Multi-view scene-oriented learning document organizing and searching method

Abstract

The invention discloses a multi-view scene-oriented learning document organizing and searching method, which comprises the following steps of: in a submitting stage, receiving a trained model submitted by a developer and a corresponding protocol by a learning document base system; the system learns a common subspace and a corresponding feature mapping function based on the protocols of all the learning items so as to align the multi-view learning items covering different view subsets; in the organization stage, the system carries out cross multiplexing evaluation and constructs an embedding space based on an evaluation result, and the distance between a task demand and a learning piece combination can reflect the cross-task multiplexing performance of the combination; in the deployment stage, the system receives task requirements submitted by a user, maps the task requirements to the embedding space, searches for a learning part combination with the optimal multiplexing performance based on distance measurement in the space, and meanwhile provides a corresponding multiplexer to guide the user to complete prediction on task data of the user. According to the method, recommendation is directly carried out according to the multiplexing performance, so that the multi-view learning piece combination with high multiplexing potential is effectively identified.

Description

TECHNICAL FIELD

[0001] The application relates to a multi-view scene-oriented artifact organization and search method, and belongs to the technical field of model reuse. BACKGROUND

[0002] Machine learning has achieved remarkable success in many real applications such as medical, robotics, finance, etc. However, training a high-quality machine learning model from scratch still faces multiple challenges, such as massive high-quality labeled data, expensive computing resources, and rich algorithm design experience, etc. At the same time, due to data privacy and ownership issues, high-quality data cannot be shared and published, limiting the application of large models in data-sensitive scenarios. To systematically solve the above challenges, the artifact paradigm aims to protect privacy and guide users to reuse these models to solve their own tasks by building an artifact base system. Artifacts are composed of trained models and specifications, which characterize the model so that the model can be accurately identified according to user task requirements. The thumbnail kernel mean embedding specification realizes the characterization of model capabilities without revealing the developer's data. Various artifacts are managed by the artifact base system.

[0003] The prior art scheme is based on the thumbnail kernel mean embedding specification, and designs a corresponding artifact search and reuse method. Specifically, the existing artifact search method matches by calculating the similarity between the artifact specification and the user task requirements, thereby recommending the model closest to the user's task. However, the existing artifact reuse algorithm is usually designed independently of the search method. In practical applications, artifact search and reuse are two interdependent processes, but the existing scheme often decouples them and only matches according to the task distribution similarity, ignoring the reuse potential of models on tasks other than the training task. This makes it difficult for existing methods to identify artifact combinations that have high reuse value, limiting the ability of the artifact base system to locate effective model combinations for users. For example, for a multi-view user task, there may be no single high-quality model in the base system that covers all views, but there are multiple models trained based on different subsets of views. Although the similarity between the original training task of these models and the user task is low, they have strong discriminative ability on the user task corresponding to their respective views, and their combined reuse can achieve excellent overall performance on the user task, demonstrating significant reuse potential. However, since the existing method does not directly rely on reuse performance when searching, such effective solutions composed of multiple low-similarity but high-reuse-value artifacts are often missed. SUMMARY

[0004] Purpose of the invention: Existing learning tool base systems typically decouple the search and reuse processes. The search phase only matches based on the distributional similarity between the original training task of the model and the user task, and the results are difficult to accurately reflect the actual reuse performance of the learning tool on the user task. This limitation is particularly prominent in multi-view task scenarios, making it difficult for the system to discover multi-view learning tool combinations consisting of multiple learning tools with low similarity but high reuse value.

[0005] Therefore, in multi-view task scenarios, it is necessary to closely link learning tool retrieval with its reuse performance, and construct a learning tool organization and retrieval mechanism driven by reuse performance. It is necessary to introduce an organization phase, in which the system evaluates the expected reuse performance of different learning tool combinations across various tasks covered by diverse learning tools, and constructs an organizational structure that reflects their reuse potential. Based on this, the retrieval process no longer relies solely on the distribution similarity between the original training task and the user task, but recommends learning tool combinations with high expected reuse performance based on this organizational structure, thereby more effectively solving user tasks and achieving model reuse beyond its original purpose. This invention provides a learning tool organization and retrieval method for multi-view scenarios. By constructing the aforementioned organizational structure within the system, the retrieval process directly recommends learning tool combinations based on their reuse performance on user tasks, effectively identifying multi-view learning tool combinations with high reuse potential.

