Collaborative classification regression decision-making method based on dynamic belief allocation and evidence fusion

Abstract

The invention provides a collaborative classification regression decision-making method based on dynamic belief allocation and evidence fusion. The method comprises the steps of obtaining test samples and training samples, and screening out K training samples closest to each test sample to obtain a neighbor set of the test samples; generating a structured evidence supporting classification decision according to the neighbor set of the test sample, calculating a belief value of each proposition according to the structured evidence, and obtaining a prediction category of the test sample according to the proposition with the maximum belief value; and counting the occurrence frequency of each category of training samples in the neighbor set according to the category to which the training samples belong, constructing a category weight according to the occurrence frequency of the training samples, carrying out belief value distribution through the category weight to obtain a belief value of each category, taking the belief value as a label weight, and generating a prediction label. By applying the method, the association strength of the test sample and each candidate category can be quantified, and false confidence misleading is avoided; and multi-source evidence is fused to generate structured decision support, so that the robustness in a high-conflict evidence scene is improved.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology in artificial intelligence, and in particular to a collaborative classification regression decision-making method based on dynamic belief assignment and evidence fusion. Background Technology

[0002] In high-risk decision-making fields such as national defense and security, medical diagnosis, and autonomous driving, the reliability of classification systems is directly related to the success or failure of missions and even personal safety. Traditional machine learning frameworks take "accurate classification" as their core goal, requiring test samples to be forcibly classified into a single training class. However, this paradigm has a fundamental flaw when dealing with uncertain samples: when the input data lacks sufficient class evidence (such as out-of-domain (OOD) samples or inter-class intersection (IM) samples), the forced classification mechanism will produce systematic misjudgments due to insufficient or conflicting evidence, leading to catastrophic consequences.

[0003] Typical cases demonstrate that the limitations of existing technologies pose substantial risks. In a 2016 autonomous driving accident, a system misidentified a white truck as a cloud; in the field of image classification, a serious ethical crisis occurred when an African American was mislabeled as a "gorilla." These misclassification cases expose significant problems with accurate classification: classifiers trained within a specific domain lack the ability to detect out-of-domain samples (such as unfamiliar truck types) and imprecise samples (such as boundary samples with interwoven features); forced classification mechanisms output high-confidence results even with insufficient evidence, creating a "wrong but certain" decision-making trap.

[0004] Furthermore, there are obvious contradictions in the existing technology system, mainly reflected in the following aspects: high-risk scenarios require zero tolerance for false judgments, but the cost of adopting high-precision detection methods in all aspects is too high; traditional classifiers output false confidence through mechanisms such as softmax, which cannot truly reflect the correlation strength between samples and categories; accurate classification focuses on classifying a single sample, ignoring the distribution relationship between categories and the expression of sample uncertainty.

[0005] Therefore, it is necessary to provide a classification and regression decision-making method that can improve the robustness of decision-making in high-risk scenarios and the accuracy of complex regression tasks. Summary of the Invention

[0006] The purpose of this invention is to provide a collaborative classification and regression decision-making method based on dynamic belief allocation and evidence fusion, so as to improve the robustness of decision-making in high-risk scenarios and the accuracy of complex regression tasks.

[0007] In a first aspect, the collaborative classification regression decision-making method based on dynamic belief allocation and evidence fusion provided by the present invention includes: acquiring test samples and training samples; calculating the distance between each test sample and all training samples in the feature space; selecting the K nearest training samples to each test sample to obtain the nearest neighbor set of the test sample; generating structured evidence supporting classification decision based on the nearest neighbor set of the test sample; calculating the belief value of each proposition corresponding to each test sample based on the structured evidence; obtaining the predicted category of the test sample based on the proposition with the largest belief value; counting the occurrence frequency of each category of training samples in the nearest neighbor set according to the category to which the training sample belongs; constructing category weights based on the occurrence frequency of training samples; allocating belief values ​​through category weights to obtain the belief value of each category as the label weight; and generating predicted labels based on the label weights.

