples correspond to the majority class, and the number of samples in the majority class is denoted as . Among them, the minority class is the class with fewer samples in the current training data, and the majority class is the class with more samples, i.e. ; Step 2-2: Calculate the average activity weight across stages, based on the observed samples in the evolutionary time series. The life cycle dynamics changes are calculated using an exponential mapping function. In the Average weight across stages in the evolutionary dimension: , in, These correspond to the three lifecycle dimensions of merging and assimilation, association closure, and configuration derivation, respectively. For the sample A set of time indices that demonstrate an effective evolutionary response; For the first Mean of global dynamics index under dimension; It is a smoothing factor; Step 2-3: Introduce category correction coefficients The average activity weight across different stages is adjusted to obtain the corrected lifecycle weight. The correction coefficient for minority class samples is set to be greater than that for majority class samples; Step 2-4: Construct evolutionary hysteresis suppression logic. Determine hysteresis behavior based on the life cycle response time sequence of the sample in the multi-stage evolutionary scan sequence. For samples that do not meet the hysteresis determination conditions, keep the weights obtained from Step 2-2 to Step 2-3 unchanged. Steps 2-5: Construct an adaptive cooperative coupling model for multidimensional lifecycle signals, treating the dynamic weights of the three dimensions of assimilation, correlation closure, and configuration derivation as independent evolutionary signal channels, and defining a single lifecycle outlier suppression weight. The result of coordinated modulation of multi-channel signals: , Among them, the weighting coefficient The algorithm is determined by particle swarm optimization under simplex constraints. The objective function is a validation metric including geometric mean (G-mean), recall (Sensitivity), F1 score (F1-score), and precision. The algorithm employs three updates: inertia, individual, and swarm. Steps 2-6: Perform global normalization on the lifetime outlier suppression weights of all samples and construct a sample-level diagonal lifetime outlier suppression matrix. Where diag represents a diagonal matrix, For the first Outlier suppression weights corresponding to each sample.
4. The method according to claim 3, characterized in that, Step 3 includes the following steps: Step 3-1: Real-time monitoring of cross-class interaction structures during the evolutionary process; during the advancement of the full life-cycle evolutionary model, when samples from different categories... In evolutionary timeline The following satisfies the heterogeneous interaction strength When a direct cross-class connection path is established, it is denoted as a set. : , If three sample points In evolutionary timeline Active connectivity pathways were established in each pair of samples, and three sample points were used. If the category labels are not completely consistent, it is determined that a higher-order hybrid entangled structure has been formed, denoted as set. : ; Step 3-2: Evaluate the lifecycle entanglement initiation index and basic purity confidence weight of the sample; track the sample. The critical evolution timing for the first time involved in constructing a cross-class interaction structure is defined as the entanglement initiation point. The basic purity confidence weights are constructed using a sigmoid mapping function with shift and scaling factors. : , in, Entanglement boundary threshold; The coefficient of discrimination is exp; exp represents the natural exponential function. Step 3-3: Calculate the inter-class conflict intensity and density suppression factor; Steps 3-4: Generate a lifecycle class overlap suppression mechanism.
5. The method according to claim 4, characterized in that, Step 3-3 includes: calculating the sample Cross-class connectivity at a specified evolutionary time sequence : , in, For the elements of the adjacency matrix of the evolutionary model, As an indicator function, a density suppression factor is constructed based on cross-class connectivity. .
6. The method according to claim 5, characterized in that, Steps 3-4 include: assigning confidence weights to the basic purity. With density inhibition factor Couple the components to generate the final lifecycle class overlap suppression weights. and to Global normalization is performed to construct a sample-level diagonal lifetime class overlap suppression matrix. ,in Indicates the first The lifetime class overlap suppression weights corresponding to each sample.
