Prototype evidence sequential regression data processing method and system based on geometric ordering manifold learning

By constructing learnable prototype vectors and evidence networks in the unit hypersphere feature space, and combining Dempster-Schoff theory and Bayesian evidence loss, the geometric interpretability and uncertainty problems of deep ordinal regression models are solved, achieving highly reliable and interpretable data processing.

CN122451844APending Publication Date: 2026-07-24BEIJING NORMAL UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610606782.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-05-06
Publication Date
2026-07-24

AI Technical Summary

Technical Problem

Existing deep ordinal regression models lack geometric interpretability, ignore the semantic continuity of labels, and cannot effectively distinguish data cognitive uncertainty, leading to misjudgment and overconfidence in application scenarios with high reliability requirements.

Method used

By constructing a prototype evidence order regression method based on geometric ordering manifold learning, a learnable prototype vector is constructed in the feature space of a unit hypersphere using a deep neural network. The Euclidean distance is calculated and mapped to the concentration parameters of the Dirichlet distribution. Combining Dempster-Schoff theory and Bayesian evidence loss, a loss function is constructed for end-to-end training to process new input data.

Benefits of technology

It improves the interpretability and semantic consistency of the model, can explicitly decouple and quantify uncertainty, provides rejection capability, and is suitable for application scenarios with high reliability requirements.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122451844A_ABST
    Figure CN122451844A_ABST
Patent Text Reader

Abstract

The present application provides a prototype evidence sequence regression data processing method and system based on geometric ordering manifold learning, comprising: constructing a unit hypersphere feature space; based on the unit hypersphere feature space, constructing a learnable prototype vector; calculating the Euclidean distance between the feature vector and the learnable prototype vector to obtain an evidence vector; mapping the evidence vector to the concentration parameter of the Dirichlet distribution, combining the Dempster-Shafer theory reasoning framework to obtain a prototype evidence network model, and outputting a prediction result; based on the prediction result, introducing a ordinal manifold constraint based on the bulldozer distance, combining the Bayesian evidence loss to construct a loss function; based on the trained prototype evidence network model, processing new input data to obtain a final prediction result. The present application not only realizes accurate prediction of ordered data and geometric interpretability of feature space, but also has automatic rejection ability for out-of-distribution samples, significantly improving the robustness of the model in long-tail distribution and high-risk scenarios.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and big data intelligent processing technology, specifically relating to a prototype evidence order regression data processing method and system based on geometric sorting manifold learning. Background Technology

[0002] With the rapid development of artificial intelligence technology, deep learning has become the core engine driving computer vision, pattern recognition, and big data intelligent analysis. Through an end-to-end feature learning paradigm, deep neural networks can automatically extract complex high-level semantic representations from massive amounts of high-dimensional data. Although traditional supervised learning tasks are mainly divided into two paradigms: discrete classification and continuous regression, a special task paradigm that lies between the two—ordinal regression—is widely used in complex application scenarios such as industrial production, medical auxiliary analysis, social computing, and security monitoring.

[0003] Currently, relying on powerful backbone architectures such as convolutional neural networks (CNNs, such as ResNet and EfficientNet) or visual transformers (ViT), deep models have demonstrated excellent feature extraction capabilities for ordinal regression tasks. To effectively utilize the ordinal information of labels and overcome the limitation of standard cross-entropy loss in ignoring inter-class relationships, mainstream solutions in academia and industry typically employ specific strategies to improve standard classifiers. These mainly include ordinal decomposition strategies and label distribution learning / soft label strategies.

[0004] Despite improvements in accuracy from existing technologies, the following common technical bottlenecks remain insurmountable in applications with high reliability requirements (such as autonomous driving, precision medicine, and financial risk control): Lacking inherent geometric interpretability and relying on "post-hoc" explanations: Most existing deep ordinal regression models are essentially "black box" systems, relying on fully connected layers of deep neural networks to directly map high-dimensional features into class probabilities. Although visualization techniques such as saliency maps exist, these are "post-hoc explanations," only showing the correlation region between input and output, but failing to reveal the actual reasoning logic or decision path within the model. The feature space often exhibits a loosely distributed isotropic structure, lacking structured geometric meaning, making it difficult for users to understand why the model makes a particular decision.

