User Mobile Cellular Network Data Synthesis Method and System Based on Deep Generative Adversarial Network

Abstract

The present invention discloses a method and system for synthesizing user mobile cellular network data based on a deep generative adversarial network. The method constructs a user multi-dimensional attribute synthesis module, a user traffic usage sequence synthesis module, and a user spatial base station connection sequence synthesis module, jointly trains the user multi-dimensional attribute synthesis module, the user traffic usage sequence synthesis module, and the user spatial base station connection sequence synthesis module, and uses the trained user multi-dimensional attribute synthesis module, the user traffic usage sequence synthesis module, and the user spatial base station connection sequence synthesis module to synthesize data. Compared with the prior art, in the present invention, three mobile network data with associated attributes are respectively constructed with data synthesis models, and then the data synthesis models are jointly trained, so that the synthesized data of the trained model not only has the attributes of the data itself, but also has the relevance with other data, greatly improving the reliability and quality of the synthesized data.

Description

Technical Field

[0001] The present invention relates to the technical field of data synthesis, and particularly to a method and system for synthesizing user mobile cellular network data based on a deep generative adversarial network. Background Art

[0002] Mobile cellular networks are widely deployed in smart cities, which is crucial for realizing wireless communication and data transmission of mobile users. At the same time, a large amount of accumulated mobile cellular data plays a crucial role in various practical applications such as cellular traffic prediction, network optimization, urban planning, and human mobility modeling.

[0003] Despite the huge potential of data, the inaccessibility of large-scale cellular network data also poses a huge challenge to data-driven research in this field. This is because network operators are usually reluctant to share data due to concerns about disclosing trade secrets or infringing on user privacy. Therefore, limited data access has become a huge obstacle for the network system community to promote data-driven research and development.

[0004] To fully utilize the research value of these data, a promising approach is to generate and share data synthesized from real data. In recent years, researchers have extensively explored data synthesis techniques, including generating single cellular traffic data, radio data, traffic of Internet of Things devices, and network packet header traffic. In addition, these techniques have been extended to synthesize various sensor data, such as millimeter-wave radar, Doppler radar, and inertial measurement unit data.

[0005] However, due to limitations in data availability and quality, most existing works mainly focus on generating single network traffic and fail to model the combination of multi-dimensional features such as entity attributes and spatial mobility. For example, in a mobile cellular network environment, different user groups show different temporal patterns in the use of application (App) traffic, and each user also shows different spatial base station association characteristics.

[0006] Therefore, although most existing methods can meet the synthesis of network and sensing data in different scenarios, they cannot be directly applied to the task of generating mobile cellular network data. This is because the synthesis of mobile cellular network data requires combining spatio-temporal sequences and user attributes to achieve different downstream applications. Summary of the Invention

[0007] The present invention provides a method and system for synthesizing user mobile cellular network data based on a deep generative adversarial network, which is used to solve the technical problem that the synthesized data of the existing mobile network data synthesis method has poor reliability and quality.

[0008] To solve the above technical problem, the technical solution proposed by the present invention is as follows:

[0009] A method for synthesizing user mobile cellular network data based on a deep generative adversarial network, comprising the following steps:

[0010] Construct a user multi-dimensional attribute synthesis module, which is used to learn the attribute characteristics of real users with different attributes and synthesize user attributes according to the attribute characteristics of real users;

[0011] Construct a user traffic sequence synthesis module, which is used to learn the traffic time series characteristics of real traffic sequences of real users with different attributes and synthesize the traffic time series of users with different attributes according to the traffic time series characteristics; the input of the user traffic sequence synthesis module includes the output of the user multi-dimensional attribute synthesis module;

[0012] Construct a user spatial base station connection sequence synthesis module, which is used to learn the spatial connection characteristics of real spatial base station connection sequences of real users with different attributes and synthesize the spatial base station connection sequences of users with different attributes according to the spatial connection characteristics; the input of the user spatial base station connection sequence synthesis module includes the output of the user multi-dimensional attribute synthesis module;

[0013] Jointly train the user multi-dimensional attribute synthesis module, the user traffic sequence synthesis module, and the user spatial base station connection sequence synthesis module, and use the trained user multi-dimensional attribute synthesis module, the user traffic sequence synthesis module, and the user spatial base station connection sequence synthesis module to synthesize data.