[0008] The beneficial effects of the collaborative classification and regression decision-making method based on dynamic belief allocation and evidence fusion provided by this invention are as follows: It quantifies the spatial association between test and training sample categories using an improved KNN (K-Nearest Neighbors) algorithm; generates differentiated evidence based on a dynamic belief allocation mechanism; and fuses multi-source evidence using structured evidence combination rules to generate belief values. Furthermore, it allocates beliefs from multiple-case and unknown classes to single-case classes through frequency statistics and category weight optimization to support specific complex regression tasks. While maintaining low complexity, this invention achieves inference speeds tens of times faster than deep learning models, and the entire decision path is interpretable, effectively solving the systematic misjudgment problem in handling uncertain samples using traditional methods.

[0009] In one possible embodiment, generating structured evidence supporting classification decisions based on the nearest neighbor set of the test sample includes: determining the distribution of nearest neighbor samples in the nearest neighbor set for each single-element category in the identification framework based on the nearest neighbor set and the identification framework, generating single-element category corresponding evidence based on the determination result; and obtaining structured evidence supporting classification decisions for the test sample based on the single-element category corresponding evidence in the identification framework combination.

[0010] In another possible embodiment, the proposition types corresponding to the test samples include singleton propositions, multi-instance propositions, and ignorance propositions;

[0011] Based on the structured evidence, the belief value for each proposition corresponding to each test sample is calculated, including:

[0012] When the proposition is a singleton or an ignorance proposition, the belief value of the proposition is calculated according to the following formula: A∈Ω or A=Ω, where, Let m represent the belief value of proposition A obtained after fusing s pieces of evidence for the i-th test sample, where s represents the number of pieces of evidence, A represents the proposition, B1 and B2 represent certain propositions in the evidence, and m represents the belief value of proposition A. is(B2) represents the belief value assigned to proposition B2 by the s-th piece of evidence in the i-th test sample, and Ω represents the identification frame; when the proposition is a multi-instance proposition, the belief value of the proposition is calculated according to the following formula: A∈Ψ, where Ψ represents the multi-instance class in the subset Ω.

[0013] In other possible embodiments, the distribution of nearest neighbor samples in the nearest neighbor set is determined, and evidence corresponding to a single-element category is generated based on the determination result. This includes: determining whether there are nearest neighbor samples in the single-element category; when there are no nearest neighbor samples in the single-element category, generating corresponding weak confidence evidence; when there are nearest neighbor samples in the single-element category, determining whether all the nearest neighbor samples belong to the single-element category; when not all the nearest neighbor samples belong to the single-element category, generating corresponding mixed nearest neighbor weighted evidence; when all the nearest neighbor samples belong to the single-element category, generating corresponding high confidence evidence; and designing belief value allocation mechanisms for weak confidence evidence, mixed nearest neighbor weighted evidence, and high confidence evidence, respectively.

[0014] Design belief value allocation mechanisms for weak-confidence evidence, mixed-nearest neighbor weighted evidence, and high-confidence evidence respectively, including: the belief value allocation for weak-confidence evidence satisfies the following formula: m(ω s )=0.1p, m(Ω)=1-0.1p, where, N s ω represents the training samples s The number of class samples, N represents the total number of training samples, and p represents ω. s The distribution ratio of class samples in the training samples, m(ω) s ) indicates that it is assigned to ω s The belief value of a class is m(Ω), which represents the belief value assigned to Ω, and Ω represents the identification frame. The belief value assignment of weighted evidence from mixed nearest neighbors satisfies the following formula: m(ω s ) = f, m(Ω) = 1 - f, where f represents the weighted evidence of the mixed nearest neighbors assigned to ω. s The belief value of a class, where 'a' represents the belief value of the test sample to class ω. s The average distance of the nearest neighbor samples, Represents ω s Intra-class tightness; the belief value assignment for high-confidence evidence satisfies the following formula: m(ω s )=0.999,m(Ω)=0.001.

[0015] Constructing class weights based on the frequency of training samples involves: adding the frequency of training samples to an optimization factor to obtain the unnormalized weights of each class of training samples; and then normalizing the unnormalized weights of each class of training samples with the sum of all class weights to obtain the class weights for each category. The normalization process satisfies the following formula: Where w(ω) s ) represents ω s Class weight, w′(ω) s ) represents ω s The unnormalized weights of the classes, w′(T) represent the sum of the unnormalized weights of all classes.

