A network service quality prediction method based on tensor Tucker decomposition

Abstract

The application provides a network service quality prediction method based on tensor Tucker decomposition, comprising the following steps: 1, constructing existing network service quality data into a three-dimensional sparse tensor model of user-service-time, used for describing multi-dimensional potential relationship; 2, calculating the similarity between users and the similarity between services respectively by using multiple similarity measurement methods, and constructing a user similarity matrix and a service similarity matrix; 3, clustering and grouping users and services according to the similarity matrix to form a user group set and a service group set, so as to alleviate the data sparsity problem; 4, selecting a target user and a target service, determining the user group and the service group to which the target user and the target service belong, extracting a corresponding sub-data set from the original data and constructing a sub-tensor; 5, performing Tucker decomposition on the sub-tensor, and completing missing data by iterative optimization to obtain a prediction result; and 6, reasonably using an error evaluation index to evaluate the prediction result, so as to verify the model performance.

Description

Technical Field

[0001] This invention relates to the field of computer network service quality prediction technology, and in particular to a network service quality prediction method based on tensor decomposition models, specifically a network service quality prediction method based on Tucker decomposition. Background Technology

[0002] With the rapid development of cloud computing, service computing, and internet technologies, various network services are widely used in e-commerce, online payment, smart manufacturing, smart cities, and other scenarios. Network service quality, as a crucial indicator for measuring network service performance, typically includes multiple dimensions such as response time, throughput, reliability, and availability, directly impacting user experience and service selection decisions. In practical applications, due to factors such as dynamic changes in the network environment, fluctuations in service load, and differences in user geographical location, network service quality data often exhibits characteristics such as high dimensionality, sparseness, and high noise. Furthermore, a significant amount of interaction data between users and services is missing, making it difficult for traditional statistical analysis-based methods to obtain accurate prediction results.

[0003] Currently, network service quality prediction mainly relies on mining potential patterns from historical network service quality data. While existing methods can achieve some predictive effectiveness when data is relatively abundant, they still have significant shortcomings. Firstly, these methods heavily depend on historical interactions between users and services. When data is sparse or new users or services emerge, it becomes difficult to accurately characterize their features, leading to the cold start problem and a significant decrease in prediction accuracy. Secondly, most existing models only model two-dimensional data structures. While this can alleviate the data sparsity problem to some extent, it struggles to simultaneously characterize the complex relationships between multiple dimensions such as users, services, time, and location. The models' expressive power is limited, making it difficult to fully mine the potential features in high-dimensional network service quality data. Furthermore, some methods improve predictive ability by constructing complex models, but this usually requires assumptions about data distribution. The model's generalization ability is greatly affected by the actual data distribution, and the inference and training processes are computationally complex, demanding high computing resources and making it difficult to meet the needs of large-scale applications. At the same time, while complex model structures possess strong nonlinear modeling capabilities, they have a large parameter scale, high training costs, and poor model interpretability, hindering practical engineering deployment and model optimization. Therefore, existing network service quality prediction methods still have shortcomings in terms of data sparsity handling, multi-dimensional feature modeling, computational efficiency, and model interpretability, and need further improvement.

[0004] In recent years, tensor decomposition-based methods have been proposed. This method constructs a multidimensional tensor structure to uniformly represent multidimensional network service quality information such as users, services, and time. It can simultaneously uncover potential correlations between multiple dimensions, effectively overcoming the limitation of traditional two-dimensional modeling methods in characterizing high-dimensional features. Compared to matrix factorization, tensor decomposition can more comprehensively describe the multidimensional characteristics of network service quality data, improving prediction accuracy while enhancing model expressiveness. Compared to deep learning methods, tensor decomposition models have a relatively simple structure, controllable parameter size, higher computational efficiency, and better interpretability and engineering applicability. Summary of the Invention

