Cross-domain migration space-time integrated network algorithm

Abstract

The invention belongs to the technical field of mobile traffic prediction, and particularly relates to a cross-domain migration space-time integrated network algorithm, which comprises the following steps: S1, constructing an integrated learning model, carrying out initial training on a basic learner by utilizing integrated learning, and capturing space-time characteristics in combination with a space-time network, the method is used for overcoming the defects of instability and overfitting in a multi-dimensional space-time cross-domain environment, and an integrated learning model comprises a traffic data set structure, spectral clustering, long-short term memory (LSTM) and a convolutional neural network (CNN). The invention provides a new migration learning algorithm, the algorithm trains a mobile traffic prediction model from source domain data with rich data, then the model is migrated to a mobile edge network environment with sparse data so as to improve the prediction precision of the model, and meanwhile, a new migration mapping framework is provided so as to improve the prediction precision of the model. That is, a reflection function is learned from empirical data and optimized to predict when or how to migrate to achieve optimal predictive performance.

Description

Technical Field

[0001] The present invention belongs to the technical field of mobile traffic prediction, and in particular relates to a cross-domain migration spatiotemporal integrated network algorithm. Background Art

[0002] Mobile edge computing (MEC) has become a hot topic in the next generation of computing architecture, which is a new computing paradigm that transfers network computing resources and task data to the edge network to improve task computing performance and reduce network bandwidth requirements. With the continuous advancement of MEC technology in the Internet industry, it has led to a surge in network traffic generation, that is, mobile traffic accounts for 20% of the total network traffic of web applications, especially when short video mobile applications such as TikTok and blibli appear, mobile devices generate 86% of the total mobile traffic.

[0003] One of the keys to achieving high-quality computing services for MEC is accurate prediction of wireless network traffic, because mobile network services such as offloading, spectrum access, and computing resource allocation require the ability to analyze and predict network traffic (which affects data rates). Although it has been studied in depth by both researchers and the industry, traffic prediction in existing mobile edge networks still faces many challenges: (1) Mobile devices have different service requirements in different locations and at different times, which makes wireless traffic difficult to predict. (2) Wireless traffic prediction for mobile devices involves temporal and spatial characteristics, and complex spatial and temporal dependencies make prediction even more difficult. (3) Due to the different levels of development of monitoring infrastructure in mobile edge networks, sensor data in some areas are relatively sparse, resulting in low accuracy in wireless traffic prediction.

[0004] Therefore, in response to the above technical problems, it is necessary to provide a cross-domain migration spatiotemporal integrated network algorithm.

[0005] The information disclosed in this background technology section is only intended to enhance the understanding of the overall background of the invention and should not be regarded as an acknowledgment or any form of suggestion that the information constitutes the prior art already known to a person skilled in the art. Summary of the invention

[0006] The purpose of the present invention is to provide a cross-domain migration spatiotemporal integrated network algorithm, which can solve the problems in the above-mentioned background technology.

[0007] In order to achieve the above object, a technical solution provided by a specific embodiment of the present invention is as follows:

[0008] A cross-domain migration spatiotemporal integrated network algorithm includes the following steps:

[0009] S1. Construct an integrated learning model, use integrated learning to perform initial training on basic learners, and combine the spatiotemporal network to capture spatiotemporal features to overcome the shortcomings of instability and overfitting in a multi-dimensional spatiotemporal cross-domain environment, wherein the integrated learning model includes a traffic dataset structure, spectral clustering, long short-term memory (LSTM) and a convolutional neural network (CNN);

[0010] S2. In view of the diversity and similarity between the target domain and the source domain, data preprocessing is performed on different cross-domain data. By summarizing historical transfer experience, an experience-based mechanism is established to design transfer reflection functions between the source and target domains, learn and parameterize the feature differences between different target domains and source domains, and optimize knowledge transfer to improve the learning performance in the cross-domain environment of the latent space.