[0016] The belief value for each category is assigned as a label weight by assigning belief values ​​based on category weights. This assignment follows the formula: Where m(ω) s ) represents ω s Class label weight, w(ω) s ) represents ω s Class weights, m i (ω s ) indicates that the i-th element contains ω s The belief value of a class of multiple instance propositions or ignorant propositions.

[0017] Secondly, the present invention also provides a collaborative classification regression decision-making device based on dynamic belief allocation and evidence fusion, comprising: a nearest neighbor calculation unit, used to acquire test samples and training samples, calculate the distance between each test sample and all training samples in the feature space, and select the K training samples closest to each test sample to obtain the nearest neighbor set of the test sample; a category prediction unit, used to generate structured evidence supporting classification decision based on the nearest neighbor set of the test sample, calculate the belief value of each proposition corresponding to each test sample based on the structured evidence, and obtain the predicted category of the test sample based on the proposition with the largest belief value; and a label generation unit, used to count the occurrence frequency of each category of training samples in the nearest neighbor set according to the category to which the training sample belongs, construct category weights based on the occurrence frequency of training samples, allocate belief values ​​through category weights to obtain the belief value of each category as the label weight, and generate predicted labels based on the label weights.

[0018] Thirdly, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described collaborative classification regression decision-making method based on dynamic belief allocation and evidence fusion.

[0019] Fourthly, the present invention also provides an electronic device, comprising: a processor and a memory; the memory being used to store a computer program; the processor being used to execute the computer program stored in the memory, so that the electronic device performs the above-described collaborative classification regression decision-making method based on dynamic belief allocation and evidence fusion.

[0020] For the beneficial effects of the second to fourth aspects mentioned above, please refer to the description of the first aspect mentioned above. Attached Figure Description

[0021] Figure 1 A flowchart illustrating a collaborative classification regression decision-making method based on dynamic belief allocation and evidence fusion, provided in an embodiment of the present invention;

[0022] Figure 2 A flowchart illustrating a collaborative classification regression decision-making method based on dynamic belief allocation and evidence fusion, provided in an embodiment of the present invention;

[0023] Figure 3 A schematic diagram of a collaborative classification regression decision-making device based on dynamic belief allocation and evidence fusion provided in an embodiment of the present invention;

[0024] Figure 4 This is a schematic diagram of an electronic device structure provided in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art. The terms "comprising" and similar expressions used herein mean that the element or object preceding the word covers the element or object listed following the word and its equivalents, but do not exclude other elements or objects.

[0026] This embodiment provides a collaborative classification regression decision-making method based on dynamic belief assignment and evidence fusion. See also Figure 1 and Figure 2 The method includes:

[0027] S101: Obtain test samples and training samples, calculate the distance between each test sample and all training samples in the feature space, and select the K training samples that are closest to each test sample to obtain the nearest neighbor set of the test sample.

[0028] In one possible embodiment, samples without corresponding category attribution information are obtained as test samples, and samples with corresponding category attribution information are obtained as training samples.

[0029] In one possible embodiment, calculating the distance between each test sample and all training samples in the feature space includes: calculating the Euclidean distance between each test sample and all training samples based on the sum of squared differences in each dimension of the feature vectors of the test sample and the training samples; and selecting the K nearest training samples in the feature space to the test sample through a distance sorting mechanism to obtain the nearest neighbor set of the test sample.

[0030] S102: Generate structured evidence supporting classification decisions based on the nearest neighbor set of the test samples, calculate the belief value of each proposition corresponding to each test sample based on the structured evidence, and obtain the predicted category of the test sample based on the proposition with the largest belief value.

[0031] In one possible embodiment, generating structured evidence supporting classification decisions based on the nearest neighbor set of the test sample includes: determining the distribution of nearest neighbor samples in the nearest neighbor set for each single-element category in the identification framework based on the nearest neighbor set and the identification framework, generating single-element category corresponding evidence based on the determination result; and obtaining structured evidence supporting classification decisions for the test sample based on the single-element category corresponding evidence in the identification framework combination.

