Multi-view credible classification method based on subspace evidence fusion

Abstract

The invention provides a subspace evidence fusion-based multi-view credible classification method, which comprises the following steps of: mapping multi-view data to a unified discrimination space by constructing a learnable discrimination matrix; inputting the feature matrix in the discrimination space into a visual angle-specific evidence deep neural network, and extracting visual angle-specific evidence; and aggregating specific evidences of all view angles based on the cross-view-angle weight matrix, and generating a fused result. According to the method, the information difference between different view angles is effectively balanced, the negative influence of the conflict view angles on the decision result is reduced, and the decision reliability is improved.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and in particular to a multi-perspective credible classification method based on subspace evidence fusion. Background Art

[0002] Multi-view data can originate from different modalities or features. Integrating heterogeneous views provides a more comprehensive representation of data instances, thus stimulating widespread interest in multi-view classification. With the development of deep learning, most multi-view classification methods rely primarily on neural networks to integrate multi-view data, achieving remarkable results. However, due to the inherent inconsistencies and differences between multiple data sources, simple fusion often introduces potential uncertainties, which can affect performance. Fortunately, evidence-based deep learning (EDL), derived from subjective logic and Dempster-Schafer theory, has emerged as a research direction to address uncertainty in multi-view classification. EDL provides a powerful framework for estimating uncertainty, enabling more accurate decision-making in various applications. Compared to traditional multi-view classification methods, EDL-based methods not only learn view-specific evidence—the amount of support for each class derived from the data—but also parameterize the learned evidence, modeling the view-specific class probability distribution via a Dirichlet distribution. EDL-based methods generate view-specific opinions to assess the uncertainty of each view, thereby improving the reliability of the fusion process. However, in practical applications, conflicting instances may exist between views, limiting the applicability of EDL-based multi-view classification in safety-critical domains.

[0003] like Figure 2 As shown in the figure, wolves and dogs have highly similar characteristics. If each view fails to accurately extract category-specific features to learn view-specific evidence, it will lead to considerable uncertainty in a single view, undermining the reliability of the final result. Most existing EDL-based methods neglect to improve the discriminative power of each view, making it difficult to accurately extract view-specific evidence, such as distinguishing between dogs and wolves, and introducing considerable uncertainty in the fusion process.

[0004] In recent years, researchers have increasingly focused on trusted multi-view learning to provide more reliable results. A groundbreaking contribution is Trusted Multi-view Classification (TMC), which leverages evidence-based deep neural networks to learn view-specific evidence and model the view-specific distribution of class probabilities via a Dirichlet distribution to accurately quantify uncertainty. TMC achieves reliable results by aggregating evidence from all viewpoints via the Dempster-Schafer theory. However, these methods primarily focus on evidence fusion strategies, often neglecting the accuracy and reliability of each viewpoint. Summary of the Invention

[0005] In view of this, the present invention provides a multi-perspective credible classification method based on subspace evidence fusion to solve the above problems.

[0006] The present invention provides a multi-perspective credible classification method based on subspace evidence fusion, comprising: mapping multi-perspective data to a unified discriminant space by constructing a learnable discriminant matrix; inputting the feature matrix in the discriminant space into a perspective-specific evidence deep neural network to extract perspective-specific evidence; and aggregating all perspective-specific evidence based on a cross-perspective weight matrix to generate a fused result.

[0007] In another implementation of the present invention, the multi-view data in the discriminant space is represented as:

[0008] H v =X v P v ,v=1,2,…,V,

[0009] in, Represents the original data of the v-th perspective; Represents the feature representation of the original data of the v-th view in the discriminant space after projection; P v Represents the discriminant matrix corresponding to the v-th view.

[0010] In another implementation of the present invention, the perspective-specific evidence is represented as:

[0011]

[0012] Among them, g v (·) represents the evidence deep neural network of the v-th view, θ v Represents the parameters of the v-th view network; is the feature vector in the discriminant space.

[0013] In another implementation of the present invention, the view-specific total loss function is expressed as:

[0014]

[0015] in, is the classification loss function; is the perspective consistency loss function; Represent the loss function for the class level.

[0016] In another implementation of the present invention, the classification loss function is expressed as:

[0017]

[0018] in, represents the corresponding Dirichlet parameter; λ t =min(1,t / T) is the annealing coefficient, and t represents the number of rounds of current training.

[0019] In another implementation of the present invention, the view consistency loss function is expressed as:

[0020]

[0021] Among them, V represents the total number of viewing angles; is the evidence matrix composed of the i-th instance from all perspectives; is the evidence matrix composed of the j-th instance from all perspectives.

[0022] In another implementation of the present invention, the class-level representation loss function is expressed as:

[0023]

[0024] in, is the intra-class consistency loss; is the inter-class separation loss.