[0011] In one or more embodiments of the present invention, in S1, for the spatiotemporal relationship, the spatial position of the mobile device is defined as X×B, and there are n mobile networks with different degrees of data sparsity, that is, v∈{1,...,n}. Given u base stations, the wireless traffic spatiotemporal sequence of each network data point can be recorded as Θ v ={Θ v，u，t |t=1,...,T}, where T is the total time interval, Θ v，u，t is the traffic flow matrix at point t, defined as:

[0012]

[0013] in, It is used to measure the traffic capacity of mobile networks related to time and space, that is, it is regarded as a tensor

[0014] In one or more embodiments of the present invention, since θ t The value at t+1 depends only on the traffic of the adjacent mobile network and is related to the distant mobile network. Therefore, the autocorrelation coefficient in the spatiotemporal domain can be defined as:

[0015]

[0016] in, is the average value that varies with time domain;

[0017] In order to solve the problem of uneven distribution of mobile network traffic, the spatial correlation is measured by the Pearson correlation coefficient, that is:

[0018]

[0019] Among them, C(.) is the covariance operator, s is the standard deviation, and the spatiotemporal data is regarded as a spatial point in a time series. The spectral clustering strategy is used for clustering, and then dij Denotes the Euclidean distance between two points, that is, it follows that:

[0020]

[0021] where ω ij is the weight between data points x i , x j .

[0022] In one or more embodiments of the present invention, by calculating the weights between data points, a graph composed of all data sets is constructed, and then the sub - graph obtained by connecting the edges with the sum of the weights of the graph as small as possible and the sum of the weights of the graph as large as possible, the measure of the graph is:

[0023]

[0024] According to KNN, the adjacency matrix M is obtained as:

[0025]

[0026] In one or more embodiments of the present invention, according to the spatio - temporal characteristics of mobile network traffic flow, three types of data input are defined: the first is that the input data is {D t-1 , …, D t-m , m ∈ N}, representing the data flow sequence before the target time interval; the second is time - related data, m - data; the third is cross - domain data sets.

[0027] In one or more embodiments of the present invention, based on the three types of data input, a deep network is established, and the deep network includes constructing a spatio - temporal model, m - data embedding, and constructing cross - domain data, where m is a type of time metadata.

[0028] In one or more embodiments of the present invention, in the S2, let E v = ((S v , T v ), l v ) represent the transfer learning memory pool, where is the source domain, is the target domain, is the feature matrix.

[0029] Due to the sparsity of the data, there is ntvl << nsvl, and the transferred knowledge can be parameterized as Wv, lv = psv / ptv represents the performance improvement rate of each transfer learning memory pool, where psv represents the learning performance on the test data set after transferring Wv from the source domain, and ptv represents the learning performance without transfer.

[0030] In one or more embodiments of the present invention, As input, the mobile network intelligent system agent learns the reflection function f(W v , S v , T v ) to approximate l v , given a pair of source domain and target domain, the function f(W v , S v , T v ) can guide how to transmit and improve the performance improvement rate, that is, when When it appears, the mobile network proxy will optimize the Knowledge.

[0031] In one or more embodiments of the present invention, the reflection function f(W v , S v , T v )From the memory pool Approximately v , l v The update is affected by two factors: the difference between the target domain and the source domain and the recognition degree of the target domain. Based on this, the difference between the target domain and the source domain is measured by the following formula:

[0032]

[0033]

[0034] in, express The i-th sample in , Θ is the mapping from the latent space to the reproducing kernel Hilbert space κ(.,.) is the kernel function, which affects the maximum average dispersion distance and thus affects the function f(W v , S v , T v ) value.

[0035] In one or more embodiments of the present invention, unlabeled samples are used to evaluate the recognition performance of the target domain. In the traffic prediction of mobile networks, the unlabeled discrimination criterion is considered:

[0036]

[0037] in:

[0038]

[0039] is the local scatter covariance matrix, and

[0040]

[0041] in, express and r-nearest neighbor. By maximizing τ v , Guaranteed first principles, and

[0042]

[0043] Compared with the prior art, the cross-domain migration spatiotemporal integrated network algorithm of the present invention proposes a new migration learning algorithm, which trains a mobile traffic prediction model from data-rich source domain data, and then migrates the model to a mobile edge network environment with sparse data to improve the prediction accuracy of the model. At the same time, a new migration mapping framework is proposed, that is, learning the reflection function from empirical data and optimizing the reflection function to predict when to migrate or how to transfer to obtain the best prediction performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0045] Figure 1 Figure 1. An ensemble learning architecture for mobile wireless traffic prediction in mobile edge networks.

[0046] Figure 2 Flowchart of spectrum clustering for mobile networks.

