Recommendation system based on dual graph representation learning pre-trained model

Abstract

A recommendation system based on double 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 the model data and dataset data of the external data source and outputs 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 dataset; the graph construction module constructs the double graph representation of the model graph and the dataset graph based on the meta feature vectors, and calculates the architecture feature similarity between the models and analyzes the label weight relationship between the datasets; the intelligent recommendation module uses a deep recommendation model based on residual graph convolution and multilayer perceptron to predict the regression accuracy of the dataset based on the double graph representation and the corresponding meta feature vectors, and then obtains the result list recommended by the model. Through meta feature extraction of the model and the dataset, and double graph representation learning based on the similarity relationship, the present application can effectively capture the complex relationship between the dataset and the model, thereby realizing efficient and accurate model recommendation.

Description

TECHNICAL FIELD

[0001] The application relates to a technology in the field of machine learning, in particular to a recommendation system based on double-graph representation learning pre-training model. BACKGROUND

[0002] With the rapid development of deep learning technology, more and more pre-training models are widely used in various tasks, and these models are mostly published in open or private model libraries. However, in the face of a large number of model libraries, how to select a suitable model for a specific task has become a challenging problem. On the one hand, model retrieval in the model library mainly relies on labels or description information to simply classify and filter the models. These methods usually only focus on the task type label of the model, do not deeply understand the specific needs of the task, and do not fully utilize the model architecture information, resulting in limited accuracy and applicability of the recommended results. On the other hand, the model screening method based on labels ignores the complex relationship between the dataset and the model, and it is difficult to effectively capture the subtle differences between the datasets and the models. SUMMARY

[0003] The application proposes a recommendation system based on double-graph representation learning pre-training model to solve the problems of insufficient recommendation accuracy and poor generalization ability of the existing model recommendation system when dealing with complex tasks. The application can effectively capture the complex relationship between the dataset and the model by extracting meta-features of the model and the dataset and performing double-graph representation learning based on the similarity relationship, thereby realizing efficient and accurate model recommendation.

[0004] The application is implemented by the following technical solutions:

[0005] The application relates to a recommendation system based on double-graph representation learning pre-training model, which comprises a data entry module, a meta-feature extraction module, a graph construction module and an intelligent recommendation module. The data entry module enters and parses the model data and dataset data of external data sources and outputs them to the meta-feature extraction module. The meta-feature extraction module extracts and aggregates the meta-feature vectors of the model and the dataset. The graph construction module constructs double-graph representation of the model graph and the dataset graph based on the meta-feature vectors, calculates the architecture feature similarity between the models and analyzes the label weight relationship between the datasets. The intelligent recommendation module uses a deep recommendation model based on residual graph convolution and multilayer perception to predict the regression accuracy of the dataset based on the double-graph representation and the corresponding meta-feature vectors, and then obtains the result list of the model recommendation.

[0006] The model data is loaded and entered by interface regular matching, and the dataset data is loaded and entered by meta-information parsing.

[0007] The data entry module regularly updates relevant data in the system through automatic data entry.

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

[0009] Technical Effects

[0010] The present invention is based on a deep neural network model recommendation system for dual-graph representation learning. It constructs dual graphs and recommends models based on input data; it uses residual graph convolution and multi-layer perceptrons to perform regression predictions on datasets and models. Compared to existing technologies, the present invention can perform model recommendations in an automated, data-driven manner, reducing the reliance on manual participation and expert experience in the model selection process. By constructing model graphs and dataset graphs, it can effectively capture the complex relationship between datasets and models, 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

[0011] Figure 1 Schematic diagram of the system of the present invention;

[0012] Figure 2 Schematic diagram of the application of the embodiment. DETAILED DESCRIPTION

[0013] 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.

[0014] The data entry module includes: a model entry unit based on call information and a dataset 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 based on 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 dataset entry unit based on meta-information loads the Python library corresponding to the image dataset platform, enters the metadata of the dataset by calling the full dataset metadata extraction function provided in the library; then, through the automated loading method in the image dataset library, the image dataset content is loaded based on the extracted dataset metadata. The finally loaded dataset content and data will be input into the meta-feature extraction module.

[0015] 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 the multivariate feature splicing method to aggregate multiple features into a complete meta-feature.

[0016] 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 from 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.

[0017] 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 .

[0018] 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 the following method: 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 f flops .

[0019] The dataset cardinality feature f 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 .

[0020] The label constitutes the feature f smp , is obtained by the following method: for a given data set D i , count all the labels it contains T={t1,t2,…,t k} and its corresponding sample size Where: k is all labels. Calculate each label t j The proportion p in the data set j , specifically Where: |D i | is the dataset 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 feature extraction, get the feature vector f of each sample smp (d j ). These eigenvectors are averaged, specifically Further, the sample feature f of the label is obtained smp (t j ), the label sample feature matrix f is obtained by calculating each label smp (D i ) = [f smp (t1), f smp (t2), …, f smp (t k )]. Finally, the label composition vector p(D i ) and the label sample feature matrix f smp (D i ) are multiplied, denoted as The label composition feature f of the data set is obtained smp .

