Recommendation system based on double-graph representation learning pre-training model

Abstract

A recommendation system based on a double-graph representation learning pre-training model comprises a data input module, a meta-feature extraction module, a graph construction module and an intelligent recommendation module, and the data input module inputs model data and dataset data of an external data source, analyzes the model data and the dataset data and then outputs the model data and the dataset data to the meta-feature extraction module; a meta-feature extraction module extracts and aggregates meta-feature vectors of the model and the data set; the atlas construction module constructs double-graph representation of a model atlas and a data set atlas on the basis of the element feature vectors, calculates architecture feature similarity between models and analyzes a label weight relationship between data sets; and the intelligent recommendation module performs regression accuracy prediction on the data set by using a deep recommendation model based on residual image convolution and a multi-layer perceptron on the basis of the double-image representation and the corresponding element feature vector so as to obtain a result list of model recommendation. According to the method, meta-feature extraction is carried out on the model and the data set, double-graph representation learning is carried out based on the similar relation, the complex relation between the data set and the model can be effectively captured, and therefore efficient and accurate model recommendation is achieved.

Description

Technical Field

[0001] The present invention relates to a technology in the field of machine learning, specifically a recommendation system based on a dual-graph representation learning pre-training model. Background Art

[0002] With the rapid development of deep learning technology, more and more pre-trained models are widely used in various tasks. Most of these models are published in open or private model libraries. However, faced with a huge model library, how to choose a suitable model for a specific task has become a very challenging problem. On the one hand, model retrieval in the model library mainly relies on labels or description information to simply classify and filter models. These methods usually only focus on the task type label of the model, do not have a deep understanding of the specific requirements of the task, and fail to make full use of the model architecture information, resulting in limited accuracy and applicability of the recommendation results. On the other hand, the label-based model filtering method ignores the complex relationship between datasets and models, and it is difficult to effectively capture the subtle differences across datasets and models. Summary of the invention

[0003] In order to solve the problems of insufficient recommendation accuracy and poor generalization ability of existing model recommendation systems when dealing with complex tasks, the present invention proposes a recommendation system based on a dual-graph representation learning pre-training model. By extracting meta-features of the model and the data set and performing dual-graph representation learning based on similarity relationships, the present invention can effectively capture the complex relationship between the data set and the model, thereby achieving efficient and accurate model recommendation.

[0004] The present invention is achieved through the following technical solutions:

[0005] The present invention relates to a recommendation system based on a dual-graph representation learning pre-training model, comprising: a data entry module, a meta-feature extraction module, a graph construction module and an intelligent recommendation module, wherein: the data entry module enters model data and data set data of an external data source and outputs them to the meta-feature extraction module after parsing; the meta-feature extraction module extracts and aggregates meta-feature vectors of the model and the data set; the graph construction module constructs a dual-graph representation of the model graph and the data set graph based on the meta-feature vector and calculates the architectural feature similarity between the models and analyzes the label weight relationship between the data sets; the intelligent recommendation module, based on the dual-graph representation and the corresponding meta-feature vector, uses a deep recommendation model based on residual graph convolution and multi-layer perceptron to predict the regression accuracy for the data set, thereby obtaining a result list of model recommendation.

[0006] The model data is loaded and entered by means of regular matching of interfaces, and the data set data is loaded and entered by means of meta-information parsing.

[0007] The data entry module periodically updates the relevant data in the system by automatic data entry.