[0005] Ignoring ordinal topology leads to semantic inconsistencies: Traditional classification loss functions (such as cross-entropy loss) are typically based on the independent and identically distributed (IOD) assumption, treating each ordinal level as a discrete and independent entity (i.e., the "closed-world assumption"). This approach severs the objectively existing semantic continuity between labels, ignoring the fact that the feature similarity between adjacent levels (such as level 1 and level 2) should be significantly higher than that between levels with larger gaps (such as level 1 and level 4). This causes the model to fail to construct a continuous topological structure in the feature space that reflects the evolution of the target, easily leading to "jump-like" misclassifications that violate ordinal logic (e.g., misclassifying mild samples as extremely severe).

[0006] Insufficient quantification of uncertainty poses a risk of "overconfidence": Most existing models only provide deterministic point estimations. When faced with out-of-distribution (OOD) samples, noisy samples, or ambiguous samples at the decision boundary, models often fail to recognize their own cognitive blind spots, tending to provide an overconfident high-probability prediction. Existing solutions such as Bayesian neural networks (BNNs) are computationally expensive and difficult to converge, while standard evidence-based deep learning (EDL) methods ignore ordinal geometry, resulting in a lack of reliable "reject option" criteria for high-risk decisions.

[0007] The disordered nature of prototype learning makes it difficult to represent evolutionary trajectories: While prototype learning provides an interpretable approach through case-based reasoning, existing prototype networks are primarily designed for discrete classification tasks (such as species identification). The distribution of prototype vectors in the feature space is typically free and disordered, lacking explicit geometric constraints. In ordinal regression tasks, this disordered distribution cannot guarantee that the learned prototypes accurately reflect the progressive trajectory of increasing degree or state evolution, causing distance metrics in the feature space to fail to truly reflect semantic differences.

[0008] In summary, existing data processing methods struggle to simultaneously ensure high prediction accuracy while also considering the interpretability of the decision-making process, the ordinal geometric structure of the feature space, and the reliable quantification of prediction uncertainty. Therefore, there is an urgent need to develop a data processing method that can simultaneously construct ordered geometric feature manifolds, provide intuitive decision-making support, and accurately decouple and quantify prediction uncertainty. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of existing deep ordinal regression techniques, such as lack of geometric interpretability, neglect of semantic continuity of labels, and inability to effectively distinguish data cognitive uncertainty, and to provide a prototype evidence ordinal regression data processing method and system based on geometric ordering manifold learning.

[0010] To achieve the above objectives, the present invention provides the following solution: Prototype evidence-based order regression data processing methods based on geometric ordering manifold learning include: Deep neural networks are used to encode features in the input data to be processed, obtain feature vectors, and construct the feature space of a unit hypersphere. Based on the unit hypersphere feature space, a learnable prototype vector representing the order level is constructed. Calculate the Euclidean distance between the feature vector and the learnable prototype vector to obtain the evidence vector; The evidence vector is mapped to the concentration parameters of the Dirichlet distribution, and combined with the Dempster-Schoff theoretical reasoning framework, a prototype evidence network model is obtained, and the prediction results are output. Based on the prediction results, an ordinal manifold constraint based on bulldozer distance is introduced, and a loss function is constructed by combining Bayesian evidence loss to achieve end-to-end training of the prototype evidence network model. The newly input data is processed based on the trained prototype evidence network model to obtain the final prediction result, thus completing the prototype evidence ordinal regression data processing.

[0011] Preferred methods for constructing the feature space of a unit hypersphere include: The deep neural network is used as a feature encoder to map the input data to be processed onto the feature space to obtain feature vectors; The feature vectors are L2 normalized and projected onto a unit hypersphere manifold with a fixed radius to obtain the unit hypersphere feature space.

[0012] Preferably, the Euclidean distance is converted into an evidence vector using an RBF kernel function with a learnable scaling parameter.

[0013] Preferably, the methods for obtaining prediction results include: The evidence vector is mapped to the concentration parameters of a Dirichlet distribution to obtain the total intensity of the Dirichlet distribution; The prediction results are obtained based on the total strength of the Dirichlet distribution; wherein, the prediction results include belief quality and uncertainty quality.

