Product grading and classifying management and control method based on multi-dimensional feature fusion algorithm
Through the algorithm product hierarchical classification method of multi-dimensional feature fusion, the complex management and search difficulties of massive intelligent algorithms are solved, and the automated hierarchical classification of algorithm products is realized, which improves management and search efficiency and accuracy.
Patent Information
- Application Number
- CN202510338034.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-21
- Publication Date
- 2025-07-25
AI Technical Summary
The existing technology cannot effectively manage and quickly retrieve massive intelligent algorithms, resulting in complex and difficult algorithm product management.
The algorithm product hierarchical classification method is adopted with multi-dimensional feature fusion, and the surface feature extraction and deep feature generation are combined with attention mechanism and multi-layer perception machine to realize the automated hierarchical classification of algorithm products.
It improves the labeling and retrieval efficiency of algorithm products, ensures the accuracy of feature extraction and the accuracy of hierarchical classification, and simplifies the management process of algorithm products.
Smart Images
Figure CN120372375A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of data processing and machine learning, and particularly relates to an algorithm product hierarchical classification control method based on multi-dimensional feature fusion. Background Art
[0002] With the rapid development of artificial intelligence technology, algorithm products have been widely applied in various industries, and various military intelligent algorithms have gradually been applied to various components. In the current information system, intelligent algorithms are being applied to more and more different application scenarios. Through the applications in aspects such as situation awareness, intelligence analysis, and command decision-making, these intelligent algorithms bring great development prospects for intelligent components. However, the continuous increase in intelligent algorithms also brings relatively severe challenges to the management of intelligent algorithms. How to construct a standard process specification for algorithm management, establish a reasonable and perfect algorithm product management system, and achieve scientific and efficient management of algorithm products is an important task currently faced. Summary of the Invention
[0003] The technical problem to be solved by the present invention is: to overcome the deficiencies of the prior art, and propose an algorithm product hierarchical classification control method based on multi-dimensional feature fusion. Aiming at the problems such as complex management and difficult retrieval of algorithm products brought by the current sharp increase trend of intelligent algorithms, the method successively adopts algorithm product hierarchical feature extraction and multi-dimensional feature fusion classification technology to realize the automatic hierarchical classification of algorithm products, provides a clear and complete label query mechanism for algorithm users, and promotes the rapid retrieval, reuse, and iteration of algorithm products.
[0004] The technical solution of the present invention is: an algorithm product hierarchical classification control method based on multi-dimensional feature fusion, including:
[0005] Extracting multi-dimensional feature data from algorithm products;
[0006] According to the obtained multi-dimensional feature data, designing and constructing a feature fusion model, and according to the constructed feature fusion model, completing the automatic hierarchical classification of algorithm products by using a multi-dimensional feature fusion method;
[0007] For the hierarchical classification labels of different dimensions, constructing several feature dimensionality reduction models; for the fused features output by the feature fusion model, using the constructed feature dimensionality reduction models to extract the key features required for classification in the current dimension, and completing the hierarchical classification of algorithm products in each dimension according to the corresponding key features.
[0008] When extracting multi-dimensional feature data from algorithm products, two methods of algorithm surface feature extraction and algorithm deep feature generation are adopted; among them,
[0009] The algorithm feature extraction method, that is, directly extracting and encoding the metadata information features of the algorithm product to obtain the surface features of the algorithm product;
[0010] An algorithm feature generation method, that is, through a large model, analyze and generate algorithm feature vectors from algorithm source code and corresponding documents to represent the deep features of algorithm products.
[0011] Directly extract and encode the metadata information features of the algorithm product to obtain the surface features of the algorithm product, including:
[0012] Sort out the existing metadata of the algorithm product, and sort out several characteristic data representing the representative features of the algorithm product; for different characteristic data, use corresponding encoding methods to obtain their feature vectors:
[0013] x = e(w);
[0014] Among them, w represents the original feature data, x represents the encoded feature vector, and e represents the corresponding encoder; the feature vectors of different dimensions are represented by x1, x2,..., x n respectively.
[0015] The method of analyzing and generating algorithm feature vectors from algorithm source code and corresponding documents through a large model to represent the deep features of algorithm products includes:
[0016] Collect the source code of multiple algorithms and their usage documents, and perform data preprocessing to obtain usage documents and function code blocks with standard specifications and unified styles;
[0017] According to the obtained usage documents and function code blocks, research and screen existing open-source basic large language pre-trained models, conduct experimental verification on the standard code dataset, sort the performance of different pre-trained models for text extraction tasks, and select the pre-trained model with the best performance to ensure the ability of the pre-trained model to capture code semantics and structural information;
[0018] For the pre-trained model, use the usage document and function code block to implement the generation of the deep features of the algorithm product.
