Privileged learning classification method based on category perception and explainability guidance

Abstract

This invention discloses a privileged learning classification method based on category awareness and interpretability guidance, belonging to the field of artificial intelligence and machine learning. The method includes: first, dividing the training data into a privileged set containing complete information and a non-privileged set containing only source domain features; for a specific category, extracting feature importance through interpretability analysis and sparsifying it to generate a category-specific weight vector; then, calculating the weighted similarity between non-privileged samples and similar samples in the privileged set based on this weight vector, retrieving and transferring the privileged features of the best-matching sample as reconstructed features, and dynamically assigning confidence coefficients negatively correlated with similarity; finally, fusing real and reconstructed privileged features to construct a joint optimization objective function with an adaptive penalty mechanism for confidence coefficients for model training. This invention effectively improves the classification accuracy and robustness of the model in scenarios where some privileged information is missing.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and machine learning, and in particular relates to a privileged learning classification method based on category awareness and interpretability guidance. Background Technology

[0002] In the fields of artificial intelligence and machine learning, data classification is one of the core tasks. With the development of data acquisition technologies, the Learning Using Privileged Information (LUPI) framework has been widely applied. This framework utilizes "privileged information" (such as clearer images, additional text descriptions, etc.) that is only visible during the training phase but not during the testing phase to assist model training, thereby improving the classification performance of the main task. Existing technical solutions typically employ Support Vector Machine Addition (SVM+) or privileged information distillation methods based on deep neural networks. These methods usually use privileged information as a regularization term or auxiliary supervision signal, correcting the decision boundary by minimizing the feature differences between the privileged space and the original space.

[0003] However, the aforementioned existing technologies have the following major drawbacks in practical applications: First, existing methods typically treat privileged information as a whole, failing to distinguish the importance of different feature dimensions within the privileged information. When the privileged information contains noisy or irrelevant features, this "whole-whole acceptance" approach can interfere with model training. Second, existing methods usually employ a uniform similarity metric or regularization strategy for all categories, ignoring the differences in feature space distribution among samples from different categories, resulting in poor generalization ability when dealing with certain hard-to-distinguish or minority class samples. Third, in scenarios where some privileged information is missing, existing methods often treat the imputed "pseudo-data" as equal to the real "data." If the imputation is incorrect, it can severely mislead the delineation of decision boundaries, lacking an effective fault tolerance mechanism for low-quality reconstructed information. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a privileged learning classification method based on category awareness and interpretability guidance, comprising: Obtain a training dataset, which includes a first training subset and a second training subset. Each sample in the first training subset contains source domain features, privileged domain features, and class labels. Each sample in the second training subset contains only source domain features and class labels. Based on the samples in the first training subset, determine the feature importance weight vector corresponding to the category; Based on the feature importance weight vector, calculate the weighted similarity between non-privileged samples in the second training subset and similar samples in the first training subset; Based on the weighted similarity, target samples are matched for the non-privileged samples from the first training subset, and the privileged domain features of the target samples are assigned to the non-privileged samples as reconstructed privileged features. Based on the weighted similarity, determine the confidence coefficient corresponding to the reconstructed privileged feature; Based on the first training subset, the second training subset carrying the reconstructed privileged features and the confidence coefficients, a joint optimization objective function is constructed and the model is trained. The confidence coefficients are used to dynamically adjust the penalty intensity of the reconstructed privileged features on model training within the joint optimization objective function.

[0005] Optionally, based on the samples in the first training subset, a feature importance weight vector corresponding to the category is determined, including: For the target category, interpretability analysis is performed on the source domain samples belonging to the target category in the first training subset to obtain an initial feature importance vector; The initial feature importance vector is normalized to obtain a normalized importance vector; The normalized importance vector is sparsified to generate a sparse feature importance weight vector corresponding to the target category.

[0006] Optionally, the normalized importance vector is sparsified to generate a sparse feature importance weight vector corresponding to the target category, including: Based on a preset screening ratio, dimensions whose importance scores meet preset conditions are selected from the normalized importance vector. The weight values ​​of the selected dimensions in the normalized importance vector are retained, and the weight values ​​of the remaining dimensions are set to zero to generate the sparse feature importance weight vector.