[0025] In another implementation of the present invention, the fused result is expressed as:

[0026]

[0027] Among them, W k is the final cross-view weight vector; Evidence matrix E k The jth evidence vector in .

[0028] On the other hand, the present invention provides a multi-perspective trusted classification system based on subspace evidence fusion, including: a unified multi-perspective discriminant embedding module: used to map multi-perspective data to a unified discriminant space by constructing a learnable discriminant matrix; a perspective-specific evidence learning module: inputting the feature matrix in the discriminant space into a perspective-specific evidence deep neural network to extract perspective-specific evidence; a cross-perspective evidence fusion module: aggregating all perspective-specific evidence based on the cross-perspective weight matrix to generate a fused result.

[0029] The multi-perspective credible classification method based on subspace evidence fusion of the present invention dynamically learns a multi-perspective discriminant matrix by combining multi-perspective discriminant analysis with deep learning, and then aligns samples across perspectives and maintains discriminant information; the learned cross-perspective weight matrix can effectively distinguish consistent and conflicting perspectives, assign optimal weights to each perspective, effectively balance the information differences between different perspectives, reduce the negative impact of conflicting perspectives on decision results, enhance the stability and reliability of the multi-perspective data fusion process, and further improve the reliability of decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. By reading the detailed description of the embodiments below, the advantages and benefits of the solutions will become clear to those skilled in the art. The drawings are only for the purpose of illustrating preferred embodiments and are not to be considered as limiting the present invention. In the drawings:

[0031] Figure 1 The figure is a flowchart of a multi-perspective credible classification method based on subspace evidence fusion according to an embodiment of the present invention.

[0032] Figure 2 The present invention is a schematic diagram illustrating problems existing in the prior art method according to an embodiment of the present invention.

[0033] Figure 3 Schematic diagram of a multi-perspective trusted classification system framework based on subspace evidence fusion according to an embodiment of the present invention.

[0034] Figure 4 FIG. 4 is a schematic diagram showing a comparison of single-view accuracy according to an embodiment of the present invention.

[0035] Figure 5 A schematic diagram of visualizing a cross-view weight matrix according to an embodiment of the present invention. DETAILED DESCRIPTION

[0036] In order to enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and detailedly described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in the embodiments of the present invention should fall within the scope of protection of the embodiments of the present invention.

[0037] Figure 1 A flow chart of a multi-perspective credible classification method based on subspace evidence fusion provided by an embodiment of the present invention is shown as follows: Figure 1 As shown, this embodiment mainly includes:

[0038] S101. Map multi-view data to a unified discriminant space by constructing a learnable discriminant matrix.

[0039] For example, multi-view discriminant analysis is combined with deep learning to learn a multi-view discriminant matrix in a unified multi-view discriminant embedding module. The discriminant matrix is used to map multi-view data to a unified discriminant space, ensuring that each sample is aligned under different viewpoints and maintains discriminative features.

[0040] S102: Input the feature matrix in the discriminant space into the perspective-specific evidence deep neural network to extract perspective-specific evidence.

[0041] Exemplarily, after enhancing the discriminability of feature representations, samples are passed to a view-specific evidence deep neural network to extract view-specific evidence—i.e., quantifying the amount of support for each class from the data.

[0042] S103. Aggregate all view-specific evidence based on the cross-view weight matrix to generate a fused result.

[0043] Exemplarily, in order to distinguish consistent and conflicting perspectives and assign optimal weights to each perspective in the fusion process, a self-updating paradigm is developed in cross-perspective evidence fusion to optimize the learning dynamics of the cross-perspective weight matrix, through which evidence from all perspectives is aggregated, effectively reducing the impact of conflicting perspectives on decision-making.

[0044] The multi-perspective credible classification method based on subspace evidence fusion of the present invention dynamically learns a multi-perspective discriminant matrix by combining multi-perspective discriminant analysis with deep learning, and then aligns samples across perspectives and maintains discriminant information; the learned cross-perspective weight matrix can effectively distinguish consistent and conflicting perspectives, assign optimal weights to each perspective, effectively balance the information differences between different perspectives, reduce the negative impact of conflicting perspectives on decision results, enhance the stability and reliability of the multi-perspective data fusion process, and further improve the reliability of decision-making.

[0045] In another implementation of the present invention, the multi-view data in the discriminant space is represented as:

[0046] H v =X v P v ,v=1,2,…,V,

[0047] in, Represents the original data of the v-th perspective; Represents the feature representation of the original data of the v-th view in the discriminant space after projection; P v Represents the discriminant matrix corresponding to the v-th view.

[0048] For example, in multi-view classification, the multi-view dataset is defined as:

[0049]

[0050] Among them, n represents the number of samples; each sample An instance containing multiple perspectives is represented as:

[0051]

[0052] Where V represents the total number of viewing angles and Represents the i-th sample of the v-th perspective, with a dimension of d v The eigenvector of .

