A method and device for determining important links of a graph neural network model based on shapley value interpretation

Abstract

The application discloses a method and device for determining important links of a graph neural network model based on Shapley value interpretation, and combines the working principle of a graph neural network model for SDN network performance prediction with game theory knowledge to provide effective and reasonable interpretation for the graph neural network model for SDN network performance prediction. The Shapley value is used as the importance score of an input node, and an important node set is selected through a trainable soft discrete mask matrix, so that the calculation amount of the Shapley value is reduced. Experimental results show that the application can provide effective interpretation for the graph neural network model, select an important link set for a prediction result, provide support for further optimizing a network topology structure, and to a certain extent, verify that the model can correctly learn the relationship between features, enhance the reliability, and provide a new scheme for promoting the application of the graph neural network model for SDN network performance prediction in engineering practice.

Description

Technical Field

[0001] This invention relates to the interpretability of graph neural networks, and more specifically to a method and apparatus for determining important links in a graph neural network model based on Shapley value interpretation. Background Technology

[0002] While graph neural networks (Graph Neural Networks) possess a superior theoretical foundation and excellent experimental performance in the networking field, the inherent characteristics of deep learning mean that Graph Neural Networks remain in a "black box" state. For users and operators, this "black box" state makes the models unreliable. Currently, many works have focused on the interpretability of Graph Neural Networks, but these methods have their own limitations when applied to Graph Neural Networks used for network system modeling. Because links in network systems have sequential relationships, Graph Neural Networks used for network system modeling are often combined with recurrent neural networks. For such complex Graph Neural Networks, gradient / feature-based interpretation methods and contribution decomposition-based interpretation methods are often unsuitable. Perturbation-based interpretation methods can only select important node sets but struggle to generate fair and reasonable importance rankings. For Graph Neural Networks used for network system modeling, this method can only provide limited interpretability and cannot generate reasonable explanations at the level of individual nodes. Proxy-based interpretation methods approximate the target model through a simple interpretable proxy model, requiring the assumption that the neighbor relationships of the interpreted target are relatively simple. This has significant limitations when applied to Graph Neural Networks used for network system modeling, easily losing spatial dependencies within the network, making it difficult to find a suitable interpretable proxy model. Due to the lack of interpretability, graph neural network technology is difficult to be widely used in engineering practice. Therefore, studying the interpretability of graph neural networks is of great practical significance for the development of new network systems. Summary of the Invention

[0003] This invention provides an effective and reasonable explanation for the graph neural network model used in SDN network performance prediction. It verifies to a certain extent that the model can correctly learn the relationships between features, enhancing its reliability and providing a new solution for promoting the application of graph neural network models for SDN network performance prediction in engineering practice. This method combines Shapley values ​​with a soft discrete mask matrix to achieve the importance ranking of topology nodes with appropriate computational resources. Applying the important link determination method to the proposed graph neural network model, experimental results show that this method can provide an effective explanation for the graph neural network model, selecting a set of important links relevant to the prediction results, and providing support for further optimization of the network topology.

[0004] The technical solution adopted by this invention to solve the technical problem is as follows: A method for determining important links in a graph neural network model based on Shapley value interpretation, the method comprising the following steps:

[0005] (1) Construct a graph neural network model, with the input parameter being the global information of the network and the output being the predicted network performance;

[0006] (2) Using a soft discrete mask matrix, select the set of input nodes that play an important role in the output of the current graph neural network model from all input nodes of the graph neural network model;

[0007] (3) Calculate the Shapley value for the selected important input nodes to obtain the importance score of the important input nodes. Based on the importance score, determine the important links that affect the prediction results of the current graph neural network model, optimize the network topology, and adjust the network traffic distribution.

[0008] Furthermore, in step (1), the global network information includes network topology, routing policy, node configuration, and traffic matrix; the predicted network performance includes latency, jitter, and packet loss rate.

[0009] Furthermore, in step (2), the soft discrete mask matrix M represents the link set, and the set element m i Indicates input node v i The corresponding mask value, which is used to determine whether a link is important, is the mask value m. i The closer a node is to 0, the lower its importance.

[0010] Furthermore, the search for important links is achieved by training a soft discrete mask matrix M. During the training process, two factors need to be considered: the number of important nodes and the output variation.