[0007] Optionally, based on the feature importance weight vector, the weighted similarity between non-privileged samples in the second training subset and similar samples in the first training subset is calculated, including: Obtain the source domain features of the non-privileged samples and the feature importance weight vector of their respective categories; Based on the feature importance weight vector, a weighted cosine similarity is calculated between the source domain features of the non-privileged sample and the source domain features of similar samples in the first training subset to obtain the weighted similarity.

[0008] Optionally, based on the weighted similarity, the confidence coefficient corresponding to the reconstructed privileged feature is determined, including: The confidence coefficient is calculated using a nonlinear mapping function based on the weighted similarity and a preset first adjustment parameter, wherein the confidence coefficient is negatively correlated with the weighted similarity.

[0009] Optionally, the confidence coefficient is calculated using the following formula: ; in, The first adjustment parameter is... The weighted similarity is denoted as .

[0010] Optionally, the joint optimization objective function includes a first error term corresponding to the first training subset and a second error term corresponding to the second training subset, wherein the second error term is constructed based on the reconstructed privileged features and their corresponding confidence coefficients.

[0011] Optionally, in the joint optimization objective function, the penalty terms of the model parameters related to the reconstructed privileged features are weighted by the confidence coefficient.

[0012] On the other hand, the present invention also provides an electronic device including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0013] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0014] Compared with the prior art, the present invention has the following advantages and technical effects: First, this invention extracts feature contribution by introducing an interpretability-guided mechanism and generates sparse category-specific importance weight vectors using a threshold truncation mechanism. This accurately filters out noisy dimensions that are not beneficial to the current category classification, allowing the model to focus on core features with high discriminative power, thus improving the purity of data processing and the final classification accuracy from the source. Second, this invention proposes a category-aware weighted similarity measurement method, directly embedding the sparse weight vector into the distance metric formula. This ensures that only sample pairs that are highly consistent in the core semantic dimension are judged as "similar," greatly improving the accuracy of retrieving matching samples between the source domain and the privileged domain. Finally, this invention designs a non-linear adaptive confidence penalty mechanism in the joint optimization objective function, dynamically binding the penalty weight of the reconstructed features to its weighted similarity. This allows the model to fully utilize some privileged information to improve performance while effectively resisting the interference of reconstruction noise, exhibiting strong engineering robustness. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall architecture of an autonomous driving embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the extraction of the features that have the greatest impact on the classification results in an embodiment of the present invention. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0018] Example 1 This embodiment provides a privileged learning classification method based on category awareness and interpretability guidance, including: like Figure 1 As shown, a training dataset is obtained, which includes a first training subset and a second training subset. Each sample in the first training subset contains source domain features, privileged domain features, and class labels, while each sample in the second training subset contains only source domain features and class labels. Based on the samples in the first training subset, determine the feature importance weight vector corresponding to the category; Based on the feature importance weight vector, calculate the weighted similarity between non-privileged samples in the second training subset and similar samples in the first training subset; Based on the weighted similarity, target samples are matched for the non-privileged samples from the first training subset, and the privileged domain features of the target samples are assigned to the non-privileged samples as reconstructed privileged features. Based on the weighted similarity, determine the confidence coefficient corresponding to the reconstructed privileged feature; Based on the first training subset, the second training subset carrying the reconstructed privileged features and the confidence coefficients, a joint optimization objective function is constructed and the model is trained. The confidence coefficients are used to dynamically adjust the penalty intensity of the reconstructed privileged features on model training within the joint optimization objective function.

[0019] The specific process includes: First, the training dataset is obtained and divided into two disjoint subsets: a first training subset and a second training subset. Each sample in the first training subset contains complete source domain features, privileged domain features, and class labels; each sample in the second training subset contains only source domain features and class labels, lacking privileged domain features.

[0020] Next, for a specific category in the training dataset, an initial feature importance vector of the source domain data for that category is extracted using a pre-defined interpretability analysis algorithm (e.g., the Locally Interpretable Model-Unknown Interpretation Algorithm, LIME). The absolute value of the initial feature importance vector is taken and normalized to obtain a normalized importance vector representing the feature contribution. Furthermore, a threshold truncation mechanism is used to sparsify the normalized importance vector, generating a sparse category-specific importance weight vector. This is done to retain the core feature dimensions with high discriminative power and suppress the weights of irrelevant or noisy feature dimensions to zero.

[0021] Specifically, the sparse category-specific importance weight vector The calculation formula is: ; in, Indicates category Normalized Dimensional feature importance score, The importance score indicates the top ranking The set of feature dimensions, This represents the initial feature importance score vector for a specific category c (e.g., c=1 represents a high-risk obstacle). This step preserves the weights of features with high discriminative power and suppresses the weights of irrelevant or noisy features to 0.