[0053] The label corresponding to the feature vector of each sample is recorded as in is the set of all sample true value labels. The goal of MVC is to learn a model Given an input with unknown label By integrating all available perspectives, its label can be accurately predicted

[0054] Due to the differences in distribution and dimensionality of multi-view data, a strategy combining traditional multi-view discriminant analysis (MvDA) with deep learning is adopted to dynamically learn the multi-view discriminant matrix to align the multi-view data into a shared discriminant space.

[0055] Specifically, the MvDA method is first used to analytically solve the initial multi-view discriminant matrix P, which serves as the starting point for mapping multi-view data to the discriminant space; then, during the training process of the deep network, backpropagation is used to dynamically optimize the multi-view discriminant matrix P, so that it can adapt to network updates, thereby improving the modeling ability and discrimination performance of complex data relationships.

[0056] In traditional MvDA, by optimizing the generalized Rayleigh quotient problem, the present invention can jointly learn multiple view-specific linear transformation matrices to maximize the inter-class differences and minimize the intra-class differences, thereby achieving cross-view data alignment and retaining discriminant information.

[0057] The optimization equation of multi-view discriminant analysis can be expressed as:

[0058]

[0059] in, is the optimization variable, representing the projection matrix of the view union; S b is the inter-class scatter matrix, which is used to measure the compactness between samples of the same category. w is the intra-class scatter matrix, which is used to measure the separation between samples of different categories.

[0060] The above optimization equation is a generalized Rayleigh quotient problem. Since it has a closed-form solution, the initial discriminant matrix P can be directly obtained by solving the optimization problem.

[0061] Inter-class scatter matrix S b and the intra-class scatter matrix S w The calculation is as follows:

[0062]

[0063] in, represents the total number of samples of category c in the vth view, is the total number of samples of category c in all views, represents the mean of all samples of category c in all view angles, Defined as the mean of all samples in all view angles, Indicates the total number of samples.

[0064] Since the discriminant matrix solved by the traditional method relies on linear assumptions, it has significant limitations when processing high-dimensional and complex distributed data. For this reason, the present invention integrates the initial projection matrix into the network optimization framework to make it dynamically adaptable. By designing a specific loss function (such as perspective consistency loss), the network can dynamically adjust the projection matrix during the optimization process, so that it not only retains the discriminant ability, but also effectively reduces the feature differences between multiple perspectives, while adapting to the consistency of multi-perspective data. This optimization method expands the projection matrix from a static linear mapping to a dynamic feature learner, which can more accurately capture the essential characteristics of the data. The discriminant matrix update formula for each batch of data is as follows:

[0065]

[0066] Where η is the learning rate, is the total loss function.

[0067] The learned discriminant matrix is divided into rows according to the dimension of each perspective to obtain the discriminant matrix of each perspective:

[0068]

[0069] Among them, d c is the dimension of the common subspace.

[0070] Multi-view data in discriminative space It is expressed as follows:

[0071] H v =X v P v ,v=1,2,…,V,

[0072] in, Represents the original data of the v-th perspective, Represents the feature representation of the original data of the v-th view in the discriminant space after projection, P v The discriminant matrix corresponding to the v-th view.

[0073] The unified multi-view discriminative embedding module achieves feature alignment and distribution consistency across view data by projecting multi-view data with heterogeneous dimensions and significant feature differences into a unified discriminant space, providing more discriminative feature representation for the subsequent specific view evidence learning module.

[0074] In another implementation of the present invention, the perspective-specific evidence is represented as:

[0075]

[0076] Among them, g v (·) represents the evidence deep neural network of the v-th view, θv Represents the parameters of the v-th view network; is the feature vector in the discriminant space.

[0077] For example, traditional neural networks typically rely on the Softmax function to convert confidence into a probability distribution. However, this approach only provides a single-point estimate, which is difficult to reflect model uncertainty and easily leads to overconfidence. To address this limitation, the Softmax activation function is replaced with Softplus, and evidence-based deep neural networks (EvidentialDNNs) are constructed to generate non-negative outputs as the amount of evidence for each category. By parameterizing the amount of evidence as the Dirichlet parameter, a Dirichlet distribution is used to construct the category probability distribution, effectively modeling second-order uncertainty.

[0078] Specifically, given a feature vector after enhancing the consistency feature representation Obtaining view-specific evidence using view-specific evidential DNNs The formula is as follows:

[0079]

[0080] Among them, g v (·) represents the evidence deep neural network of the v-th view, θ v Represents the parameters of the v-th view network.

[0081] The Dirichlet density function is used to establish the category probability distribution, and the Dirichlet parameter is calculated by α=e+1 to ensure Dirichlet non-sparseness.

[0082] In another implementation of the present invention, the fused result is expressed as:

[0083]

[0084] Among them, W k is the final cross-view weight vector; Evidence matrix E k The jth evidence vector in .