[0011] Output change (L(O) I O M Given the original input I of the graph neural network model and the graph neural network model f(·), the prediction result generated by the original input is O. I The prediction result after masking is O M The output change L(O) is measured by the similarity between the prediction results of the original image and the new image. I O M Consider it as O I With O M The mean square error;

[0012] Number of important nodes (||M||): Based on the obtained soft discrete mask matrix M, the set of elements m i The sum of the absolute values ​​represents the number of important nodes.

[0013] Furthermore, a penalty factor Q(M) is designed, using the entropy value used in information theory to calculate information uncertainty to constrain the elements m of the soft discrete mask matrix M. iTry to get as close to 0 or 1 as possible; according to the penalty factor formula Q(M), the objective function of the optimization process is l. function (M) is:

[0014] l function (M)=L(O I O M )+α||M||+λQ(M)

[0015] Where α is a hyperparameter controlling the number of important nodes, and λ is a hyperparameter controlling the intensity of discreteness penalty;

[0016] The final set of important nodes V Selected Represented as:

[0017] V Selected =V*M′

[0018] Where V represents the set of input nodes of the graph neural network; and m′ is an element of the discrete mask matrix M′. i The value is 0 or 1.

[0019] Using the important node set V Selected The Shapley score is calculated instead of the input node set V.

[0020] Furthermore, the process of calculating the Shapley value for the selected important input nodes is as follows:

[0021] Important node set V Selected There are n nodes in total, of which the input node v i As the target node, the remaining input nodes {v1, v2, v3, ..., v...} i-1 v i+1 , ..., v n Let P be a set, and C be the power set of P. Consider the element c in C. j For a federation, the input node v i Join the alliance c j The marginal contribution function m(v) i c j ) is represented as:

[0022] m(v i c j )=f(v i ∪c j )-f(c j )

[0023] Input node v i The contribution to the entire graph is correspondingly represented as v. i The mathematical expectation of the gains after joining the alliance, i.e., the input node v iThe corresponding Shapley value S(v) i The details are as follows:

[0024]

[0025] in Indicates input node v i Join the current alliance c j The probability of.

[0026] Furthermore, according to the Shapley value S(v) i ) Calculate the input node v i The specific calculation process for the importance score is as follows, based on the marginal contribution formula m(v i c j The average contribution value generated by performing T random samplings on the coalition set C is used as an approximation of the importance score. Furthermore, the original Shapley value formula S(v) can be used to approximate this. i Replace with the following formula:

[0027]

[0028] Furthermore, for those not belonging to v i ∪c j For the input nodes, select to retain the original topology and fill them with only 0 values.

[0029] Secondly, the present invention also provides an apparatus for determining important links in a graph neural network model based on Shapley value interpretation, comprising a memory and one or more processors, wherein the memory stores executable code, and when the processor executes the executable code, it is used to implement the steps of the method for determining important links in a graph neural network model based on Shapley value interpretation.

[0030] Thirdly, the present invention also provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of the method for determining important links in a graph neural network model based on Shapley value interpretation.

[0031] The beneficial effects of this invention are as follows: With the increasing scale, complexity, and rapid growth of network services, the demand for reliable performance evaluation of complex networks is growing. For graph neural network models with promising network performance prediction results, due to the black-box nature of deep learning models, these models may encounter unpredictable errors during training and prediction. For example, they may learn incorrect patterns and relationships, leading to incorrect predictions, which makes it difficult to generalize the model in practical applications. Taking an intuitive image classification task as an example, when training a model to determine the presence of a dog in an image, if all dog images in the training set have grass as a background, the trained model may incorrectly classify grass features as indicating the presence of a dog—an undesirable error. Therefore, this invention constructs a method for determining important links in graph neural network models based on Shapley value interpretation. This method can interpret the prediction process of graph neural network models with promising network performance prediction results, select the set of important links relevant to the prediction results, and provide support for further optimization of the network topology. This method improves the reliability and robustness of the model, facilitating its use and promotion in practical applications. Attached Figure Description

[0032] Figure 1 This is a diagram illustrating the overall framework of the method for determining important links in a graph neural network model based on Shapley value interpretation proposed in this invention.

[0033] Figure 2 The cumulative distribution function curve of the elements in the soft discrete mask matrix M;

[0034] Figure 3 Correlation analysis of actual traffic and Shapley value for important links;

[0035] Figure 4 For global analysis of the correlation between normalized real flow and Shapley value;

[0036] Figure 5 Correlation analysis between actual traffic flow and Shapley value under multi-objective tasks.

