Intelligent labeling method and system in distributed data platform based on differential communication

Abstract

The application provides an intelligent labeling method and system in a distributed data middle platform based on differential communication, and the method comprises the following steps: building a distributed data middle platform framework; the distributed data middle platform comprises a total data middle platform and a sub data middle platform; a graph convolution network model of each sub data middle platform is established for intelligent labeling, a preset number of labels are taken as nodes, and the connection relationship between each node is obtained; based on the private data of each sub data middle platform, a private model is established, the private model is trained by using public data, and node interaction is performed based on the connection relationship between each node by judging the feature matrix similarity of the private model; wherein, the node interaction is a private model parameter; the public data and the trained private model are compared and differential analysis is performed, the node interaction time is obtained, and the intelligent labeling of the distributed data middle platform based on differential communication is completed. The technical scheme of the application can improve the consistency of intelligent labeling.

Description

Technical Field

[0001] This invention belongs to the field of intelligent annotation technology, specifically relating to an intelligent annotation method and system for a distributed data platform based on differential communication. Background Technology

[0002] Labeling is one of the core functions of a data platform. In a distributed data platform, there exist primary and secondary platforms. Despite continuous advancements in intelligent labeling technology, maintaining consistency in distributed intelligent labeling across a distributed platform remains a challenging problem that has not yet been adequately studied. This paper addresses the need for consistency in intelligent labeling within a distributed data platform. It also addresses the issues of limited computing resources and high communication overhead leading to low transmission efficiency due to the large volume and rapid updates of private data in each secondary data platform. Therefore, a method and system for intelligent labeling in a distributed data platform based on differential communication is proposed. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes an intelligent annotation method and system for a distributed data platform based on differential communication. This method solves the annotation consistency problem during the intelligent annotation process in the distributed data platform, while reducing the consumption of computing resources and communication overhead.

[0004] To achieve the above objectives, the present invention provides the following solution:

[0005] The intelligent annotation method for distributed data platforms based on differential communication includes the following steps:

[0006] A distributed data platform framework is established; the distributed data platform includes a main data platform and sub-data platforms.

[0007] Build graph convolutional network models for each sub-data platform for intelligent labeling, use a preset number of labels as nodes, and obtain the connection relationships between each node;

[0008] Based on the private data of each sub-data platform, a private model is established, public data is obtained to train the private model, and node interaction is performed by judging the similarity of the feature matrix of the private model based on the connection relationship between each node; wherein, the node interaction is the private model parameters.

[0009] By comparing the public data with the trained private model and performing a difference analysis, the node interaction time is obtained, and intelligent labeling of the distributed data platform based on difference communication is completed.

[0010] Preferably, a distributed data middleware framework is built using a cloud-edge-device three-layer structure;

[0011] The cloud layer consists of large cloud servers, serving as the central data platform;

[0012] The edge layer consists of edge servers, which serve as a sub-data platform.

[0013] The end layer consists of sensors and mobile devices, providing raw data.

[0014] Preferably, an interpretable distributed semi-supervised feature similarity fusion mechanism based on reputation value is incorporated into the graph convolutional network model to perform node classification prediction and obtain the connection relationship between each node; wherein, the graph convolutional network model includes a two-layer structure, one layer takes a preset number of nodes as input; the other layer takes the parameters of the private model as input.

[0015] The preferred method for node classification prediction is as follows:

[0016] In graph data of a graph convolutional network model, nodes represent labels, and edges represent the relationships between nodes;

[0017] Using state vectors to represent the state of nodes, and normalizing and aggregating the degrees of the node representations of the current node's neighbors, updating and training at each layer, we obtain the node representation that captures the features and structural information related to the task.

[0018] The graph convolutional network model is trained by inputting the feature weights of the node representations that capture task-related features and structural information. The graph convolutional network model is then interpreted by explaining the node relationships between the input feature weights, thus completing the classification and prediction of the nodes.