[0032] In one possible embodiment, the distribution of nearest neighbor samples in the nearest neighbor set is determined, and evidence corresponding to a single-element category is generated based on the determination result. This includes: determining whether there are nearest neighbor samples in the single-element category; when there are no nearest neighbor samples in the single-element category, generating corresponding weak confidence evidence; when there are nearest neighbor samples in the single-element category, determining whether all the nearest neighbor samples belong to the single-element category; when not all the nearest neighbor samples belong to the single-element category, generating corresponding weighted evidence of mixed nearest neighbors; when all the nearest neighbor samples belong to the single-element category, generating corresponding high confidence evidence; and designing belief value allocation mechanisms for weak confidence evidence, weighted evidence of mixed nearest neighbors, and high confidence evidence, respectively.

[0033] In a specific embodiment, belief value allocation mechanisms are designed for weak-confidence evidence, mixed nearest neighbor weighted evidence, and high-confidence evidence, respectively, including: the belief value allocation for weak-confidence evidence satisfies the following formula: m(ω s )=0.1p, m(Ω)=1-0.1p, where, N s ω represents the training samples s The number of class samples, N represents the total number of training samples, and p represents ω. s The distribution ratio of class samples in the training samples, m(ω) s ) indicates that it is assigned to ω s The belief value of a class is m(Ω), which represents the belief value assigned to Ω, and Ω represents the identification frame. The belief value assignment of weighted evidence from mixed nearest neighbors satisfies the following formula: m(ω s) = f, m(Ω) = 1 - f, where f represents the weighted evidence of the mixed nearest neighbors assigned to ω. s The belief value of a class, where 'a' represents the belief value of the test sample to class ω. s The average distance of the nearest neighbor samples, Represents ω s Intra-class tightness; the belief value assignment for high-confidence evidence satisfies the following formula: m(ω s )=0.999,m(Ω)=0.001.

[0034] In one possible embodiment, the proposition types corresponding to the test samples include singleton propositions, multi-instance propositions, and ignorance propositions; the belief value of each proposition corresponding to each test sample is calculated based on structured evidence, including: when the proposition is a singleton proposition or an ignorance proposition, the belief value of the proposition is calculated according to the following formula: A∈Ω or A=Ω, where, Let m represent the belief value of proposition A obtained after fusing s pieces of evidence for the i-th test sample, where s represents the number of pieces of evidence, A represents the proposition, B1 and B2 represent certain propositions in the evidence, and m represents the belief value of proposition A. is (B2) represents the belief value assigned to proposition B2 by the s-th piece of evidence in the i-th test sample, and Ω represents the identification frame; when the proposition is a multi-instance proposition, the belief value of the proposition is calculated according to the following formula: A∈Ψ, where Ψ represents the multi-instance class in the subset Ω.

[0035] In one possible embodiment, ω is calculated. s The intra-class tightness of a class is calculated as follows: The average distance from each training sample within that class to its K nearest neighbors is calculated, and the mean of these distances is taken as the intra-class tightness of that class. Specifically, this includes: firstly, in ω... s Within a class, calculate the Euclidean distance between each training sample and all other training samples, and sort them by distance to obtain the K nearest neighbor set for each training sample. Calculate the average distance between each training sample and its K nearest neighbors based on the K nearest neighbor set to obtain the average nearest neighbor distance. Then calculate ω. s ω is obtained by taking the mean of the nearest neighbor distances of all training samples within the class. s Intra-class tightness. Intra-class tightness is calculated according to the following formula: Where, N s Represents ω s The number of training samples in the class Represents ω s Training samples in the class, x i,j (j=1,…,K) represents The K nearest neighbors.