[0022] Then, for each non-privileged sample in the second training subset, a feature reconstruction operation is performed, specifically including the following steps: First, the weighted retrieval step. This involves introducing a category-specific importance weight vector corresponding to the category to which the non-privileged sample belongs. Calculate the weighted similarity between the non-privileged sample and similar samples in the first training subset.

[0023] Second, the feature transfer step. In the first training subset, the best matching sample with the highest weighted similarity is retrieved, and the privileged domain features of the best matching sample are transferred to the non-privileged sample as reconstructed privileged features.

[0024] Third, the confidence calculation step. The weighted similarity score between the non-privileged sample and the best-matching sample is used as the confidence coefficient of the reconstructed privileged feature. .

[0025] Specifically, in weighted retrieval, the non-privileged samples Similar samples to those in the first training subset The formula for calculating the weighted similarity between them is: ; in, Indicates weighted similarity. This represents the source domain feature vector of the target sample (i.e., the sample with missing privileged information). This represents the source domain feature vector of the candidate matching sample (i.e., the historical / source sample containing complete privileged information). Indicates the category to which the target sample belongs. A sparse importance weight vector. This vector is extracted using an interpretable algorithm to strengthen feature dimensions that play a key role in classification decisions and suppress redundant or noisy features. This represents the total number of dimensions of the sample feature vectors, where k is the index of the feature dimension, ranging from 1 to d. Indicates target sample The feature value in the k-th dimension, Indicates candidate samples The feature value in the k-th dimension, Represents the weight vector The weight value in the k-th dimension.

[0026] In confidence level assignment, the confidence coefficient The calculation formula is: in, The first preset adjustment parameter (i.e., the adjustment weight of real privileged information) is used. This formula constructs a non-linear adaptive adjustment mechanism: by introducing weighted similarity into the denominator, the confidence coefficient is... The value of increases as similarity decreases. In subsequent joint optimization, the model will assign a larger ... The value is increased to enhance the penalty, thereby effectively shielding the negative interference of reconstruction noise on the decision boundary of the main task.

[0027] Finally, the first training subset is merged with the second training subset that has completed the privileged information reconstruction to construct a unified training dataset containing the reconstructed privileged features and their corresponding confidence coefficients. Based on this unified training dataset, a joint optimization objective function with an adaptive penalty mechanism is constructed. In this objective function, the confidence coefficients are used to optimize the performance of the reconstructed privileged features. The dynamic adjustment of the impact of reconstructed privileged features on the model's error tolerance makes reconstructed features with high confidence (i.e., high similarity) more effective in guiding the decision boundary and enhances the model's robustness to low-confidence reconstructed noise.

[0028] Specifically, the joint optimization objective function with adaptive penalty mechanism and its constraints are as follows: ; The optimization problem satisfies the following constraints: ; ; ; ; in, and The weight vector and bias of the source domain data; and The weight vector and bias for the privileged domain data; For regularization parameters; The number of samples in the first training subset. This represents the number of samples in the second training subset; This is the first adjustment parameter, used to control the impact of real privileged information on the model; This is the second adjustment parameter. ( ) represents the similarity propagation algorithm. Representation similarity transitive algorithm All parameters in parentheses are used to control the impact of reconstructing privileged information on the model; the... The value of is dynamically determined by the weighted similarity between the non-privileged sample and the retrieved privileged sample; the higher the similarity, the higher the value. The smaller the value, the stronger the guiding role of reconstructed privileged features with high similarity in the model's classification boundary.

[0029] This embodiment provides a similarity transfer algorithm. ( ), used to reconstruct privileged information and assign confidence weights to non-privileged samples in scenarios where some privileged information is missing. The specific execution process is as follows: First, define the input and output data for the algorithm.

[0030] Input data includes: a non-privileged dataset lacking privileged information. A privileged dataset containing complete privileged information ; and category label set The output data is an augmented dataset with confidence weights. .

[0031] The specific execution process of the algorithm is divided into two stages.

[0032] Phase 1: Feature Importance Extraction; For category label set Each specific category in Perform the following operations: First, from the privileged dataset Extracting categories Data subset .

[0033] Subsequently, the importance vector of aggregated locally interpretable model unexplained (LIME) features for this subset is calculated using an interpretability algorithm. .