[0019] Preferably, the calculation of the node includes four parts: the node's feature vector, the feature vectors of neighboring nodes, the state vectors of neighboring nodes, and the feature vectors of the edges.

[0020] Preferably, the method for calculating the reputation value is as follows:

[0021] Calculate the differences between the private models and the differences between the public data;

[0022] The private data of the preset sub-data platform is substituted into the private model of other sub-data platforms for detection to obtain the difference value;

[0023] Based on all the obtained discrepancy values, a global reputation value is formed.

[0024] Preferably, the similarity of feature matrices is determined using minimum hashing. The specific method is as follows:

[0025] Establish a private model feature matrix, rearrange the feature matrix, and obtain the minimum hash signature composed of the minimum hash value;

[0026] Based on the minimum hash signature, obtain the minimum hash signature matrix;

[0027] Based on the minimum hash signature matrix, the similarity of the feature matrices is obtained.

[0028] This invention also provides an intelligent annotation system for a distributed data middleware platform based on differential communication, used to implement an intelligent annotation method, including:

[0029] The data middle platform framework building module is used to build a distributed data middle platform framework; the distributed data middle platform includes a main data middle platform and sub-data middle platforms.

[0030] The graph convolution module is used to build graph convolutional network models for each sub-data platform and perform intelligent labeling. It uses a preset number of labels as nodes to obtain the connection relationships between each node.

[0031] The private model building module is used to build a private model based on the private data of each sub-data platform, obtain public data to train the private model, and perform node interaction based on the connection relationship between each node by judging the similarity of the feature matrix of the private model; wherein, the node interaction is the private model parameters.

[0032] The difference analysis module is used to compare the public data and the trained private model and perform difference analysis to obtain the node interaction time and complete the intelligent labeling of the distributed data platform based on difference communication.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] 1) This invention proposes a method for intelligent annotation based on differential communication conditions through non-timed communication, which reduces communication overhead and computational consumption.

[0035] 2) This invention proposes a method combining GCN with matrix similarity, using an attention approach to solve dynamic node classification problems. Attention weights are learned as weights of a linear combination of node representations, while also considering attribute dynamics. An interpretable distributed semi-supervised feature similarity fusion mechanism is proposed, its most significant feature being interpretability, and this fusion mechanism is integrated into GCN.

[0036] 3) This invention proposes a novel metric: utilizing public data to evaluate the correlation between the importance at different time steps and the importance indicated by the learned attention weights, thereby quantifying the interpretability of a given classification model. This invention also proposes a reputation-based distributed semi-supervised fusion mechanism that leverages feature interpretability to fuse GCNs based on interpretable feature similarity. Attached Figure Description

[0037] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a flowchart of the intelligent annotation method in a distributed data middleware based on differential communication in an embodiment of the present invention;

[0039] Figure 2 This is a flowchart illustrating the information comparison process between the secondary middleware platforms in this embodiment of the invention.

[0040] Figure 3 This is a GCN network fusion diagram in an embodiment of the present invention;

[0041] Figure 4 This is a flowchart illustrating the process of obtaining the weights of each secondary middle platform in an embodiment of the present invention;

[0042] Figure 5 This is a graph showing the fusion optimization accuracy results in an embodiment of the present invention;

[0043] Figure 6 This is a trend graph of the loss function in an embodiment of the present invention;

[0044] Figure 7 This is a diagram illustrating the consistency results of timed communication in a distributed scenario according to an embodiment of the present invention.

[0045] Figure 8 This is the fused consensus graph in an embodiment of the present invention;

[0046] Figure 9 This is a graph showing the relationship between the number of updates and iterations after comparing the four datasets in this embodiment of the invention.

[0047] Figure 10 As in the embodiments of the present invention Figure 9 A diagram showing the corresponding number of interaction iterations;

[0048] Figure 11 The figure shows the comparative experimental results in the embodiments of the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] Example 1

[0052] like Figure 1 As shown, the intelligent annotation method in a distributed data middleware platform based on differential communication includes the following steps:

[0053] Build a distributed data platform framework; the distributed data platform includes a main data platform and sub-data platforms.