[0008] The extraction and aggregation means that the data of the data set and the model will be respectively subjected to multivariate feature extraction. After extracting several key features of the model and the data set, these features will be aggregated based on dimensionality scaling, and the meta-feature vector will be output to the graph construction module. Technical Effects

[0009] The deep neural network model recommendation system based on dual-graph representation learning of the present invention constructs dual graphs and recommends models according to the input data; and uses residual graph convolution and multi-layer perceptron to perform regression prediction on the data set and model. Compared with the prior art, the present invention can automatically and data-drivenly perform model recommendations, reducing the reliance on manual participation and expert experience in the model selection process. By constructing model maps and data set maps, the complex relationship between the data set and the model can be effectively captured, significantly improving the accuracy and efficiency of model recommendations; while being able to dynamically adapt to changes in the model library, it can demonstrate good performance, efficiency, and generalization in applications in different fields and different scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a schematic diagram of the system of the present invention;

[0011] Figure 2 The following is a schematic diagram of an application example. DETAILED DESCRIPTION

[0012] like Figure 1 As shown, this embodiment relates to a recommendation system based on a dual-graph representation learning pre-training model, including: a data entry module, a meta-feature extraction module, a graph construction module and an intelligent recommendation module.

[0013] The data entry module includes: a model entry unit based on call information and a data set entry unit based on meta information, wherein: the model entry unit based on call information constructs a corresponding regular matching expression to match the call interface in the document according to the model call interface of the known pre-trained model library; by entering the model call interface information in each pre-trained model library, and performing regular matching on the model description document based on the previously defined regular pattern, the corresponding model call interface in the model description is obtained; then the library type used in the specific call interface is identified, and the corresponding Python library is used to load the model content and extract metadata. The model content and data finally loaded will be input into the meta feature extraction module; the data set entry unit based on meta information loads the Python library corresponding to the image data set platform, and enters the meta information of the data set by calling the full data set meta information extraction function provided in the library; then the content of the image data set is loaded based on the extracted data set meta information through the automatic loading method in the image data set library. The data set content and data finally loaded will be input into the meta feature extraction module.

[0014] The meta-feature extraction module includes: a pre-trained model multivariate feature extraction unit, a data set multivariate feature extraction unit and a feature aggregation unit based on dimensionality scaling, wherein: the pre-trained model multivariate feature extraction unit extracts three types of features of the pre-trained model according to the model file, including model architecture features, model parameter quantities and inference complexity; the data set multivariate feature extraction unit extracts two types of features of the data set according to the data set file, including data set cardinality features, category cardinality features and label composition features; the feature aggregation unit based on dimensionality scaling performs feature weight calculation on the extracted multiple features to obtain the weight of each numerical feature, and performs weight-based dimensionality scaling on the numerical feature vector, and finally uses a multivariate feature splicing method to aggregate multiple features into a complete meta-feature.

[0015] The model architecture features are extracted in the following way: i , select a small batch image dataset D s As a sample data set. Based on the pre-trained weights, the model is used to extract high-dimensional feature vectors in the sample data set sample by sample, and the feature matrix F = [f1, f2, ..., f n ], where n is the number of samples in the sample data set. The sample dimensions of the feature matrix are then averaged to compress the dimensions, and the feature vectors of all samples are averaged. Get the average eigenvector of the sample data set Finally, for the high-dimensional feature vectors output by different models M, one-dimensional average pooling is used to reduce all feature vectors to the same dimension. The feature vectors output by one-dimensional pooling As architectural meta-features of the model.

[0016] The model parameters are extracted in the following way: i , calculate the total number of parameters θ by counting the number of all parameters in the model i , where: the total number of parameters of the model θ i is the sum of all parameters in the model, including weights and bias parameters in all layers. i As the parameter element feature f of the model params .

[0017] The inference complexity is expressed in floating point operations FLOPS, which represents the computational load of the model in actual application. It is obtained by: i , analyze the structure and operation of each layer of the model, accumulate the number of floating-point operations required for each layer, and thus obtain the total inference complexity of the model Calculated model inference complexity As the complexity meta-feature of the model flops .

[0018] The cardinality feature f of the dataset size That is, the image dataset D i The total number of images in; the category cardinality feature f class That is, the image dataset D i The number of all image categories in .

