Industrial internet system based on data mining technology

Abstract

The application relates to the technical field of data mining, in particular to an industrial internet system based on a data mining technology, which comprises a data acquisition module, a node construction module, a node quality analysis module and a model training module, wherein the data acquisition module acquires equipment data and sensor data in an industry; the node construction module constructs nodes according to the acquired data, performs data enhancement on the nodes, forms positive sample pairs and negative sample pairs; the node quality analysis module calculates feature differences of the positive sample pairs, repairs the nodes according to the feature difference results, generates complete positive sample pairs; calculates feature differences of the negative sample pairs, repairs the nodes according to the feature difference results, and generates complete negative sample pairs; and the model training module initializes parameters in a graph neural network model, and calculates the similarity of the complete positive sample pairs and the similarity of the complete negative sample pairs, so as to optimize the parameters in the graph neural network model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data mining, and particularly to an industrial internet system based on data mining technology. BACKGROUND

[0002] Traditional industrial internet systems often face multiple technical problems when processing device data and sensor data, which directly affect the training effect and actual application performance of the model.

[0003] Firstly, the problem of data quality is particularly prominent. In traditional systems, device data and sensor data are usually directly used for model training without detailed quality assessment and repair. Due to the presence of noise, missing values and other abnormal situations in the data collection process in industrial environments, the node feature information in the positive sample pair and the negative sample pair is not consistent or the difference is not obvious. This data quality problem will cause the model to fail to accurately capture the internal real behavior pattern of the device or sensor during the training process, thereby affecting the precision and robustness of the model. For example, in anomaly detection, low-quality positive samples may cause the model to misreport normal behavior as abnormal, increasing unnecessary maintenance costs and production interruptions; in fault prediction, the model may delay fault detection because it cannot accurately identify the device state, missing the best preventive maintenance opportunity.

[0004] Secondly, the traditional system lacks effective data augmentation methods in the generation and processing of sample pairs. Due to the diversity and complexity of industrial data, lack of diverse sample pairs will make the model have weak generalization ability when facing new data, making it difficult to cope with various situations in actual production. For example, in state recognition, the model may fail to accurately determine the true working state of the device due to the lack of diversity in the training data, resulting in poor production monitoring results; in device optimization scheduling, the model may produce mis-scheduling due to its inability to effectively distinguish the states of different devices, reducing production efficiency.

[0005] In addition, the traditional system often does not fully consider the contrast balance problem of positive sample pairs and negative sample pairs in the design of loss function and parameter optimization process, which may lead to inaccurate distinction between positive samples and negative samples during the training process, affecting the final performance of the model. For example, in state recognition, the imbalance of similarity will cause the model to produce deviation when judging the device state; in fault prediction, the inaccurate similarity will make the model unable to effectively distinguish between normal and fault states, thereby affecting the reliability of the prediction.

[0006] In summary, the traditional industrial internet system has technical problems in data quality, sample pair generation and loss function optimization, which limits the training effect and actual application performance of the model, and urgently needs to be improved through more detailed and scientific methods. SUMMARY

[0007] The present application aims to provide an industrial internet system based on data mining technology to solve the problems raised in the background art.

[0008] To achieve the above-mentioned purpose, the present application provides the following technical solutions: an industrial internet system based on data mining technology, comprising a data acquisition module, a node construction module, a node quality analysis module and a model training module, wherein:

[0009] The data acquisition module acquires device data and sensor data in the industry;

[0010] The node construction module constructs nodes according to the device data and the sensor data, performs different data augmentation on the same nodes to form positive sample pairs, and performs the same data augmentation on different nodes to form negative sample pairs;

[0011] The node quality analysis module acquires the features of the nodes in the positive sample pairs and the negative sample pairs, calculates the feature difference of the features of the same nodes in the positive sample pairs, determines the node quality of the same nodes in the positive sample pairs according to the result of the feature difference, and generates complete positive sample pairs according to the node quality; the feature difference of the features of the different nodes in the negative sample pairs is calculated, the node quality of the different nodes in the negative sample pairs is determined according to the result of the feature difference, and complete negative sample pairs are generated according to the node quality;

[0012] The model training module initializes the parameters in the graph neural network model, calculates the similarity of the complete positive sample pairs, and calculates the similarity of the complete negative samples; through back propagation and gradient descent, the similarity between the complete positive sample pairs is maximized, and the similarity between the complete negative sample pairs is minimized, so as to optimize the parameters in the graph neural network model.