[0054] A further implementation method involves building a distributed data middleware framework with a cloud-edge-device three-layer structure.

[0055] The cloud layer consists of large cloud servers, which have abundant computing resources and data scheduling capabilities, and serve as the overall data platform;

[0056] The edge layer consists of edge servers, which serve as a sub-data platform.

[0057] The end layer consists of sensors and mobile devices, which are widely distributed and numerous, providing a large amount of raw data.

[0058] In this invention, each sub-data platform trains a private model for a subsidiary and shares a portion of its data as public data, which is then uploaded to the head office's data platform. To obtain a more accurate model, the accuracy and generalization of a model trained solely on such a small amount of real data are difficult to guarantee. Therefore, the framework utilizes the private models and public data of each subsidiary within each secondary platform. However, to reduce computational resources, each sub-platform does not exchange raw data with neighboring platforms, only exchanging private models and public data. First, the data within the secondary platforms is populated using interactive information to obtain relatively accurate populated data. From a communication perspective, the biggest problem in a distributed data platform is the continuous data sharing between secondary platforms, which incurs significant computational and communication overhead. While timed communication can ensure data label consistency, it easily leads to resource waste. Therefore, non-timed communication based on differential communication is used for information exchange between subsidiaries within each secondary platform, ensuring the consistency of data labeling among distributed subsidiaries.

[0059] So-called differential communication refers to the situation where communication between subsidiaries is typically timed. Although regular, this information exchange occurs even when the differences between models are not significant, leading to a serious waste of computing resources and increased communication overhead under bandwidth constraints. Therefore, this invention chooses to use public data as the design standard, comparing each model against it. Information exchange occurs only when the differences reach a certain threshold. This differential-based, non-timed communication can both ensure the consistency of intelligent annotation among subsidiaries in a distributed training system and reduce communication overhead, saving computing resources.

[0060] By combining prediction problems with node classification, a graph convolutional network model is built for each sub-data platform for intelligent labeling. A preset number of labels are used as nodes to obtain the connection relationships between each node.

[0061] A further implementation involves integrating a reputation-based, interpretable, distributed, semi-supervised feature similarity fusion mechanism into a graph convolutional network model to perform node classification prediction and obtain the connection relationships between nodes. The graph convolutional network model includes a two-layer structure: one layer takes a preset number of nodes as input, and the other layer takes the parameters of a private model as input.

[0062] In this embodiment, 10 feature weights are first trained using training data to represent the input part, and these 10 feature weights form a feature map. Each input weight is different. The training data is used to train the global reputation of a single company's global model (private model) by combining different proportions of the input feature weights, which can solve the problem of model type saturation. At the same time, since spatial information is preserved through convolutional layers, model mapping can be generated directly by inputting the 10 feature weights during update iterations. Feature maps with different model structures are formed by different proportions of weights. However, since the feature weights of the six subsidiaries are different, and the connection relationships between nodes are also different, the feature maps are not entirely the same, so the obtained trained model is not accurate. At the same time, there are information interaction and fusion relationships between the subsidiaries. Therefore, a fusion mechanism can be established by using the node interaction information of the models of different subsidiaries to improve the accuracy and consistency of the trained model.

[0063] A further implementation method involves performing node classification prediction as follows:

[0064] like Figure 3 As shown, in the graph data of the graph convolutional network model, node n represents the label, and the edge represents the relationship between nodes;

[0065] Using state vector Xn The state of a node is represented by the degree of the node representations of the current node's neighbors. The normalized aggregation is performed in each layer for updating and training to obtain a node representation that captures the feature and structural information related to the task. A further implementation method is that the computed node includes four parts: the feature vector of the node, the feature vectors of the neighboring nodes, the state vectors of the neighboring nodes, and the feature vectors of the edges (connected to n).