[0006] Technical Solution: A learning tool organization and retrieval method for multi-view scenarios, applied to a learning tool base system. "Multi-view" refers to multiple views within the same data type, composed of different feature subsets. The learning tool consists of a machine learning model and its corresponding specification, where the specification characterizes the tasks the model excels at. Since the specification can represent the statistical characteristics of its corresponding real-world task without accessing the original data, the system can use the specification of one learning tool to evaluate the reusability performance of other learning tool combinations (i.e., a set consisting of one or more learning tools) on the task characterized by that specification. Based on this, the invention conducts cross-reusability evaluation within the system, relying on the extensive task coverage provided by a large number of high-quality, diverse learning tools to construct an embedding space oriented towards reusability performance. The distance between task requirements and learning tool combinations reflects the cross-task reusability performance of the combination. This embedding space serves as the organizational structure of the learning tool base system, supporting efficient reusability-driven retrieval.

[0007] The method comprises the following three stages: In the submission stage, the learning material base system receives the trained model and corresponding reduction submitted by the developer; the system learns a common subspace and corresponding feature mapping function based on the reduction of all learning materials to align multi-view learning materials covering different view subsets; in the organization stage, the system conducts cross-reuse evaluation and constructs an embedding space based on the evaluation results, where the distance between the task requirements and the learning material combination reflects the cross-task reuse performance of the combination; in the deployment stage, the system receives the task requirements submitted by the user, maps them to the above embedding space, and searches for the learning material combination with the best reuse performance based on the distance metric in the space, while providing a corresponding multiplexer (a model for performing combined prediction) to guide the user to complete the prediction on their task data.

[0008] The specific steps in the submission phase are as follows, designed to address representation differences between different learning materials caused by inconsistencies in feature spaces: Step 101: The developer trains the model on the local dataset and generates a quasi-reduction. The quasi-reduction is in the form of labeled abbreviated kernel mean embedding (RKME), consisting of a set of weighted labeled sample points. Here, the quasi-reduction refers to the original reduction that has not yet been transformed to a common subspace.

[0009] Step 102: Developers upload the model and specifications to the learning platform system.

[0010] Step 103: The learning base system learns the mapping function from each view feature space to the common subspace, and the reconstruction mapping function from the common subspace to each view feature space, based on the quasi-reduction of all learning components.

[0011] Step 104: The learning tool base system uses the learned mapping function to project the sample points in the quasi-reduction to the common subspace, while keeping their labels and weights unchanged, thus obtaining the reduction. Subsequently, the system combines the reduction with the corresponding model to form learning tools, which are then incorporated into the learning tool base system for unified management.

[0012] The organization phase, also known as the learning material base system's learning material organization, involves evaluating the learning reuse performance-driven embedding space through cross-reuse to achieve learning material organization. The specific steps are as follows: Step 201: Initialize a parameterized distance metric function for the learning base system.

[0013] Step 202: For each learning object in the system, the system takes the task represented by its quasi-specification as the target task and performs the following operations.

[0014] Step 203: Based on the current distance metric function, the system calculates the similarity score between the requirements of the target task and the specifications of the other learning materials in the system, and selects several learning materials with high similarity as a candidate set.

[0015] Step 204: Based on the candidate set and the current distance metric function, the system generates multiple candidate learning combinations. Then, using the quasi-reduction of the current target task, a corresponding multiplexer is constructed for each candidate combination, and its reuse performance on the target task is evaluated. Based on the evaluation results, combinations with better reuse performance are identified as positive examples, and combinations with poor reuse performance are identified as negative examples.

[0016] Step 205: Based on the target task requirements and the specifications of the members in each learning component combination, the system calculates the overall similarity of the positive example combination and each negative example combination with respect to the target task, and constructs a contrastive loss to optimize the parameters of the distance metric function so that the overall similarity of the positive example combination is higher than that of the negative example combination.

[0017] Step 206: After completing the traversal of all learning materials, an embedding space and a distance metric on it are obtained, where the distance between the task requirement and any combination of learning materials can reflect the reusability performance of the combination on the corresponding task.

[0018] During the deployment phase, the learning-based system performs searches based on the distance metric function learned in the organization phase and guides users through the model reuse process. The specific steps are as follows: Step 301: The user computes a labeled, abbreviated kernel mean embedding on their task data as a quasi-requirement for the task. Here, quasi-requirements refer to the original requirements that have not yet been transformed into the common subspace.