[0034] Furthermore, applying this importance vector The threshold filtering mechanism truncates redundant features to obtain a sparse class-specific mask weight vector. .

[0035] Phase Two: Weighted Imputation and Confidence Calculation Initialize the final augmented dataset .

[0036] The first step is to perform complete data initialization. This involves iterating through the privileged dataset. All samples (index) to The confidence weights of the real privileged samples are set to a fixed hyperparameter, i.e. and the sample Add directly to augmented dataset middle.

[0037] The second step is to execute a missing data imputation loop, iterating through the non-privileged dataset. All samples (index) to ), perform the following sub-operations: In the weighted similarity retrieval sub-operation, the current sample is identified. Category tags And retrieve the corresponding mask weight vector. Calculate the target sample Calculate the weighted cosine similarity between the sample and similar samples in the privileged dataset, and find the index of the best matching sample with the highest similarity. The calculation formula is as follows: ; In the privileged information transfer and confidence allocation sub-operation, the privileged features of the best-matching sample are transferred to the target sample to complete the reconstruction imputation: Simultaneously, based on the weighted similarity value of the best match, an adaptive penalty coefficient is calculated for the reconstructed feature using a nonlinear function. The calculation formula is as follows: ; In the dataset update sub-operation, new samples will be completed with privileged information reconstruction and confidence assignment. Add to augmented dataset middle.

[0038] After the loop ends, the algorithm outputs the constructed augmented dataset. This is used for training subsequent joint optimization models (such as support vector machine plus (SVM+) solvers).

[0039] Example 2 This embodiment provides a privileged learning classification method based on category awareness and interpretability guidance, including: In this embodiment, the mathematical calculation object of the present invention is the real physical signal collected by the sensors of the autonomous vehicle, and the specific meanings of the variables are as follows: Source domain feature vector This represents the physical features of the road ahead image captured by the vehicle's 2D vision camera. Specifically, the forward field of view is divided into... Each grid area This is the optical feature vector containing the edge gradients and textures of these 81 regions.

[0040] Privileged domain feature vector : Represents the spatial depth physical features collected by the vehicle-mounted 3D LiDAR. This corresponds to the mean vector of laser reflection point cloud density and depth distance for the aforementioned 81 fields of view. This sensor is expensive and is only installed on some advanced data acquisition vehicles, constituting "missing privileged information."

[0041] Category Tags : Indicates the hazard category of obstacles in the physical space ahead. (1 represents high-risk dynamic obstacles such as pedestrians / vehicles, -1 represents a safe static environment).

[0042] Decision function The classification results output by the model are directly converted into electrical control commands for the vehicle's underlying electronic control unit (ECU) to trigger the braking hardware.

[0043] Implementation steps based on specific physical scenarios: The control method in this embodiment operates in the vehicle's domain controller and strictly corresponds to the above technical solution, including the following steps: Step S1: Sensor dataset construction and partitioning; The fleet acquired real-world scenario datasets through road tests and divided them into two subsets: First training subset (Advanced Data Acquisition Vehicle data, total) Each scene): contains complete (visual features) LiDAR characteristics Hazard categories ).

[0044] Second training subset (data from ordinary mass-produced vehicles, total) (Single scenario): Due to hardware cost limitations, only (visual features) are included. Hazard categories (The depth features of the lidar are missing.)

[0045] Step S2: Interpretability-guided visual feature sparsification; against For this specific hazard category (i.e., high-risk obstacles), the onboard computing unit uses an interpretability analysis algorithm to extract the physical contribution of 81 visual grids to "obstacle identification". .

[0046] To eliminate physical interference from background meshes such as the sky and road surface on subsequent similarity calculations, a preset Top- (e.g., Top-40%) Threshold truncation mechanisms generate sparse class-specific importance weight vectors. : ; in, Indicates the first The optical weights of each visual grid are determined by this formula. The system retains only the grid weights covering the core physical contours of pedestrians or vehicles, forcing the weights of irrelevant background grids to zero.

[0047] Step S3: Reconstruction and confidence assignment of missing LiDAR depth information; For ordinary vehicle scenarios where LiDAR data is missing in the second training subset Introducing the aforementioned visual weights Search for the most similar historical advanced vehicle scene in the first training subset. The system calculates the physical visual similarity between the two in the core contour region using the following formula: ; Find the best matching scenario with the highest similarity Subsequently, the system transfers the actual 3D LiDAR depth features of the scene to ordinary vehicles as the reconstructed privileged features. .