[0066] The trained graph convolutional network (GCN) model is fed with feature weights representing nodes that capture task-related features and structural information. The model interprets the node relationships between these feature weights to complete node classification predictions. GCN means that for a given point, the convolution of that point can be viewed as a weighted sum of its neighbors. Given a pre-trained model and 10 input feature weights, this invention interprets the model by explaining the node relationships between the input features. Intuitively, the explanations should capture features for both the model's predictions and the label predictions. Thus, when the image interpretations are provided to the pre-trained model, it should make predictions that are the same as or similar to those of the original image.

[0067] The graph convolutional network model predicts output Pr. This invention considers both labeled and unlabeled data predictions simultaneously, and calculates the loss using the mean absolute error (MAE) as the minimized loss function.

[0068] Regarding model accuracy, in this invention, intelligent annotation ultimately achieves accurate model annotation. In the distributed model, it is required that each subsidiary achieve consistency of intelligent annotation among the sub-models through non-timed communication of information exchange, thereby making the global model annotation more accurate.

[0069] Regarding regularization terms, their introduction can improve the model's generalization ability, enabling it to perform better on unseen data. It can effectively control model complexity, avoid overfitting, and balance the model's fitting and generalization abilities. Regularization measures are incorporated in this invention to ensure that the learned masks remain sparse and have better interpretability.

[0070] like Figure 4 As shown, the reputation (similar to weight) of each local model (secondary middle platform) is calculated, and then the 1-6 subsidiaries finally form the reputation of a global model.

[0071] Regarding smoothing, firstly, oversmoothing manifests as similarity between node embeddings. The goal of learning embeddings is to ultimately provide them to a classifier to predict their labels. Considering this oversmoothing effect, nodes without the same label will end up with similar embeddings, leading to mislabeling. Since convolution is an aggregation operation, taking specific values ​​for the convolution kernel constitutes a smoothing operation. Parameter sharing is crucial for convolution operations, and even more so in GCNs because the degree of each vertex on a graph is different. Research has revealed that the oversmoothing problem ultimately stems from the convolution kernel. This can be addressed by modifying the convolution kernel regularization to handle high-order operations, or by preserving the lower-level results and concatenating them into features, thus mitigating the oversmoothing problem. This can be achieved by performing convolutions at different neighbor scales and then fusing the convolution results from all scales to obtain the node's feature representation.

[0072] Based on the private data of each sub-data platform, a private model is established, and public data is obtained to train the private model. Based on the connection relationship between each node, node interaction is carried out by judging the similarity of the feature matrix of the private model. Among them, the node interaction is the private model parameters. Private data refers to the data that is only available in the local (secondary) data platform and is used to train the model locally.

[0073] A further implementation method involves using minimum hashing to determine the similarity of feature matrices. The specific method is as follows:

[0074] Establish a private model feature matrix, rearrange the feature matrix, and obtain the minimum hash signature composed of the minimum hash value;

[0075] Obtain the minimum hash signature matrix based on the minimum hash signature;

[0076] The similarity of feature matrices is obtained based on the minimum hash signature matrix.

[0077] By comparing public data with pre-trained private models and performing differential analysis, the node interaction time is obtained, and intelligent labeling of the distributed data platform based on differential communication is completed.

[0078] Example 2

[0079] Specifically, the process of determining the conditions for information exchange is as follows:

[0080] In this embodiment, the private data collected by the subsidiary has a local correlation with the private model and has a spatial range. To achieve model awareness and improve data relevance, model similarity is considered during reputation calculation in the process of data sharing between models. The higher the model similarity, the more relevant the data sharing from the data provider is, resulting in higher quality, more accurate, and more reliable data sharing. The subsidiary's reputation value is defined as {attention, parameter, MAE}, with the weights of the corresponding coefficients in reputation being ψ1, ψ2, and ψ3, where ψ1 + ψ2 + ψ3 = 1. For subsidiary 1, the similarity between the subsidiary and its own model in the previous time step is:

[0081] SIM(L1,J1)=1-DISS=reputation1,

[0082] Where L1 represents subsidiary 1, and J1 represents the previous state of the subsidiary;

[0083] 1) attention1: Based on the 10 feature weights formed by private data, which are equivalent to the topological structure formed by the connection between 10 nodes, attention is generated. The higher the attention, the better the reputation.