[0019] Step 302: The user uploads the tagged preliminary requirements to the learning software base system.

[0020] Step 303: The learning base system uses the mapping function from the feature space of each view to the common subspace learned in the submission phase to project the sample points in the quasi-requirements to the common subspace to form requirements, while keeping the coefficients and labels unchanged during the projection process.

[0021] Step 304: Based on the requirements, the learning material base system generates multiple candidate learning material combinations under the distance metric function learned in the organization phase, and selects the learning material combination with the best expected reuse performance as the recommendation result based on the overall similarity between each combination and the task requirements.

[0022] Step 305: The learning base system provides the user with the following: (1) the recommended learning combination; (2) the relevant mapping functions learned in the submission phase, including the mapping function from each view to the common subspace, and the reconstruction mapping function from the common subspace to each view; (3) the reusable built based on the learning combination and user requirements.

[0023] Step 306: The user uses the provided mapping function to complete the feature space of its original task data: First, the observed view features are transformed into the common subspace through the corresponding view-to-subspace mapping function; then, the feature representation of the missing view is generated through the subspace-to-missing view mapping function.

[0024] Step 307: The user inputs the completed multi-view task data into the multiplexer to perform prediction, wherein the multiplexer is used to call one or more learning components from the recommended learning component combination for prediction, such as a model selector, a weighted integrator, or a routing network.

[0025] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the learning object organization and retrieval method for multi-view scenes as described above.

[0026] A computer-readable storage medium storing a learning object organization and retrieval method for a multi-view scene as described above.

[0027] Compared with existing technologies, this invention provides a learning tool organization and retrieval method for multi-view scenarios, enabling the learning tool base system to recommend learning tool combinations with better reuse performance for the target task. Specifically, through the organization scheme of this invention, the learning tool base system can uniformly organize learning tools uploaded by developers from different view feature spaces based on reuse performance; and through its retrieval and reuse scheme, the learning tool base system can search beyond the original purpose of the model, select learning tool combinations with better reuse performance for the user's task according to the user's needs, and build corresponding reusers for the user to complete the prediction. Thus, the learning tool retrieval process is directly guided by reuse performance, ensuring that the recommendation results are highly matched with the actual reuse needs of the user's task. This invention can be widely applied to various practical application scenarios involving multi-view data, effectively improving the actual application efficiency of the system without accessing the original training data or the original user task data, and fully leveraging the potential reuse value of existing machine learning models. Attached Figure Description

[0028] Figure 1 This is a complete flowchart of the learning device base system described in the embodiments of the present invention; Figure 2 This is a flowchart of the student application base system during the submission phase as described in this embodiment of the invention. Figure 3 This is a flowchart of the learning material base system during the organization phase as described in this embodiment of the invention; Figure 4 This is a flowchart of the learning base system during the deployment phase as described in this embodiment of the invention. Detailed Implementation

[0029] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0030] Taking intelligent medical assistance systems as an example, clinical data often exhibits multi-view characteristics. For the same prediction task, a patient's input can come from multiple heterogeneous feature views, such as laboratory test results, medication records, and medical history. In practice, different researchers often train learning tools based on a single view and submit them to the learning tool base system. For example, one learning tool might only use test indicators to predict diabetes risk, while another might only use medication records for the same prediction task. When the user task is based on test indicators, existing search mechanisms typically match based on reduction similarity, recommending only learning tools that are also based on test indicators, making it difficult to effectively recommend complementary learning tools trained on other views such as medication records. However, combining and reusing these learning tools, which are geared towards the same task but originate from different views, can often significantly improve prediction performance. Existing methods, due to the decoupling of the search and reuse processes, struggle to effectively identify such combinations of learning tools with superior reuse performance, limiting the collaborative potential of multi-view learning tools in practical applications.