[0019] The label constitutes the feature f smp , obtained by: For a given data set D i , count all the labels it contains T = {t1, t2, ..., t k} and its corresponding sample number Where: k is all labels. Calculate each label t j The proportion p in the data set j , specifically Where: |D i | is the data set D i The number of samples in the final dataset D i The label constitutes a vector p(D i )=[p1,p2,…,p k ]. By using the pre-trained model ResNet for each label t j All samples d under the above conditions are subjected to feature extraction to obtain the feature vector f of each sample. smp (d j ). These eigenvectors are averaged, specifically: Then we get the sample feature f of the label smp (t j ), by calculating each label, we get the label sample feature matrix f smp (D i )=[f smp (t1),f smp (t2),…,f smp (t k )]. Finally, the labels are constructed into a vector p(D i ) and the label sample feature matrix f smp (D i ) and we can write it as Get the label of the data set to form the feature f smp .

[0020] The feature weight calculation specifically includes: all numerical features h1, h2, ..., h n Use Min-Max normalization to scale it to the range of [0,1], specifically: Where: min(h i ) and max(h i ) represent the features h i The minimum and maximum values ​​of i and mean Calculate the weight α i , specifically Weights determine how much influence each feature has in the final vector representation.

[0021] The weight-based dimensional scaling specifically includes: assigning smaller dimensions to numerical features with larger weights, and assigning larger dimensions to numerical features with smaller weights, so that the contribution of numerical features to the whole in the final concatenated high-dimensional vector is relatively balanced, and the scaled dimensions are Among them: Dimension expansion operation The numerical feature h' i The dimension expands to Weight α i The larger the size, the smaller the expanded dimension, and the weight α i The smaller it is, the larger the dimension after expansion.

[0022] The multi-dimensional feature concatenation is specifically: concatenating the high-dimensional features and the numerical features after dimension scaling to form the final high-dimensional feature vector h concat , the specific splicing method is

[0023] The dataset element feature vector f dataset , specifically: high-dimensional labels constitute features fsmp And the numerical feature f size 、f class Multiple features of splicing composition.

[0024] The model element feature vector f model , specifically: high-dimensional model architecture features f arch And the numerical feature f params 、f flops Multiple features of splicing composition.

[0025] The graph construction module includes: a model relationship calculation unit based on architectural cosine similarity, a data set relationship calculation unit based on label weights, and a dual-graph representation construction unit based on entity relationships, wherein: the model relationship calculation unit based on architectural cosine similarity calculates the similarity relationship between all models in the model graph; the data set relationship calculation unit based on label weights calculates the similarity relationship between all data sets in the data set graph and associates the relationship between the data sets with additional label nodes; the dual-graph representation construction unit based on entity relationships constructs a model graph based on the calculated model similarity relationship.

[0026] The similarity relationship between the models is obtained in the following way: for each model M i , based on the model meta-feature f extracted by the meta-feature module model (M i ), reversely extract the architectural features of the model f arch (M i ). For all models M in the model library, select the 2-combination of all models (M i ,M j ), according to the architectural features f of the two models in the 2-combination arch (M i ) and f arch (M j ), calculate the cosine similarity between the two Among them: "·" operation represents the dot product operation of two vectors, |f arch (M i )| and|f arch (M j )|represents the model M i and M j The norm of the architecture feature vector. For any 2-combination of the model library (M i ,M j ), establish e(M i ,M j ) and use the similarity as the edge weight w ij =sim(M i ,M j). Cosine similarity sim(M i ,M j ) is between -1 and 1. The closer the value is to 1, the more similar the two models are, and the closer the value is to -1, the less similar they are.

[0027] The additional label nodes are: for each label t i , for each data set D j Calculate: label t i The proportion This ratio represents the label t i In the dataset D j The frequency of occurrence in , where: Denotes the dataset D j Contains the label t i The number of samples, |D j | represents the data set D j The total number of samples. Then, the connection between the dataset and the label is established based on the label frequency. If the dataset D j Contains tag t i , then establish e(D j ,t i ) and use the calculated ratio as the edge weight Label frequencies are used to model the similarities between different datasets due to shared labels.