[0021] The feature weight calculation specifically includes: all numerical features h1, h2, …, h n Using Min-Max standardization processing, it is scaled to the range of [0, 1], specifically Where: min(h i ) and max(h i ) represent the minimum and maximum values of the feature h i , respectively. Then all weights α i are calculated according to the dimension d i and the mean value , specifically The weight is the influence degree of each feature in the final vector representation.

[0022] The weight-based dimension scaling specifically includes: numerical features with larger weights are given smaller dimensions, while numerical features with smaller weights are given larger dimensions, so that the contributions of numerical features in the final high-dimensional vector are relatively balanced. The dimension after scaling Where: the dimension expansion operation The dimension of the numerical feature h' i is expanded to The larger the weight α i , the smaller the expanded dimension, and the smaller the weight α i , the larger the expanded dimension.

[0023] The multi-element feature splicing specifically includes: splicing the high-dimensional feature and the numerical feature after dimension scaling to form the final high-dimensional feature vector h concat , and the specific splicing method is

[0024] The data set meta feature vector f dataset specifically includes: a high-dimensional label composition feature fsmp and numerical features f size , f class .

[0025] The model meta-feature vector f model , specifically: high-dimensional model architecture features f arch and numerical features f params , f flops .

[0026] The atlas construction module comprises a model relationship calculation unit based on architecture cosine similarity, a dataset relationship calculation unit based on label weight, and a double-graph representation construction unit based on entity relationship, wherein: the model relationship calculation unit based on architecture cosine similarity calculates the similarity relationship between all models in the model atlas; the dataset relationship calculation unit based on label weight calculates the similarity relationship between all datasets in the dataset atlas and associates the relationship between the datasets with additional label nodes; and the double-graph representation construction unit based on entity relationship constructs the model atlas based on the calculated model similarity relationship.

[0027] The similarity relationship between the models is specifically obtained by: for each model M i , according to the model meta-feature f model (M i ) extracted by the meta-feature module, the architecture feature f arch (M i ) of the model is extracted in reverse. For all models M in the model library, select the 2-combination (M i , M j ) of all models, according to the architecture features f arch (M i ) and f arch (M j ) of the two models in the 2-combination, calculate the cosine similarity wherein the “·” operation represents the dot product operation of two vectors, |f arch (M i )| and |f arch (M j )| represent the norms of the architecture feature vectors of the models M i and M j , respectively. For any 2-combination (M i , M j ) of the model library, a connection relationship e(M i , M j ) is established, and the similarity is taken as the weight w ij of the edge. i 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.

[0028] The additional label nodes are: for each label t i , for each dataset D j Calculate: label t i proportion This ratio represents the label t i In the dataset D j The frequency of occurrence in , where: Represents the dataset D j Contains the label t i The number of samples, |D j | represents the dataset 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 uses the calculated ratio as the edge weight. Label frequencies are used to model the similarity between different datasets due to shared labels.

[0029] 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 label t i In the dataset D j Then, the model graph G is transformed into M and dataset graph GD The joint is formed into a dual graph representation, each model M i is represented by its performance on the dataset D j as a relationship between the model and the dataset. The weight w ij of the model-dataset relationship is calculated by the prediction accuracy of the model M i on the dataset D j wherein: is the accuracy of the model M i on the dataset D j , 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 vectors f model and f dataset are used as the embedding vectors of the nodes, respectively.

[0030] The dual graph representation includes a dataset graph and a model graph, and specifically includes three types of nodes: model nodes M i , dataset nodes D j , and label nodes t i ; and three types of edges: architecture similarity edges e(M i , M j ) between models, relationship edges e(D j , t i ) between datasets and labels, and fine-tuning relationship edges e(M i , D j ) between models and datasets.

[0031] The intelligent recommendation module includes a graph convolution model unit based on residual connection and a regression score calculation model unit based on multilayer perceptron, wherein the graph convolution model unit based on residual connection extracts features from the dual graph representation while retaining original feature information through residual connection, and a total of K layers are stacked; the regression score calculation model unit based on multilayer perceptron performs final regression score calculation using multilayer perceptron (MLP) after obtaining the hidden layer graph representation output by the graph convolution model, and predicts the accuracy of each model on a specific dataset; the graph convolution model unit based on residual connection and the regression score calculation model unit based on multilayer perceptron together constitute a deep recommendation model based on the dual graph representation.