[0084] 2) parameter1: The parameters of the private model trained from the private data of subsidiary 1. The higher the parameter, the better the reputation.

[0085] 3) Substitute the validation set data into the trained private model and calculate the difference value (MAE). 11, Therefore, DISS(L1,J1) is the normalized difference between subsidiary 1 and its previous state, defined as:

[0086] DISS(L1,J1)=ψ1*attention1+ψ2*parameter1+ψ3*MAE 11, =reputation1

[0087] reputation2 now consists of the following parts:

[0088] 1) Substitute private data 1 into private model 2 → Calculate the MAE of the private data d12

[0089] 2) MMD between public data 1 and public data 2 1&2

[0090] 3) Differences between parameter1 and parameter2 (MAE) p1&2

[0091] Similarly, the similarity between subsidiary 1 and subsidiary 2 is calculated as SIM(L1,J2) = 1 - DISS(L1,J2).

[0092] Therefore, DISS(L1,J2) is the normalized difference between subsidiary 1 and subsidiary 2, defined as: DISS(L1,J2)=ψ1*MAE d12 +ψ2*MMD 1&2 +ψ3*MAE p1&2 =reputation2

[0093] Similarly, reputation6 now consists of the following parts:

[0094] 1) Substitute private data 1 into private model 6 → calculate the MAE d16

[0095] 2) MMD between public data 1 and public data 6 1&6

[0096] 3) Differences between parameter1 and parameter6 (MAE) p16

[0097] 4) Therefore, DISS(L1,J6) is the normalized difference between subsidiary 1 and subsidiary 6, defined as:

[0098] DISS(L1,J6)=ψ1*MAE d16 +ψ2*MMD 1&6 +ψ3*MAE p16

[0099] Each subsidiary interacts with Company 1 to recommend information. Therefore, subjective opinions from different subsidiaries (adjacent nodes) are integrated into a global opinion model, which is named the global model of the subsidiary based on the weight of each opinion.

[0100]

[0101] For subsidiary 1, the final reputation is:

[0102] Global1*finalrep1=reputation1*mod1+reputation2*mod2+...+rep6*mod6

[0103]

[0104] Because private data is constantly updated and changing during the integration and interaction between subsidiaries, the private models trained from this data are also constantly evolving. Reputation is affected by both time t and public data. Therefore, when calculating reputation values, it is necessary to determine the differences between private models and between public data. Furthermore, the private data must be substituted into the private models of other interacting subsidiaries for testing, and the resulting differences are then combined to form a global reputation value. This proposed method fully considers all aspects of the factors influencing reputation, making the calculated reputation value more convincing, improving the integration effect between models, and ultimately enhancing the consistency of labeling across subsidiaries.

[0105] like Figure 2 As shown, information comparison between secondary platforms involves evaluating the error between two local models based on several reference metrics, such as MAE and MMD. If the error reaches 5%, information exchange occurs. This error comparison is performed by selecting a portion of well-trained data as public data and then comparing them.

[0106] For the six subsidiaries, at time t:

[0107] The first row contains private data models.

[0108] The second row contains public data (the global model built from the first two rows).

[0109] The third line contains the private data at time t.

[0110] At time t+n, subsidiary 1 has private models 1-6 (only knowing the information after the interaction at time t, and determining when to interact at the next time, i.e., determining what n equals before updating), public data 1-6, and its own private data.

[0111] Information exchange conditions:

[0112] At time t+1, Company 1's private data has been updated, its public data has also been updated (but without interaction), and its private model has been updated. At this time, reputation1 has also been updated (because the MAE (Mean Absolute Error) is calculated by substituting private data into other companies' private models, and the minhash is calculated using Company 1's own private model and other companies' private models to calculate the difference, and the MMD (Maximum Mean Difference) is calculated using Company 1's own public data and other companies' public data, therefore its reputation is updated in real time. It's just that there's no interaction yet; now we need to determine how much reputation has changed before interacting). — The condition for determining when to interact.