[0047] Figure 3 It is a performance comparison chart of the proposed algorithm and the baseline algorithm on the MAE indicator.

[0048] Figure 4 It is a performance comparison chart of the proposed algorithm and the baseline algorithm in terms of RMSE indicator.

[0049] Figure 5 It is a performance comparison chart of the proposed algorithm and the baseline algorithm on the R2 indicator. DETAILED DESCRIPTION

[0050] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0051] like Figure 1 to Figure 2As shown, a cross-domain migration spatiotemporal integrated network algorithm in one embodiment of the present invention includes the following steps:

[0052] S1. Construct an integrated learning model, use integrated learning to initially train the basic learners, and combine the spatiotemporal network to capture spatiotemporal features to overcome the shortcomings of instability and overfitting in a multi-dimensional spatiotemporal cross-domain environment. The integrated learning model includes traffic dataset structure, spectral clustering, long short-term memory (LSTM) and convolutional neural network (CNN);

[0053] S2. In view of the diversity and similarity between the target domain and the source domain, data preprocessing is performed on different cross-domain data. By summarizing historical transfer experience, an experience-based mechanism is established to design transfer reflection functions between the source and target domains, learn and parameterize the feature differences between different target domains and source domains, and optimize knowledge transfer to improve the learning performance in the cross-domain environment of the latent space.

[0054] In S1, for the spatiotemporal relationship, the spatial location of the mobile device is defined as X × B, and there are n mobile networks with different degrees of data sparsity, that is, v∈{1,...,n}. Given u base stations, the spatiotemporal sequence of wireless traffic at each network data point can be recorded as 2v={Θ v,u，t |t=1,...,T}, where T is the total time interval, Θ v，u，t is the traffic flow matrix at point t, defined as:

[0055]

[0056] in, It is used to measure the traffic capacity of mobile networks related to time and space, that is, it is regarded as a tensor

[0057] Since the analysis in the spatiotemporal domain is for all similar mobile network traffic flows, for convenience, the data type of the traffic flow is omitted, i.e., Θ v，u，t =Θ t At the same time, due to Θ t The value at t+1 depends only on the traffic of the adjacent mobile network and is related to the distant mobile network. Therefore, the autocorrelation coefficient in the spatiotemporal domain can be defined as:

[0058]

[0059] in, is the average value that varies with time domain;

[0060] In order to solve the problem of uneven distribution of mobile network traffic, the spatial correlation is measured by the Pearson correlation coefficient, that is:

[0061]

[0062] Among them, C(.) is the covariance operator, s is the standard deviation, and the spatiotemporal data is regarded as a spatial point in a time series. The spectral clustering strategy is used for clustering, and then d ij Expressing the Euclidean distance between two points gives:

[0063]

[0064] Among them, ω ij is the data point x i , x j The weight between .

[0065] By calculating the weights between data points, a graph consisting of all data sets is constructed, and then the sum of the weights of the graph connected to the edges is made as small as possible and the sum of the weights is made as large as possible. The measure of the graph is:

[0066]

[0067] According to KNN, the adjacency matrix M is obtained as:

[0068]

[0069] The spectrum clustering flow chart of mobile network is as follows: Figure 2 As shown, according to the spatiotemporal characteristics of mobile network traffic flow, three types of data input are defined: the first is that the input data is {D t-1 , …, D t-m , m∈N}, representing the data stream sequence before the target time interval; the second is the time-related data, m-data, such as the 5 hours on Friday, February 10; the third is the cross-domain dataset.

[0070] Based on three types of data input, a deep network is established. The deep network includes building a spatiotemporal model, m data embedding, and building cross-domain data, where m is a temporal metadata.

[0071] Among them, data D t The input format of is similar to video stream data, with M data frames. In order to process input in this format, a convolutional neural network (CNN) is introduced. Although CNN has strong spatial data modeling performance and can effectively extract the features of local area information, it cannot capture the feature information of time series. In order to capture the temporal features, we fuse LSTM and CNN and design a multi-layer CNN-LSTM spatiotemporal network.