[0048] Meanwhile, to ensure the driving safety of autonomous driving and prevent erroneous braking caused by incorrect depth interpolation, the system assigns an adaptive penalty coefficient (confidence level) based on visual similarity. : ; Step S4: Constructing a unified dataset and adaptive joint optimization By aggregating all the vehicle scenario data mentioned above, the domain controller constructs a joint optimization objective function with an adaptive penalty mechanism: ; in, This item plays a core role in security control. When visual matches are extremely dissimilar ( (approaching -1) The penalty coefficient increases sharply. When defining the decision boundary, the Optimization Solver (QP) will force the model to ignore these poorly reconstructed fake LiDAR data, thus ensuring the model's robustness to environmental noise.

[0049] Step S5: Model-level control and inference application; The main task decision function obtained by solving It is deployed in the onboard chips of ordinary mass-produced vehicles. In actual road driving, the vehicle only needs to use a regular 2D camera to collect optical features ahead in real time. The input is given to the decision function. When the prediction result... When the obstacle is a high-risk obstacle, the system immediately generates a high-level trigger signal to drive the vehicle to perform deceleration or emergency braking (AEB) operation, thus completing closed-loop control.

[0050] like Figure 2 The diagram shown illustrates the overall data flow and framework of an autonomous driving obstacle recognition and braking control system according to an embodiment of the present invention. The specific physical meanings of each variable node and processing stage in the diagram are as follows: 1. The meaning of the mapping of physical variable nodes: Square node ( ): Represents source domain features. In autonomous driving scenarios, this corresponds to the physical feature vector of the image captured by the front-facing 2D vision camera in a regular mass-produced vehicle (such as the edge and texture features of the 81 grid regions divided in the front field of view).

[0051] Circular node ( ): Represents the privileged domain features. Corresponds to the spatial depth and point cloud density feature vectors acquired by the 3D LiDAR (LiDAR) on the advanced data acquisition vehicle.

[0052] Rhombus node ( ): Represents a category label. It corresponds to the actual hazard category of obstacles in the physical space in front of the autonomous driving system (such as high-risk dynamic obstacles or safe static environment).

[0053] 2. Explanation of the data flow stages in the control system: Phase ① (Sensor Data Segmentation): The left side of the diagram shows the separation status of the road test data. The top represents the first training subset collected by the advanced data acquisition vehicle. Each scene possesses a complete (2D visual) 3D LiDAR Hazard categories The image below represents the second training subset collected from ordinary mass-produced vehicles. (In one scenario), limited by hardware costs, only visual features are available. and tags Its lidar depth features are missing.

[0054] Phase ② (Spatial Depth Physical Information Reconstruction): For ordinary mass-produced vehicle scenarios lacking LiDAR data below, the vehicle domain controller performs the weighted similarity retrieval proposed in this invention. By searching for historical scenes with the most similar visual core contours in the advanced vehicle scene library above, its 3D LiDAR depth data is extracted and knowledge transferred to generate reconstructed depth features (circular nodes) as shown in the dotted area below the figure. ).

[0055] Phase ③ (Building a unified traffic data set): As shown in the splicing symbols in the figure. As shown, the system merges the data streams of scenes containing real LiDAR data and scenes containing reconstructed LiDAR data at the hardware level to construct a unified road condition training dataset with a complete structure.

[0056] Phase 4 (Joint Optimization and Underlying Hardware Control): The merged dataset is input into a solver with an adaptive penalty mechanism (output direction of the rightmost arrow in the diagram). In this phase, the main task decision function is solved. It will be burned into the computing chip of the mass-produced vehicle; during actual driving, the model only needs to receive square nodes (2D camera signals) to output prediction commands and directly drive the vehicle's underlying electronic control unit (ECU) to perform safety braking control.

[0057] Figure 3 This is a schematic diagram of the process for extracting key features (i.e., interpretability-guided feature sparsification) in an embodiment of the present invention.

[0058] This diagram visually illustrates, from left to right, the four stages of data transformation from a high-dimensional redundant space to a sparse critical space: Input and Original Feature Space: The leftmost part acquires the original sample image (such as an animal, camera, water glass, etc.), and initially extracts and maps it into a high-dimensional original feature space. The colored dots in the figure represent samples of different categories (or feature dimensions). At this point, the feature space is relatively dense and contains all background noise and redundant information.