[0036] For example, when the training samples contain five classes ω1, ω2, ω3, ω4, and ω5, the identification framework is Ω = {ω1, ω2, ω3, ω4, ω5}. Based on the nearest neighbor set, the existence of nearest neighbor samples for each single-element class in the identification framework is determined, resulting in the distribution of nearest neighbor samples (i.e., training samples in the nearest neighbor set) in each single-element class. The distribution of nearest neighbor samples may result in the following situations: all nearest neighbor samples belong to the same single-element class, or the nearest neighbor samples are distributed across multiple single-element classes. For example, the training samples may include five classes, while the nearest neighbor samples are distributed across three classes: ω1, ω3, and ω4. In this case, based on the distribution of nearest neighbor samples, the structured evidence supporting the classification decision of the test sample can be derived, including: weighted evidence of mixed nearest neighbors corresponding to classes ω1, ω3, and ω4, and weak confidence evidence corresponding to classes ω2 and ω5. The test sample can support all propositions including singleton, multi-instance, and unknown propositions. Singleton propositions refer to sets in the identification frame containing only a single element, indicating that the specific category of the test sample can be inferred based on the current information. Multi-instance propositions refer to sets in the identification frame containing two or more elements, indicating that the most likely category of the test sample can be inferred based on the current information. For example, if the nearest neighbor samples are distributed across categories ω1, ω3, and ω4, multi-instance propositions include (ω1, ω3), (ω1, ω4), (ω3, ω4), and (ω1, ω3, ω4). The unknown category is the entire set of identification frames, indicating that the possible category of the test sample cannot be inferred based on the current information. For each proposition supported by the test sample, its belief value is calculated based on structured evidence. The proposition with the highest belief value is selected as the final category prediction result for the test sample to obtain the predicted category.

[0037] In a specific embodiment, to facilitate a deeper understanding of the application of structured evidence combination rules, a detailed calculation process using a specific numerical case is provided as an example: The identification framework includes three classes: Ω = {ω1, ω2, ω3}. In generating structured evidence supporting classification decisions based on the nearest neighbor set of the test sample, the basic belief values ​​constructed for each piece of evidence are: m1(ω1) = 0.6, m1(Ω) = 0.4, m2(ω2) = 0.7, m2(Ω) = 0.3, m3(ω3) = 0.15, m3(Ω) = 0.85, where m1(ω1) represents the belief value assigned to evidence m1 for ω1, i.e., the belief value assigned to the test sample. The support for class ω1 is represented by m1(Ω), which represents the belief value assigned to evidence m1 as Ω, i.e., the support for the test sample belonging to the ignorance class Ω. The support for class ω2 is represented by m2(Ω), which represents the belief value assigned to evidence m2 as ω2, i.e., the support for the test sample belonging to class ω2. The support for class ω3 is represented by m3(Ω), which represents the belief value assigned to evidence m3 as ω3, i.e., the support for the test sample belonging to class ω3. The support for class ω3 is represented by m3(Ω), which represents the belief value assigned to evidence m3 as Ω, i.e., the support for the test sample belonging to the ignorance class Ω. If proposition D is a singleton or ignorance class proposition, the belief value of proposition D can be calculated using structured evidence combination: m 1,3 (D)=m1(d)m2(Ω)m3(Ω)=m 1,2 (D)m3(Ω)=m1(D)m 2,3 (Ω)=m 1,3 (D). If proposition D is a multi-case proposition, such as D = ω1∪ω2, the belief value of proposition D can be calculated by applying structured evidence combination: The calculation process is similar when D = ω1∪ω3 or D = ω2∪ω3 to that when D = ω1∪ω2. Taking the singleton class ω1, the multi-case class ω1∪ω2, and the ignorant class Ω as examples, the specific belief values ​​are calculated as follows: m(ω1) = m1(ω1)m2(Ω)m3(Ω) = 0.1530; m(Ω) = m1(Ω)m2(Ω)m3(Ω) = 0.1020; m(ω1∪ω2) = m1(ω1)m2(ω2)m3(Ω) = 0.3570.

[0038] In one possible embodiment, when applying the method of obtaining the predicted category of a test sample based on structured evidence of the test sample to a high-risk misclassification scenario, the misclassification rate and inaccuracy rate can be used as evaluation metrics to assess classification accuracy. These high-risk misclassification scenarios include, but are not limited to, medical diagnosis, financial fraud detection, and autonomous driving safety, characterized by the potential for serious consequences from misclassification. The misclassification rate and inaccuracy rate are typically inversely related: a decrease in the misclassification rate is often accompanied by an increase in the inaccuracy rate. In application, acceptable thresholds for the misclassification rate and inaccuracy rate can be set according to the actual risk tolerance, and a balance between the two can be achieved by adjusting model parameters (such as classification threshold, validation strength, etc.) to train a classification model that meets the needs of high-risk scenarios. Here, misclassification refers to misclassifying a test sample with the true category A into ω. s Class, and In this case, the error rate is calculated according to the following formula: Where, N m This represents the number of misclassified test samples, where T represents the total number of test samples. Inaccurate classification refers to classifying a test sample whose true class is A as a composite class ω. i In the middle, and In this case, the inaccuracy rate is calculated according to the following formula: Where, N n This indicates the number of test samples that were not precisely classified.