[0113] The two secondary middleware platforms are measured using the Maximum Mean Median Discrepancy (MMD) and the Mean Absolute Error (MAE).

[0114]

[0115]

[0116]

[0117] This invention utilizes prediction accuracy to quantify the performance of a well-trained local model. More specifically, in classification during training, accuracy is represented by the score of correctly classified records. In regression tasks, accuracy is measured by the mean absolute error (MAE).

[0118] Among them, y i For model m i The predicted value, x i The recorded true value. Model m i The lower the MAE, the more m i The higher the accuracy, the better.

[0119] f i This is a quantitative formula for measuring the quality of a model. The first term... Model accuracy, also known as model precision, represents the ability of a predictive model to correctly assign instances to their respective classes. It quantifies the closeness between the model's predictions and the true values. Higher accuracy indicates higher predictive precision, while lower accuracy signifies lower reliability. (Second term) This is the regularization term, a term typically appended to the loss function and directly linked to the model parameters. Its main goal is to reduce overfitting, which occurs when a model performs well on training data but poorly on test data. Commonly used regularization techniques include L1 regularization and L2 regularization, which penalize model parameters using the L1 and L2 norms, respectively. L1 regularization promotes feature selection by encouraging some model parameters to reach zero, while L2 regularization reduces model complexity by driving model parameters closer to zero. The optimal selection and fine-tuning of the regularization term should be based on the specific problem and characteristics of the dataset at hand. To ensure sparsity and improve interpretability, this invention integrates regularization measurements. Smoothing. The problem of oversmoothing manifests as similarity between observed node embeddings. The ultimate goal of embedding learning is to use them for classification tasks to predict labels. However, this oversmoothing phenomenon can lead to the generation of similar embeddings for nodes that do not share the same labels, resulting in misclassification. Since convolutions inherently involve aggregation, they produce a smoothing effect when the convolutional kernel assumes specific values. Given the crucial role parameter sharing plays in convolutional operations, its importance becomes even more pronounced within the framework of Graph Convolutional Networks (GCNs) due to variations in vertex degree in the graph.

[0120] At this time, reputation t+1 You can add a similar reputation to the original one. t The degree of difference assessment. Could it be that there are three differences: 1) the difference between the subsidiary and the original, 2) the difference with reputation2, and 3) the difference with reputation3?

[0121] Δ=Δreputation 1(t+1) With reputation 1t +Δreputation 2t With reputation t +Δreputation 3t With reputation 1t .

[0122] Example 3

[0123] Experimental process using the technical solution of this invention:

[0124] Dataset Acquisition: The NGSIM (Next Generation Simulation) dataset is US highway traffic data collected by FHWA, including driving conditions of all vehicles traveling on US101, I-80, and I-80 over a period of time. The data was acquired via cameras and then processed into trajectory points. This embodiment of the invention selects three well-known trajectory prediction datasets: NGSIM I-80, US-101, and the Apollo Landscape Trajectory Dataset.

[0125] Data processing: First, divide the dataset into 10 parts. For example, set n subsidiaries (local models) as 10 local models (secondary middleware). Record the differences between the datasets at time t. This difference is viewed using the reputation based on attention mentioned earlier.

[0126] The global parameters are then fed into P1 (parameter1, the parameters trained by the private model 1) in the sub-network, while P2 is still optimized using its own local data → P2'. This P2' will definitely have higher prediction accuracy than the previous P2 because it is a fused result. This demonstrates the feasibility of the experimental method of this invention. The reputation-based fusion mechanism proposed in this invention can indeed improve the accuracy of intelligent model annotation. Figure 5 As shown.

[0127] like Figure 6 As shown, the loss function decreases continuously with the increase of the number of iterations, and eventually tends to stabilize.