[0037] Figure 6 This is a structural diagram of the important link determination device for graph neural network models based on Shapley value interpretation proposed in this invention. Detailed Implementation

[0038] like Figure 1As shown, this invention provides a method for determining important links in a graph neural network model based on Shapley value interpretation. The graph neural network model is an SDN network performance prediction model; the specific input parameters of the network performance prediction model are path features, link features, and node features; the output is the predicted network performance; the graph neural network consists of a message passing process and a graph reading process. The method for determining important links in a graph neural network model based on Shapley value interpretation consists of two parts: important node set generation and Shapley value scoring. An important node set is selected using a trainable soft discrete mask matrix, and then the Shapley value is used as the importance score of the input nodes, thereby reducing the computational cost of the Shapley value and making this method more suitable for graph neural network models with a large number of input nodes. The specific process is described below:

[0039] (1) Construct a graph neural network model. The input parameters are global network information, and the output is the predicted network performance. The global network information is network topology, routing strategy, node configuration and traffic matrix. The network performance is latency, jitter and packet loss rate.

[0040] (2) Using a soft discrete mask matrix, select the set of input nodes that play an important role in the output of the current graph neural network model from all input nodes of the graph neural network model; specifically as follows:

[0041] Using the Shapley value can fairly score the importance of each input node. However, since the Shapley value calculation requires traversing all input node rankings, it becomes computationally expensive when there are many input nodes. To address this limitation, this invention designs and trains a soft discrete mask matrix to select the set of important nodes that represent the original graph output, and considers calculating only important nodes, thus greatly reducing computational complexity and increasing the applicability of the method. To address the approximate failure problem caused by an excessive number of input nodes, the design goal of this invention is to search for input nodes that play a crucial role in the current output, i.e., to represent the current graph input with a set of important nodes. Formally, the set of important nodes can be represented by a trainable soft discrete mask matrix M, i.e., the set of important links, as follows:

[0042] M = (m1, m2, m3, ..., m...) n )m i ∈(0,1)

[0043] Where m i Indicates input node v i The corresponding mask value, which is used to determine whether a link is important, is the mask value m. i The closer a node is to 0, the lower its importance.

[0044] The search process is implemented by training a soft discrete mask matrix M. During training, two factors need to be considered: the number of important nodes and output variation. Retaining more important nodes results in a stronger representation of the original image and smaller output variation, but significantly increases computational cost. Conversely, retaining fewer important nodes reduces computational cost but weakens the representation of the original image and increases output variation. Specifically, this invention quantifies these two factors: output variation and the number of important nodes.

[0045] (1) Output change (L(O) I O M Important nodes should be input nodes that have a significant impact on the predicted output; therefore, the set of important nodes needs to have a strong representational ability of the original graph output. Given the original input I and the graph neural network model f(·), the predicted result O generated by the original input is... I The prediction result O after processing with the mask matrix M The calculation process can be derived from formula O. I Formula O M To illustrate, taking the time delay model as an example, the original input path traffic set and the important link set after mask matrix processing are used as input parameters of the graph neural network model f(·), respectively, to obtain the corresponding time delay prediction results. Specifically:

[0046] O I =f(I)

[0047] O M =f(M*I)

[0048] Furthermore, this invention proposes using the similarity between the prediction results of the original image and the new image to measure the output change L(O). I O M In deep learning, this can be considered as O(n). I With O M The mean square error, as shown in the formula L(O) I O M As shown in the image:

[0049] L(O I O M )=||O I -O M || 2

[0050] (2) Number of Important Nodes (||M||): To achieve the design goals, the soft discrete mask matrix M obtained by this invention needs to retain as few important nodes as possible while maintaining good representational capabilities. Based on the obtained soft discrete mask matrix M, i.e., the set of important links after mask matrix processing, the number of important nodes can be represented by the formula ||M||:

[0051]

[0052] According to the output change formula L(O) I O M The formula for the number of important nodes is ||M||, and the objective function of the optimization process can be expressed by formula l. function (M) is used to represent:

[0053] l function (M)=L(O I O M )+α||M||

[0054] Where α is a hyperparameter controlling the number of important nodes, and M is a soft discrete mask matrix.