[0039] S103: Count the frequency of occurrence of training samples of each category in the nearest neighbor set according to the category to which the training sample belongs, construct the category weights according to the frequency of occurrence of training samples, assign belief values ​​to each category through the category weights to obtain the belief value of each category as the label weight, and generate the predicted label according to the label weight.

[0040] In one possible embodiment, the frequency of occurrence of training samples of each category in the nearest neighbor set is counted according to the category to which the training sample belongs, that is, the number of samples belonging to each single element category in the nearest neighbor set is counted.

[0041] In one possible embodiment, constructing class weights based on the frequency of training samples includes: adding the frequency of training samples to an optimization factor to obtain the unnormalized weights of each class of training samples; and normalizing the unnormalized weights of each class of training samples with the sum of all class weights to obtain the class weights for each class. The normalization process satisfies the following formula: Where w(ω) s ) represents ω s Class weight, w′(ω) s ) represents ω s The unnormalized weights of the classes, w′(T) represent the sum of the unnormalized weights of all classes.

[0042] In one specific embodiment, the value of the optimization factor is dynamically adjusted through the cross-validation training process. Specifically, the optimization factor is initialized to a preset small value (e.g., ε = 0.1), and the factor is updated through multiple cross-validation iterations to ensure that the class weights remain effectively non-zero during training, thereby improving the model's classification performance for minority classes.

[0043] In one possible embodiment, assigning belief values ​​to each category as label weights through category weights includes: assigning belief values ​​through category weights according to the following formula: Where m(ω) s ) represents ω s Class label weight, w(ω) s ) represents ω s Class weights, m i (ω s ) indicates that the i-th element contains ω s The belief value of a class of multiple instance propositions or ignorant propositions.

[0044] The predicted label is generated based on the label weights. This involves calculating the label weights for each category of the training samples appearing in the nearest neighbor set, and recording the category and corresponding weight in the predicted label.

[0045] This invention provides a collaborative classification and regression decision-making method based on dynamic belief allocation and evidence fusion. Addressing the problem of misjudgment due to insufficient evidence in traditional precise classification methods in high-risk scenarios, this invention proposes a modular collaborative architecture: It quantifies the spatial association between test and training sample classes using an improved KNN (K-Nearest Neighbors) algorithm, calculating intra-class density and multi-class distribution; it generates differentiated evidence based on a dynamic belief allocation mechanism (covering prior probability evidence of samples without nearest neighbors, weighted distance evidence of mixed nearest neighbors, and high-confidence preserved evidence of pure nearest neighbors), and fuses multi-source evidence using structured evidence combination rules, outputting belief values ​​for single-class, multi-class, and unknown classes; through frequency statistics and class weight optimization, it achieves the allocation of beliefs from multi-class and unknown classes to single-class, supporting specific complex regression tasks. This invention maintains low complexity while achieving inference speeds tens of times faster than deep learning models, and the decision path is fully interpretable. It is particularly suitable for real-time decision-making scenarios requiring a balance between accuracy and cost, such as national defense security and medical diagnosis, effectively solving the systematic misjudgment problem of traditional methods in handling uncertain samples.

[0046] This invention presents a collaborative classification and regression decision-making method based on dynamic belief allocation and evidence fusion. By constructing a dynamic belief allocation mechanism, it quantifies the association strength between test samples and each candidate category, avoiding misleading false confidence levels. Furthermore, by integrating structured evidence combination rules, it fuses multi-source evidence to generate structured decision support, improving robustness in high-conflict evidence scenarios. The constructed classification-regression collaborative architecture, through the dynamic belief allocation mechanism and structured evidence combination rules, achieves cross-task knowledge transfer and performance complementarity, addressing the limitation of existing mainstream classification frameworks, which are constrained by task isolation design and cannot directly support regression analysis.

[0047] This invention has been verified through numerous experiments to have advantages over current research methods, such as faster operation speed, stronger interpretability, and lower complexity.