[0059] Base classifier processing: The raw features are input into the baseline classification model for initial learning. The brain and gear icon in the diagram represents the classifier's broad compatibility; it can employ models such as Support Vector Machines (SVM), XGBoost, or various neural networks.

[0060] Feature Attribution Analysis: An interpretability-guided mechanism is introduced (symbolized by a "magnifying glass" icon in the diagram). Feature importance attribution analysis is performed on the baseline model (e.g., using the LIME algorithm). Different colored geometric shapes under the "magnifying glass" represent the quantification of the contribution of features in different dimensions, thereby accurately identifying the features that truly play a role in classification decisions.

[0061] Constructing a sparse key feature space: After filtering out redundant features through a truncation mechanism, the data is reconstructed into a sparse key feature space. The rightmost part of the figure shows the local topology in this space: the dashed elliptical regions with connecting lines represent the similarity between the central sample (cyan dot) and its neighboring samples after feature sparsification. This greatly improves the noise resistance of subsequent information retrieval and reconstruction.

[0062] On the other hand, this embodiment also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0063] On the other hand, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0064] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A privileged learning classification method based on category awareness and interpretability guidance, characterized in that, include: Obtain a training dataset, which includes a first training subset and a second training subset. Each sample in the first training subset contains source domain features, privileged domain features, and class labels. Each sample in the second training subset contains only source domain features and class labels. Based on the samples in the first training subset, determine the feature importance weight vector corresponding to the category; Based on the feature importance weight vector, calculate the weighted similarity between non-privileged samples in the second training subset and similar samples in the first training subset; Based on the weighted similarity, target samples are matched for the non-privileged samples from the first training subset, and the privileged domain features of the target samples are assigned to the non-privileged samples as reconstructed privileged features. Based on the weighted similarity, determine the confidence coefficient corresponding to the reconstructed privileged feature; Based on the first training subset, the second training subset carrying the reconstructed privileged features and the confidence coefficients, a joint optimization objective function is constructed and the model is trained. The confidence coefficients are used to dynamically adjust the penalty intensity of the reconstructed privileged features on model training within the joint optimization objective function.

2. The method according to claim 1, characterized in that, Based on the samples in the first training subset, determine the feature importance weight vector corresponding to the category, including: For the target category, interpretability analysis is performed on the source domain samples belonging to the target category in the first training subset to obtain an initial feature importance vector; The initial feature importance vector is normalized to obtain a normalized importance vector; The normalized importance vector is sparsified to generate a sparse feature importance weight vector corresponding to the target category.

3. The method according to claim 2, characterized in that, The normalized importance vector is sparsified to generate a sparse feature importance weight vector corresponding to the target category, including: Based on a preset screening ratio, dimensions whose importance scores meet preset conditions are selected from the normalized importance vector. The weight values ​​of the selected dimensions in the normalized importance vector are retained, and the weight values ​​of the remaining dimensions are set to zero to generate the sparse feature importance weight vector.

4. The method according to claim 1, characterized in that, Based on the feature importance weight vector, calculate the weighted similarity between non-privileged samples in the second training subset and similar samples in the first training subset, including: Obtain the source domain features of the non-privileged samples and the feature importance weight vector of their respective categories; Based on the feature importance weight vector, a weighted cosine similarity is calculated between the source domain features of the non-privileged sample and the source domain features of similar samples in the first training subset to obtain the weighted similarity.

5. The method according to claim 1, characterized in that, Based on the weighted similarity, the confidence coefficient corresponding to the reconstructed privileged feature is determined, including: The confidence coefficient is calculated using a nonlinear mapping function based on the weighted similarity and a preset first adjustment parameter, wherein the confidence coefficient is negatively correlated with the weighted similarity.

6. The method according to claim 1, characterized in that, The confidence coefficient is calculated using the following formula: ； in, This is the first adjustment parameter. For weighted similarity.

7. The method according to claim 1, characterized in that, The joint optimization objective function includes a first error term corresponding to the first training subset and a second error term corresponding to the second training subset, wherein the second error term is constructed based on the reconstructed privileged features and their corresponding confidence coefficients.

8. The method according to claim 1, characterized in that, In the joint optimization objective function, the penalty terms of the model parameters related to the reconstructed privileged features are weighted by the confidence coefficient.

9. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, When the processor executes the computing program, it implements the method of any one of claims 1-8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1-8.

Images