[0055] Furthermore, this invention aims for the importance of each input node to be deterministic, i.e., discretely distributed as 0 or 1; otherwise, preserving the uncertainty of the nodes would significantly interfere with the interpretation results. The elements m of the soft discrete mask matrix M... i Since the data is continuous within (0, 1), errors inevitably arise due to the non-discrete data. To address this issue, this invention designs a penalty factor Q(M), using the entropy value, used in information theory to calculate information uncertainty, to constrain the elements m of the soft discrete mask matrix M. i Try to get as close as possible to 0 or 1, as follows:

[0056]

[0057] Based on the penalty factor formula Q(M), the objective function formula for the optimization process can be further derived. function (M) rewritten as:

[0058] l function (M)=L(O I O M )+α||M||+λQ(M)

[0059] Where α is a hyperparameter controlling the number of important nodes, and λ is a hyperparameter controlling the intensity of the discreteness penalty.

[0060] In summary, the final set of important nodes V is generated. Selected It can be represented as:

[0061] V Selected =V*M′

[0062] The discrete mask matrix M can be represented as:

[0063]

[0064] To address the issue of an excessive number of input nodes, this invention considers using a set of important nodes V. Selected The Shapley score is calculated instead of the input node set V.

[0065] (3) Calculate the Shapley value for the selected important input nodes to obtain their importance scores. Based on these scores, identify the key links affecting the prediction results of the current graph neural network model, optimize the network topology, and adjust the network traffic distribution. Details are as follows:

[0066] Given a graph G with n input nodes and a corresponding graph neural network model f(·), where the input nodes v i As the target node, the remaining input nodes {v1, v2, v3, ..., v...} i-1 v i+1 , ..., v n Let P be a set, and C be the power set of P. Consider the element c in C. j For a federation, the input node v i Join the alliance c j The marginal contribution function m(v) i c j This can be represented as:

[0067] m(v i c j )=f(v i ∪c j )-f(c j )

[0068] It should be noted that for those that do not belong to v i ∪c j For the input nodes, select to retain the original topology and fill them with only 0 values ​​to ensure that the interpretation results conform to the original topology.

[0069] Input node v i The contribution to the entire graph can be represented accordingly as v i The mathematical expectation of the gains after joining the alliance, i.e., the input node v i The corresponding Shapley value S(v) i The details are as follows:

[0070]

[0071] in Indicates input node v i Join the current alliance c j The probability of.

[0072] According to the Shapley value formula S(v) i ) Calculate the input node vi The importance score requires iterating through all alliance cases, resulting in a time complexity of O(n^2). n The computational efficiency is low and difficult to implement, therefore this invention chooses to use Monte Carlo sampling for approximate calculation. Monte Carlo sampling greatly alleviates the computational complexity. The specific calculation process is as follows: based on the marginal contribution formula m(v... i c j The average contribution value generated by performing T random samplings on the coalition set C is used as an approximation of the importance score. Furthermore, the original Shapley value formula S(v) can be used to approximate this. i Replace with the following formula:

[0073]

[0074] This invention selects a set of important nodes using a trainable soft discrete mask matrix, and then uses the Shapley value as the importance score of the input nodes. This reduces the computational cost of the Shapley value, making this important link determination method more suitable for graph neural network models with a large number of input nodes and compatible with network systems with complex topologies. For network systems, path traffic is determined by user-issued service requests; different service requests generate different amounts of path traffic, resulting in different path-level network performance indicators. This invention treats the path traffic of input nodes as the feature to be interpreted. Taking a latency model as an example, it selects important links using a soft discrete mask matrix and then performs importance analysis on them, i.e., calculates the Shapley value corresponding to the path traffic feature, and then analyzes the impact of path traffic on path latency.

[0075] The present invention will be further described below with reference to the embodiments.

[0076] The experiments in this invention were conducted on a GPU server equipped with an NVIDIA TITAN XP graphics card. The system environment consisted of a 64-bit operating system (Ubuntu 18.04.1), a computing architecture (CUDA 10.0), a deep learning framework (Tensorflow 1.4), and a programming language (Python 3.6.9). The model used in the experiments was an intelligent prediction model for network performance based on an SDN architecture. The dataset used in the experiments was an open-source dataset consistent with the RouteNet model. This dataset was generated based on an NSFnetwork network topology with 14 nodes and 21 links, and included network information such as path traffic, link bandwidth, and routing policies, as well as network performance parameters such as latency, jitter, and packet loss rate.