[0005] To address the aforementioned unresolved key issues, this invention proposes a network service quality prediction method based on tensor Tucker decomposition. Tucker decomposition, a classic high-order tensor decomposition model, effectively extracts latent features from various dimensions by decomposing the original tensor into a product of a core tensor and multiple factor matrices, while flexibly controlling model complexity. Tucker decomposition not only captures high-order interactions between different dimensions but also achieves dimensionality reduction by adjusting the size of the core tensor, thus reducing computational overhead while maintaining prediction accuracy. Therefore, introducing Tucker decomposition into the field of network service quality prediction helps to fully explore the latent structural features in multidimensional network service quality data, overcoming the shortcomings of existing methods in multidimensional modeling capabilities and prediction accuracy.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A network service quality prediction method based on tensor Tucker decomposition includes the following steps:

[0008] Step 1: Propose a network service quality prediction method based on tensor Tucker decomposition. Complete the result to obtain the prediction tensor This enables prediction of network service quality.

[0009] Step 2: Construct a sparse three-dimensional tensor model from existing network service quality data. ,in , , These represent the user, service, and time dimensions, respectively, to embed multidimensional structural information and characterize high-dimensional potential relationships.

[0010] Step 3: Calculate the similarity between users and services using multiple similarity measurement methods to construct a user similarity matrix. Service similarity matrix .

[0011] Step 4: Based on the similarity matrix and Users and services are clustered and grouped to form user group sets. Service group collection To reduce the impact of data sparsity.

[0012] Step 5: Randomly select target users and target services and determine the user groups to which they belong. Service Group Based on these two sets of data, corresponding subsets are extracted from the original dataset, and subtensors are constructed based on the network service quality data of the extracted subsets. .

[0013] Step 6: Convert the subtensor Construct a three-dimensional structure of user-service-time and perform Tucker decomposition. The prediction results are obtained through iterative optimization. .

[0014] Step 7: Calculate the true value using multiple error indices. With prediction tensor The error between the two is used to evaluate the model performance on real datasets to determine the quality of the prediction.

[0015] To reduce the sparsity of raw network service quality data and enhance the correlation between local samples, this invention clusters users and services, grouping objects with similar characteristics into the same group. This increases the similarity of samples within the group, thereby reducing the interference of irrelevant samples on the subsequent tensor modeling process and improving the stability and prediction accuracy of the Tucker decomposition model.

[0016] Preferably, in step 3, the present invention selects Pearson correlation coefficient, cosine similarity, Euclidean distance similarity and Manhattan distance similarity as similarity measurement methods, and uses K-means clustering algorithm to group users and services. The specific steps are as follows.

[0017] Step 3-1: First, calculate the similarity between users and services using four similarity measurement methods: Pearson correlation coefficient, cosine similarity, Euclidean distance similarity, and Manhattan distance similarity. Let the user or service feature vectors be respectively... , ,in, and Representing the eigenvector and In the The element values ​​on each feature dimension correspond to the observed attributes of a user or service in network service quality data. In this invention, these represent throughput or response time values.

[0018] The four similarity calculation formulas are as follows:

[0019] Pearson correlation coefficient is defined as

[0020]

[0021] in and These represent the vector mean;

[0022] Cosine similarity is defined as:

[0023]

[0024] Euclidean distance similarity is defined as

[0025]

[0026] The smaller the value, the higher the similarity;

[0027] Manhattan distance similarity is defined as

[0028]

[0029] The smaller the value, the higher the similarity.

[0030] Step 3-2: After completing the similarity calculation, the K-means clustering algorithm is used to group users and services. Let the sample set be... The number of clusters is The K-means objective function is defined as:

[0031]

[0032] in Indicates the first Clusters, This corresponds to the cluster center. The cluster center update formula is:

[0033]

[0034] The final user group set is obtained by iteratively updating the cluster centers until the objective function converges. Service group collection .