[0014] Preferably, methods for constructing the loss function include: A Bayesian risk loss based on the sum of squares error is constructed by combining a KL divergence regularization term. Based on the difference between the predicted belief quality distribution and the actual label distribution at the cumulative distribution function level, the ordinal manifold constraint is obtained. Based on the Bayesian evidence loss and the ordinal manifold constraint, the loss function is constructed by pre-setting a balance coefficient.

[0015] Preferably, the final prediction result includes a prediction level and an uncertainty score, wherein out-of-distribution samples and risky ambiguous samples are determined by comparing a preset uncertainty threshold with the uncertainty score.

[0016] This invention also provides a prototype evidence order regression data processing system based on geometric ordering manifold learning, for implementing the method, comprising: The module for constructing the feature space of a unit hypersphere is used to encode the features of the input data to be processed using a deep neural network, obtain feature vectors, and construct the feature space of a unit hypersphere. A learnable prototype vector construction module is used to construct learnable prototype vectors representing ordered levels based on the unit hypersphere feature space. An evidence vector construction module is used to calculate the Euclidean distance between the feature vector and the learnable prototype vector to obtain the evidence vector. The model building module is used to map the evidence vector to concentration parameters of the Dirichlet distribution, combine it with the Dempster-Schoff theoretical reasoning framework to obtain a prototype evidence network model, and output the prediction results. The model training module is used to introduce ordinal manifold constraints based on bulldozer distance based on the prediction results, combine Bayesian evidence loss, construct a loss function, and realize end-to-end training of the prototype evidence network model. The inference module is used to process new input data based on the trained prototype evidence network model, obtain the final prediction result, and complete the prototype evidence ordinal regression data processing.

[0017] Preferably, the model training module includes: The Bayesian evidence loss construction unit is used to construct the Bayesian evidence loss by using Bayesian risk loss based on the sum of squares error and combining it with the KL divergence regularization term. Ordinal manifold constraint building unit is used to calculate the difference between the predicted belief quality distribution and the true label distribution at the cumulative distribution function level based on bulldozer distance, and obtain ordinal manifold constraints; The loss function construction unit is used to construct the loss function based on the Bayesian evidence loss and the ordinal manifold constraint, using a preset balance coefficient.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: High interpretability: Through prototype-based distance measurement, the model is no longer a black box, and the geometric proximity of features to the prototype directly explains the classification basis.

[0019] Geometric ordering: The proposed ordinal manifold constraint forces the prototypes to be distributed along a continuous trajectory in the feature space, which solves the problem that traditional classifiers ignore the ordinal relationship of labels and improves semantic consistency.

[0020] Reliable uncertainty estimation: By combining evidence-based deep learning, it can distinguish between "cognitive uncertainty caused by data noise" and "random uncertainty of the model itself", and effectively detect abnormal samples through uncertainty scores.

[0021] De-blackboxing and rejection capability: When the data deviates from the learning manifold, the model will automatically output high uncertainty, thereby achieving a safe automatic rejection function, which is suitable for application scenarios with high reliability requirements. Attached Figure Description

[0022] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the prototype evidence order regression data processing method based on geometric sorting manifold learning according to an embodiment of the present invention; Figure 2 This is a structural diagram of a model according to an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0026] The following supplementary explanation is provided regarding the ordinal regression task involved in this invention: The core characteristic of ordinal regression lies in the inherent natural ordering property and semantic continuity of its target label space. Unlike unordered nominal classification, ordinal labels exhibit a clear relative hierarchical relationship (i.e., ...). However, the absolute distance or metric difference between different levels is often unknown, non-linear, or difficult to define. The application scenarios for this type of task are extremely diverse, including but not limited to: Image and multimedia quality assessment: rating the aesthetic quality and sharpness of images or videos (e.g., a 5-level rating from "poor" to "perfect"). Industrial intelligent inspection: fine-grained classification of surface defects or damage levels of industrial products (e.g., "no damage", "minor scratches", "moderate wear", "severe damage"). Biometrics and behavioral analysis: progressive analysis of the evolutionary stages of organisms (such as age estimation and growth cycle prediction) or the intensity of human facial expressions; Risk assessment and decision-making: such as credit rating classification in the financial sector and early warning level determination for natural disasters.