[0085] For example, we introduce subjective logic theory into the framework, establish multiple opinions to quantify uncertainty, and evaluate the reliability of each perspective. Given a C-classification problem, for the instance of the v-th perspective Evidence Expressed as multiple opinions in, represents the belief quality vector, that is, the belief value assigned to each category based on the evidence; u is the uncertainty, which measures the gaps and missing evidence. Calculate the Dirichlet parameter α for a specific perspectivev =e v +1. The mapping between multiple opinions and Dirichlet distribution is as follows:

[0086]

[0087] in, is the Dirichlet length. The uncertainty u is inversely proportional to the total amount of evidence, indicating that as evidence increases, uncertainty decreases.

[0088] Multi-view classification requires integrating all available viewpoints to obtain a more reliable result, but the existence of conflicting viewpoints often leads to uncertainty in the integration process. To solve this problem, self-representation learning, l 1, Norm and self-supervised learning are used to learn a cross-view coefficient matrix to accurately identify consistent and conflicting perspectives. Since the model training adopts a mini-batch strategy, that is, the number of samples in each batch may be inconsistent, this makes it impossible for the cross-view relationship matrix to be globally iteratively updated with the entire network. To this end, a self-update mechanism is introduced for the cross-view relationship matrix of each batch, so that it can be dynamically adjusted based on the data of the current batch. Specifically, the cross-view coefficient matrix will be locally updated with the input features of each mini-batch, so as to effectively adapt to the changes in the number of samples in the current batch and the perspective relationship in different batches. At the beginning of each round, the learnable cross-view coefficient matrix is first initialized, and the L1, L2 and L3 loss functions are based on self-representation learning, l 1, Norms and self-supervised technique design.

[0089] n b The number of samples fed into the network for the current batch. Self-representation learning is used to model the dependencies between perspectives and a loss function is designed. as follows:

[0090]

[0091] in, is the evidence matrix composed of the i-th instance from all perspectives. i is the identity matrix, and a restriction term diag(Z i 0=0.

[0092] By minimizing Obtained the cross-view coefficient matrix This matrix effectively models the cross-view dependencies between samples in the current batch. However, due to Only focusing on minimizing feature reconstruction, it has limited ability to distinguish consistent and conflicting perspectives, which may cause conflicting perspectives to obtain a larger weight in the cross-perspective relationship matrix. To this end, we introduce l 1, norm and self-supervision mechanism to guide the learning of the cross-view relation matrix. i , get its pseudo label matrix Designed and It is expressed as follows:

[0093]

[0094] in, Represents the pseudo label vector of the i-th node and the j-th node in the k-th evidence matrix.

[0095] By minimizing To increase row sparsity, elements in each row compete with each other during the optimization process, retaining only a small number of significant non-zero values, thereby filtering out some weak connections. Furthermore, a self-supervised learning mechanism is introduced to guide the optimization of the weight matrix, so that the weight of each sample is primarily assigned to samples with the same label or higher similarity, ensuring that the cross-view weight distribution is more inclined towards credible information sources.

[0096] The final loss function for optimizing the cross-view relationship matrix is:

[0097]

[0098] Among them, α, β, and γ are all hyperparameters.

[0099] Calculate the evidence matrix E k The similarity between the i-th evidence vector and the j-th evidence vector in is expressed as follows:

[0100]

[0101] Among them, ReLU is the activation function to ensure that the similarity is non-negative.

[0102] Calculate the similarity matrix By the following formula:

[0103]

[0104] The similarity matrix and the learned cross-view relationship matrix Perform element-by-element dot multiplication to obtain the final cross-view weight matrix as follows:

[0105] W k =M k ⊙Zk

[0106] Evidence after integration Calculated by the following formula:

[0107]

[0108] The cross-perspective weight matrix accurately identifies consistent and conflicting perspectives and significantly enhances their dominant role in the fusion process by assigning primary weight to consistent perspectives. The final fusion result is more strongly driven by consistent perspectives, effectively weakening the influence of conflicting perspectives and ensuring a more stable and reliable fusion process.

[0109] A self-updating mechanism is implemented for the cross-view weight matrix within each batch, dynamically adjusting it based on the current batch data. Evidence from all views is aggregated based on the cross-view weight matrix, suppressing the influence of conflicting views. Compared to current methods, our aggregation paradigm is more efficient and conflict-resistant during cross-view fusion. Finally, a class-level representation learning loss is introduced to bring samples closer to their corresponding class prototypes while maximizing the distance between class prototypes, thereby enhancing the discriminative power of each view.

[0110] In another embodiment of the present invention, for a specific example The total loss function is expressed as:

[0111]

[0112] in, is the classification loss function; is the perspective consistency loss function; Represent the loss function for the class level.