[0031] This invention discloses a method for organizing and searching learning materials in multi-view scenarios, which effectively addresses the aforementioned problems. The core of this method lies in introducing consideration of final reuse performance during the organization phase. By using reuse performance as a supervisory signal to optimize the distance metric upon which the search depends, the search process is explicitly aligned with reuse performance. For example... Figure 1 As shown, the overall workflow of this invention comprises three stages. In the submission stage (e.g.) Figure 2 As shown), the learning material base system receives learning materials submitted by researchers. Each learning material includes a trained model and its corresponding quasi-reduction. The system learns a bidirectional mapping function from each view feature space to a common subspace based on all quasi-reductions, and projects each quasi-reduction onto this common subspace to obtain a unified reduction representation. In the organization phase (e.g.) Figure 3 As shown in the diagram, the system traverses each learning tool, takes the task represented by its quasi-reduction as the target task, and performs a complete search and reuse process on that task. Specifically, the system uses a combinatorial search strategy to generate candidate learning tool combinations, constructs a reuser for each combination, and evaluates its reuse performance, thereby determining positive and negative combinations. Based on this, the system constructs a contrastive loss to optimize the distance metric function, ensuring that the similarity of positive combinations to the target task requirement is higher than that of negative combinations. Finally, an embedding space is obtained, where the distance between the task requirement and the learning tool combination reflects the reuse performance of that combination on the corresponding task. During the deployment phase (e.g., ...),Figure 4 As shown, when a user submits a new prediction request, the learning tool base system maps it to the embedding space and searches for the combination of learning tools with the best reuse performance based on the optimized distance metric (e.g., the two learning tools based on test indicators and medication records mentioned above). At the same time, it provides the corresponding multiplexer so that the user can complete the prediction on multi-view data.

[0032] Figure 2 This demonstrates the workflow of the submission phase in a medical document base system. This phase includes the following steps: Step 101: For example, researchers in the laboratory use their local patient datasets Training models on (e.g., laboratory tabular data) And generate the corresponding marked standard specification. The reduction is a labeled, abbreviated kernel mean embedding generated by optimizing an objective function that aims to approximate both the marginal distribution of the original data and the conditional distribution under various diagnostic labels.

[0033] Step 102: N researchers from multiple departments, including the laboratory and clinical information department, will use their respective trained models. and its corresponding quasi-convention Upload to the medical device base system.

[0034] Step 103: The medical device base system learns a common subspace based on all uploaded specifications using optimization techniques. And from each view feature space to The bidirectional mapping function. For example, this learning process can be achieved by minimizing a composite loss function, which includes reconstruction loss, contrastive loss, and supervision loss. The reconstruction loss minimizes the difference between each quasi-reduction and its reconstructed representation in the common subspace to preserve the original structural information; the contrastive loss brings the representations of the same sample closer together in different views while pushing the representations of different samples further apart to achieve cross-view alignment; the supervision loss is constructed as follows: the system averages the multi-view representations of each quasi-reduction in the common subspace as its fused representation, and trains a classifier based on these fused representations and the original labels. The prediction error of this classifier is used as the supervision loss term to enhance the separability between different class representations.

[0035] Step 104: The medical component base system uses the learned mapping function to project the quasi-reduction of each model onto a common subspace. A unified specification is obtained. and with the model Shared storage for learning materials .

[0036] Figure 3This demonstrates the workflow of the organization phase in a medical device-based system. This phase includes the following steps: Step 201: Initialize a distance metric for the medical device base system. For example, it can be represented in the form of a neural network.

[0037] Step 202: For each learning device in the system, the medical learning device base system takes the task represented by its quasi-specification as the current target task and executes steps 203 to 205.

[0038] Step 203: Based on the current distance metric function, the system calculates the similarity score between the requirements of the target task and the specifications of the other learning components in the system, and selects several learning components with high similarity to form a candidate set.

[0039] Step 204: Based on the candidate set, the system generates multiple candidate learning component combinations. Then, using the quasi-reduction of the target task, a corresponding reusable is constructed for each combination, and its reuse effect on the target task is evaluated. Based on the evaluation results, combinations with better reuse effects are identified as positive examples, and combinations with poor reuse effects are identified as negative examples. The system employs a combination search strategy to generate the candidate combinations. This strategy is based on the current distance metric function and uses the similarity between task requirements and learning component reduction to guide the construction process. For example, the combination search strategy greedily adds highly similar learning components sequentially and retains members that improve reuse effects to form positive examples. Simultaneously, it selects learning components skipped during the search process, alternative combinations randomly sampled from the candidate set, or combinations obtained through difficult negative example mining as negative examples.