[0027] In these tasks, the cost of prediction error is usually class-sensitive, meaning that the greater the deviation of the prediction from the true label, the more severe the decision loss (for example, the severity of misclassifying "level 1" as "level 5" is much greater than misclassifying it as "level 2").

[0028] This invention aims to overcome the fundamental shortcomings of traditional deep learning classification methods (such as classifiers based on cross-entropy loss) in handling such tasks. Traditional methods are typically based on the "closed-world assumption," treating ordered labels as orthogonal and independent discrete categories (e.g., assuming the difference between "Level 1" and "Level 2" is equivalent to the difference between "Level 1" and "Level 10"). This approach severs the underlying semantic continuity of the data, ignoring the objective fact that the feature similarity between adjacent levels should be higher than that between levels with larger differences. This invention, by mining the latent category boundaries and ordinal topology in the feature space, enables accurate modeling and prediction of ordered data.

[0029] Example 1: like Figure 1 As shown, the prototype evidence order regression data processing method based on geometric ordering manifold learning includes: S1: A deep neural network is used to encode features of the input data (e.g., image data) to obtain feature vectors, constructing a unit hypersphere feature space. Let the training dataset be... ,in This represents the input data (such as images, signals, or feature vectors). This represents the corresponding real ordered labels, totaling [number]. An ordered hierarchy.

[0030] A further implementation method involves constructing the feature space of a unit hypersphere, including: Using deep neural networks (such as backbone networks based on CNN or Transformer architectures) as feature encoders Input data to be processed Mapping to a low-dimensional latent feature space to obtain feature vectors ; The eigenvectors are L2 normalized and projected onto a unit hypersphere manifold of fixed radius to obtain an isotropic unit hypersphere feature space. Specifically, to ensure the stability of distance measurements, the eigenvectors are... L2 normalization is performed to obtain the eigenvectors on the unit hypersphere. : in, .

[0031] S2: Based on the unit hypersphere feature space, construct learnable prototype vectors representing ordered levels. Specifically, in the unit hypersphere feature space, initialize and define a set of learnable prototype vectors. ,in It represents the first A parameterized vector of ordered levels, satisfying Ordered classification tasks include... Each level corresponds to a definition. There are 10 prototype vectors. Each prototype vector represents the semantic center of a specific ordered level (e.g., level 0 to level K-1). During training, these prototype vectors are updated synchronously with the network parameters and are also subject to the same conditions. Normalization constraints ensure that they are always distributed on the unit hypersphere. Through subsequent constraints, these prototypes will no longer be loosely distributed, but will form continuous trajectories along the manifold surface reflecting hierarchical evolution.

[0032] In this metric and prototype learning, the system no longer simply learns the relationship between samples and classification boundaries, but rather the geometric distances between samples and a set of learnable prototypes. These prototype vectors, serving as "semantic centers" for each ordered level, are also constrained to the unit hypersphere.

[0033] S3: Calculate the Euclidean distance between the feature vector and the learnable prototype vector to obtain the evidence vector. A further implementation uses an RBF kernel function with a learnable scaling parameter to transform the Euclidean distance into an evidence vector. This invention abandons fully connected layers and directly derives the evidence from the geometric structure. Specifically, in the feature space, the Euclidean distance between the normalized input sample feature vector and each prototype vector is calculated. That is, the input features are calculated... With each prototype Euclidean distance between : .

[0034] Based on the principle that "distance represents inverse evidence," the geometric distance described above is transformed into non-negative evidence values ​​supporting various levels using a radial basis function (RBF) kernel. In this process, a learnable parameter—the "Evidence Amplifier"—is introduced. This parameter adaptively adjusts the sensitivity of the RBF kernel, dynamically enhancing the evidence response strength of typical samples while maintaining high sensitivity to out-of-distribution samples, thus addressing the common confidence undercalibration problem in evidence networks on complex samples. Specifically, based on the principle that "the closer the distance, the stronger the evidence," the distance is transformed into an evidence vector using an RBF kernel function with a learnable scaling parameter. For the first Class, evidence The calculation is as follows: , in, It is a learnable evidence amplifier (Evidence Scale) used to adaptively adjust the amplitude of evidence; It is a hyperparameter that controls the kernel width.