[0048] See the instruction manual appendix Figure 3 This embodiment also provides a collaborative classification regression decision-making device based on dynamic belief allocation and evidence fusion, which is used to implement the above-described method embodiment. The device includes:

[0049] The nearest neighbor calculation unit 201 is used to obtain test samples and training samples, calculate the distance between each test sample and all training samples in the feature space, and select the K training samples that are closest to each test sample to obtain the nearest neighbor set of the test sample.

[0050] The category prediction unit 202 is used to generate structured evidence supporting classification decisions based on the nearest neighbor set of the test sample, calculate the belief value of each proposition corresponding to each test sample based on the structured evidence, and obtain the predicted category of the test sample based on the proposition with the largest belief value.

[0051] The label generation unit 203 is used to count the frequency of occurrence of training samples of each category in the nearest neighbor set according to the category to which the training sample belongs, construct category weights according to the frequency of occurrence of training samples, allocate belief values ​​through category weights to obtain the belief value of each category as the label weight, and generate predicted labels according to the label weights.

[0052] All relevant content of each step involved in the above method embodiments can be referenced from the functional description of the corresponding functional module, and will not be repeated here.

[0053] In other embodiments of this application, an electronic device is disclosed, such as... Figure 4As shown, the electronic device 300 may include: one or more processors 301; a memory 302; a display 303; one or more application programs (not shown); and one or more computer programs 304. These devices can be connected via one or more communication buses 305. The one or more computer programs 304 are stored in the memory and configured to be executed by the one or more processors 301. The one or more computer programs 304 include instructions that can be used to perform actions such as... Figure 1 And the steps in the corresponding embodiments.

[0054] Through the above description of the embodiments, those skilled in the art will clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device, and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0055] In the embodiments of this application, the functional units can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0056] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, essentially, or the parts that contribute to the prior art, or all or part of the technical solutions, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as flash memory, portable hard disk, read-only memory, random access memory, magnetic disk, or optical disk.

[0057] The above description is merely a specific implementation of the embodiments of this application, but the protection scope of the embodiments of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in the embodiments of this application should be covered within the protection scope of the embodiments of this application. Therefore, the protection scope of the embodiments of this application should be determined by the protection scope of the claims.

Claims

1. A collaborative classification regression decision-making method based on dynamic belief assignment and evidence fusion, characterized in that, include: Obtain test samples and training samples, calculate the distance between each test sample and all training samples in the feature space, and select the K nearest training samples to obtain the nearest neighbor set of the test sample; Structured evidence supporting classification decisions is generated based on the nearest neighbor set of the test sample. The belief value of each proposition corresponding to each test sample is calculated based on the structured evidence. The predicted category of the test sample is obtained based on the proposition with the largest belief value. The frequency of occurrence of training samples of each category in the nearest neighbor set is counted according to the category to which the training sample belongs. Category weights are constructed based on the frequency of occurrence of training samples. Belief values ​​are assigned to each category using the category weights as label weights. Predicted labels are generated based on the label weights.

2. The method according to claim 1, characterized in that, Generate structured evidence supporting classification decisions based on the nearest neighbor set of the test samples, including: Based on the nearest neighbor set and the identification framework, the distribution of nearest neighbor samples in the nearest neighbor set is judged for each single-element category in the identification framework, and evidence corresponding to the single-element category is generated based on the judgment result. Structured evidence supporting the classification decision of test samples is obtained based on the single-element category correspondence evidence in the identification framework combination.

3. The method according to claim 1, characterized in that, The test samples correspond to proposition types including singleton propositions, multi-instance propositions, and ignorance propositions; Based on the structured evidence, the belief value for each proposition corresponding to each test sample is calculated, including: When the proposition is a singleton or an ignorance proposition, the belief value of the proposition is calculated according to the following formula: Or A = Ω, where, Let m represent the belief value of proposition A obtained after fusing s pieces of evidence for the i-th test sample, where s represents the number of pieces of evidence, A represents the proposition, B1 and B2 represent certain propositions in the evidence, and m represents the belief value of proposition A. is (B2) represents the belief value assigned to proposition B2 by the s-th piece of evidence in the i-th test sample, and Ω represents the identification frame; When the proposition is a polynomial proposition, the belief value of the proposition is calculated according to the following formula: Here, Ψ represents the multi-instance class in the Ω subset.