[0013] As a further improvement of the present technical solution, the process of forming positive sample pairs and negative sample pairs by the node construction module specifically comprises:

[0014] The devices and sensors in the industry are all regarded as a node in the graph, and each node is represented by a feature vector containing all features extracted from the device data and the sensor data;

[0015] different data augmentations are performed on the same node in the graph to generate different versions of the node; different data augmentations are performed on the same node in the graph through the data augmentation methods of feature masking and feature perturbation; for feature masking, the proportion of features that need to be masked is determined, and a proportion of features are randomly selected and the values of the features are set to random values to simulate data noise; for feature perturbation, the distribution parameters of data noise are determined, and random noise of the distribution is added to each feature value to simulate errors in the data acquisition process; the positive sample pair is formed by different data augmentation versions of the same node;

[0016] two different nodes are selected from the graph, and the two nodes are not directly connected neighbor nodes in the original graph; the same data augmentation method is used on the two different nodes to generate data augmentation versions of the two different nodes, wherein the data augmentation method includes feature masking; the data augmentation versions of the two different nodes are used as a negative sample pair.

[0017] As a further improvement of the technical solution, the node quality analysis module includes a positive sample pair analysis unit, and the process of generating a complete positive sample pair specifically includes:

[0018] The feature vectors of the two different versions of the same node in the positive sample pair are compared, and different distance measurement methods are used to calculate the feature difference between the feature vectors;

[0019] According to the result of the feature difference, the quality of the same node in the positive sample pair is evaluated; wherein the smaller the feature difference is, the higher the node quality is; by setting a positive sample pair node threshold, it is determined whether the node in the positive sample pair is a low-quality node, and the node with a feature difference greater than the positive sample pair node threshold is regarded as a low-quality node;

[0020] The low-quality node in the positive sample pair is repaired to generate a complete positive sample pair, different methods are selected by re-augmenting the data of the node until the feature difference meets the positive sample pair node threshold requirement; the repaired low-quality node in the positive sample pair is filled into the positive sample pair to generate a complete positive sample pair.

[0021] As a further improvement of the technical solution, the node quality analysis module includes a negative sample pair analysis unit, and the process of generating a complete negative sample pair specifically includes:

[0022] The feature vectors of the two augmented versions of the different nodes in the negative sample pair are compared, and the distance measurement method is used to calculate the feature difference between the feature vectors;

[0023] According to the result of the feature difference, the quality of different nodes in the negative sample pair is evaluated, the greater the feature difference, the higher the node quality; by setting a negative sample pair node threshold, whether the node in the negative sample pair is a low-quality node is determined, and the node with a feature difference less than the negative sample pair node threshold is regarded as a low-quality node;

[0024] The low-quality nodes in the negative sample pair are repaired to generate a complete negative sample pair, the feature values are adjusted by increasing noise to increase the feature difference, so as to repair; the repaired low-quality nodes in the negative sample pair are filled into the negative sample pair to generate a complete negative sample pair.

[0025] As a further improvement of the technical solution, the process of optimizing the parameters of the graph neural network model by the model training module specifically includes:

[0026] A graph neural network model is selected, and the parameters in the graph neural network model are initialized, wherein the initialization method includes random initialization;

[0027] The similarity of the complete positive sample pair and the similarity of the complete negative sample are calculated according to the distance measurement method used in the node quality analysis module; the calculated distance is taken as the similarity of the corresponding sample pair;

[0028] A loss function is defined, the loss value under the parameters of the graph neural network model is calculated according to the loss function, the similarity of the complete positive sample pair is divided by an adjustment parameter, and an exponential value is taken; the similarity of the complete negative sample pair is divided by the adjustment parameter, and an exponential value is taken; and the exponential value sum is obtained by adding the above exponential values; the sum of the exponential value of the similarity of the complete positive sample and the exponential value of the similarity of the complete negative sample is added to obtain the denominator, and the proportion of the exponential value of the similarity of the complete positive sample in the exponential value sum is calculated, and the final loss value is obtained by taking the negative logarithm; wherein the adjustment parameter is used to adjust the distribution of the similarity, so as to balance the contrast between the positive sample and the negative sample;

[0029] The gradient of the loss value to the parameters of the graph neural network model is calculated, the model parameters are updated using the gradient descent algorithm, so that the similarity between the complete positive sample pairs is maximized and the similarity between the complete negative sample pairs is minimized, so as to optimize the parameters in the graph neural network model.