[0035] S4: Map the evidence vectors to concentration parameters of a Dirichlet distribution, and combine this with the Dempster-Schoff theoretical inference framework to obtain a prototype evidence network model, outputting the prediction results. A further implementation method for obtaining the prediction results includes: The evidence vectors are mapped to concentration parameters of a Dirichlet distribution to obtain the total strength of the Dirichlet distribution. Based on the total strength of the Dirichlet distribution, the prediction results are obtained, including belief mass and uncertainty mass. Specifically, an evidence reasoning framework is constructed based on Dempster-Shafer Theory (DST). The evidence vectors generated in S3 are mapped to concentration parameters of a Dirichlet distribution, i.e., the evidence value of each class is incremented by 1 to obtain the corresponding parameter. Based on this, the model's prediction output is explicitly decoupled into two parts: one is the belief mass belonging to each ordered level, used for the final classification decision; the other is the overall uncertainty mass, used to characterize the cognitive uncertainty caused by the lack of evidence (i.e., the sample is far from all prototypes). This mechanism allows the model to output a high uncertainty score when encountering unknown or ambiguous data.

[0036] This invention can explicitly decouple and quantify two types of uncertainty in model prediction results: Epistemic uncertainty stems from the model's lack of knowledge about the data (i.e., "never seen it before"), and usually corresponds to out-of-distribution (OOD) samples. Aleatoric uncertainty arises from the inherent noise or ambiguity of the data itself (such as samples located at the boundary between two levels).

[0037] Specifically, using the Dempster-Schaugh theory (DST), the evidence vector is... Parameters associated with the Dirichlet distribution : , Define the total intensity (Precision) of the Dirichlet distribution as: .

[0038] Therefore, the model's predictive output is decoupled from belief quality. and uncertainty quality : , Satisfying constraints .

[0039] The final classification prediction probability is the expected value of the Dirichlet distribution: .

[0040] like Figure 2 As shown, Figure 2 Shared encoder Using EfficientNet-B4 as the backbone network, the input fundus image is processed. Encode as a feature map, then normalize using L2 ( Project it onto the unit hypersphere. Geometrically ordered prototype. The nodes are visualized as arranged along a continuous ordinal manifold, and used to calculate the Euclidean distance. The anchor points. These geometric distances are determined by a learnable scale. Parameterized RBF kernels transformed into evidence Ultimately, DST-based evidence reasoning aggregates this evidence to generate decoupled belief quality. and uncertainty quality (satisfies the law of conservation) The optimization process of the model is achieved through Bayesian evidence loss. and EMD-based ordinal manifold constraints Joint guidance.

[0041] S5: Based on the prediction results, an ordinal manifold constraint based on bulldozer distance is introduced, and a loss function is constructed by combining Bayesian evidence loss to achieve end-to-end training of the prototype evidence network model.

[0042] A further implementation method involves constructing the loss function by: Bayesian evidence loss is constructed by employing Bayesian risk loss based on sum of squared errors and combining it with KL divergence regularization. The aim is to make the predicted distribution approximate the one-hot encoding of the true label and push evidence of non-target categories toward 0.

[0043] in It is the actual label vector. The weight coefficients representing the KL divergence regularization term. The representative parameter is a uniform Dirichlet distribution of a vector consisting of all 1s. This represents the Dirichlet distribution parameterized by the prediction parameter $\tilde{\alpha}_i$.

[0044] Based on the difference between the predicted belief quality distribution and the actual label distribution calculated from bulldozer distance at the cumulative distribution function level, ordinal manifold constraints are obtained. It is mandatory that the prototype must strictly follow the semantic order of the labels in the feature space (e.g., level 0). Level 1 ... The rank (K-1) is arranged along a continuous manifold. This ensures that the geometric topology of the feature space is aligned with the semantic ordinality of the data, thereby suppressing outlier predictions with large spans.

[0045] in, , . The predicted probability distribution is at the th Cumulative probability at each category The true label distribution is in the th order. The cumulative probability of each category.