[0035] As a preferred embodiment, the sample selection mechanism and tensor model construction method in step 5 include the following:

[0036] Step 5-1: In the sample selection phase, first select from the entire user set. Randomly select target users and from the full set of services Randomly select target service Then query the target users respectively. User group and target services Service Group Next, the user group All users With the service team All services Perform pairwise pairings to construct a set of user-service pairs. Then, each pair is retrieved sequentially from the original data. The network service quality observations at different time dimensions are analyzed, and the extracted data are summarized and integrated to form a data set including user groups. With the service team Local subset of all interactive information .

[0037] Step 5-2: When constructing the tensor model, the dimension of the tensor is first determined based on the characteristics of the original network service quality data. Since the data used in this invention includes three dimensions: user, service, and time, a three-dimensional tensor is constructed. ,in , , These represent the number of users, the number of services, and the number of time slices, respectively. During the actual construction process, the user index is used. Service Index Time Index As a tensor The three-dimensional coordinate index, and the corresponding network service quality observation values. Fill to tensor elements In, that is:

[0038]

[0039] When there is no observation at the corresponding position, it is marked as a missing value, thus constructing a sparse three-dimensional tensor model with user-service-time as the index and network service quality value as the coefficient element.

[0040] As a preferred method, the Tucker decomposition of the tensor in step 6 is as follows:

[0041] Let the original sparse tensor be... in, Indicates the size of the user set. Indicates the size of the service set. Indicates the number of time slices, tensor elements Indicates the first The user in the first The time slice for the first The network service quality observation for a service is empty if no observation exists at the corresponding location.

[0042] Step 6-1: In this invention, the core tensor dimension is set to... and for the three factor matrices The core tensor is randomly initialized. The iteration counter is also initialized. Loss function variables Temporary variables Learning rate and regularization parameters .

[0043] Step 6-2: Based on the tensor Tucker decomposition model, reconstruct the complete tensor according to the following formula:

[0044] in, The tensor represents the tensor in the first place. Modular multiplication, This is the prediction tensor for the current iteration round.

[0045] Step 6-3: For the original sparse tensor All non-zero observation samples In predicting tensors Find the element at the corresponding position in the middle And calculate the prediction error:

[0046]

[0047] Step 6-4: Iteratively update the factor matrix and core tensor using stochastic gradient descent, with the following update rules:

[0048]

[0049]

[0050]

[0051]

[0052] in, , , These represent the corresponding row vectors in the factor matrix. This indicates the outer product operation.

[0053] Step 6-5: After completing one round of sample traversal, calculate the current loss function value:

[0054]

[0055] And assign the previous round's loss value to The iteration terminates when the following condition is met:

[0056]

[0057] in, This is the error threshold.

[0058] After the iteration is complete, output the padded tensor: .

[0059] As a preferred option, the error index selected in step 7 is detailed below:

[0060] Step 7-1: Select the Mean Absolute Error (MAE) as the evaluation index for model prediction accuracy. Its calculation process and meaning are as follows:

[0061] Calculation formula:

[0062]

[0063] in, This represents the actual quality of network service (QoS) value. This represents the predicted network service quality value. This indicates the number of samples involved in the evaluation. The Mean Absolute Error (MAE) measures the average deviation between predicted and actual values. A smaller MAE value indicates a smaller average error between the predicted and actual values, and higher model prediction accuracy.

[0064] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0065] Existing network service quality prediction methods suffer from several drawbacks. Collaborative filtering exhibits significant accuracy degradation in scenarios with sparse data and cold start conditions; matrix factorization only models two-dimensional relationships, failing to capture multi-dimensional features; probabilistic models rely on distribution assumptions and lack generalization ability; and deep learning models are complex and costly to train. Centralized methods incur high communication and computational overhead, while distributed methods have slow convergence speeds, making them unsuitable for large-scale applications. This invention proposes a network service quality prediction method based on tensor Tucker decomposition. This method constructs a three-dimensional tensor structure of users, services, and time to uniformly model multi-dimensional feature information and utilizes Tucker decomposition to mine latent feature relationships, achieving high-precision completion and prediction of missing network service quality data. Furthermore, this invention introduces an iterative modeling mechanism based on similarity, grouping users and services before modeling to reduce the impact of data sparsity on prediction performance. The contributions of the proposed method are summarized as follows: 1. This invention proposes a prediction model based on tensor decomposition, which can complete network service quality prediction using only local similar user and service data, reducing the dependence on global data; 2. The iterative optimization mechanism proposed in this invention can jointly update the core tensor and factor matrix without introducing an external proxy model, and has good scalability; 3. Experimental results on real network service datasets show that the method of this invention is superior to existing technologies in prediction accuracy. Attached Figure Description