[0019] The collection of source code of multiple algorithms and their usage documents, and the performance of data preprocessing to obtain usage documents and function code blocks with standard specifications and unified styles includes:
[0020] Code cleaning: Remove non-code parts from the code;
[0021] Code formatting: Unify the code style and reduce the impact of format differences on the model;
[0022] Code splitting: Split long code files into smaller functions or code blocks.
[0023] The implementation of generating the deep features of the algorithm product using the usage document and function code block includes:
[0024] Divide the source code file into a code file and a document description file;
[0025] For the code file, parse the source code into several syntax units in a way based on the abstract syntax tree, input them into the code large model respectively, and input the prompt words to generate the function descriptions of each code block;
[0026] For the document description file, extract the required algorithm function descriptions from the document description file by means of regular expressions, and sort out and expand the function descriptions through the language large model.
[0027] For the generated function descriptions of each code block and the extracted algorithm function descriptions, convert the description information into a feature vector representation through token encoding and token embedding in sequence, perform text encoding through the Transformer model, and perform knowledge distillation on the two types of features through the contrast loss to obtain more accurate algorithm features. Finally, splice the two types of features to obtain the final deep features of the algorithm product.
[0028] The automatic grading and classification of the algorithm product is completed by using the multi-dimensional feature fusion method, including:
[0029] According to the obtained surface features and deep features of the algorithm product, use the attention mechanism to realize the weight correction of the surface features and deep features:
[0030] x′ i =attention(x i ,y)
[0031] Among them, the surface feature is x i ,i = 1,2,...,n, n is the number of existing surface features; y is the deep feature, and the corrected feature is x′ i ,i = 1,2,..,n; attention represents the attention mechanism;
[0032] According to the obtained corrected feature and deep feature, splice them to obtain the fusion feature:
[0033] z i =concat(x′ i ,y)
[0034] Among them, z i represents the fusion feature, and concat represents the splicing operation; construct N multi-layer perceptrons to realize feature dimensionality reduction, and map the fusion feature into the feature spaces where different classification dimensions are located respectively:
[0035] z′ i =MLP(z i );
[0036] Among them, z′ i represents the features of different classification dimensions, and MLP represents a multi-layer perceptron, which consists of two fully connected layers;
[0037] A classifier is used to achieve hierarchical classification of the algorithm product in each dimension for the features of different classification dimensions.
[0038] The use of a classifier to achieve hierarchical classification of the algorithm product in each dimension for the features of different classification dimensions includes:
[0039] Adopt a fully connected layer + non-linear layer + fully connected layer structure to construct a classifier;
[0040] Collect algorithm product data and the labels of the algorithm product in each dimension as training data, and complete the training of the classifier through supervised learning methods;
[0041] Use the trained classifier to achieve hierarchical classification of the algorithm product in each dimension.
[0042] The beneficial effects of the present invention are:
[0043] (1) The present invention proposes an automatic hierarchical classification method for algorithm products, which alleviates the problems of complex management and difficult retrieval of a large number of intelligent algorithms, and improves the annotation and retrieval efficiency of algorithm products;
[0044] (2) The present invention proposes a hierarchical feature extraction method for algorithm products, and the extracted algorithm product features are more accurate;
[0045] (3) The present invention proposes a multi-dimensional feature fusion classification technology for algorithm products, which realizes further refinement of algorithm product features and improves the accuracy of hierarchical classification of algorithm products in each dimension. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 It is a schematic diagram for generating deep features of algorithm products.
[0047] Figure 2 It is a schematic diagram of the algorithm product feature fusion process.
[0048] Figure 3 It is a schematic diagram of the algorithm product hierarchical classification process based on multi-dimensional feature fusion. DETAILED DESCRIPTION OF THE INVENTION
[0049] As Figure 3 shown, it is a schematic diagram of the algorithm product hierarchical classification process based on multi-dimensional feature fusion of the present invention, and the specific implementation scheme is as follows:
[0050] 1. Algorithm product hierarchical feature extraction
[0051] As Figure 1As shown in the figure, an algorithm product hierarchical feature extraction module is designed and implemented to extract multi-dimensional feature data from algorithm products. This module consists of two parts: the extraction of surface features of algorithm products and the generation of deep features of algorithm products. The specific scheme is as follows:
[0052] 1.1 Extraction of surface features of algorithm products
[0053] Generally, when an algorithm product is uploaded, it needs to carry various identification information in addition to entity data such as algorithm source code and algorithm models, such as algorithm name, algorithm version, general algorithm type, algorithm usage description, training dataset description, training environment description, usage instructions and other types of meta-information. Based on this, this module sorts and classifies the existing meta-information of algorithm products, sorts out several representative data of algorithm products, and uses encoding methods such as One-Hot encoding, word2vec, and bert to implement feature encoding for different data types. The meanings and encoding methods of some feature data are as follows:
[0054] Table 1 Algorithm feature data table
[0055]
[0056]
[0057] For different feature data, the corresponding encoding method is used to obtain its feature vector:
[0058] x = e(w),
[0059] where w represents the original feature data, x represents the encoded feature vector, and e represents the corresponding encoder. Feature vectors of different dimensions are represented by x1, x2,..., x n respectively.