[0128] Figure 7 This is a consistency result diagram for timed communication in a distributed scenario. From the diagram, we can conclude that 5 updates are required, the consistency probability is not high, and it does not meet the condition of being no less than 95%.

[0129] Figure 8 After integration, the information exchange between subsidiaries can maintain consistency of over 95%. Timed communication updates only 4 times, all of which meet the consistency condition of over 95%, improving consistency accuracy, reducing update time, and reducing computational overhead.

[0130] Figure 9Comparing the update counts and iteration relationships of the four datasets, it is evident that these four datasets maintain an accuracy consistency of over 95% while reducing the number of information interactions and computational communication overhead. Surprisingly, the fourth dataset only requires three interactions. Training within the architecture model designed in this invention, and using the technical solution of this invention, it was found that not only is flexible information interaction achieved, enabling retraining even when the difference reaches 5%, but the fourth subsidiary, in particular, only requires three interactions to achieve both non-timed information interaction based on differences, further reducing the number of interactions, saving bandwidth resources, and reducing computational overhead.

[0131] exist Figure 10 It can be seen from this that, with Figure 9 The total number of interactions and the number of iterations at which interactions occur are corresponding. Figures (a), (b), (c), and (d) show the relationship curves between the number of interaction iterations and the loss rate for the four secondary middleware platforms. For example, in Figure (a), information interaction occurs in the 4th iteration. Figure 9 The first "+" sign in the diagram corresponds to the previous one, and so on. Meanwhile, it can be observed that as the number of training iterations increases, the model continuously improves. As shown in Figure (d), in the fourth experiment, only three interactions are needed to achieve the desired experimental results, effectively saving computational resources and communication overhead, and improving the value of data utilization.

[0132] like Figure 11 As shown in the comparative experiments, the results trained on different datasets all show an upward trend, while the results trained on other datasets tend to be consistent. The method proposed in this paper is significantly better than the results trained by other methods. It not only achieves a consistency of over 95% in the annotation training of the distributed data platform, but also exceeds the expected results, proving the feasibility of the method in this paper.

[0133] Example 4

[0134] This invention also provides an intelligent annotation system for a distributed data middleware platform based on differential communication, used to implement an intelligent annotation method, including:

[0135] The data middle platform framework building module is used to build a distributed data middle platform framework; the distributed data middle platform includes a main data middle platform and sub-data middle platforms.

[0136] The graph convolution module is used to build graph convolutional network models for each sub-data platform and perform intelligent labeling. It uses a preset number of labels as nodes to obtain the connection relationships between each node.

[0137] The private model building module is used to build private models based on the private data of each sub-data platform, train private models with public data, and perform node interaction based on the connection relationship between each node by judging the similarity of the feature matrix of the private model; among them, the node interaction is the private model parameters.

[0138] The difference analysis module is used to compare public data and trained private models and perform difference analysis to obtain node interaction time and complete intelligent labeling of the distributed data platform based on difference communication.