[0072] At the same time, in order to save historical information, each unit in the CNN-LSTM network is designed to have a storage unit Using Input Gate Forget the door and output gate When new traffic flow data is input into the CNN-LSTM spatiotemporal network, if is activated, the data information carried will be stored in If is activated, the attribute information of the unit can be forgotten or discarded. Output hidden state h according to cell state t The main calculation method of CNN-LSTM spatiotemporal network is as follows:

[0073]

[0074] Among them, B(·) is the layer weight, b(·) is the bias, * is the convolution operation, σ is the activation function, and ⊙ is the Hadamard product. and h t In CNN-LSTM space-time, the network is a three-dimensional tensor. Let M represent the number of feature maps, and the output is

[0075] Among them, in the time-related data, for example, the t-th time point is 15:00:00 on December 10, 2021. We extracted four metadata: hours (15), days (12), months (10), and years (2021), and generated a feature vector m. The feature vector m is input into the CNN-LSTM network, and its dimension changes from 4 to M×X×B. Let fom denote the output of m-data embedding:

[0076]

[0077] in, and are the weights and biases of layer I. After repeating the operation, the new output looks like this:

[0078]

[0079] In order to model the impact of different domains on traffic flow and learn the data features of cross-domain datasets, a two-layer CNN network model can be used. In this network model, datasets from different domains are concatenated to generate a tensor Θ c Then, after Θ c Perform nonlinear transformation operations to obtain Θ c Initial characteristics:

[0080]

[0081] Among them, Bc is the weight, is a concatenation operation. The execution order of function f(·) is BN, ReLU and convolution operations. Using density network, the above three outputs are fused through concatenation operation and written as

[0082]

[0083] Where fo is the initial feature map. Take fo as the input of density. The network architecture has I layers, each of which contains a function f(·), that is, the execution order of function f(·) is BN, ReLU and convolution operations. In order to greatly capture the spatiotemporal characteristics of traffic flow, a connectivity rejection pattern structure can be designed. Therefore, layer I accepts the feature maps of all previous network layers

[0084]

[0085] in, and the final output is The final prediction is

[0086]

[0087] Finally, the objective function is obtained:

[0088]

[0089] Among them, W θ are neural network parameters.

[0090] In S2, let E v =((S v , T v ), l v ) represents the transfer learning memory pool, where is the source domain, is the target domain, is the feature matrix.

[0091] Due to the sparsity of data, there is The transferred knowledge can be parameterized as W v , represents the performance improvement rate of each transfer learning memory pool, represents the learning performance on the test dataset after transferring Wv from the source domain, Indicates the learning performance after no transmission.

[0092] As input, the mobile network intelligent system agent learns the reflection function f(W v , S v , T v ) to approximate lv , given a pair of source domain and target domain, the function f(W v , S v , T v ) can guide how to transmit and improve the performance improvement rate, that is, when When it appears, the mobile network proxy will optimize the Knowledge.

[0093] Using the reflection function f(W v , S v , T v )From the memory pool Approximately v , l v The update is affected by two factors: the difference between the target domain and the source domain and the recognition degree of the target domain. Based on this, the difference between the target domain and the source domain is measured by the following formula:

[0094]

[0095] in, express The i-th sample in , Θ is the mapping from the latent space to the reproducing kernel Hilbert space, κ(.,.) is the kernel function, which affects the maximum average dispersion distance and thus affects the function f(W v , S v , T v ) value. Therefore, the function f(W v , S v , T v ) is essentially to optimize the kernel function κ(.,.), so that the maximum average dispersion can characterize the lv of all target domains and source domains. Based on this, κ(.,.) is parameterized as

[0096]

[0097] Among them, N κ is the number of PSD kernels, δ is the learning coefficient, then

[0098]

[0099] in, and By the kth kernel κ κ Calculation. The kernel function we use is as follows

[0100] κ κ (x, y) = exp(-||xy|| 2 / ω κ )

[0101] where ω κis the bandwidth of the change. Since the maximum average dispersion is difficult to measure the difference between different domains, we need to measure the variance between cross-domain data pairs. The difference between the target domain and the source domain can be written as

[0102]

[0103]

[0104] To ensure consistency with the maximum average dispersion, rewrite in

[0105]

[0106]

[0107] In addition, unlabeled samples are used to evaluate the recognition performance of the target domain. In the traffic prediction of mobile networks, the unlabeled discrimination criterion is considered:

[0108]

[0109] in:

[0110]

[0111] is the local scatter covariance matrix, and

[0112]