[0014] Preferably, the user multi-dimensional attribute synthesis module is a generative adversarial network, including an attribute generator and an attribute discriminator;

[0015] The attribute generator includes an attribute encoder and an attribute decoder. The attribute encoder generates an encoded feature h according to the input attribute noise vector e , and the attribute decoder is used to decode the encoded feature h e to synthesize a user attribute vector A generated ;

[0016] The attribute discriminator is used to compare the synthesized attribute vector A generated with the real attribute vector A real to judge the authenticity of the synthesized attribute vector A generated , and adjust the network parameters of the attribute generator according to the judgment result until the attribute discriminator discriminates that the synthesized attribute vector A generated is true.

[0017] Preferably, the user multi-dimensional attribute synthesis module encodes the input real user attributes using one-hot encoding to obtain a real encoding vector, and constructs a real attribute vector A based on the real encoding vectors of multiple different users. real ;

[0018] The attribute encoder includes a first Linear layer, a Leaky ReLU layer, and a BatchNorm layer connected in sequence. The attribute decoder includes a second Linear layer and a Softmax layer connected in sequence. The output end of the BatchNorm layer is connected to the input end of the second Linear layer.

[0019] The attribute discriminator includes a third Linear layer and a second Leaky ReLU layer connected in sequence. The input of the third Linear layer is the synthetic attribute vector A generated concatenated with the real attribute vector A real .

[0020] Preferably, the user usage traffic sequence synthesis module is a conditional temporal generative adversarial network, including: a temporal generator and a temporal discriminator. The temporal generator is used to synthesize the input synthetic attribute vector A generated into a time series vector T of the usage traffic of the corresponding user generated ; The temporal discriminator is used to compare the time series vector T generated with the real usage traffic sequence T real , and judge the authenticity of the time series vector T generated according to the comparison result, and adjust the network parameters of the temporal generator according to the judgment result until the temporal discriminator judges that the synthesized time series vector T generated is true.

[0021] Preferably, the input of the temporal generator includes a temporal noise vector and the synthetic attribute vector A generated , and the temporal generator includes:

[0022] A concatenation layer for concatenating the temporal noise vector and the synthetic attribute vector A generated ;

[0023] A TCN block for capturing the short-term temporal features of the usage traffic of the synthetic attribute vector A based on the temporal noise vector generated ;

[0024] A Transformer encoding block for synthesizing the time series vector T based on the short-term temporal features generated ;

[0025] The timing discriminator includes a fourth Linear layer and a third Leaky ReLU layer connected in sequence; the input of the fourth Linear layer is the time series vector T generated and the concatenated vector with the real usage traffic sequence T real .

[0026] Preferably, the TCN block satisfies:

[0027] h TCN = Leaky ReLU(Dropout(TCN([T nosie , A generated )))

[0028] where h TCN is the short-term timing feature; TCN is the temporal convolutional network, Dropout is the overfitting function, and LeakyReLU is the rectified linear activation function;

[0029] The Transformer encoding block satisfies:

[0030] T generated = Norm(TransEncoder(h TCN ))

[0031] where TransEncoder is the autoencoder and Norm is the normalization function;

[0032] and / or

[0033] The real usage traffic sequence T real is obtained through the following method:

[0034] Obtain the usage traffic sequences of multiple real users, perform normalization processing on the usage traffic sequences, and convert the multiple normalized usage traffic sequences into a three-dimensional matrix, where the dimensions of the three-dimensional matrix include user attributes, traffic time steps, and traffic types.

[0035] Preferably, the user space base station connection sequence synthesis module is a conditional space generative adversarial network, including: a space generator and a space discriminator. The space generator is used to synthesize a user space base station connection sequence according to the input space noise vector and user attribute vector; the space discriminator is used to identify the authenticity of the synthesized user space base station connection sequence and adjust the network parameters of the space generator according to the identification result until the space discriminator identifies the synthesized user space base station connection sequence as true.