[0040] Step 205: Based on the target task requirements and the specifications of the members in each learning component combination, the system calculates the overall similarity of each positive example combination and each negative example combination relative to the target task, and constructs a contrastive loss to optimize the parameters of the distance metric function, so that the overall similarity of the positive example combination is higher than that of the negative example combination. For example, a triplet loss can be used, where the target task is the anchor point, the positive example combination is the positive sample, and the negative example combination is the negative sample. By optimizing the distance metric function, the system ensures that the distance between the anchor point and the positive sample is smaller than its distance with the negative sample, thereby ensuring that the overall similarity corresponding to the positive sample is higher.

[0041] Step 206: After completing the traversal of all learning materials, the system obtains an embedding space and a distance metric on it, where the distance between the task requirement and any combination of learning materials can reflect the reusability performance of the combination on the corresponding task.

[0042] Figure 4 This demonstrates the workflow of the deployment phase in a medical device base system. This phase includes the following steps: Step 301: The doctor generates task requirements by calculating labeled abbreviated kernel mean embeddings based on data from their new patients (e.g., data containing only laboratory test results).

[0043] Step 302: The doctor uploads the generated task requirements to the medical software base system.

[0044] Step 303: The learning base system utilizes the information learned during the submission phase, from the feature spaces of each view to the common subspace. The mapping function projects the quasi-requirements onto a common subspace to form task requirements. During the projection process, the coefficients and labels in the quasi-requirements remain unchanged.

[0045] Step 304: The learning material base system constructs recommended combinations using the same combinatorial search strategy described in Step 204. This strategy guides combination construction within the embedding space learned during the organization phase, leveraging the similarity between task requirements and the specifications of each learning material. It also evaluates the reusability of each candidate combination based on the quasi-task requirements to determine the recommendation result. For example, learning materials from both the laboratory and clinical information departments can be recommended simultaneously.

[0046] Step 305: The learning material base system provides the doctor with the following: (1) the recommended learning material combination; (2) the mapping function learned in the submission phase, including the mapping function from each view to the common subspace and from the common subspace to each view; (3) a reusable device constructed based on the learning material combination and the doctor's task requirements. For example, the reusable device can be implemented as a model selector, which is trained through supervised learning, using sample points in the specification of the recommended learning material as input features, and the serial number of the corresponding learning material as a label. The classification model used includes support vector machine, gradient boosting tree, etc. The construction method of this reusable device is consistent with the reusable device used in the organization phase to evaluate the combined reuse effect.

[0047] Step 306: The doctor uses the mapping function provided by the system to complete the feature space of the local task data. First, the observed view features are transformed to the common subspace using the corresponding view-to-subspace mapping function. Then, the feature representation of the missing view is generated using the mapping function from the common subspace to the missing view.

[0048] Step 307: The doctor uses the multiplexer to make predictions on the completed multi-view task data. For example, when the multiplexer is a model selector, the doctor can use the selector to judge the samples in the task data, determine which learning software model to use for prediction, and then call the corresponding learning software to complete the diagnostic prediction on the sample.

[0049] Obviously, those skilled in the art should understand that the steps of the learning object organization and search method for multi-view scenarios described in the above embodiments of the present invention can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using device-executable program code, thereby storing them in a storage device for execution by the computing device. Furthermore, in some cases, the steps shown or described can be performed in a different order than presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of the present invention are not limited to any particular hardware and software combination.

Claims

1. A method for organizing and searching learning materials in multi-view scenarios, characterized in that, The method is applied to a learning software base system and includes the following three stages: In the submission stage, the learning software base system receives the trained model and corresponding specification submitted by the developer; the system learns a common subspace and corresponding feature mapping function based on the specifications of all learning software to align multi-view learning software covering different view subsets; During the organization phase, the system conducts cross-reuse evaluation and constructs an embedding space based on the evaluation results. The distance between the task requirements and the learning component combinations reflects the cross-task reuse performance of the combinations. During the deployment phase, the system receives the task requirements submitted by the user, maps them to the above embedding space, and searches for the learning component combinations with the best reuse performance based on the distance metric in the space. At the same time, it provides the corresponding multiplexer to guide the user to complete the prediction on their task data.

2. The learning object organization and retrieval method for multi-view scenarios according to claim 1, characterized in that, The multiple views refer to multiple views composed of different feature subsets within the same data type; the learning tools consist of a machine learning model and its corresponding specifications, wherein the specifications are used to characterize the task scenarios that the model is good at, and the system uses the specifications of a learning tool to evaluate the reusability of other learning tool combinations on the task characterized by the specifications.