[0066] Figure 1 This is a flowchart of the network service quality prediction method based on tensor Tucker decomposition in this invention;

[0067] Figure 2 This is a schematic diagram of sparse tensor modeling in this invention;

[0068] Figure 3 This is a schematic diagram illustrating the prediction of a complete tensor in this invention;

[0069] Figure 4 This is a schematic diagram of Tucker decomposition and iteration in this invention;

[0070] Figure 5 This is a graph showing the model error of the Tucker model after clustering at different similarity levels in this invention;

[0071] Figure 6 This is a graph showing the model error performance of the Tucker model under different training ratios in this invention;

[0072] Figure 7 The image shows the error performance of the Tucker model in this invention under different core tensor sizes. Detailed Implementation

[0073] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings and examples. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them.

[0074] Figure 1 The technical approach of this invention mainly consists of the following steps: calculating the similarity between users and services using a similarity metric; clustering users and services using K-Means clustering; selecting new datasets and constructing tensors; performing multiple rounds of tensor Tucker decomposition and iteration until convergence is achieved; and verifying the model accuracy using error metrics.

[0075] This example uses the WS-Dream-2 dataset for evaluation and validation. Provided by a project at the Chinese University of Hong Kong, this dataset is a typical dataset for network service quality assessment. The WS-Dream-2 dataset covers 145 users, 4500 network services, and 64 time slices (each time slice is 15 minutes long), containing over 35 million data samples. The dataset uses two independent network service quality metrics to characterize network service performance: response time (RT): in seconds; throughput (TP): in kb / s. This invention employs the tensor-based Tucker decomposition method to solve the prediction problem. The following sections will introduce this example from three aspects: user service similarity calculation and clustering, tensor model construction and decomposition prediction, and model performance evaluation.

[0076] 1) User service similarity calculation and clustering

[0077] Step 1-1: Calculate the similarity between users and services using four similarity measurement methods: Pearson correlation coefficient, cosine similarity, Euclidean distance similarity, and Manhattan distance similarity. Let the user or service feature vectors be respectively... , ,in, and Representing the eigenvector and In the The element values ​​on each feature dimension correspond to the observed attributes of a user or service in network service quality data. In this invention, these represent values ​​for throughput or response time. This represents the dimension of the feature vector.

[0078] The four similarity calculation formulas are as follows:

[0079] The Pearson correlation coefficient is defined as:

[0080]

[0081] in and These represent the vector mean;

[0082] Cosine similarity is defined as:

[0083]

[0084] Euclidean distance similarity is defined as

[0085]

[0086] The smaller the value, the higher the similarity;

[0087] Manhattan distance similarity is defined as

[0088]

[0089] The smaller the value, the higher the similarity.

[0090] Steps 1-2: After completing the similarity calculation, the K-means clustering algorithm is used to group users and services. Let the sample set be... in, Represents the sample set The total number of samples, i.e., the number of users or services participating in the clustering. Let the number of clusters be . The K-means objective function is defined as:

[0091]

[0092] in Indicates the first Clusters, This corresponds to the cluster center. The update formula for the cluster center is:

[0093]

[0094] set up Indicates the first During the nth iteration, the 1st The cluster center vectors are used to determine the cluster center vectors. When the change in the cluster center vectors between two consecutive iterations is less than a preset threshold... ,Right now: The algorithm is considered to have converged, thus obtaining the user group set. Service group collection .