[0036] Preferably, the space generator includes a plurality of sub-generators, and the space discriminator includes a plurality of sub-discriminators. The plurality of sub-generators, the plurality of sub-discriminators, and the plurality of users correspond to each other one by one. Each sub-generator is used to capture the correlation feature of the spatial base station connection sequence of its corresponding user, and synthesize the user spatial base station connection sequence of the corresponding user according to the correlation feature of the spatial base station connection sequence. Each sub-discriminator is used to identify whether the synthesized user spatial base station connection sequence of its corresponding sub-generator is true, and adjust the network parameters of the corresponding sub-generator according to the identification result until the corresponding sub-generator identifies that the synthesized user spatial base station connection sequence is true.

[0037] Preferably, each sub-generator includes an LSTM network, a Linear network, and a Softmax network connected in sequence;

[0038] The LSTM network is used to capture the correlation feature of the spatial base station connection sequence of its corresponding user;

[0039] The Linear network is used to calculate the probability matrix of the user spatial base station connection sequence according to the correlation feature of the spatial base station connection sequence;

[0040] The probability matrix satisfies:

[0041]

[0042] where, P i represents the probability matrix of the i-th user u i , and are the weight matrix and bias vector of the Linear layer of user u i respectively; represents the feature vector of the correlation of the spatial base station connection sequence of user u i ; represents the spatial noise vector of user u i , A generated represents the synthetic user attribute of user u i ;

[0043] Each sub-discriminator includes a fifth Linear layer and a fourth Leaky ReLU layer connected in sequence. The input of the fifth Linear layer is the concatenated vector of the synthesized user spatial base station connection sequence and the real user spatial base station connection sequence.

[0044] Preferably, the loss function of the joint training is as follows:

[0045]

[0046] where, G ais an attribute generator, D a is an attribute discriminator, L(G a , D a ), D t is a temporal generator, D t is a temporal discriminator, L(G t , D t ), D s is a spatial generator, D s is a spatial discriminator, L(G s , D s ), D is the generated sample, represents G a , G t , G s generator distribution, E represents the expectation function, represents the gradient descent function.

[0047] A computer system includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the above method are implemented.

[0048] The present invention has the following beneficial effects:

[0049] 1. In the present invention, by separately constructing data synthesis models for three types of mobile network data with associated attributes (that is, designing a new framework CellSyn, which consists of three modules based on generative adversarial networks: the user App traffic sequence synthesis module, the user spatial base station connection sequence synthesis module, and the user multi-dimensional attribute synthesis module, which respectively learn dynamic time patterns, heterogeneous spatial patterns, and complex multi-dimensional relationships with user attributes), and then jointly training the data synthesis models, the synthesized data of the trained models not only has the attributes of the data itself but also has associations with other data, greatly improving the reliability and quality of the synthesized data. In addition, the present invention, by perceiving the behavior patterns of users' associated mobile networks, learns and synthesizes large-scale multi-dimensional mobile cellular network data. The data synthesized by the present invention has important value for mobile network computing communities lacking real data.

[0050] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The following will refer to the accompanying drawings to further elaborate on the present invention in detail. Brief Description of the Drawings

[0051] The accompanying drawings, which form a part of this application, are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0052] Figure 1 is the framework diagram of the data synthesis framework CellSyn for user-associated mobile network behaviors

[0053] Figure 2(a) is the framework diagram of the user multi-dimensional attribute synthesis module

[0054] Figure 2(b) is the framework diagram of the user App usage traffic sequence synthesis module

[0055] Figure 2(c) is the framework diagram of the user spatial base station connection sequence synthesis module

[0056] Figure 3(a) is the CDF graph of the EMD values of the uplink traffic

[0057] Figure 3(b) is the CDF graph of the EMD values of the downlink traffic

[0058] Figure 3(c) is the CDF graph of the EMD values of the base station ID sequence

[0059] Figure 3(d) is the CDF graph of the JSD values of gender

[0060] Figure 3(e) is the CDF graph of the JSD values of age

[0061] Figure 4(a) is the graph of the average EMD values of the App usage traffic

[0062] Figure 4(b) is the graph of the average EMD values of the base station ID

[0063] Figure 4(c) is the graph of the average EMD values of the user attributes

[0064] Figure 5 is the Acc@3 graph for predicting the next connected base station. Detailed implementation manners

[0065] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings. However, the present invention can be implemented in many different ways defined and covered by the claims.