[0139] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for intelligent annotation in a distributed data middleware platform based on differential communication, characterized in that: Includes the following steps: A distributed data platform framework is established; the distributed data platform includes a main data platform and sub-data platforms. Build graph convolutional network models for each sub-data platform for intelligent labeling, use a preset number of labels as nodes, and obtain the connection relationships between each node; Based on the private data of each sub-data platform, a private model is established, public data is obtained to train the private model, and node interaction is performed by judging the similarity of the feature matrix of the private model based on the connection relationship between each node; wherein, the node interaction is the private model parameters. By comparing the public data with the trained private model and performing a difference analysis, the node interaction time is obtained, and intelligent labeling of the distributed data platform based on differential communication is completed. Build a distributed data middleware framework using a cloud-edge-device three-layer structure; The cloud layer consists of large cloud servers, serving as the central data platform; The edge layer consists of edge servers, which serve as a sub-data platform. The edge layer consists of sensors and mobile devices, providing raw data; The interpretable distributed semi-supervised feature similarity fusion mechanism based on reputation value is incorporated into the graph convolutional network model to perform node classification prediction and obtain the connection relationship between each node. The method for calculating the reputation value is as follows: The reputation value of the sub-data platform is calculated as reputation = {attention, parameter, MAE}, where the weights of the corresponding coefficients in reputation are... ,in For sub-data platform 1, the similarity between the sub-data platform and its own model in the previous time step is: ， in, Representative data platform 1, This represents the state of the sub-data platform at a previous moment; Attention, in other words, is generated based on 10 feature weights formed from private data, which are equivalent to the topological structure formed by the connection methods between 10 nodes. In other words, parameters: private model parameters trained from the private data of sub-data platform 1; The validation set data is fed into the trained private model to calculate the discrepancy value. ; therefore, The normalized difference between sub-data platform 1 and its previous state is defined as: It consists of the following parts: Substituting private data 1 into private model 2 → deriving the private data ; Between public data 1 and public data 2 ; and Differences between Similarly, the similarity between sub-data platform 1 and sub-data platform 2: ， The normalized difference between sub-data platform 1 and sub-data platform 2 is defined as: ； It consists of the following parts: Substituting private data 1 into private model 6 → the result ; Between public data 1 and public data 6 ; and Differences between ; The normalized difference between sub-data platform 1 and sub-data platform 6 is defined as: ； Each sub-data platform interacts with sub-data platform 1 to recommend information. Therefore, subjective opinions from different sub-data platforms are integrated into a global opinion model, which is named the global model of the sub-data platform according to the weight of each opinion. For sub-data platform 1, the final global reputation value is: 。 2. The intelligent annotation method for a distributed data middleware platform based on differential communication according to claim 1, characterized in that, The graph convolutional network model includes a two-layer structure: one layer takes a preset number of nodes as input, and the other layer takes the parameters of the private model as input.

3. The intelligent annotation method for a distributed data middleware platform based on differential communication according to claim 1, characterized in that, The specific method for node classification prediction is as follows: In graph data of a graph convolutional network model, nodes represent labels, and edges represent the relationships between nodes; Using state vectors to represent the state of nodes, and normalizing and aggregating the degrees of the node representations of the current node's neighbors, updating and training at each layer, we obtain the node representation that captures the features and structural information related to the task. The graph convolutional network model is trained by inputting the feature weights of the node representations that capture task-related features and structural information. The graph convolutional network model is then interpreted by explaining the node relationships between the input feature weights, thus completing the classification and prediction of the nodes.

4. The intelligent annotation method for a distributed data middleware platform based on differential communication according to claim 3, characterized in that, The calculation of a node includes four parts: the node's feature vector, the feature vectors of its neighboring nodes, the state vectors of its neighboring nodes, and the feature vectors of its edges.

5. The intelligent annotation method for a distributed data middleware platform based on differential communication according to claim 3, characterized in that, The similarity of feature matrices is determined using minimum hashing. The specific method is as follows: Establish a private model feature matrix, rearrange the feature matrix, and obtain the minimum hash signature composed of the minimum hash value; Based on the minimum hash signature, obtain the minimum hash signature matrix; Based on the minimum hash signature matrix, the similarity of the feature matrices is obtained.

6. An intelligent annotation system for a distributed data middleware platform based on differential communication, characterized in that: To implement the intelligent annotation method according to any one of claims 1-5, comprising: The data middle platform framework building module is used to build a distributed data middle platform framework; the distributed data middle platform includes a main data middle platform and sub-data middle platforms. The graph convolution module is used to build graph convolutional network models for each sub-data platform and perform intelligent labeling. It uses a preset number of labels as nodes to obtain the connection relationships between each node. The private model building module is used to build a private model based on the private data of each sub-data platform, obtain public data to train the private model, and perform node interaction based on the connection relationship between each node by judging the similarity of the feature matrix of the private model; wherein, the node interaction is the private model parameters. The difference analysis module is used to compare the public data and the trained private model and perform difference analysis to obtain the node interaction time and complete the intelligent labeling of the distributed data platform based on difference communication.

Images