[0113] For example, different loss functions are designed to improve the discriminative ability of each viewpoint and optimize the projection matrix P. Different loss functions are designed to meet different requirements: classification loss, viewpoint consistency loss, and class-level representation loss. Combining the class-level representation learning loss function brings samples closer to their corresponding class prototypes while simultaneously moving different class prototypes further away from each other, thereby sharpening class boundaries and enhancing the discriminative ability of each viewpoint.

[0114] In another implementation of the present invention, the classification loss function is expressed as:

[0115]

[0116] in, represents the corresponding Dirichlet parameter; λ t =min(1,t / T) is the annealing coefficient, and t represents the number of rounds of current training.

[0117] For example, since the Dirichlet distribution is introduced for each instance to model the category probability distribution and uncertainty, the traditional cross entropy loss function is no longer directly applicable. Therefore, an improved cross entropy loss function is introduced to better integrate evidence information and enhance the classification ability and robustness of the model. The improved cross entropy loss function is as follows:

[0118]

[0119] Where ψ(·) is the Digamma function, represents the prediction evidence vector, represents the corresponding Dirichlet parameter.

[0120] We hope that the evidence for the wrong class is at a low level, so we introduce an additional term into the loss function, namely Kullback-Leibler (KL) divergence:

[0121]

[0122] Among them, D(p n |1) is a uniform Dirichlet distribution, is obtained by taking the original parameter α n The Dirichlet parameter after the evidence of the true label is removed. t =min(1,t / T) is the annealing coefficient, where t represents the number of epochs in the current training and T represents the total number of annealing steps. As the training process progresses, the influence of the KL divergence is gradually increased to prevent premature convergence to a uniform distribution.

[0123] For the nth instance of the vth view with parameter a n The Dirichlet distribution of , the classification loss for a specific perspective is:

[0124]

[0125] In another implementation of the present invention, the view consistency loss function is expressed as:

[0126]

[0127] Among them, V represents the total number of viewing angles; is the evidence matrix composed of the i-th instance from all perspectives; is the evidence matrix composed of the j-th instance from all perspectives.

[0128] For example, in order to eliminate the heterogeneity between perspectives and encourage the evidence vectors generated by different perspectives for the same instance to be consistent, a perspective consistency constraint is introduced:

[0129]

[0130] in, The evidence matrix composed of the i-th instance from all perspectives

[0131] In another implementation of the present invention, the class-level representation loss function is expressed as:

[0132]

[0133] in, is the intra-class consistency loss; is the inter-class separation loss.

[0134] For example, in order to ensure that each perspective can accurately extract evidence of a specific category and increase the discriminability of each category, a class-level representation loss function is designed. Specifically, the present invention aims to make samples of the same category clustered tightly in the feature space, while maintaining a clear separation between samples of different categories. However, directly calculating intra-class and inter-class losses on multi-view data will lead to a significant increase in computational complexity. To this end, a prototype-based method is introduced. By determining the central prototype of each category, the samples are clustered towards the center of the category to which they belong, while ensuring that the centers of different categories maintain a sufficient distance, thereby reducing the computational complexity while improving the discriminative ability between categories. It is the intra-class consistency loss. It aims to minimize the average distance between the samples in each category and the center of the category under all views, which is expressed as follows:

[0135]

[0136] Among them, C is the total number of categories, M c is the number of samples in category c, The set of samples belonging to category c among all samples of the v-th perspective, c c represents the center of category c, which is defined as:

[0137]

[0138] is the inter-class separation loss. It aims to maximize the distance between the centers of different categories and is expressed as follows:

[0139]

[0140] in, is the total number of category pairs. The class level represents the learning loss function It can be defined as:

[0141]

[0142] The class-level representation learning loss function strengthens the discernibility of each view's category representation and, through multi-view collaborative optimization, ensures the consistency and completeness of category information across all views. This effectively reduces confusion between categories, thereby improving the classification performance and reliability of each view in complex data scenarios.

[0143] In another implementation of the present invention, the proposed multi-view credible classification method based on subspace evidence fusion was benchmarked against 10 state-of-the-art methods on six real-world datasets. Furthermore, the reliability and accuracy of each view were evaluated, and the effectiveness of the cross-view weight matrix in identifying conflicting views was analyzed.

[0144] Specifically, as shown in Table 1, for the normal dataset, PIE contains 680 samples divided into 68 categories. Three different perspectives are extracted from intensity, LBP, and Gabor features. The HandWritten dataset includes 2,000 handwritten digit samples ranging from 0 to 9, with 200 samples per category. Each sample is described by six sets of features: Fourier transform (FOU), contour correlation (FAC), Karhunen–Loeve coefficient (KAR), pixel average (ZER), Zernike moment (ZER), and morphological feature (MOR).