[0066] The technical solutions of this application will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0067] Glossary:

[0068] One-Hot Encoding: A commonly used data encoding technique for representing discrete features in the form of binary vectors.

[0069] GAN: It is the abbreviation of Generative Adversarial Network, which is a deep learning model composed of a Generator and a Discriminator. Its goal is to generate realistic new samples similar to real data by training the Generator and the Discriminator against each other.

[0070] Linear layer: It is used to perform a linear transformation on the input, thereby introducing non-linear relationships and learning appropriate weight and bias parameters to achieve more complex function mappings.

[0071] Leaky ReLU layer: It is a variant of the rectified linear unit and is used for the non-linear transformation of the activation function.

[0072] BatchNorm layer: It is a commonly used neural network layer and a technique for normalizing input data in a deep learning model.

[0073] Softmax layer: It is a commonly used activation function in neural networks and is used to convert the output of the network into a probability distribution.

[0074] Temporal Convolutional Network (TCN): It is a model based on the structure of Convolutional Neural Network (CNN) and is specifically used for the modeling and prediction of time series data.

[0075] Transformer Encoder Block: It is the basic building block in the Transformer model and is used for the encoding and feature extraction of input sequence data.

[0076] LSTM network: The LSTM (Long Short-Term Memory) network is a variant of the Recurrent Neural Network (RNN) and is used to process and model sequence data, especially sequence data with long-term dependencies.

[0077] To address the challenges of multi-scale time variations, complex multi-dimensional relationships, and heterogeneous user-base station association patterns in the process of synthesizing cellular traffic data, the present invention proposes a method for synthesizing user mobile cellular network data based on a deep generative adversarial network, as Figure 1 shown, which is divided into the following steps:

[0078] S1 To synthesize user attributes, App usage traffic sequences, and spatial base station connection sequences, the present invention designs three modules to achieve this, and the following is a description of each module:

[0079] 1) User multi-dimensional attribute synthesis module: This module uses an Attribute Generative Adversarial Network (AT-GAN) to capture and synthesize different user attributes. As shown in Figure 2(a), the AT-GAN consists of an attribute generator and an attribute discriminator: The attribute generator includes an encoder and a decoder. The encoder includes a Linear layer, a Leaky ReLU layer, and a BatchNorm layer. The decoder includes a Linear layer and a Softmax layer. The attribute discriminator includes a Linear layer and a Softmax layer. The attribute generator adopts an encoder-decoder structure. The encoder maps the noise vector to the latent space through a series of layers (including a Linear layer, a Leaky ReLU layer, and a BatchNorm layer) and passes it to the decoder. The decoder decodes the vector in the latent space into user attributes, including an anonymous user identifier (uid), age, gender, and App preference categories. The attribute discriminator is used to judge the authenticity of the generated user attributes.

[0080] 2) User App usage traffic sequence synthesis module: This module uses a Conditional Temporal Generative Adversarial Network (CT-GAN) to synthesize the App usage traffic sequence according to user attributes. As shown in Figure 2(b), the CT-GAN consists of two parts: a temporal generator and a temporal discriminator. The temporal generator consists of a TCN (Teporal Convolutional Network) and a Transformer encoding block. The TCN is used to capture short-term (e.g., daily and weekly) time patterns, and the Transformer encoding block is used to capture long-term (e.g., monthly granularity) dependencies in sequential data. The temporal discriminator consists of a Linear layer and a Leaky ReLU layer. The temporal generator uses user attributes as conditions to synthesize the App usage traffic sequence, and the temporal discriminator is used to distinguish the synthesized traffic sequence from the real traffic sequence.