[0028] Each node v in the model graph i Represents a model M i , the edge e(M i ,M j ) represents the architectural similarity of the two models, and the weight w ij Representative Model M i and M j The cosine similarity sim(M i ,M j ); Then, the dataset graph G is constructed based on the calculated label weight relationship D =(V D ,E D ), each node D in the dataset graph j Represents a data set, each t i A node represents a label, and the edge e(D j ,t i ) represents the relationship between the data set and the label, and the weight w ji Represents the label t i In the dataset D j Then, the model graph G is transformed into M And the dataset graph GD Combined to form a dual graph representation, each model M i Through its data set D j The performance on the model is the relationship between the model and the dataset. The weight of the relationship between the model and the dataset is w ij By Model M i In the dataset D j The prediction accuracy is calculated based on in: It is model M i In the dataset D j The accuracy on the dataset is obtained by fine-tuning the pre-trained model on the dataset. Finally, for the model nodes and dataset nodes in the dual graph representation, the meta-feature vector f is used respectively. model and f dataset as the embedding vector of the node.

[0029] The dual graph representation includes two parts: data set graph and model graph, specifically including: three types of nodes model node M i , Dataset node D j and label node t i ; and three types of edge model-model architecture similarity edges e(M i ,M j ), the relationship edge e between datasets and labels (D j ,t i ) and the fine-tuning relationship edge e(M i ,D j ).

[0030] The intelligent recommendation module includes: a graph convolution model unit based on residual connection and a regression score calculation model unit based on multi-layer perceptron, wherein: the graph convolution model unit based on residual connection extracts features from the dual-graph representation, and maintains the original feature information through residual connection, with a total of K layers stacked; after obtaining the hidden layer graph representation output by the graph convolution model, the regression score calculation model unit based on multi-layer perceptron uses a multi-layer perceptron (MLP) to perform the final regression score calculation to predict the accuracy of each model on a specific data set; the graph convolution model unit based on residual connection and the regression score calculation model unit based on multi-layer perceptron constitute a deep recommendation model based on dual-graph representation.

[0031] The feature extraction specifically includes: first, the graph convolution model accepts the input dual-graph representation, including the dataset graph G D and model graph G M ; Then, K layers of graph convolution layers with residuals are used to aggregate the features of the nodes in the graph. The graph convolution operation of each layer is in: is the feature vector of node i after the k-th layer of graph convolution, represents the k-th layer graph convolution operation, A is the adjacency matrix of the graph, r k is the residual weight of the kth layer. After stacking K layers of residual graph convolution connections, the final output latent graph representation and Among them: The high-dimensional feature vector of node i is

[0032] The accuracy is obtained by first taking the hidden layer representation obtained from the graph convolutional network as the input of a multilayer perceptron, which contains multiple fully connected layers, performing feature transformation and nonlinear mapping layer by layer, and finally obtaining the regression score in: It is model M j In the dataset D i The prediction accuracy on MLP is a multi-layer perceptron regression model, h pair The dataset D i and Model M j The joint feature vector obtained by concatenating the high-dimensional latent vectors For the recommended dataset D i , by traversing all models in the model library M, calculate the prediction accuracy of each model on the data set Use accuracy to sort the recommended models and output the Top-k recommended model list

[0033] The training process of the deep recommendation model based on dual-graph representation is as follows: first, according to the joint relationship in the established dual-graph representation, that is, the dataset D j In model M i Fine-tuning accuracy on As the model fitting target, the residual connection-based graph convolution model unit and the multi-layer perceptron-based regression model are trained. During the training process, Min-Max normalization is used to process the fitting target value to expand the accuracy difference. Specifically, in: It is the dataset D j The minimum fine-tuning accuracy in the model library M, It is the dataset D j The maximum fine-tuning accuracy in the model library M. The mean square error is used as the loss function during training, that is, in: is the standardized value of the model prediction accuracy. The actual fine-tuning accuracy is standardized. The model parameter update process uses a dynamic learning rate to ensure that the model can converge faster. The learning rate lr is dynamically adjusted based on the loss changes during the training process. new =lr current ×γ, where: γ is the learning rate reduction coefficient, between 0.1 and 1.