[0030] Compared with the prior art, the beneficial effects of the present application are:

[0031] The industrial internet system based on data mining technology significantly improves the data quality for training the graph neural network model through meticulous quality evaluation and repair of positive sample pairs and negative sample pairs. Specifically, for low-quality nodes in positive sample pairs, the system compares the feature differences of the same node under different data enhancement versions; if the feature difference exceeds the preset positive sample pair node threshold, the node will be marked as a low-quality node and re-enhanced through data until the quality requirements are met; this process ensures the consistency and reliability of the nodes in the positive sample pairs, enabling the model to more accurately capture the internal real behavior patterns of devices or sensors during the training process; for example, in anomaly detection, high-quality positive sample pairs enable the model to more accurately identify device behavior in normal states, significantly reducing false positive rates; in fault prediction, potential faults can be detected earlier, allowing preventive measures to be taken in advance; overall, the repair process of the node quality analysis module effectively improves the quality of the data set, providing a reliable data foundation for model training.

[0032] In addition, by initializing the parameters in the graph neural network model and calculating the similarity of complete positive sample pairs and negative sample pairs, and then optimizing the model parameters through backpropagation and gradient descent algorithms to maximize the similarity between positive sample pairs and minimize the similarity between negative sample pairs, this process significantly enhances the model's discrimination ability and learning efficiency; in practical applications, maximizing the similarity of positive sample pairs ensures that the model can better understand the consistency of the same device or sensor under different conditions, improving the model's accuracy; for example, in state recognition, the model can more accurately determine whether a device is in normal working condition, thereby improving the monitoring effect of the production process; minimizing the similarity of negative sample pairs helps the model to distinguish different devices or sensors, enhancing the system's robustness and generalization ability; for example, in device optimization scheduling, the model can more accurately allocate resources and tasks, avoiding mis-scheduling due to similar device states, significantly improving production efficiency; overall, the optimization process of the model training module not only improves the real-time responsiveness and accuracy of the industrial internet system, but also enhances the system's prediction and decision support capabilities, having a significant impact on the optimization of the entire industrial production process. BRIEF DESCRIPTION OF DRAWINGS

[0033] Fig. 1 is the overall module schematic diagram of the present application;

[0034] Fig. 2 is the node quality analysis module unit schematic diagram of the present application.

[0035] In the figure: 100, data acquisition module; 200, node construction module; 300, node quality analysis module; 301, positive sample pair analysis unit; 302, negative sample pair analysis unit; 400, model training module. DETAILED DESCRIPTION

[0036] The technical solutions in the embodiments of the present application will be apparently and completely described with reference to the drawings in the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all the other embodiments obtained by a person of ordinary skill in the art without any creative work fall within the protection scope of the present application.

[0037] Next, please refer to Figs. 1-2 The present application provides a technical solution: an industrial internet system based on data mining technology, comprising a data acquisition module 100, a node construction module 200, a node quality analysis module 300 and a model training module 400.

[0038] The data acquisition module 100 acquires corresponding equipment data and sensor data from equipment and sensors in the industry, the equipment data includes running state data based on time series (such as pressure and vibration at each time point) and the sensor data includes environmental information (such as temperature and humidity), and the equipment running state information and environmental information in the equipment data and sensor data are respectively standardized to ensure that the data has a uniform scale, thereby providing a guarantee for constructing reliable nodes and sample pairs.

[0039] The node construction module 200 constructs nodes according to the equipment data and sensor data, performs different data augmentation on the same node to form a positive sample pair, and performs the same data augmentation on different nodes to form a negative sample pair, wherein the process of forming the positive sample pair and the negative sample pair specifically comprises:

[0040] The equipment and sensors in the industry are all regarded as a node in the graph, which has equipment data and sensor data related thereto; features are extracted from the equipment data and sensor data, for time series data therein, statistical features (such as mean, variance, maximum, minimum) and time and frequency domain features (such as Fourier transform, wavelet transform) are used; each node is represented by a feature vector, which contains all the features extracted from the equipment data and sensor data;

[0041] Different data augmentations are performed on the same node in the graph to generate different versions of the node. The purpose of data augmentation is to increase the diversity of data, enabling the model to better generalize. Different data augmentations are performed on the same node in the graph through feature masking and feature perturbation data augmentation methods. For feature masking, first determine the proportion of features that need to be masked, for example 20%, then randomly select the proportion of features and set the value of the feature to 0 or a random value to simulate data missing or noise. For feature perturbation, first determine the distribution parameters of the noise, for example the mean of Gaussian noise is 0 and the standard deviation is 0.1, then add random noise of the distribution to each feature value to simulate errors in the data acquisition process.