[0081] 3) User Space Base Station Connection Sequence Synthesis Module: In this module, a Conditional Spatial Generative Adversarial Network (CS-GAN) is used to generate personalized base station connection sequences for each user. As shown in Figure 2(c), the CS-GAN consists of n (n represents the number of users in the real dataset) spatial generators and n spatial discriminators. Each spatial generator includes an LSTM network, a Linear layer, and a Softmax layer, which are used to learn the association relationship between user attributes and spatial base station connection sequences and synthesize spatial base station connection sequences for each user. Each spatial discriminator includes a Linear layer and a Leaky ReLU layer, which are used to judge the authenticity of the synthesized spatial base station connection sequences.

[0082] S2 Use the user multi-dimensional attribute synthesis module to synthesize user attributes

[0083] Based on the observation of the impact of user attributes on the traffic characteristics of App usage, for example, young users show more traffic consumption than older users. Therefore, it is particularly important to design a modeling mechanism for the joint distribution of user spatio-temporal behavior and attributes. To effectively address this challenge, the present invention uses a user multi-dimensional attribute synthesis module to synthesize user attributes such as age, gender, and other relevant factors that affect the temporal App usage traffic and spatial base station association. Specifically as follows:

[0084] S2.1 First, perform attribute deduplication operations on all users according to the user ID. Through this step, ensure that each unique user ID has its corresponding attribute information such as gender and age. The purpose is to synthesize diverse and non-repetitive user attributes. By removing duplicate user attributes, ensure that each user's attributes only appear once during the synthesis of user attributes, providing more variations and diversities for the synthesis of the model, enabling the model to learn from various different combinations of user attributes and synthesize more diverse and realistic data.

[0085] User attributes are discrete categorical values. These values are preprocessed by converting the discrete variables into one-hot encoding. Specifically, given a user attribute variable x with a value range of [1, n], the formula for one-hot encoding is as follows:

[0086] OneHot(x) = [0, 0, …, 1, …, 0]

[0087] Among them, the length of the vector is n, and the element at the x-th position is 1, and the elements at other positions are 0. For example, if the user attribute variable x has 3 possible values, the formula for one-hot encoding is as follows:

[0088] OneHot(x = 1) = [1, 0, 0]

[0089] OneHot(x = 2) = [0, 1, 0]

[0090] OneHot(x = 3) = [0, 0, 1]

[0091] Through one - hot encoding, the original user attribute variables are represented as multiple binary features, which are convenient for input into the user multi - dimensional attribute synthesis module for processing.

[0092] S2.3 Stack the one - hot encoded vectors of all users' attributes row - by - row to form a matrix vector of real attributes, and define this matrix vector as A real . The number of rows of this matrix corresponds to the number of users n, and the number of columns corresponds to the vector dimension d = n * c after one - hot encoding, where n represents the dimension after one - hot encoding of each user attribute variable, and c represents the number of user attribute variables. The finally obtained matrix vector has dimension (n, d), where n represents the number of users and d represents the one - hot encoding dimension of each user's attributes. Each row represents the attribute encoding information of a user. This matrix vector is provided as input data to the user multi - dimensional attribute synthesis module for training and sampling to achieve the modeling of user behavior, preferences, and other features.

[0093] S2.4 Define an attribute noise vector where R represents the real number field, and n z is the dimension of the noise vector. This vector contains random attribute information and can be used to synthesize vectors that conform to the user attribute distribution.

[0094] Through the attribute generator of AT - GAN in the user multi - dimensional attribute synthesis module, using the attribute noise vector A noise sample the distribution characteristics of the user attribute latent space to synthesize an attribute vector A that fits the user attribute distribution generated .

[0095] S2.4.1 Specifically, the attribute generator G a uses the noise vector A noise as input to synthesize different user attributes. G a employs an encoder and a decoder to achieve this task. The encoder consists of several stacked layers and outputs an encoded feature h e , which is defined as follows:

[0096] h e = BatchNorm(Leaky ReLU(W e A noise + b e ))

[0097] where We and b e are the weight matrix and bias vector of the Linear layer in the encoder, respectively. Leaky ReLU and BatchNorm are activation functions and batch normalization operations, which help the model converge and generalize faster.