[0046] This constraint forces the feature space to form a continuous and ordered manifold, so that the distance between the input sample and the prototype naturally reflects the degree of data evolution.

[0047] Based on Bayesian evidence loss and ordinal manifold constraints, a loss function is constructed by pre-setting balance coefficients: .

[0048] in This is the balance coefficient.

[0049] S6: Based on the trained prototype evidence network model, process the new input data to obtain the final prediction result and complete the prototype evidence ordinal regression data processing.

[0050] A further implementation method involves the final prediction result including a prediction level. and uncertainty score Among them, by setting a preset uncertainty threshold With uncertainty score The system compares the predicted probabilities to determine whether a sample is out of distribution or a high-risk ambiguous sample. Specifically, the trained network processes the new input data. The model outputs a probability distribution (based on the Dirichlet distribution mean) for each level and a scalar uncertainty score. The system outputs the final level label based on the predicted probability while monitoring the uncertainty score. When the uncertainty score exceeds a preset threshold, the system triggers a "Reject Option" mechanism to determine whether the sample is out of distribution or a high-risk sample, thus ensuring the reliability of the system.

[0051] This characteristic makes the present invention particularly suitable for long-tailed distribution data or complex scenarios with severe class imbalance. It effectively solves the "overconfidence" problem that traditional deep learning models often make high-confidence erroneous predictions when faced with unknown or anomalous data, thereby significantly improving the robustness, interpretability, and security of artificial intelligence systems in open environments.

[0052] In summary, this invention proposes a prototype evidence network (PEN) data processing method and system based on geometrically ordered manifold learning. This invention overcomes the limitations of traditional deep ordinal regression models, which lack geometric interpretability and uncertainty quantification capabilities, by innovatively constructing a "Prototype Evidence Network (PEN)" framework. By introducing ordinal manifold constraints based on bulldozer distance (EMD) into the unit hypersphere metric space, this invention successfully forces learnable prototype vectors to exhibit a continuous and ordered distribution along the semantic trajectory of data evolution, thereby achieving a high degree of unification between the geometric structure of the feature space and the ordinal logic of the label space.

[0053] Furthermore, this invention utilizes the Dempster-Schoff theory (DST) to transform geometric distance into evidential belief, achieving explicit decoupling of belief quality and cognitive uncertainty in the model's prediction results. This mechanism endows the system with the ability to automatically perceive and reject out-of-distribution samples or high-risk ambiguous data, effectively solving the "overconfidence" problem of traditional deep learning models in open environments. Both experimental results and theoretical analysis demonstrate that this invention significantly improves the interpretability, robustness, and safety of ordinal regression tasks while maintaining high accuracy.

[0054] Example 2: This invention also provides a prototype evidence order regression data processing system based on geometric ordering manifold learning, for implementing the method of Embodiment 1, including: The module for constructing the feature space of a unit hypersphere is used to encode the features of the input data to be processed using a deep neural network, obtain feature vectors, and construct the feature space of a unit hypersphere. The learnable prototype vector construction module is used to construct learnable prototype vectors representing ordered levels based on the unit hypersphere feature space. The evidence vector construction module is used to calculate the Euclidean distance between the feature vector and the learnable prototype vector to obtain the evidence vector; The model building module is used to map evidence vectors to concentration parameters of a Dirichlet distribution, and combined with the Dempster-Schoff theoretical reasoning framework, obtain a prototype evidence network model and output prediction results. The model training module is used to construct a loss function based on the prediction results, by introducing ordinal manifold constraints based on bulldozer distance and combining them with Bayesian evidence loss, so as to achieve end-to-end training of the prototype evidence network model. The inference module is used to process new input data based on the trained prototype evidence network model, obtain the final prediction result, and complete the prototype evidence ordinal regression data processing.

[0055] A further implementation method includes a model training module comprising: The Bayesian evidence loss construction unit is used to construct the Bayesian evidence loss by using Bayesian risk loss based on the sum of squares error and combining it with the KL divergence regularization term. Ordinal manifold constraint building unit is used to calculate the difference between the predicted belief quality distribution and the true label distribution at the cumulative distribution function level based on bulldozer distance, and obtain ordinal manifold constraints; The loss function construction unit is used to construct a loss function based on Bayesian evidence loss and ordinal manifold constraints, using preset balance coefficients.