[0042] The positive sample pair is formed by different data augmentation versions of the same node, where the data augmentation versions are different in features but still represent the same physical device or sensor in essence.

[0043] Two different nodes are selected from the graph, ensuring that the two nodes are not directly connected neighbor nodes in the original graph or that they have significant differences in feature space. The nodes are selected to ensure that there is sufficient difference between them, so as to form an effective negative sample pair in the subsequent steps.

[0044] The same data augmentation method is used on the two different nodes to generate data augmentation versions of the two different nodes, where the data augmentation method includes feature masking. Although the two nodes have undergone the same data augmentation method, due to their different original features or topological structures, there are still significant differences between the generated augmented versions. Therefore, the data augmentation versions of the different nodes are considered as negative sample pairs. Negative sample pairs are used in contrastive learning to emphasize the differences between different nodes, helping the model better learn node representation and generalization ability.

[0045] The positive sample pair analysis unit 301 in the node quality analysis module 300 calculates the feature difference of the features of the same node in the positive sample pair, determines the node quality of the same node in the positive sample pair according to the result of the feature difference, and performs positive sample pair node repair according to the node quality to generate a complete positive sample pair, which specifically includes:

[0046] Compare the feature vectors of the two different versions of the same node in the positive sample pair, and use different distance measurement methods such as Euclidean distance, cosine similarity, etc. to calculate the feature difference between the feature vectors.

[0047] According to the result of the feature difference, the quality of the same node in the positive sample pair is evaluated; the smaller the feature difference is, the higher the node quality is, because it means that the enhanced version of the data can still retain the main feature information of the original node; by setting a positive sample pair node threshold, it is determined whether the node in the positive sample pair is a low-quality node, for example, a node with a feature difference greater than the positive sample pair node threshold 0.2 is considered a low-quality node;

[0048] The low-quality nodes in the positive sample pair are repaired to generate a complete positive sample pair, by re-enhancing the data of the nodes and selecting different methods until the feature difference meets the positive sample pair node threshold requirement; if the feature difference is caused by the absence or abnormality of certain feature values, the quality of the nodes is improved by supplementing the missing data or correcting the abnormal data; the repaired low-quality nodes in the positive sample pair are filled into the positive sample pair to generate a complete positive sample pair.

[0049] The positive sample pair analysis unit 301 ensures that the feature difference of the nodes in the positive sample pair is small by evaluating and repairing the nodes in the positive sample pair, and retains the main feature information of the nodes; this helps to more accurately learn the similarity of the positive sample pair in the subsequent model training process, improves the training effect and reliability of the model, solves the problem of excessive feature difference of the nodes in the positive sample pair, ensures the quality of the positive sample pair, and thus helps the model to better learn the feature representation of the positive sample pair.

[0050] The negative sample pair analysis unit 302 in the node quality analysis module 300 calculates the feature difference of the features of different nodes in the negative sample pair, determines the node quality of the different nodes in the negative sample pair according to the result of the feature difference, and repairs the negative sample pair nodes according to the node quality to generate a complete negative sample pair, which specifically includes:

[0051] The feature vectors of the two enhanced versions of different nodes in the negative sample pair are compared, and a distance measurement method (such as Euclidean distance, cosine similarity) is used to calculate the feature difference between the feature vectors;

[0052] According to the result of the feature difference, the quality of the different nodes in the negative sample pair is evaluated; the greater the feature difference is, the higher the node quality is, because it means that the feature information of the two different nodes is significantly different; by setting a negative sample pair node threshold, it is determined whether the node in the negative sample pair is a low-quality node, for example, a node with a feature difference less than 0.8 is considered a low-quality node;

[0053] The low-quality nodes in the negative sample pair are repaired to generate a complete negative sample pair by adjusting the feature values to increase the feature difference; for example, the adjustment can be made by increasing the noise or selecting more representative features; the repaired low-quality nodes in the negative sample pair are filled into the negative sample pair to generate a complete negative sample pair.