[0077] The validity of the set of important nodes is the foundation for generating valid interpretations. To verify the effectiveness of the proposed method, this invention proposes the following evaluation metrics for the set of important nodes:

[0078] (1) Fidelity: This invention uses the Fidelity metric to represent the rate of change of the output. Given a set of input nodes V, a set of important nodes Vi Selected The graph neural network model f(·) and Fidelity can be represented as:

[0079]

[0080] This invention imposes restrictions on output variation. Since the numerical value of output variation cannot intuitively reflect the effect of the restrictions, this invention uses the Fidelity index to evaluate the output variation caused by the set of important nodes. The closer the Fidelity is to 1, the smaller the output variation, and the more faithful the output is to the original graph.

[0081] (2) Remain: The Remain index is obtained by summing the elements of the discrete mask matrix M′. Given the discrete mask matrix M′, m′ i If a given element is a given element, then the Remain index can be expressed as:

[0082]

[0083] This invention imposes a limit on the number of important nodes ||M||. The effect of limiting the number of important nodes ||M|| can be evaluated by the Remain index. The smaller the Remain index, the fewer nodes are retained, and the better the limiting effect.

[0084] (3) Distribution: The elements in the soft discrete mask matrix M generated by the set of important nodes should be as close as possible to a 0 and 1 distribution. Therefore, this invention uses the Distribution index to evaluate the discreteness of the soft discrete mask matrix M. Specifically, the Distribution index is composed of elements m in the soft discrete mask matrix M. i The cumulative distribution function is obtained as follows:

[0085]

[0086] Where ε is the distribution threshold, and Distribution(ε) calculates the cumulative distribution probability of the elements of the soft discrete mask matrix M under the current distribution threshold ε. Plotting the cumulative distribution function using the Distribution index can help analyze the elements m in the soft discrete mask matrix M. i The discreteness is evaluated.

[0087] Experiments and Results Analysis. For SDN network systems, link bandwidth is a necessary element for successful data packet transmission. The bandwidth of each link segment in the path is indispensable; therefore, discussing the importance of link bandwidth characteristics lacks practical significance. Path traffic is determined by the service demands issued by users. Different service demands will generate different amounts of path traffic, thus producing different path-level network performance indicators. Therefore, this invention uses the path traffic of the input node as the feature to be interpreted. Taking the latency model as an example, the proposed important link determination method is used to interpret and analyze the network performance prediction model. It should be noted that the selected NSF network topology contains 182 paths, and the model interpretation is generated for a single path output in the output graph. Taking path 27 as an example, using path 27 in the output graph as the interpretation target, the experimental interpretation process is as follows: Figure 1 As shown.

[0088] To evaluate the effectiveness of the interpretation results, this invention calculated and analyzed the evaluation metrics of the important node set through experiments. Table 1 shows the evaluation metrics obtained for some interpretation targets. Based on the analysis of the Fidelity and Remain index parameters in Table 1, the obtained important node set can still have a good expressive ability of the original image while retaining a relatively small number of nodes. The training results of the soft discrete mask matrix can effectively balance the two factors of output variation and the number of important nodes, which meets the design expectations.

[0089] Table 1 partially explains the set of key nodes for evaluating the objectives.

[0090] Explain the goal Fidelity Remain Path 27 0.994 5 Path 51 0.861 10 Path 77 0.949 10 Path 98 0.871 7 Path 108 0.869 7 Path 125 0.968 4 Path 142 0.969 7 Path 161 0.873 11

[0091] Furthermore, due to the elements m in the soft discrete mask matrix M i Since the input nodes are not strictly discrete, there is some uncertainty in their retention, inevitably introducing errors. Therefore, this invention further analyzes the discreteness of the trained soft discrete mask matrix M. Based on the index formula Distribution(ε), ε is used as the x-axis, and Distribution(ε) as the y-axis. Figure 2 The element m is given i The cumulative distribution function graph.

[0092] The cumulative distribution function curve shows the elements m in the soft discrete mask matrix M. i A large number of particles cluster around 0, and some cluster around 1. This indicates that the soft discrete mask matrix M has strong discreteness, which can alleviate the error caused by "introducing evidence" to a certain extent, thus proving the effectiveness of the present invention.

[0093] Based on the above evaluation and verification results, this invention uses the proposed method for determining important links in a graph neural network model based on Shapley value interpretation to analyze the importance of the path traffic characteristics of input nodes in the intelligent prediction model of network performance. Some interpretation results are given in Table 2, where the important links retained in the ranking results are the set of selected important links.