[0145] The Scene15 dataset contains 4,485 images distributed across 15 indoor and outdoor scene categories, and the extracted features include GIST, PHOG, and LBP. The ALOI dataset contains 10,158 images of small objects, covering a total of 1,000 categories, and provides four sets of deep feature representations. The NUS-WIDE-OBJECT (NUS) dataset contains 30,000 images covering 31 categories. Each sample is represented by features from five different perspectives, including color histogram, block-wise color moments, color correlogram, edge direction histogram, and wavelet texture.

[0146] For the conflict dataset, we followed the method described by Professor Xu Si's team to construct a conflict variant of the aforementioned dataset. Each experimental method was evaluated 10 times, and the average accuracy and standard deviation were recorded.

[0147] Table 1: Dataset overview

[0148]

[0149] The comparison methods include baseline methods based on feature fusion and decision fusion. Specifically, the baseline methods based on feature fusion include:

[0150] (1) EDL (Evidential Deep Learning) uses Dirichlet distribution to explicitly model uncertainty and capture the confidence of predictions to achieve robust classification.

[0151] (2) DCCAE (Deep Canonically Correlated AutoEncoders) is a classic feature fusion method that uses autoencoders to learn shared representations of multiple views.

[0152] (3) CPM-Nets (Cross Partial Multi-view Networks) is an advanced feature fusion framework that handles the complex associations of multi-view data by learning diverse representations applicable to different datasets.

[0153] (4) DUA-Nets (Dynamic Uncertainty-Aware Networks) is an uncertainty-based framework that forms a unified feature representation by dynamically integrating intrinsic information from different perspectives.

[0154] Baseline methods based on decision fusion include:

[0155] (1) TMC (Trusted Multi-view Classification) is an uncertainty-aware classification method that aims to estimate the uncertainty of multi-view data to provide reliable and robust classification results.

[0156] (2) TMDL-OA (Trusted Multi-View Deep Learning with Opinion Aggregation) is a decision fusion method based on evidence-based deep neural networks. It introduces consistency metric loss to ensure credible learning and reliable predictions.

[0157] (3) RMVC (Robust Multi-view Classification) is a robust classification method that uses evidence inconsistency measurement to evaluate view quality and enhances the model's robustness against adversarial attacks through inconsistency-aware belief integration.

[0158] (4) CCML (Consistent and Complementary-aware trusted Multi-view Learning) adopts a dynamic evidence decoupling strategy to separate consistency evidence and complementary evidence to effectively deal with the semantic ambiguity problem in multi-view data.

[0159] (5) ECML (Evidential Conflictive Multi-view Learning) is a method specifically designed for reliable conflicting multi-view learning. It constructs perspective-specific opinions and integrates them using conflict strategies.

[0160] (6) TUNED (Trusted Unified Feature-Neighborhood Dynamics) combines the neighborhood structure with the feature structure and introduces the selective Markov random field (S-MRF) to effectively handle conflicts in multi-view classification.

[0161] As shown in Tables 2 and 3, the best and second best results are highlighted by bold and underline respectively, and Δ% represents the performance improvement of the present application over the best baseline. NUS refers to “NUS-WIDE-OBJECT”.

[0162] (1) On the normal test set, except for the PIE dataset, the multi-view credible classification based on subspace evidence fusion performs well compared with all other baseline methods. Notably, on the ALOI and NUS datasets, the multi-view credible classification based on subspace evidence fusion significantly surpasses the second-ranked baseline method, with improvements of 10.29% and 6.64%, respectively. This result can be attributed to its effectiveness in enhancing the feature representation of a single sample and improving the ability to extract cross-view evidence.

[0163] (2) On the conflict test set, although the accuracy of all compared methods dropped significantly, the multi-view credible classification based on subspace evidence fusion still performed well on all datasets, especially on the Scene15, ALOI, and NUS datasets, achieving significant improvements of 10.87%, 10.86%, and 10.33%, respectively. This shows that the learned cross-view weight matrix can effectively identify consistent and conflicting views and assign higher weights to consistent views, thereby significantly reducing the impact of conflicting views on decision making.

[0164] In addition, the accuracy of the conflict test set of the ALOI dataset is higher than that of its normal test set, which may be related to the sensitivity of the parameters. In summary, the multi-view credible classification based on subspace evidence fusion shows excellent performance in classification tasks and has excellent conflict resistance in the decision-making process.

[0165] Table 2: Accuracy on the normal test set (%)

[0166]

[0167] Table 3: Accuracy on the conflict test set (%)

[0168]

[0169]

[0170] This application compares the single-view accuracy of multi-view credible classification based on subspace evidence fusion with RCML and TUNED on the HandWritten, Scene and ALOI datasets under normal and conflict conditions, respectively. Figure 4 As shown in the figure, “fusion” refers to the accuracy after integrating all perspectives. From left to right, the methods are TUNED, RCML and Ours.