[0054] The negative sample pair analysis unit 302 ensures that the node feature difference in the negative sample pair is significant by evaluating and repairing the nodes in the negative sample pair, and enhances the contrast between different nodes; this helps the model to learn the difference of the negative sample pair more accurately during the training process, improves the training effect and generalization ability of the model, solves the problem of insignificant node feature difference in the negative sample pair, ensures the quality of the negative sample pair, and thus helps the model to better learn the difference of the negative sample pair.

[0055] The model training module 400 initializes the parameters in the graph neural network model, and calculates the similarity of the complete positive sample pairs and the similarity of the complete negative samples; through back propagation and gradient descent, the similarity between the complete positive sample pairs is maximized, and the similarity between the complete negative sample pairs is minimized, to optimize the parameters in the graph neural network model, specifically including:

[0056] Select a graph neural network model, such as a graph convolution network and an attention network; initialize the parameters in the graph neural network model, including weight matrices and bias terms, where the initialization method includes random initialization;

[0057] According to the distance measurement method used in the node quality analysis module 300, the similarity of the complete positive sample pairs and the similarity of the complete negative samples are calculated; the calculated distance is taken as the similarity of the corresponding sample pair;

[0058] Define a loss function, calculate the loss value under the graph neural network model parameters according to the loss function, divide the similarity of the complete positive sample pairs by an adjustment parameter, and take the exponential value; divide the similarity of the complete negative sample pairs by the adjustment parameter, and take the exponential value; and add the above exponential values to obtain the total sum of exponential values; add the exponential value of the complete positive sample similarity and the exponential value of the complete negative sample similarity to obtain the denominator, and calculate the proportion of the exponential value of the complete positive sample similarity in the total sum of exponential values, and take the negative logarithm to obtain the final loss value; where the adjustment parameter is to adjust the distribution of the similarity, so as to better balance the contrast between the positive samples and the negative samples;

[0059] Calculate the gradient of the loss value to the graph neural network model parameters, and update the model parameters using the gradient descent algorithm (such as SGD, Adam, etc.) to maximize the similarity between the complete positive sample pairs and minimize the similarity between the complete negative sample pairs, to optimize the parameters in the graph neural network model;

[0060] The model is optimized by maximizing the similarity between complete positive sample pairs and minimizing the similarity between complete negative sample pairs, by calculating the similarity between a node and multiple negative sample nodes and comparing it with the similarity of the positive sample nodes, so as to ensure that the feature representations of the complete positive sample pairs are closer, and the feature representations of the complete negative sample pairs are farther away; so that the graph neural network model can better distinguish the feature representations of different nodes in the training process, and the optimized model has stronger generalization ability and robustness, and can perform well on new data; solve the problem that the model cannot effectively distinguish positive sample pairs and negative sample pairs in the training process, and by optimizing the loss function, the model can more accurately learn the feature representation of the node, improve the performance and generalization ability of the model.

[0061] The basic principles, main features and advantages of the present application are shown and described above. Those skilled in the art should understand that the present application is not limited by the above examples, and the above examples and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. An industrial internet system based on data mining technology, characterized in that: It includes a data acquisition module (100), a node construction module (200), a node quality analysis module (300), and a model training module (400), wherein: The data acquisition module (100) acquires equipment data and sensor data in the industry; The node construction module (200) constructs nodes based on device data and sensor data, performs different data augmentations on the same node to form positive sample pairs, and performs the same data augmentation on different nodes to form negative sample pairs; specifically including: The equipment and sensors in the industry are all regarded as nodes in the graph, and each node is represented by a feature vector containing all the features extracted from the equipment data and sensor data. Different data augmentations are applied to the same node in the graph to generate different versions of that node. Data augmentation methods, including feature masking and feature perturbation, are used to perform different data augmentations on the same node. For feature masking, the proportion of features to be masked is determined, and features of that proportion are randomly selected and their values ​​are set to random values ​​to simulate data noise. For feature perturbation, the distribution parameters of the data noise are determined, and random noise of that distribution is added to each feature value to simulate errors during data acquisition. Positive sample pairs are formed from the different data augmentation versions of the same node. Select two distinct nodes from the graph that are not directly connected neighbors in the original graph; apply the same data augmentation method to these two distinct nodes to generate two different data-augmented versions of the nodes, where the data augmentation method includes feature masking; use these two different data-augmented versions of the nodes as negative sample pairs. The node quality analysis module (300) acquires the features of nodes in positive and negative sample pairs, calculates the feature differences of the same node in the positive sample pair, determines the node quality of the same node in the positive sample pair based on the result of the feature differences, and repairs the nodes in the positive sample pair based on the node quality to generate a complete positive sample pair; it also calculates the feature differences of different nodes in the negative sample pair, determines the node quality of different nodes in the negative sample pair based on the result of the feature differences, and repairs the nodes in the negative sample pair based on the node quality to generate a complete negative sample pair. The model training module (400) initializes the parameters in the graph neural network model and calculates the similarity between complete positive sample pairs and the similarity between complete negative samples. Through backpropagation and gradient descent, the similarity between complete positive sample pairs is maximized and the similarity between complete negative sample pairs is minimized to optimize the parameters in the graph neural network model.