[0095] 2) Construction and decomposition prediction of tensor models

[0096] Step 2-1: In the sample selection phase, first start with the entire user set. Randomly select target users and from the full set of services Randomly select target service Then query the target users respectively. User group and target services Service Group Next, the user group All users With the service team All services Perform pairwise pairings to construct a set of user-service pairs. Then, each pair is retrieved sequentially from the original data. The network service quality observations at different time dimensions are analyzed, and the extracted data are summarized and integrated to form a data set including user groups. With the service team Local subset of all interactive information This ensures that the constructed sample dataset maintains the correlation between the target users and the target services while also having sufficient data scale and statistical representativeness, thereby providing a reliable data foundation for subsequent tensor modeling and predictive analysis.

[0097] Step 2-2: When constructing the tensor model, the dimension of the tensor is first determined based on the characteristics of the original network service quality data. Since the data used in this invention includes three dimensions: user, service, and time, a three-dimensional tensor is constructed. ,in , , These represent the number of users, the number of services, and the number of time slices, respectively. During the actual construction process, the user index is used. Service Index Time Index As a tensor The three-dimensional coordinate index, and the corresponding network service quality observation values. Fill to tensor elements In, that is:

[0098]

[0099] When an observation is missing at a corresponding location, it is marked as a missing value, thus constructing a sparse three-dimensional tensor model with user-service-time as the index and network service quality values ​​as coefficient elements. The constructed sparse tensor is as follows: Figure 2 As shown.

[0100] Steps 2-3: Let the original sparse tensor be... in, Indicates the size of the user set. Indicates the size of the service set. Indicates the number of time slices, tensor elements Indicates the first The user in the first The time slice for the first The network service quality observations for each service are defined; if no observation exists at a given location, the element is considered empty. The core tensor dimension is set to [value missing]. and for the three factor matrices and core tensor Perform random initialization. Simultaneously initialize the iteration counter. Loss function variables Temporary variables Learning rate and regularization parameters Among them, the iteration counter It can represent the number of model iterations, with an initial value of 0. This represents the loss function value of the model in each iteration; a temporary variable. This records the loss function value and learning rate from each previous iteration. The step size parameter represents the step size during gradient updates and is used to control the magnitude of each parameter update. In this invention, it is set to 0.01. The regularization parameter... The penalty coefficient used to control model complexity suppresses overfitting and improves the model's generalization ability. In this invention, it is set to a value of [value missing]. .

[0101] Steps 2-4: Based on the tensor Tucker decomposition model, reconstruct the complete tensor according to the following formula:

[0102] in, The tensor represents the tensor in the first place. Modular multiplication, This is the prediction tensor for the current iteration round.

[0103] Steps 2-5: For the original sparse tensor All non-zero observation samples In predicting tensors Find the element at the corresponding position in the middle And calculate the prediction error. The calculation method is as follows:

[0104]

[0105] Steps 2-6: Algorithm flow as follows Figure 4 As shown, the factor matrix and core tensor are iteratively updated using stochastic gradient descent, and the update rules are as follows:

[0106]

[0107]

[0108]

[0109]

[0110] in, , , These represent the corresponding row vectors in the factor matrix. This indicates the outer product operation.

[0111] Steps 2-7: After completing one round of sample traversal, calculate the current loss function value:

[0112]

[0113] in, and Indicates user In service and time slices The actual network service quality observation values ​​and the predicted network service quality values ​​for the corresponding locations obtained by the method proposed in this invention;

[0114] The calculated value is obtained Then assign its value to The iteration terminates when the following condition is met:

[0115]

[0116] in, As the error threshold, in this invention, we take... If the value is 0.0001, the iteration ends when the difference between two consecutive errors is less than 0.0001.