4. The method according to claim 2, characterized in that, Determine the distribution of nearest neighbor samples in the nearest neighbor set, and generate single-element category correspondence evidence based on the determination results, including: Determine whether there are nearest neighbor samples in a single-element category; When there are no nearest neighbor samples in a single element category, generate corresponding weak confidence evidence; When there are nearest neighbor samples in a single element category, determine whether all the nearest neighbor samples belong to that single element category; When the nearest neighbor samples do not all belong to the single-element category, a weighted evidence of the corresponding mixed nearest neighbors is generated; When all neighboring samples belong to the same single-element category, corresponding high-confidence evidence is generated. We designed belief value allocation mechanisms for weak confidence evidence, mixed nearest neighbor weighted evidence, and high confidence evidence, respectively.

5. The method according to claim 4, characterized in that, Design belief value allocation mechanisms for weak-confidence evidence, mixed-nearest neighbor weighted evidence, and high-confidence evidence respectively, including: The belief value assignment for weak confidence evidence satisfies the following formula: m(ω s )=0.1p, m(Ω)=1-0.1p, where, N s ω represents the training samples s The number of class samples, N represents the total number of training samples, and p represents ω. s The distribution ratio of class samples in the training samples, m(ω) s ) indicates that it is assigned to ω s The belief value of the class, m(Ω) represents the belief value assigned to Ω, and Ω represents the identification frame; The belief value assignment of weighted evidence from mixed nearest neighbors satisfies the following formula: m(ω s ) = f, m(Ω) = 1 - f, where f represents the weighted evidence of the mixed nearest neighbors assigned to ω. s The belief value of a class, where 'a' represents the belief value of the test sample to class ω. s The average distance of the nearest neighbor samples, Represents ω s Intra-class tightness; The belief value assignment for high-confidence evidence satisfies the following formula: m(ω) s )=0.999,m(Ω)=0.

001.

6. The method according to claim 1, characterized in that, The class weights are constructed based on the frequency of occurrence of training samples, including: The unnormalized weights of training samples for each category are obtained by adding the frequency of occurrence of training samples to the optimization factor. The class weights for each category are obtained by normalizing the unnormalized weights of the training samples in each category and the sum of the weights of all categories. The normalization process satisfies the following formula: Where w(ω) s ) represents ω s Class weights, w′(ω) s ) represents ω s The unnormalized weights of the classes, w′(T) represent the sum of the unnormalized weights of all classes.

7. The method according to claim 1, characterized in that, The belief value for each category is assigned as a label weight by assigning belief values ​​based on the category weights, including: The assignment of belief values ​​based on the category weights satisfies the following formula: Where m(ω) s ) represents ω s Class label weight, w(ω) s ) represents ω s Class weights, m i (ω s ) indicates that the i-th element contains ω s The belief value of a class of multiple instance propositions or ignorant propositions.

8. A collaborative classification and regression decision-making device based on dynamic belief allocation and evidence fusion, characterized in that, The device includes: The nearest neighbor calculation unit is used to obtain test samples and training samples, calculate the distance between each test sample and all training samples in the feature space, and select the K training samples that are closest to each test sample to obtain the nearest neighbor set of the test sample. The category prediction unit is used to generate structured evidence supporting classification decisions based on the nearest neighbor set of the test sample, calculate the belief value of each proposition corresponding to each test sample based on the structured evidence, and obtain the predicted category of the test sample based on the proposition with the largest belief value. The label generation unit is used to count the frequency of occurrence of training samples of each category in the nearest neighbor set according to the category to which the training sample belongs, construct category weights according to the frequency of occurrence of training samples, allocate belief values ​​to each category through the category weights to obtain the belief value of each category as the label weight, and generate predicted labels according to the label weights.

9. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is executed by the processor, it implements the collaborative classification regression decision-making method based on dynamic belief assignment and evidence fusion as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, include: Processor and memory; The memory is used to store computer programs; The processor is used to execute the computer program stored in the memory to cause the electronic device to perform the collaborative classification regression decision method based on dynamic belief assignment and evidence fusion as described in any one of claims 1 to 7.

Images