2. The industrial internet system based on data mining technology according to claim 1, characterized in that, The node quality analysis module (300) includes a positive sample pair analysis unit (301), and the process of generating complete positive sample pairs by the positive sample pair analysis unit (301) specifically includes: Compare the feature vectors of two different versions of the same node in a positive sample pair, and use different distance metrics to calculate the feature differences between the feature vectors; Based on the results of feature differences, the quality of the same node in the positive sample pair is evaluated; the smaller the feature difference, the higher the node quality; by setting a node threshold for the positive sample pair, it is determined whether the node in the positive sample pair is a low-quality node, and nodes with feature differences greater than the node threshold for the positive sample pair are regarded as low-quality nodes. Low-quality nodes in positive sample pairs are repaired to generate complete positive sample pairs. Data augmentation is performed on the nodes again using different methods until the feature differences meet the threshold requirements for positive sample pair nodes. The low-quality nodes in the repaired positive sample pairs are then filled into the positive sample pairs to generate complete positive sample pairs.

3. The industrial internet system based on data mining technology according to claim 2, characterized in that, The node quality analysis module (300) includes a negative sample pair analysis unit (302), and the process by which the negative sample pair analysis unit (302) generates complete negative sample pairs specifically includes: Compare the feature vectors of two enhanced versions of different nodes in a negative sample pair, and use a distance metric to calculate the feature differences between the feature vectors. Based on the results of feature differences, the quality of different nodes in the negative sample pair is evaluated. The greater the feature difference, the higher the node quality. By setting a threshold for nodes in the negative sample pair, it is determined whether a node in the negative sample pair is a low-quality node. Nodes with feature differences less than the threshold for nodes in the negative sample pair are regarded as low-quality nodes. To repair low-quality nodes in negative sample pairs and generate complete negative sample pairs, noise is added to adjust feature values ​​to increase feature differences. The repaired low-quality nodes are then filled into the negative sample pairs to generate complete negative sample pairs.

4. The industrial internet system based on data mining technology according to claim 1, characterized in that, The process of optimizing the graph neural network model parameters by the model training module (400) specifically includes: Choose a graph neural network model and initialize the parameters in the graph neural network model, where the initialization method includes random initialization; The similarity of complete positive sample pairs and the similarity of complete negative samples are calculated based on the distance metric method used in the node quality analysis module (300); the calculated distance is used as the similarity of the corresponding sample pairs. Define a loss function and calculate the loss value under the parameters of the graph neural network model based on the loss function. This is done by dividing the similarity of complete positive sample pairs by the adjustment parameter and taking the exponential value; dividing the similarity of complete negative sample pairs by the adjustment parameter and taking the exponential value; and summing these exponential values ​​to obtain the total exponential value. Then, sum the exponential values ​​of the complete positive sample similarity and the complete negative sample similarity to obtain the denominator, calculate the proportion of the exponential value of the complete positive sample similarity in the total exponential value, and take the negative logarithm to obtain the final loss value. The adjustment parameter is used to adjust the distribution of similarity to balance the contrast between positive and negative samples. Calculate the gradient of the loss value with respect to the parameters of the graph neural network model, and use the gradient descent algorithm to update the model parameters, so as to maximize the similarity between complete positive sample pairs and minimize the similarity between complete negative sample pairs, thereby optimizing the parameters in the graph neural network model.

Images