[0171] As can be seen, multi-view credible classification based on subspace evidence fusion achieves the highest accuracy across all viewpoints. Particularly noteworthy is that in the ALOI dataset, regardless of the presence of conflict, multi-view credible classification based on subspace evidence fusion significantly outperforms other methods across all individual viewpoints, resulting in overall accuracy improvements of 10.29% (normal test set) and 10.33% (conflict test set), respectively.

[0172] These results indicate that multi-perspective credible classification based on subspace evidence fusion not only improves the discriminative ability of each perspective, enabling it to accurately extract classification-oriented perspective-specific evidence, but also promotes cross-perspective fusion more effectively.

[0173] In order to verify whether the learned cross-view weight matrix can accurately identify consistent and conflicting perspectives and assign optimal weights, this paper visualizes the cross-view weight matrix W under normal and conflicting conditions on the HandWritten and ALOI datasets, as shown in the following figure: Figure 5 As shown, parts (a) and (b) in the figure correspond to the HandWritten and ALOI datasets respectively. On the left side of (a) and (b), the visualization of the cross-view weight matrix W of the HandWritten and ALOI test sets under normal conditions is shown, and the right side shows the visualization results of each test set under conflict conditions. Under conflict conditions, the 6th view in the HandWritten dataset and the 3rd view in the ALOI dataset are set as conflicting views.

[0174] The following points were observed:

[0175] When all views describe the same instance, i.e., view consistency, the cross-view weight matrix assigns a non-zero weight to each view, as shown in (a) left and (b) left.

[0176] When a conflict occurs, the cross-perspective weight matrix (as shown in Figures (a) right and (b) right) can effectively identify consistent and conflicting perspectives and assign most weights to the consistent perspective, thereby effectively suppressing the influence of conflicting perspectives on decision-making.

[0177] Experimental results on six real-world datasets demonstrate that our method consistently outperforms the best baseline method in terms of accuracy and reliability. In particular, on the ALOI dataset, our method achieved classification accuracies of 98.08% and 98.10% under normal and conflict conditions, respectively, outperforming the best baseline method by 10.29% and 10.86%, respectively.

[0178] It should be understood that obtaining reliable results through cross-view fusion involves two key points:

[0179] The first key point is the reliability of each perspective, specifically, whether each perspective can effectively extract evidence related to a specific category.

[0180] The second is whether the fusion strategy takes into account the correlation between all perspectives.

[0181] However, existing methods ignore the first part, resulting in insufficient feature representation capabilities for each perspective. Furthermore, cross-perspective evidence fusion strategies in current methods fail to fully consider the relationship between perspectives, making it difficult to capture conflicting perspectives. Once conflicting perspectives emerge, uncertainty in the fusion process increases, ultimately leading to a breakdown in the decision-making process.

[0182] Therefore, the present invention focuses on improving the discriminative ability of each individual perspective and developing a robust cross-perspective fusion strategy to further promote the reliability of decision-making in conflicting situations. Specifically, the method of the present invention includes two key aspects: first, by designing a perspective-specific feature enhancement module, the discriminative information in each perspective is fully mined to improve the classification performance of the model under a single perspective; second, a fusion mechanism based on a cross-perspective weight matrix is introduced to effectively balance the information differences between different perspectives and reduce the negative impact of conflicting perspectives on decision-making results. Experimental results show that the method of the present invention has achieved significant performance improvements on multiple benchmark datasets, especially in processing complex scenarios with conflicting information, demonstrating superior reliability and robustness.

[0183] Another aspect of the present invention, as Figure 3 As shown, a multi-perspective credible classification system based on subspace evidence fusion is provided, including:

[0184] Unified Multi-view Discriminative Embedding (UMDE) module: used to map multi-view data into a unified discriminant space by constructing a learnable discriminant matrix.

[0185] View-specific Evidence Learning (VSEL) module: Inputs the feature matrix in the discriminant space into the view-specific evidence deep neural network to extract view-specific evidence.

[0186] Cross-view Evidence Fusion (CVEF) module: Aggregates all view-specific evidence based on the cross-view weight matrix to generate a fused result.

[0187] The multi-perspective credible classification system based on subspace evidence fusion of the present invention dynamically learns a multi-perspective discriminant matrix by combining multi-perspective discriminant analysis with deep learning, and then aligns samples across perspectives and maintains discriminant information; the learned cross-perspective weight matrix can effectively distinguish consistent and conflicting perspectives, assign optimal weights to each perspective, effectively balance the information differences between different perspectives, reduce the negative impact of conflicting perspectives on decision results, enhance the stability and reliability of the multi-perspective data fusion process, and further improve the reliability of decision-making.

[0188] In another aspect of the present invention, an electronic device includes a processor, a memory, a communication bus, and a communication interface.

[0189] in:

[0190] The processor, memory and communication interface communicate with each other through a communication bus.

[0191] Communication interface, used to communicate with other electronic devices or servers.