[0094] Table 2. Interpretation of Path Traffic Characteristics

[0095]

[0096] For network systems, the reason why path traffic has a significant impact on path latency can generally be considered to be the following two:

[0097] (1) This path has a large path traffic;

[0098] (2) The traffic in this path flows through more links in the target path.

[0099] According to the analysis in Table 2, the interpretation ranking results of paths 27 and 98 are basically consistent with the actual traffic ranking results. For path 98, its traffic is less than that of path 127, but because the traffic of path 98 flows through all links within the interpretation target, it has a greater impact on the latency of the interpretation target. This indicates that when the intelligent network performance prediction model uses path traffic characteristics to predict path latency, it can, to some extent, follow the actual impact of path traffic on path latency. Experiments show that the important link determination method based on the Shapley value interpretation graph neural network model explores the degree of influence of important links on predicted latency, providing data support for achieving network load balancing, thereby adjusting network traffic distribution to improve traffic transmission efficiency and network resource utilization efficiency, and ultimately enhance service performance.

[0100] Based on the experimental results, this invention further explores the learning effect of the intelligent network performance prediction model on path traffic characteristics when traffic flows through the same links. Based on a selected interpretation target, this invention applies different amounts of path traffic to paths containing the same links within the interpretation target and measures the relationship between the actual traffic of each important link and its Shapley value, such as... Figure 3 , where r represents the Pearson correlation coefficient. Experimental results show that there is a certain positive linear correlation between the actual traffic of critical links and the Shapley value, which preliminarily demonstrates that the interpretation results provided by the graph neural network model for determining critical links based on Shapley value interpretation are consistent with the actual patterns to a certain extent.

[0101] Because the flow of traffic within a single path is relatively small, Figure 3The results obtained in the previous paper cannot fully explain the relationship between the Shapley value and the actual traffic. To strengthen the persuasiveness of the conclusions, this invention not only analyzes each interpretation target individually but also performs a global analysis of all interpretation targets. Since network performance metrics within a path are affected not only by traffic volume but also by link bandwidth, this invention normalizes the actual traffic and Shapley value within each interpretation target to ensure that the results from different interpretation targets are on the same scale. N represents the number of important links possessed by the interpretation target, and t... i s represents the actual traffic of the important link i. i The Shapley value representing the value generated by the important link i is the normalized real traffic S. i Shapley value T i They can be represented as follows:

[0102]

[0103]

[0104] Figure 4 The graph shows the correlation between the normalized real traffic and the Shapley value of important links. A large number of scatter points are clustered near the regression line, and the Pearson correlation coefficient is about 0.95. This further proves that there is a certain degree of positive linear correlation between the Shapley value generated by the interpretation and the real traffic. It shows that the network performance intelligent prediction model of this experiment can effectively learn the real rules between path traffic and path delay, and the model structure meets the requirements of network modeling.

[0105] Based on data normalization and combined with the additivity characteristic of Shapley values, this invention also obtains importance analysis based on multiple explanatory objectives. Specifically, the normalized traffic flowing through the same segment of important links and passing through different explanatory objectives is accumulated to obtain the accumulated true traffic on the important links for multiple explanatory objectives. Similarly, by accumulating the normalized Shapley values ​​generated by important links based on different explanatory objectives, the accumulated Shapley value generated by the important links for multiple explanatory objectives can be obtained. Furthermore, a correlation analysis between the accumulated true traffic and the accumulated Shapley value can be obtained, such as... Figure 5 As shown, the calculated Pearson correlation coefficient is 0.977. Experimental results show that the accumulated actual traffic and the accumulated Shapley value are statistically highly correlated. This indicates that the important link determination method for graph neural network models based on Shapley value interpretation proposed in this invention can not only effectively explain single-objective tasks, but also provide reasonable explanations for multi-objective tasks. It can select the set of important links for the prediction results, providing support for further optimization of network topology, and further verifying the effectiveness of the intelligent network performance prediction model in network modeling tasks.

[0106] Corresponding to the aforementioned embodiments of the method for determining important links in a graph neural network model based on Shapley value interpretation, the present invention also provides embodiments of a device for determining important links in a graph neural network model based on Shapley value interpretation.

[0107] See Figure 6 The present invention provides an apparatus for determining important links in a graph neural network model based on Shapley value interpretation, comprising a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it is used to implement the method for determining important links in a graph neural network model based on Shapley value interpretation in the above embodiments.