[0034] like Figure 2 As shown, the specific application scenario of this embodiment includes: the above-mentioned recommendation system based on the dual-graph representation learning pre-training model and the upper-layer application connected thereto through an interface (RESTful API), including: a module for submitting a dataset to be recommended, a module for feedback of recommendation results, a graph visualization module, a module for retrieving similar datasets and models, and a module for uploading datasets and models, wherein: the model and dataset files are stored using the distributed file storage system HDFS, the model and dataset meta-feature vectors are stored and retrieved using the vector database Milvus, and the graph representation of the model and dataset is stored and retrieved through the graph database Neo4j. The Web end uses the React+SpringBoot development framework for front-end and back-end construction, which specifically includes: a module for submitting recommended datasets, a module for feedback on recommended results, a graph visualization module, a module for retrieving similar datasets and models, and a module for uploading models and datasets. Each module on the Web end interacts with the recommendation system through a RESTful API, including: the dataset submission module, the recommendation result module, and the intelligent recommendation module. The graph visualization module subsequently reads and writes the database through Neo4j's graph data transmission interface py2neo. The similar dataset and model retrieval module searches the vector database through a similar vector search based on KNN.

[0035] After specific practical experiments, the following experiments were conducted in the hardware environment of RTX-4090GPU: three different types of user task sets were generated using the CIFAR-10 dataset, including 45 2-category tasks, 252 5-category tasks, and 297 mixed category tasks. The candidate models were from 17 pre-trained image classification models in the HuggingFace model library, including mainstream architecture models such as ResNet, ViT, and Swin. In the experiment, 200 rounds of training, the parameters of the Adam optimizer and the dynamic learning rate decay strategy were fine-tuned to obtain the fine-tuning accuracy of each candidate model on different task sets as the experimental benchmark. The experimental data are as follows: in the Top-3 model recommendation task, the average ratio of the recommendation accuracy and the optimal accuracy of this system on the three types of task sets reached 99.87%; in the 2-model selection task, this system achieved a selection accuracy of 85.71%; in the time efficiency evaluation of the recommendation, the total time consumption of feature extraction and inference time was about 76.48 seconds, which was significantly lower than the traditional fine-tuning method.

[0036] The comparison between this system and the existing technology is shown in Table 1.

[0037] Table 1 Comparison of technical characteristics

[0038] Compared with the prior art, the data entry module in the present invention adopts an automated parsing and entry method, parses and obtains information through model description documents and data set metadata, and supports distributed storage and real-time data updating, with higher processing efficiency and automation; the feature extraction and aggregation module introduces a multi-feature aggregation algorithm based on dimensionality scaling, which can deeply extract and weight balance multiple features of the model and data set, and generate meta-feature vectors with high expressiveness; in terms of graph construction and representation, by constructing a dual-graph representation learning based on model architecture cosine similarity and label weights, it can capture the complex multi-dimensional relationship between the model and the data set, and improve the accuracy of the recommendation system; the intelligent recommendation module adopts a graph convolutional network based on residual connections, combined with a multi-layer perceptron for regression prediction, and makes recommendations based on prediction accuracy to ensure recommendation accuracy and interpretability of results. The model training optimization module trains the regression model based on the accuracy of the fine-tuning data, and uses Min-Max normalization to optimize the accuracy differences during the training process and improve the generalization ability of the training model. The system is also seamlessly integrated into upper-level applications through the RESTful API, supports dynamic expansion of multiple data sources, and is based on distributed storage and graph databases to ensure good scalability and portability of the system. The dynamically updated graph structure supports the processing of dynamically updated data sets and model libraries, and has a wider applicability.

[0039] The above-mentioned specific implementation can be partially adjusted in different ways by those skilled in the art without departing from the principle and purpose of the present invention. The protection scope of the present invention shall be based on the claims and shall not be limited by the above-mentioned specific implementation. Each implementation scheme within its scope shall be subject to the constraints of the present invention.