7. The method according to claim 6, characterized in that, Step 4 includes the following steps: Step 4-1: Construct a sample quantity balancing mechanism based on lifecycle evolution abundance. First, calculate the imbalance ratio between majority and minority class samples, perform a square root operation on the imbalance ratio, and set an upper limit threshold to obtain a smoothed basic quantity factor. Then, calculate the ratio of the number of effective evolutionary responses to the total number of samples in the three lifecycle evolution dimensions of merging and assimilation, association closure, and configurational derivation for both majority and minority class samples. Take the arithmetic mean of the ratios for the three lifecycle dimensions to obtain the comprehensive evolutionary abundance ratios for both the minority and majority classes. Finally, calculate the ratio of the comprehensive evolutionary abundance ratio of the majority class to the comprehensive evolutionary abundance ratio of the minority class, define this ratio as the evolutionary adjustment coefficient, and multiply it by the basic quantity factor as the final weighting multiple for the minority class samples, constructing a diagonal quantity weight matrix. ; Step 4-2: Construct a convex combination fusion model with multidimensional constraint weights; combine the diagonal weight matrix Sample-level diagonal lifetime outlier suppression matrix and sample-level diagonal lifetime class overlap suppression matrix The fusion is performed using a convex combination method to obtain the comprehensive weight vector W: , Among them, the fusion coefficient , , satisfy and ; Step 4-3: Establish a surrogate model based on Gaussian process regression with fusion parameters; use the fusion coefficients to adjust the weights for sample size, lifecycle outlier suppression, and lifecycle class overlap suppression. The search problem is transformed into a hyperparameter optimization problem for imbalanced classification performance; a Bayesian optimization method is introduced, and Gaussian process regression is used as a surrogate model to achieve adaptive fusion of multi-source lifecycle weights; Step 4-4: In the Bayesian optimization process, the expected improvement criterion is used as the sampling strategy. The classification performance indicators, including G-mean, Sensitivity, F1-score, and Precision, are used as the objective function. The surrogate model is iteratively evaluated and updated to finally obtain the optimal combination of fusion coefficients. Thus, the comprehensive weight vector X1 is determined; These represent the optimal sample size weight, the optimal lifetime outlier suppression weight, and the optimal lifetime class overlap suppression weight, respectively.
8. The method according to claim 7, characterized in that, Step 5 includes the following steps: Step 5-1: Initialize the parameter system of the evolution-constrained stochastic network; set the maximum hidden layer node capacity of the network. Maximum number of candidate searches generated by the node The system expects to tolerate residuals. Random weight generation interval and contractile factor ,in This represents the upper bound of the magnitude of the random weights of hidden layer nodes in a randomly configured network; Step 5-2: During the iterative addition of nodes, generate the interval from the random weights. A set of input weights and biases for candidate hidden layer nodes are randomly generated, and the output vectors of the candidate hidden layer nodes are calculated. Step 5-3: Perform supervised constraint-based node selection; calculate the effectiveness index of candidate nodes according to the supervised inequality constraint criterion, select effective evolving hidden layer nodes that can reduce the current residual and satisfy the linear independence constraint to add to the network structure, and update the hidden layer output matrix synchronously. ; Step 5-4: Embed the lifecycle comprehensive weight matrix to solve the output layer; construct the comprehensive weight vector X1 into a sample-level diagonal weighted matrix. The output layer weights are analytically solved using the weighted least squares method. ,in, For the training set label matrix, Represents the generalized inverse of a matrix. Indicates transpose; Step 5-5: Evaluate the model's evolution convergence state; calculate the current model's training residuals and determine whether the training residuals are less than the expected tolerance error. Or has the current number of hidden layer nodes reached the capacity limit? If any stopping condition is met, the model construction is considered complete; otherwise, the current residual is updated and the process returns to step 5-2 to continue incrementally adding nodes. Steps 5-6: Use the trained model parameters to build the final classifier, perform classification prediction on the test set, and calculate the classification performance index.
9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 8.
10. A storage medium, characterized in that, It stores a computer program or instructions that, when run on a computer, perform the steps of the method as described in any one of claims 1 to 8.