[0032] The feature extraction specifically includes: first, the graph convolution model accepts the input dual graph representation, including a dataset graph G D and a model graph G M ; then the nodes in the graph are aggregated by K layers of graph convolution layers with residual connection, and the graph convolution operation of each layer is wherein: ​is the feature vector of node i after k-th layer graph convolution, denotes the k-th layer graph convolution operation, A is the adjacency matrix of the graph, r k is the residual weight of the k-th layer, and the final output hidden graph representation is obtained after the residual graph convolution connection of the stacked K layers and wherein the high-dimensional feature vector of node i is

[0033] The accuracy is obtained by the following method: first, the hidden layer representation obtained from the graph convolution network is taken as the input of the multilayer perceptron, the multilayer perceptron includes multiple fully connected layers, and the feature conversion and nonlinear mapping are performed layer by layer to finally obtain the regression score wherein: is the model M j The prediction accuracy on the data set D i , f MLP is a multilayer perceptron regression model, h pair is the joint feature vector obtained by splicing the high-dimensional hidden vector of the data set D i and the model M j For the recommended data set D i , the prediction accuracy of each model on the data set is calculated by traversing all models in the model library M The accuracy is used to sort the recommended models, and a Top-k recommended model list can be output

[0034] The training process of the deep recommendation model based on the double graph representation is specifically: first, according to the joint relationship in the established double graph representation, that is, the data set D j The fine-tuning accuracy of the model M i on the data set D is taken as the fitting target, and the training of the graph convolution model unit based on residual connection and the regression model based on multilayer perceptron is carried out. In the training process, Min-Max standardization is used to process the target value of fitting to expand the difference of the accuracy, specifically wherein: is the minimum fine-tuning accuracy of the data set D j in the model library M, is the maximum fine-tuning accuracy of the data set D j in the model library M. In the training process, the mean square error is taken as the loss function, that is wherein: is the value of the model prediction accuracy after standardization, ​is the actual fine-tuning accuracy after standardization. The parameter update process of the model uses a dynamic learning rate to ensure that the model can converge more quickly, dynamically adjusting the learning rate lr based on the loss changes during training new = lr current x γ, where γ is the learning rate reduction factor, between 0.1 and 1.

[0035] As shown in Figure 2 , this embodiment is applied to a specific scenario, including the above-mentioned recommendation system based on double-graph representation learning pre-trained model and the upper-layer application connected through an interface (RESTful API), including: a to-be-recommended dataset submission module, a recommendation result feedback module, a graph visualization module, a similar dataset and model retrieval module, and a dataset and model upload module, wherein: the model and dataset files are stored using a distributed file storage system HDFS, the model and dataset meta-feature vectors are stored and retrieved using a vector database Milvus, and the graph representation of the model and dataset is stored and retrieved through a graph database Neo4j. The Web side uses the development framework of React+SpringBoot to build the front and back ends, specifically including: a to-be-recommended dataset submission module, a recommendation result feedback module, a graph visualization module, a similar dataset and model retrieval module, a model and dataset upload module, and each module of the Web side interacts with the recommendation system through RESTful API, wherein: the dataset submission module and the recommendation result module interact with 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 KNN-based similar vector retrieval.

[0036] Through specific actual experiments, under the hardware environment of RTX-4090 GPU, the following experiments are performed: using the CIFAR-10 dataset to generate three different types of user task sets, including 45 2-class tasks, 252 5-class tasks, and 297 mixed-class tasks, and the candidate models come 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, the parameters of 200 rounds of training, Adam optimizer, and dynamic learning rate decay strategy are used for fine-tuning, and the fine-tuning accuracy of each candidate model on different task sets is obtained as the experimental benchmark. The experimental data is as follows: in the Top-3 model recommendation task, the average ratio of the recommendation accuracy of the system to the optimal accuracy on the three task sets reaches 99.87%; in the 2-model selection task, the system reaches an accuracy of 85.71%; in the time efficiency evaluation of the recommendation, the total time consumption of feature extraction and inference is about 76.48 seconds, which is significantly lower than that of the traditional fine-tuning method.

[0037] The comparison of the present system with the prior art is shown in Table 1.

[0038] Table 1 Comparison of technical characteristics

[0039]

[0040]

[0041] Compared with the prior art, the data entry module in the present application adopts an automatic parsing entry mode, obtains information by parsing the model specification document and dataset meta information, and supports distributed storage and real-time data update, with higher processing efficiency and automation degree; the feature extraction and aggregation module introduces a multi-element feature aggregation algorithm based on dimension scaling, which can deeply extract and weight balance various features of the model and dataset, and generate meta feature vectors with high expression capacity; in terms of graph construction and representation, by constructing a double graph representation learning based on model architecture cosine similarity and label weight, the complex multi-dimensional relationship between the model and the dataset can be captured, and the accuracy of the recommendation system can be improved; the intelligent recommendation module adopts a graph convolution network based on residual connection, combines a multilayer perceptron for regression prediction, and simultaneously performs recommendation by predicting the accuracy rate, to ensure the recommendation accuracy and the interpretability of the results. The model training optimization module trains the regression model by the accuracy rate based on fine-tuning data, and adopts Min-Max standardization processing to optimize the accuracy difference in the training process and improve the generalization ability of the trained model; the system is also seamlessly integrated into the upper-layer application through RESTful API, supports dynamic expansion of multiple data sources, is based on distributed storage and graph database, ensures good scalability and portability of the system, and the dynamically updated graph structure supports processing of dynamically updated dataset and model library, with wider applicability.