[0113] in, express and r-nearest neighbor. By maximizing τ v , The first principle is guaranteed, and

[0114]

[0115] Implement the second principle. Since τ v Depends on the kernel function, according to τ v , from τ v，k The obtained κ k , τ v Then written as:

[0116]

[0117] Among them, τ v =[τ e，1 ,…,τ v (N k )]. Based on the above analysis, we can get the following optimization problem

[0118]

[0119] in,

[0120]

[0121]

[0122]

[0123]

[0124]

[0125] Based on the above formula, we can see that the loss function consists of five parts. is the Huber regression loss, b is the bias. L1 is the training error of the target domain, G t is a feature extractor. L2 is used to constrain the feature distance between the target domain and the source domain. L3 is used to ensure the distinguishability between the target domain and the source domain. L4 is used to prevent overfitting. L0 can be optimized by optimizing the reflection function f(W v , S v , T v , δ * ,λ * , μ * , b * ) to calculate it using the latest feature factor matrix W to optimize the transfer method, that is, W Nv+1 f(W v , S v , T v , δ * ,λ * , μ * , b * ) is maximized. Therefore,

[0126]

[0127] Where φ2 is the complexity factor of W, is the matrix Frobenius norm of W. The first and second terms of the equation are obtained by:

[0128]

[0129] Finally, the problem is optimized using the conjugate gradient method. Stochastic gradient descent (SGD) is used to solve L1-L4:

[0130]

[0131] have

[0132]

[0133] Among them, C 180 is the inverse convolution kernel 180°, and ⊙ is the element-wise product.

[0134]

[0135] have

[0136]

[0137] Experimental example:

[0138] Two real-world datasets are used, which contain mobile network data on telecommunications, weather, news, social networks, and electricity for the city of Milan and the province of Trentino. Since these datasets come from different companies using different standards, their spatial distribution irregularities are clustered in a grid containing 1000 city cells. In the process of optimizing traffic flow, data from multiple time periods and the previous week are used to estimate the missing current traffic parameter data. The simulation uses 8×NVIDIATESLAV100 32G GPUs for Linux simulation. The hyperparameters used in the experiment are shown in the following table:

[0139] parameter value epoch 100 Training batch size 32 Learning Rate 0.001 LSTM Neuron 64 LSTM Activation Function Tanh Convolutional layer filters 16 Kernel size 3*3 CNN Activation Function ReLU

[0140] Root mean square error (RMSE), mean absolute error (MAE) and R-square error (R2) were used as evaluation indicators to verify the convergence performance of the model.

[0141] In order to evaluate the superiority and accuracy of the algorithms, the following traffic prediction algorithms are selected as baselines: including: LSTM, Support Vector Regression (SVR), Temporal Graph Convolutional Network (T-GCN), Spatiotemporal Graph Convolutional Network (ST-GCN), and Densely Connected Network (DenseNet).

[0142] This application adopts the framework of ensemble learning, uses clustering algorithm to divide mobile edge networks into different categories according to the feature similarity between different mobile target domains, and uses CNN-LSTM hybrid model to fit the actual wireless traffic of mobile edge networks. In view of the different data sparsity in different mobile network areas, a transfer learning algorithm based on experience memory pool and reflection function is proposed to achieve cross-domain traffic prediction. Finally, we use two real datasets from Milan and Trentino Province to evaluate the prediction performance of the proposed TSTEN. Figure 3 , Figure 4 and Figure 5 The three indicators of MAE, RMSE and R2 were analyzed, and the experiments showed that TSTEN outperformed the existing baselines in the accuracy of cross-domain traffic prediction.

[0143] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

[0144] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A cross-domain migration spatiotemporal integrated network algorithm, characterized in that: The following steps are involved: S1. Construct an integrated learning model, use integrated learning to perform initial training on basic learners, and combine the spatiotemporal network to capture spatiotemporal features to overcome the shortcomings of instability and overfitting in a multi-dimensional spatiotemporal cross-domain environment, wherein the integrated learning model includes a traffic dataset structure, spectral clustering, long short-term memory (LSTM) and a convolutional neural network (CNN); S2. In view of the diversity and similarity between the target domain and the source domain, data preprocessing is performed on different cross-domain data. By summarizing historical transfer experience, an experience-based mechanism is established to design transfer reflection functions between the source and target domains, learn and parameterize the feature differences between different target domains and source domains, and optimize knowledge transfer to improve the learning performance in the cross-domain environment of the latent space.