[0060] 1.2 Generation of deep features of algorithm products
[0061] In addition to various metadata carried by algorithm products, the information contained in algorithm source code and various documents is more abundant. Therefore, it is necessary to deeply analyze the algorithm product source code and various documents to obtain more representative algorithm features. The main implementation scheme is as follows:
[0062] (1) Data collection and processing
[0063] First, collect the source code and its usage documents of multiple algorithms and perform data preprocessing, including:
[0064] Code cleaning: Remove non-code parts in the code (such as HTML tags and non-text information in comments).
[0065] Code formatting: Unify the code style, such as indentation, spaces, etc., to reduce the impact of format differences on the model.
[0066] Code splitting: Split long code files into smaller functions or code blocks for better model processing.
[0067] (2) Model selection and design
[0068] Research and screen existing open-source large language pre-trained models, conduct experimental verification on standard code datasets, rank the performance of different pre-trained models for text extraction tasks, select the pre-trained model with the best performance, and ensure the ability of the pre-trained model to capture code semantic and structural information.
[0069] (3) Feature generation
[0070] The feature generation scheme is as follows:
[0071] First, divide the source code files into two categories: code files and document description files:
[0072] For code files, use the abstract syntax tree-based method to parse the source code into several syntax units, input them into the code large model respectively, and input prompt words to generate the function descriptions of each code block;
[0073] For document description files, extract the required algorithm function descriptions from the document description files through regular expressions, and use the language large model to sort out and expand the function descriptions.
[0074] For the generated function descriptions of each code block and the extracted algorithm function descriptions, convert the description information into feature vector representations through tokenization encoding (Tokenizer) and token embedding (Embedding) in sequence, perform text encoding through the Transformer model, and perform knowledge distillation on the two types of features through contrastive loss to obtain more accurate algorithm features. Finally, splice the two types of features to obtain the final deep features of the algorithm product.
[0075] 2. Multi-dimensional feature fusion classification of algorithm products
[0076] As Figure 2 shown, after obtaining the multi-dimensional feature data of the algorithm product, use the method based on multi-dimensional feature fusion to realize the automatic hierarchical classification of the algorithm product, which is mainly divided into three parts: feature weight correction, feature fusion and dimensionality reduction, and algorithm hierarchical classification. The specific scheme is as follows:
[0077] (1) Feature weight correction
[0078] Use the attention mechanism to realize the weight correction of surface features and deep features:
[0079] x′i = attention(x i , y)
[0080] where the surface feature is x i , i = 1, 2, ..., n, where n is the number of existing surface features; y is the deep feature, and the corrected feature is x' i , i = 1, 2, .., n; attention represents the attention mechanism.
[0081] (2) Feature fusion and dimensionality reduction
[0082] Concatenate the corrected feature and the deep feature to obtain the fused feature:
[0083] z i = concat(x' i , y)
[0084] where z i represents the fused feature, and concat represents the concatenation operation. Construct N multi-layer perceptrons to implement dimensionality reduction, and map the fused feature into the feature spaces corresponding to different classification dimensions respectively:
[0085] z' i = MLP(z i )
[0086] where z' i represents the features of different classification dimensions, MLP represents the multi-layer perceptron, which consists of 2 fully connected layers.
[0087] (3) Algorithm hierarchical classification
[0088] Finally, use a classifier to obtain the hierarchical classification results of the algorithm product in each dimension:
[0089] 1) Use a structure of fully connected layer + non-linear layer + fully connected layer to construct a classifier;
[0090] 2) Collect the algorithm product data and the labels of the algorithm product in each dimension as training data, and complete the training of the classifier through the supervised learning method;
[0091] 3) Use the trained classifier to implement the hierarchical classification of the algorithm product in each dimension.
Claims
1. An algorithm product grading and classification control method based on multi-dimensional feature fusion, characterized in that Including: Extract multi-dimensional feature data from algorithm products; According to the obtained multi-dimensional feature data, design and construct a feature fusion model, and according to the constructed feature fusion model, complete the automatic grading and classification of algorithm products by using a multi-dimensional feature fusion method; For the grading and classification labels of different dimensions, construct several feature dimensionality reduction models; for the fused features output by the feature fusion model, use the constructed feature dimensionality reduction models to extract the key features required for classification in the current dimension, and complete the grading and classification of algorithm products in each dimension according to the corresponding key features.