[0042] The above specific implementations can be adjusted in different ways by those skilled in the art without departing from the principles and purposes of the present application, the protection scope of the present application is subject to the claims and is not limited by the above specific implementations, and each implementation within the scope is subject to the constraints of the present application.

Claims

1. A recommendation system based on a dual-graph representation learning pre-training model, characterized in that: include: The module consists of a data entry module, a meta-feature extraction module, a graph construction module, and an intelligent recommendation module. The data entry module inputs model data and dataset data from an external data source, parses them, and outputs them to the meta-feature extraction module. The meta-feature extraction module extracts and aggregates meta-feature vectors of the model and dataset. The graph construction module constructs a dual-graph representation of the model graph and dataset graph based on the meta-feature vectors, calculates the architectural feature similarity between models, and analyzes the label weight relationship between datasets. The intelligent recommendation module uses a deep recommendation model based on residual graph convolution and multi-layer perceptron to predict the regression accuracy of the dataset based on the dual-graph representation and the corresponding meta-feature vectors, and then obtains a list of model recommendation results. 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 final loaded model content and data 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 dataset platform, and enters the metadata of the dataset by calling the full dataset metadata extraction function provided in the library; then, through the automatic loading method in the image dataset library, the image dataset content is loaded based on the extracted dataset metadata, and the final loaded dataset content and data will be input into the meta feature extraction module; 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 the multivariate feature splicing method to aggregate multiple features into a complete meta-feature.

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 dataset 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 is characterized in that: The extraction and aggregation mentioned above means that the data of the dataset and model will be subjected to multivariate feature extraction respectively. After extracting several key features of the model and dataset, 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 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 of 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.

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 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, while maintaining the original feature information through residual connection, with a total of K layers stacked; the regression score calculation model unit based on multi-layer perceptron, after obtaining the hidden layer graph representation output by the graph convolution model, 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.

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 feature extraction specifically includes: first, the graph convolution model accepts the input dual-graph representation, including the dataset graph and model graph ; Then, the nodes in the graph are aggregated through K layers of residual graph convolution layers. The graph convolution operation of each layer is It is Node after layer graph convolution The eigenvector of represents the k-th layer graph convolution operation, is the adjacency matrix of the graph, It is The residual weights of the layer are connected through stacking K layers of residual graph convolution to finally output the latent graph representation and , where: node The high-dimensional feature vector of ; 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. It's a model In the dataset The prediction accuracy on is a multi-layer perceptron regression model, The dataset and model The joint feature vector obtained by concatenating the high-dimensional latent vectors ; For the dataset to be recommended , by traversing the model library All models in , calculate the prediction accuracy of each model on the dataset , use accuracy to sort the recommended models, and output the Top-k recommended model list .

7. 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 In the model 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, It is a dataset In the model library The minimum fine-tuning accuracy in , It is a dataset In the model library The maximum fine-tuning accuracy in the training process uses the mean square error as the loss function, that is, ,in: is the standardized value of the model prediction accuracy, In order to actually fine-tune the accuracy after normalization, the model parameter update process uses a dynamic learning rate to ensure that the model can converge more quickly. The learning rate is dynamically adjusted based on the loss changes during training. ,in: is the learning rate reduction coefficient.

8. An application device based on the system according to any one of claims 1 to 7, characterized in that: include: The recommendation system based on a dual-graph representation learning pre-trained model and its upper-layer applications connected to it via a RESTful API interface include: a module for submitting recommended datasets, a module for feedback on recommendation results, a graph visualization module, a module for retrieving similar datasets and models, and a module for uploading datasets and models. Model and dataset files are stored using the distributed file storage system HDFS, model and dataset meta-feature vectors are stored and retrieved using the vector database Milvus, and the graph representations of the models and datasets are stored and retrieved using the graph database Neo4j. The web-based system uses the React+SpringBoot development framework for front-end and back-end construction, specifically including: a module for submitting recommended datasets, a module for feedback on recommendation 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-based system interacts with the recommendation system via a RESTful API. The dataset submission module interacts with the recommendation result module and the intelligent recommendation module. The graph visualization module subsequently reads and writes to the database using Neo4j's graph data transmission interface py2neo. The similar dataset and model retrieval module searches the vector database using a KNN-based similarity vector search.

Images