2. According to the cross-domain migration spatiotemporal integrated network algorithm of claim 1, it is characterized in that: In S1, for the spatiotemporal relationship, the spatial position of the mobile device is defined as X×B, and there are n mobile networks with different degrees of data sparsity, i.e., v∈{t,...,n}. Given u base stations, the spatiotemporal sequence of wireless traffic at each network data point can be recorded as Where T is the total time interval, Θ v，t is the traffic flow matrix at point t, defined as: in, It is used to measure the traffic capacity of mobile networks related to time and space, that is, it is regarded as a tensor 3. The cross-domain migration spatiotemporal integrated network algorithm according to claim 2 is characterized in that: Since Θ t The value at t+t depends only on the traffic of the adjacent mobile network and is related to the distant mobile network. Therefore, the autocorrelation coefficient in the spatiotemporal domain can be defined as: in, is the average value that varies with time domain; In order to solve the problem of uneven distribution of mobile network traffic, the spatial correlation is measured by the Pearson correlation coefficient, that is: Among them, C(.) is the covariance operator, s is the standard deviation, and the spatiotemporal data is regarded as a spatial point in a time series. The spectral clustering strategy is used for clustering, and then d ij Expressing the Euclidean distance between two points gives: Among them, ω ij is the data point x i , x j The weight between .

4. The cross-domain migration spatiotemporal integrated network algorithm according to claim 3 is characterized in that: By calculating the weights between data points, a graph consisting of all data sets is constructed, and then the sum of the weights of the graph connected to the edges is made as small as possible and the sum of the weights is made as large as possible. The measure of the graph is: According to KNN, the adjacency matrix M is obtained as:

5. The cross-domain migration spatiotemporal integrated network algorithm according to claim 4 is characterized in that: According to the spatiotemporal characteristics of mobile network traffic flow, three types of data input are defined: the first is the input data {D t-t , …, D t-m , m∈N}, represents the data stream sequence before the target time interval; the second is the time-related data, m-data; the third is the cross-domain dataset.

6. The cross-domain migration spatiotemporal integrated network algorithm according to claim 5, characterized in that: Based on three types of data input, a deep network is established, which includes building a spatiotemporal model, m data embedding, and building cross-domain data, where m is a temporal metadata.

7. The cross-domain migration spatiotemporal integrated network algorithm according to claim 6, characterized in that: In S2, let E v =((S v , T v ), l v ) represents the transfer learning memory pool, where is the source domain, is the target domain, is the feature matrix. Due to the sparsity of data, there is The transferred knowledge can be parameterized as W v , represents the performance improvement rate of each transfer learning memory pool, represents the learning performance on the test dataset after transferring Wv from the source domain, Indicates the learning performance after no transmission.

8. The cross-domain migration spatiotemporal integrated network algorithm according to claim 7, characterized in that: As input, the mobile network intelligent system agent learns the reflection function f(W v , S v , T v ) to approximate l v , given a pair of source domain and target domain, the function f(W v , S v , T v ) can guide how to transfer and improve the performance improvement rate, that is, when a pair of new domains When it appears, the mobile network proxy will optimize the Knowledge.

9. The cross-domain migration spatiotemporal integrated network algorithm according to claim 8, characterized in that: Using the reflection function f(W v , S v , T v )From the memory pool Approximately v , l v The update is affected by two factors: the difference between the target domain and the source domain and the recognition degree of the target domain. Based on this, the difference between the target domain and the source domain is measured by the following formula: in, express The i-th sample in , Θ is the mapping from the latent space to the reproducing kernel Hilbert space, κ(.,.) is the kernel function, which affects the maximum average dispersion distance and thus affects the function f(W v , S v , T v ) value.

10. The cross-domain migration spatiotemporal integrated network algorithm according to claim 9, characterized in that: Unlabeled samples are used to evaluate the recognition performance of the target domain. In the traffic prediction of mobile networks, the unlabeled discrimination criterion is considered: in: is the local scatter covariance matrix, and in, express and r-nearest neighbor. By maximizing τ v , It ensures the first principle, and

Images