[0192] The processor is used to execute the program, and specifically can execute the steps of any one of the multi-view credible classification methods based on subspace evidence fusion in the above embodiments.

[0193] Specifically, the program may include program codes including computer operation instructions.

[0194] The processor may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application. The one or more processors included in the smart device may be processors of the same type, such as one or more CPUs; or different types of processors, such as one or more CPUs and one or more ASICs.

[0195] Memory is used to store programs. The memory may include high-speed RAM memory, and may also include non-volatile memory (non-volatile memory), such as at least one disk storage.

[0196] The program can be specifically used to enable the processor to execute the steps to implement any one of the multi-perspective credible classification methods based on subspace evidence fusion described in the embodiments. The specific implementation of each step in the program can refer to the corresponding descriptions in the steps and units executed by any one of the multi-perspective credible classification methods based on subspace evidence fusion in the above steps, and will not be repeated here. Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working process of the above-described devices and modules can refer to the corresponding process description in the aforementioned method embodiment.

[0197] The above method according to the embodiment of the present invention can be implemented in a server equipped with a central processing unit (CPU) and an image processing unit (COU).

[0198] Thus far, specific embodiments of the present invention have been described. Other embodiments are within the scope of the appended claims. In some cases, the actions recited in the claims can be performed in a different order and still achieve the desired results. Additionally, the processes depicted in the accompanying drawings do not necessarily require the specific order shown, or sequential order, to achieve the desired results.

[0199] It should be noted that all directional indications in the embodiments of the present invention (such as up, down, left, right, back, etc.) are only used to explain the relative position relationship between the components in a certain specific order (as shown in the accompanying drawings). If the specific order changes, the directional indication will also change accordingly.

[0200] In the description of the present invention, the terms "first" and "second" are used solely to facilitate description of different components or names and should not be construed as indicating or implying a sequential relationship, relative importance, or implicitly specifying the quantity of the technical features being described. Therefore, features specified as "first" or "second" may explicitly or implicitly include at least one of such features.

[0201] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0202] It should be noted that although the specific embodiments of the present invention are described in detail in conjunction with the accompanying drawings, this should not be construed as limiting the scope of protection of the present invention. Within the scope described by the claims, various modifications and variations that can be made by those skilled in the art without creative effort still fall within the scope of protection of the present invention.

[0203] The examples of the embodiments of the present invention are intended to briefly illustrate the technical features of the embodiments of the present invention so that those skilled in the art can intuitively understand the technical features of the embodiments of the present invention, and are not intended to improperly limit the embodiments of the present invention.

[0204] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A multi-view credible classification method based on subspace evidence fusion, characterized by: include: By constructing a learnable discriminant matrix, multi-view data is mapped to a unified discriminant space; Inputting the feature matrix in the discriminant space into a view-specific evidence deep neural network to extract view-specific evidence; All view-specific evidence is aggregated based on the cross-view weight matrix to generate a fused result.

2. The method according to claim 1, characterized in that The multi-view data in the discriminant space is expressed as: H v =X v P v ,v=1,2,…,V, in, Represents the original data of the v-th perspective; Represents the feature representation of the original data of the v-th view in the discriminant space after projection; P v Represents the discriminant matrix corresponding to the v-th view.

3. The method according to claim 1, characterized in that The perspective-specific evidence is represented by: Among them, g v (·) represents the evidence deep neural network of the v-th view, θ v Represents the parameters of the v-th view network; is the feature vector in the discriminant space.

4. The method according to claim 3, characterized in that The view-specific total loss function is expressed as: in, is the classification loss function; is the perspective consistency loss function; Represent the loss function for the class level.

5. The method according to claim 4, characterized in that The classification loss function is expressed as: in, represents the corresponding Dirichlet parameter; λ t =min(1,t / T) is the annealing coefficient, and t represents the number of rounds of current training.

6. The method according to claim 4, characterized in that The perspective consistency loss function is expressed as: Among them, V represents the total number of viewing angles; is the evidence matrix composed of the i-th instance from all perspectives; is the evidence matrix composed of the j-th instance from all perspectives.

7. The method according to claim 4, characterized in that The class-level representation loss function is expressed as: in, is the intra-class consistency loss; is the inter-class separation loss.

8. The method according to claim 1, characterized in that The fusion result is expressed as: Among them, W k is the final cross-view weight vector; Evidence matrix E k The jth evidence vector in .

9. A multi-perspective credible classification system based on subspace evidence fusion, characterized by: include: Unified multi-view discriminant embedding module: used to map multi-view data into a unified discriminant space through the constructed discriminant matrix; Viewpoint-specific evidence learning module: inputs the feature matrix in the discriminant space into a viewpoint-specific evidence deep neural network to extract viewpoint-specific evidence; Cross-view evidence fusion module: Aggregates all view-specific evidence based on the cross-view weight matrix to generate a fused result.

Images