Claims

1. A recommendation system based on a dual-graph representation learning pre-training model, characterized in that: include: Data entry module, meta-feature extraction module, graph construction module and intelligent recommendation module, wherein: the data entry module enters the model data and data set data of the external data source and outputs them to the meta-feature extraction module after parsing; the meta-feature extraction module extracts and aggregates the meta-feature vectors of the model and the data set; the graph construction module constructs a dual-graph representation of the model graph and the data set graph based on the meta-feature vector and calculates the architectural feature similarity between the models and analyzes the label weight relationship between the data sets; the intelligent recommendation module, based on the dual-graph representation and the corresponding meta-feature vector, uses a deep recommendation model based on residual graph convolution and multi-layer perceptron to predict the regression accuracy of the data set, and then obtains a list of model recommendation results.

2. The recommendation system based on the dual graph representation learning pre-training model according to claim 1 is characterized in that: The model data is loaded and entered by means of regular matching of interfaces, and the data set data is loaded and entered by means of meta-information parsing.

3. The recommendation system based on the dual graph representation learning pre-training model according to claim 1, characterized in that: The extraction and aggregation means that the data of the data set and the model will be respectively subjected to multivariate feature extraction. After extracting several key features of the model and the data set, these features will be aggregated based on dimensionality scaling, and the meta-feature vector will be output to the graph construction module.

4. The recommendation system based on the dual graph representation learning pre-training model according to any one of claims 1 to 3, characterized in that: The data entry module includes: a model entry unit based on call information and a data set entry unit based on meta information, wherein: the model entry unit based on call information constructs a corresponding regular matching expression to match the call interface in the document according to the model call interface of the known pre-trained model library; by entering the model call interface information in each pre-trained model library, and performing regular matching on the model description document based on the previously defined regular pattern, the corresponding model call interface in the model description is obtained; then the library type used in the specific call interface is identified, and the corresponding Python library is used to load the model content and extract metadata, and the model content and data finally loaded will be input into the meta feature extraction module; the data set entry unit based on meta information loads the Python library corresponding to the image data set platform, and enters the metadata of the data set by calling the full data set metadata extraction function provided in the library; then the image data set content is loaded based on the extracted data set metadata through the automatic loading method in the image data set library, and the data set content and data finally loaded will be input into the meta feature extraction module.

5. The recommendation system based on the dual graph representation learning pre-training model according to any one of claims 1 to 3, characterized in that: The meta-feature extraction module includes: a pre-trained model multivariate feature extraction unit, a data set multivariate feature extraction unit and a feature aggregation unit based on dimensionality scaling, wherein: the pre-trained model multivariate feature extraction unit extracts three types of features of the pre-trained model according to the model file, including model architecture features, model parameter quantities and inference complexity; the data set multivariate feature extraction unit extracts two types of features of the data set according to the data set file, including data set cardinality features, category cardinality features and label composition features; the feature aggregation unit based on dimensionality scaling performs feature weight calculation on the extracted multiple features to obtain the weight of each numerical feature, and performs weight-based dimensionality scaling on the numerical feature vector, and finally uses a multivariate feature splicing method to aggregate multiple features into a complete meta-feature.

6. The recommendation system based on the dual graph representation learning pre-training model according to any one of claims 1 to 3, characterized in that: The graph construction module includes: a model relationship calculation unit based on architectural cosine similarity, a data set relationship calculation unit based on label weights, and a dual-graph representation construction unit based on entity relationships, wherein: the model relationship calculation unit based on architectural cosine similarity calculates the similarity relationship between all models in the model graph; the data set relationship calculation unit based on label weights calculates the similarity relationship between all data sets in the data set graph and associates the relationship between the data sets with additional label nodes; the dual-graph representation construction unit based on entity relationships constructs a model graph based on the calculated model similarity relationship.