[0117] After the iteration is complete, output the padded tensor: ,like Figure 3 The complete tensor shown corresponds to the element at the missing position in the original tensor, which is the network service quality value predicted by this invention.

[0118] 3) Model Performance Evaluation

[0119] Step 3-1: Select the Mean Absolute Error (MAE) as the evaluation index for model prediction accuracy. Its calculation process and meaning are as follows:

[0120] Calculation formula:

[0121]

[0122] in, This represents the actual quality of network service (QoS) value. This represents the predicted network service quality value. This indicates the number of samples involved in the evaluation. The Mean Absolute Error (MAE) measures the average deviation between predicted and actual values. A smaller MAE value indicates a smaller average error between the predicted and actual values, and higher model prediction accuracy.

[0123] Table 1 shows the basic setup for the experiment:

[0124] Table 1. Experimental parameter settings

[0125]

[0126] from Figure 5 As can be seen, the X-axis represents the dataset used by the Tensor Tucker decomposition model, and the Y-axis represents the error of the model in this invention. In the network service quality prediction task of the WS-Dream dataset, the method in this invention achieves the best clustering performance in both response time (RT) and throughput (TP) based on the Manhattan distance similarity metric. This is because Manhattan distance measures distance by accumulating absolute differences, effectively reducing the impact of outliers, improving clustering robustness and stability, and has low computational complexity and high scalability, making it particularly advantageous in large-scale network service quality dataset applications.

[0127] from Figure 6 As can be seen, the X-axis represents the proportion of the dataset used for training the Tensor Tucker decomposition model, and the Y-axis represents the error of the model in this invention. This is because a larger scale of data samples during training helps the model learn the hidden patterns and associations between users and services more effectively, thereby reducing prediction bias and improving generalization ability. At the same time, sufficient data samples can also alleviate the overfitting problem to some extent, enabling the model to adapt more stably to unseen test data.

[0128] from Figure 7 As can be seen, the X-axis represents the method of this invention and other methods, while the Y-axis represents the error of the method. The method of this invention has the lowest error value, with errors of 0.3798 and 1.5806 in throughput (TP) and response time (RT), respectively. This is because the method models network service quality data in the form of a three-dimensional tensor, effectively extracting latent features through low-rank decomposition while maintaining multi-dimensional structural information, thereby more accurately depicting user behavior patterns and service performance change patterns.

[0129] In summary, this invention effectively alleviates the data sparsity problem by constructing a user-service-time three-dimensional tensor model and combining similarity calculation and clustering strategies. Furthermore, it utilizes Tucker decomposition to mine multi-dimensional potential relationships, achieving accurate completion and prediction of missing network service quality data. This method significantly improves prediction accuracy and stability, is suitable for large-scale network service scenarios, and has good engineering application value and promising prospects for widespread adoption.

[0130] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art should understand that certain modifications and variations can be made to the present invention without departing from the spirit and substance of the present invention, but these modifications and variations should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting network service quality based on tensor Tucker decomposition, characterized in that... Includes the following steps: Step 1: Decompose iteratively Complete the result to obtain the prediction tensor This enables prediction of network service quality. Step 2: Construct a sparse three-dimensional tensor model from existing network service quality data. ,in , , These represent the user, service, and time dimensions, respectively, to embed multi-dimensional structural information and characterize high-dimensional potential relationships; Step 3: Calculate the similarity between users and services using multiple similarity measurement methods to construct a user similarity matrix. Service similarity matrix ; Step 4: Based on the similarity matrix and Users and services are clustered and grouped to form user group sets. Service group collection ; Step 5: Randomly select target users and target services and determine the user groups to which they belong. Service Group Based on these two sets of data, corresponding subsets are extracted from the original dataset, and subtensors are constructed based on the network service quality data of the extracted subsets. ; Step 6: Subtensor Construct a three-dimensional structure of user-service-time and perform Tucker decomposition. Prediction results are obtained through iterative optimization. .