[0056] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A prototype evidence-based ordinal regression data processing method based on geometric ordering manifold learning, characterized in that, include: Deep neural networks are used to encode features in the input data to be processed, obtain feature vectors, and construct the feature space of a unit hypersphere. Based on the unit hypersphere feature space, a learnable prototype vector representing the order level is constructed. Calculate the Euclidean distance between the feature vector and the learnable prototype vector to obtain the evidence vector; The evidence vector is mapped to the concentration parameters of the Dirichlet distribution, and combined with the Dempster-Schoff theoretical reasoning framework, a prototype evidence network model is obtained, and the prediction results are output. Based on the prediction results, an ordinal manifold constraint based on bulldozer distance is introduced, and a loss function is constructed by combining Bayesian evidence loss to achieve end-to-end training of the prototype evidence network model. The newly input data is processed based on the trained prototype evidence network model to obtain the final prediction result, thus completing the prototype evidence ordinal regression data processing.

2. The method according to claim 1, characterized in that, Methods for constructing the feature space of a unit hypersphere include: The deep neural network is used as a feature encoder to map the input data to be processed to the feature space to obtain feature vectors; The feature vectors are L2 normalized and projected onto a unit hypersphere manifold with a fixed radius to obtain the unit hypersphere feature space.

3. The method according to claim 1, characterized in that, The Euclidean distance is transformed into an evidence vector using an RBF kernel function with learnable scaling parameters.

4. The method according to claim 1, characterized in that, Methods for obtaining prediction results include: The evidence vector is mapped to the concentration parameters of a Dirichlet distribution to obtain the total intensity of the Dirichlet distribution; The prediction results are obtained based on the total strength of the Dirichlet distribution; wherein, the prediction results include belief quality and uncertainty quality.

5. The method according to claim 1, characterized in that, Methods for constructing loss functions include: A Bayesian risk loss based on the sum of squares error is constructed by combining a KL divergence regularization term. Based on the difference between the predicted belief quality distribution and the actual label distribution at the cumulative distribution function level, the ordinal manifold constraint is obtained. Based on the Bayesian evidence loss and the ordinal manifold constraint, the loss function is constructed by pre-setting a balance coefficient.

6. The method according to claim 1, characterized in that, The final prediction result includes a prediction level and an uncertainty score. The determination of out-of-distribution samples and risky ambiguous samples is made by comparing the preset uncertainty threshold with the uncertainty score.

7. A prototype evidence order regression data processing system based on geometric ordering manifold learning, used to implement the method described in any one of claims 1-6, characterized in that, include: The module for constructing the feature space of a unit hypersphere is used to encode the features of the input data to be processed using a deep neural network, obtain feature vectors, and construct the feature space of a unit hypersphere. A learnable prototype vector construction module is used to construct learnable prototype vectors representing ordered levels based on the unit hypersphere feature space. An evidence vector construction module is used to calculate the Euclidean distance between the feature vector and the learnable prototype vector to obtain the evidence vector. The model building module is used to map the evidence vector to concentration parameters of the Dirichlet distribution, combine it with the Dempster-Schoff theoretical reasoning framework to obtain a prototype evidence network model, and output the prediction results. The model training module is used to introduce ordinal manifold constraints based on bulldozer distance based on the prediction results, combine Bayesian evidence loss, construct a loss function, and realize end-to-end training of the prototype evidence network model. The inference module is used to process new input data based on the trained prototype evidence network model, obtain the final prediction result, and complete the prototype evidence ordinal regression data processing.

8. The system according to claim 7, characterized in that, The model training module includes: The Bayesian evidence loss construction unit is used to construct the Bayesian evidence loss by using Bayesian risk loss based on the sum of squares error and combining it with the KL divergence regularization term. Ordinal manifold constraint building unit is used to calculate the difference between the predicted belief quality distribution and the true label distribution at the cumulative distribution function level based on bulldozer distance, and obtain ordinal manifold constraints; The loss function construction unit is used to construct the loss function based on the Bayesian evidence loss and the ordinal manifold constraint, using a preset balance coefficient.