7. The recommendation system based on the dual graph representation learning pre-training model according to any one of claims 1 to 3, characterized in that: The intelligent recommendation module includes: a graph convolution model unit based on residual connection and a regression score calculation model unit based on multi-layer perceptron, wherein: the graph convolution model unit based on residual connection extracts features from the dual-graph representation, and maintains the original feature information through residual connection, with a total of K layers stacked; after obtaining the hidden layer graph representation output by the graph convolution model, the regression score calculation model unit based on multi-layer perceptron uses a multi-layer perceptron (MLP) to perform the final regression score calculation to predict the accuracy of each model on a specific data set; the graph convolution model unit based on residual connection and the regression score calculation model unit based on multi-layer perceptron constitute a deep recommendation model based on dual-graph representation.

8. The recommendation system based on the dual graph representation learning pre-training model according to any one of claims 1 to 3, characterized in that: The feature extraction specifically includes: first, the graph convolution model accepts the input dual-graph representation, including the dataset graph G D and model graph G M ; Then, K layers of graph convolution layers with residuals are used to aggregate the features of the nodes in the graph. The graph convolution operation of each layer is in: is the feature vector of node i after the k-th layer of graph convolution, represents the k-th layer graph convolution operation, A is the adjacency matrix of the graph, r k is the residual weight of the kth layer. After stacking K layers of residual graph convolution connections, the final output latent graph representation and Among them: The high-dimensional feature vector of node i is The accuracy is obtained by first taking the hidden layer representation obtained from the graph convolutional network as the input of a multilayer perceptron, which contains multiple fully connected layers, performing feature transformation and nonlinear mapping layer by layer, and finally obtaining the regression score in: It is model M j In the dataset D i The prediction accuracy on MLP is a multi-layer perceptron regression model, h pair The dataset D i and Model M j The joint feature vector obtained by concatenating the high-dimensional latent vectors For the recommended dataset D i , by traversing all models in the model library M, calculate the prediction accuracy of each model on the data set Use accuracy to sort the recommended models and output the Top-k recommended model list 9. The recommendation system based on the dual graph representation learning pre-training model according to claim 1, characterized in that: The training process of the deep recommendation model based on dual-graph representation is as follows: first, according to the joint relationship in the established dual-graph representation, that is, the dataset D j In model M i Fine-tuning accuracy on As the model fitting target, the graph convolution model unit based on residual connection and the regression model based on multi-layer perceptron are trained. During the training process, Min-Max normalization is used to process the fitting target value to expand the difference in accuracy. Specifically, in: It is the dataset D j The minimum fine-tuning accuracy in the model library M, It is the dataset D j The maximum fine-tuning accuracy in the model library M, the mean square error is used as the loss function during training, that is, in: is the standardized value of the model prediction accuracy. In order to actually fine-tune the standardized value of the accuracy, the model's parameter update process uses a dynamic learning rate to ensure that the model can converge more quickly. The learning rate lr is dynamically adjusted based on the loss changes during training. new =lr current ×γ, where γ is the learning rate reduction coefficient.

10. An application device based on the system according to any one of claims 1 to 9, characterized in that: include: The recommendation system based on the dual-graph representation learning pre-training model and the upper-layer application connected to it through an interface (RESTful API) include: a module for submitting a dataset to be recommended, a module for feedback of recommendation results, a graph visualization module, a module for retrieving similar datasets and models, and a module for uploading a dataset and a model, wherein: the model and dataset files are stored using the distributed file storage system HDFS, the model and dataset meta-feature vectors are stored and retrieved using the vector database Milvus, the graph representation of the model and dataset is stored and retrieved through the graph database Neo4j, the web end uses the React+SpringBoot development framework for front-end and back-end construction, specifically including: a module for submitting a dataset to be recommended, a module for feedback of recommendation results, a graph visualization module, a module for retrieving similar datasets and models, and a module for uploading a model and dataset, and each module of the web end interacts with the recommendation system through a RESTful API, wherein: the dataset submission module interacts with the recommendation result module and the intelligent recommendation module, the graph visualization module subsequently reads and writes the database through the graph data transmission interface py2neo of Neo4j, and the similar dataset and model retrieval module retrieves the vector database through a similar vector retrieval based on KNN.

Images