2. The method according to claim 1, characterized in that, Step 3 is as follows; Step 3-1: First, calculate the similarity between users and services using four similarity measurement methods: Pearson correlation coefficient, cosine similarity, Euclidean distance similarity, and Manhattan distance similarity. Let the user or service feature vectors be respectively... , ,in, and Representing the eigenvector and In the The element values ​​on each feature dimension correspond to the observed attributes of a user or service in the network service quality data, that is, the values ​​representing throughput or response time. The four similarity calculation formulas are as follows: Pearson correlation coefficient is defined as in and These represent the vector mean; Cosine similarity is defined as: Euclidean distance similarity is defined as The smaller the value, the higher the similarity; Manhattan distance similarity is defined as The smaller the value, the higher the similarity; Step 3-2: After completing the similarity calculation, the K-means clustering algorithm is used to group users and services. Let the sample set be... The number of clusters is The K-means objective function is defined as: in Indicates the first Clusters, For the corresponding cluster center; the cluster center update formula is: The user group set is obtained by iteratively updating the cluster centers until the objective function converges. Service group collection .

3. The method according to claim 1, characterized in that, Step 5 is as follows; The sample selection mechanism and tensor model construction method in step 5 are as follows: Step 5-1: In the sample selection phase, first select from the entire user set. Randomly select target users and from the full set of services Randomly select target service Then query the target users respectively. User group and target services Service Group Then the user group All users With the service team All services Perform pairwise pairings to construct a set of user-service pairs. Then, each pair is retrieved sequentially from the original data. The network service quality observations at different time dimensions are analyzed, and the extracted data are summarized and integrated to form a data set including user groups. With the service team Local subset of all interactive information ; Step 5-2: When constructing the tensor model, first determine the dimension of the tensor based on the characteristics of the original network service quality data. Since the data used contains three dimensions—user, service, and time—a three-dimensional tensor is constructed. ,in , , These represent the number of users, the number of services, and the number of time slices, respectively. During the actual construction process, the user index is used. Service Index Time Index As a tensor The three-dimensional coordinate index, and the corresponding network service quality observation values. Fill to tensor elements In, that is: When there is no observation at the corresponding position, it is marked as a missing value, thus constructing a sparse three-dimensional tensor model with user-service-time as the index and network service quality value as the coefficient element.

4. The method according to claim 1, characterized in that, Step 5 is as follows; Let the original sparse tensor be... in, Indicates the size of the user set. Indicates the size of the service set. Indicates the number of time slices, tensor elements Indicates the first The user in the first The time slice for the first The network service quality observation for each service is empty if no observation is found at the corresponding location. Step 6-1: Set the core tensor dimension to 1. and for the three factor matrices and core tensor Perform random initialization; simultaneously initialize the iteration counter. Loss function variables Temporary variables Learning rate and regularization parameters ; Step 6-2: Based on the tensor Tucker decomposition model, reconstruct the complete tensor according to the following formula: in, The tensor represents the tensor in the first place. Modular multiplication, This is the prediction tensor for the current iteration round; Step 6-3: For the original sparse tensor All non-zero observation samples In predicting tensors Find the element at the corresponding position in the middle. And calculate the prediction error: Step 6-4: Iteratively update the factor matrix and core tensor using stochastic gradient descent, with the following update rules: in, , , These represent the corresponding row vectors in the factor matrix. Indicates the outer product operation; Step 6-5: After completing one round of sample traversal, calculate the current loss function value: And assign the previous round's loss value to The iteration terminates when the following conditions are met: in, This is the error threshold; After the iteration is complete, output the padded tensor: .

5. The method according to claim 1, characterized in that, The mean absolute error (MAE) is selected as the evaluation index for model prediction accuracy. The calculation process and meaning are as follows: Calculation formula: in, This represents the actual quality of network service (QoS) value. This represents the predicted network service quality value. This indicates the number of samples participating in the evaluation.

Images