2. A method for hierarchical classification and control of algorithm products based on multi-dimensional feature fusion according to claim 1, characterized in that: When extracting multi-dimensional feature data from algorithm products, two methods are adopted: algorithm surface feature extraction and algorithm deep feature generation; Among them, The algorithm feature extraction method, that is, directly extract and encode the metadata information features of the algorithm product to obtain the surface features of the algorithm product; The algorithm feature generation method, that is, analyze and generate algorithm feature vectors from the algorithm source code and the corresponding documents through a large model to represent the deep features of the algorithm product.
3. A method for hierarchical classification and control of algorithm products based on multi-dimensional feature fusion according to claim 2, characterized in that: The direct extraction and encoding of the metadata information features of the algorithm product to obtain the surface features of the algorithm product includes: Sort out the existing meta-information of the algorithm product, and sort out several feature data with representative representations of the algorithm product; for different feature data, use the corresponding encoding method to obtain their feature vectors: x = e(w); Among them, w represents the original feature data, x represents the encoded feature vector, and e represents the corresponding encoder; the feature vectors of different dimensions are represented by x1, x2,..., x n respectively.
4. The algorithm product classification and control method based on multi-dimensional feature fusion according to claim 3, characterized in that: The analysis and generation of algorithm feature vectors from the algorithm source code and the corresponding documents through a large model to represent the deep features of the algorithm product includes: Collect the source codes and their usage documents containing multiple algorithms, and perform data preprocessing to obtain usage documents and function code blocks with standard specifications and unified styles; According to the obtained usage documents and function code blocks, investigate and screen existing open-source basic large language pre-trained models, conduct experimental verification on the standard code dataset, rank the performance of different pre-trained models for text extraction tasks, and select the pre-trained model with the best performance to ensure the ability of the pre-trained model to capture code semantics and structural information; For the pre-trained model, use the usage documents and function code blocks to implement the generation of deep features of the algorithm product.
5. A classification and control method for algorithm products based on multi-dimensional feature fusion according to claim 4, characterized in that: The collection of source codes and their usage documents containing multiple algorithms, and the performance of data preprocessing to obtain usage documents and function code blocks with standard specifications and unified styles includes: Code cleaning: Remove non-code parts from the code; Code formatting: Unify the code style and reduce the impact of format differences on the model; Code splitting: Split long code files into smaller functions or code blocks.
6. The algorithm product grading and classification control method based on multi-dimensional feature fusion according to claim 4, characterized in that: The implementation of the generation of deep features of the algorithm product by using the usage documents and function code blocks includes: Divide the source code file into a code file and a document description file; For the code file, use the abstract syntax tree-based method to parse the source code into several syntax units, input them into the code large model respectively, and input the prompt words to generate the function descriptions of each code block; For the document description file, extract the required algorithm function descriptions from the document description file through regular expressions, and sort out and expand the function descriptions through the language large model. For the functional descriptions of the generated code blocks and the extracted algorithm functional descriptions, the description information is successively converted into feature vector representations through word segmentation encoding and word segmentation embedding, encoded by the Transformer model, and knowledge distillation is performed on the two types of features through contrastive loss to obtain more accurate algorithm features. Finally, the two types of features are concatenated to obtain the final deep features of the algorithm product.
7. A method for hierarchical classification and control of algorithm products based on multi-dimensional feature fusion according to claim 4, characterized in that: The automatic hierarchical classification of algorithm products is completed by using the multi-dimensional feature fusion method, including: According to the obtained surface features and deep features of the algorithm product, the attention mechanism is used to realize the weight correction of the surface features and the deep features: x i ′ = attention(x i , y) Among them, the surface feature is x i , i = 1, 2,..., n, where n is the number of existing surface features; y is the deep feature, and the corrected feature is x i ′ , i = 1, 2,.., n; attention represents the attention mechanism; According to the obtained corrected features and the deep features, the fused features are obtained by concatenation: z i = concat(x i ′ , y) Among them, z i represents the fused feature, and concat represents the concatenation operation; N multi-layer perceptrons are constructed to achieve feature dimensionality reduction, and the fused features are respectively mapped into the feature spaces where different classification dimensions are located: z i ′ = MLP(z i ); Among them, z i ′ represents features of different classification dimensions, and MLP represents a multi-layer perceptron, which consists of two fully connected layers; A classifier is used to realize the hierarchical classification of algorithm products in each dimension for different classification dimension features.
8. A method for hierarchical classification and control of algorithm products based on multi-dimensional feature fusion according to claim 7, characterized in that: The use of a classifier to realize the hierarchical classification of algorithm products in each dimension for different classification dimension features includes: Construct a classifier using a fully connected layer + non-linear layer + fully connected layer structure; Collect algorithm product data and algorithm product labels in each dimension as training data, and complete the training of the classifier through the supervised learning method; Use the trained classifier to realize the hierarchical classification of algorithm products in each dimension.