[0108] The embodiment of the important link determination device for graph neural network models based on Shapley value interpretation of this invention can be applied to any device with data processing capabilities, such as a computer. The device embodiment can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 6 The diagram shown is a hardware structure diagram of any data processing-capable device where the graph neural network model important link determination device based on Shapley value interpretation of the present invention is located. (Except for...) Figure 6 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0109] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0110] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0111] This invention also provides a computer-readable storage medium storing a program that, when executed by a processor, implements the method for determining important links in a graph neural network model based on Shapley value interpretation described in the above embodiments.

[0112] The computer-readable storage medium can be an internal storage unit of any data processing device as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0113] The above embodiments are used to explain and illustrate the present invention, but not to limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A method for determining critical links in a graph neural network model based on Shapley value interpretation, characterized in that, The method includes the following steps: (1) Construct a graph neural network model, with the input parameters being global network information and the output being the predicted network performance; (2) Using a soft discrete mask matrix, select the set of input nodes that play an important role in the output of the current graph neural network model from all input nodes of the graph neural network model; specifically: search for important links by training the soft discrete mask matrix M. During the training process, it is necessary to consider two factors: the number of important nodes and the output change. Output changes Given the original input of the graph neural network model Graph Neural Network Model The prediction result generated by the original input is O I The prediction result after masking is O M The output change is measured by the similarity between the prediction results of the original image and the new image. Consider it as O I With O M The mean square error; Number of important nodes Based on the obtained soft discrete mask matrix set element m i The sum of the absolute values ​​represents the number of important nodes; (3) Calculate the Shapley value for the selected important input nodes to obtain the importance score of the important input nodes. Based on the importance score, determine the important links that affect the prediction results of the current graph neural network model, optimize the network topology, and adjust the network traffic distribution.

2. The method for determining important links in a graph neural network model based on Shapley value interpretation according to claim 1, characterized in that, In step (1), the global network information includes network topology, routing policy, node configuration and traffic matrix; the predicted network performance includes latency, jitter and packet loss rate.

3. The method for determining important links in a graph neural network model based on Shapley value interpretation according to claim 1, characterized in that, In step (2), the soft discrete mask matrix M represents the link set, and the set element m i Indicates input node v i The corresponding mask value, which is used to determine whether a link is important, is the mask value m. i The closer a node is to 0, the lower its importance.

4. The method for determining important links in a graph neural network model based on Shapley value interpretation according to claim 3, characterized in that, Design penalty factors The elements m of the soft discrete mask matrix M are constrained by using the entropy value, which is used in information theory to calculate information uncertainty. i Try to get as close as possible to 0 or 1; according to the penalty factor formula The objective function of the optimization process is... for: in, To control the hyperparameters for the number of important nodes, Hyperparameters for controlling the intensity of discrete penalty; The final set of important nodes V Selected Represented as: Where V represents the set of input nodes of the graph neural network; the discrete mask matrix elements The value is 0 or 1; Using the important node set V Selected The Shapley score is calculated instead of the input node set V.

5. The method for determining important links in a graph neural network model based on Shapley value interpretation according to claim 4, characterized in that, The process of calculating the Shapley value for the selected important input nodes is as follows: Important node set V Selected The CCP There are nodes, where the input node v i As the target node, the remaining input nodes Form a set , It is a set power set, consider element c j For a federation, the input node v i Join the alliance c j marginal contribution function Represented as: Input node v i The contribution to the entire graph is correspondingly represented as v. i The mathematical expectation of the gains after joining the alliance, i.e., the input node v i Corresponding Shapley value The details are as follows: in Indicates input node v i Join the current alliance c j The probability of.

6. The method for determining important links in a graph neural network model based on Shapley value interpretation according to claim 5, characterized in that, According to the Shapley value Calculate input node The specific calculation process for the importance score is based on the marginal contribution formula. For the alliance collection conduct The average contribution value generated by the random sampling is used as an approximation of the importance score, which can further refine the original Shapley value formula. Replace with the following formula: 。 7. The method for determining important links in a graph neural network model based on Shapley value interpretation according to claim 5, characterized in that, For those that do not belong to For the input nodes, select to retain the original topology and fill them with only 0 values.

8. A device for determining critical links in a graph neural network model based on Shapley value interpretation, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that, When the processor executes the executable code, it implements the steps of the method for determining important links in a graph neural network model based on Shapley value interpretation as described in any one of claims 1-7.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method for determining important links in a graph neural network model based on the interpretation of Shapley values ​​as described in any one of claims 1-7.

Images