3. The learning object organization and retrieval method for multi-view scenarios according to claim 1, characterized in that, The specific steps for the submission phase are as follows: Step 101: The developer trains the model on the local dataset and generates a quasi-reduction; the quasi-reduction is a labeled, abbreviated kernel mean embedding form, consisting of a set of weighted labeled sample points; here, the quasi-reduction refers to the original reduction that has not yet been transformed to the common subspace. Step 102: Developers upload the model and specifications to the learning platform system; Step 103: The learning base system learns the mapping function from each view feature space to the common subspace, and the reconstruction mapping function from the common subspace to each view feature space, based on the quasi-reduction of all learning components. Step 104: The learning base system uses the learned mapping function to project the sample points in the quasi-reduction to the common subspace, while keeping their labels and weights unchanged, to obtain the reduction; Subsequently, the system combines the specifications and corresponding models into learning materials, which are then incorporated into the learning material base system for unified management.

4. The learning object organization and retrieval method for multi-view scenarios according to claim 1, characterized in that, During the organization phase, the learning material base system learns the reuse performance-driven embedding space through cross-reuse evaluation to realize the learning material organization process. The specific steps are as follows: Step 201: Initialize a parameterized distance metric function for the learning base system; Step 202: For each learning object in the system, the system takes the task represented by its quasi-specification as the target task and performs the following operations; Step 203: Based on the current distance metric function, the system calculates the similarity score between the requirements of the target task and the specifications of the other learning materials in the system, and selects several learning materials with high similarity as a candidate set; Step 204: Based on the candidate set and combined with the current distance metric function, the system generates multiple candidate learning combinations; then, using the quasi-reduction of the current target task, it constructs a corresponding multiplexer for each candidate combination and evaluates its reuse effect on the target task; according to the evaluation results, the combination whose reuse effect meets the preset standard is determined as a positive example combination, and the combination whose reuse effect does not meet the preset standard is determined as a negative example combination. Step 205: Based on the target task requirements and the specifications of the members in each learning component combination, the system calculates the overall similarity of the positive example combination and each negative example combination with respect to the target task, and constructs a contrastive loss to optimize the parameters of the distance metric function so that the overall similarity of the positive example combination is higher than the overall similarity of the negative example combination. Step 206: After completing the traversal of all learning materials, an embedding space and a distance metric on it are obtained, where the distance between the task requirement and any combination of learning materials can reflect the reusability performance of the combination on the corresponding task.

5. The learning object organization and retrieval method for multi-view scenarios according to claim 1, characterized in that, During the deployment phase, the learning-based system performs searches based on the distance metric function learned during the organization phase and guides users through the model reuse process. The specific steps are as follows: Step 301: The user calculates a labeled abbreviated kernel mean embedding on their task data as a quasi-requirement for the task; here, quasi-requirement refers to the original requirement that has not yet been transformed into the common subspace. Step 302: The user uploads the tagged preliminary requirements to the learning material base system; Step 303: The learning base system uses the mapping function from the feature space of each view to the common subspace learned in the submission phase to project the sample points in the quasi-requirements to the common subspace to form requirements, while keeping the coefficients and labels unchanged during the projection process; Step 304: Based on the requirements, the learning material base system generates multiple candidate learning material combinations under the distance metric function learned in the organization phase, and selects the learning material combination with the best expected reuse performance as the recommendation result based on the overall similarity between each combination and the task requirements. Step 305: The learning material base system provides the user with the following: (1) the recommended learning material combination; (2) the relevant mapping functions learned during the submission phase, including the mapping function from each view to the common subspace, and the reconstruction mapping function from the common subspace to each view; (3) the reusable built based on the learning material combination and user requirements; Step 306: The user uses the provided mapping function to complete the feature space of its original task data: First, the observed view features are transformed into the common subspace through the corresponding view-to-subspace mapping function; Subsequently, a feature representation of the missing view is generated using a mapping function from the subspace to the missing view; Step 307: The user inputs the completed multi-view task data into the multiplexer to perform prediction, wherein the multiplexer is used to call one or more learning materials from the recommended learning material combination for prediction.

6. A computer device, characterized in that: The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the learning object organization and search method for multi-view scenes as described in any one of claims 1-5.

7. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that: When the computer program / instruction is executed by the processor, it implements the steps of the learning object organization and search method for multi-view scenes as described in any one of claims 1-5.

Images