[0098] S2.4.2 Subsequently, the encoded feature h e is fed into the decoder, and the output of the decoder is the synthetic attribute vector where R represents the real number field, and n α is the dimension of the attribute vector, which is given by:

[0099] A generated = Softmax(W d h e + b d )

[0100] where W d and b d are the weight matrix and bias vector of the Linear layer in the decoder, respectively.

[0101] S2.4.3 Then, the synthetic attribute vector A generated is concatenated with the real attribute matrix vector A real and provided to the attribute discriminator D a , and D a is responsible for distinguishing the synthetic attribute vector from the real attribute matrix vector. It receives the real attribute matrix vector and the attribute vector synthesized by the attribute generator and tries to distinguish their sources. By continuously optimizing the attribute vector synthesized by the attribute generator, it makes it difficult for the attribute discriminator to distinguish the real attribute matrix vector from the synthesized attribute vector, thereby improving the synthesis ability of the attribute generator.

[0102] The attribute discriminator D a uses a Linear layer and a Softmax layer to give the difference between A generated and A real . Its output is a binary label y ∈ {0, 1}, indicating whether the input attribute vector is real (y = 1) or generated (y = 0), which are defined as D a (R) and D a (F), respectively.

[0103] S2.4.4 By continuously iteratively optimizing the attribute generator G a and the attribute discriminator D a , the user multi-dimensional attribute synthesis module can learn the distribution characteristics of the user attribute latent space and synthesize the attribute vector A generated that conforms to this distribution.

[0104] S3 passes through A generated Use the user App to synthesize the App usage traffic sequence with the traffic sequence synthesis module

[0105] S3.1 The data scales of the App usage traffic sequences in the real data are inconsistent, including some very large traffic values (such as 20000MB) and some very small traffic values (such as 0.03MB). Such a data form is not conducive to the training and learning of the user App usage traffic sequence synthesis module. Therefore, the App usage traffic sequences in the real data are first preprocessed to eliminate the data size differences.

[0106] Normalization is a commonly used data preprocessing technique for mapping data in different ranges to a unified range to eliminate the scale differences between data. When processing the App usage traffic sequences in the real data, normalization can improve the training effect and generalization ability of the model.

[0107] Common normalization methods include min-max scaling and standardization. The min-max scaling adopted in the present invention is a linear transformation method that scales the data to a specified range, usually [0,1]. The specific normalization formula is as follows:

[0108]

[0109] where X is the App usage traffic sequence in the real data, X min is the sequence minimum value, and X max is the sequence maximum value. By this method, the minimum value in the App usage traffic sequence is mapped to 0, the maximum value is mapped to 1, and other values are mapped to this range proportionally. Min-max scaling is applicable to the case where the data distribution does not deviate significantly from the normal distribution, and it retains the relative relationship and distribution shape of the real data.

[0110] Through min-max scaling processing, the App usage traffic sequences in the real data are converted into a standardized sequence distribution, making the data more in line with the requirements of the user App usage traffic sequence synthesis module.

[0111] S3.2 Convert the standardized sequence distribution into a three-dimensional matrix T of shape (n, t, c) real , where n represents the number of users, t represents the time step, and c represents the number of traffic types.

[0112] Through this matrix form representation, each user App usage traffic sequence is represented as a three-dimensional matrix T real, where every t rows represent the App usage traffic sequence of a user, and each column represents a traffic type. Specifically, for each user, the traffic data used in hourly time granularity within a month is arranged in chronological order to form the App usage traffic sequence of a user. Then, the uplink traffic and the downlink traffic are used as different traffic types respectively to construct a matrix with two traffic types.

[0113] Illustrated by an example, assume there are 1000 users, with data for 31 days in a month, and there are two types of traffic: uplink traffic and downlink traffic. Then, the processed user App usage traffic sequence after min-max scaling is represented as a three-dimensional matrix with a shape of (1000, 31 * 24, 2).

[0114] In this matrix, the first dimension represents the number of users, which is 1000. The second dimension represents the time step, which is 744, representing each hour of each day. The third dimension represents the number of traffic types, which is 2, namely uplink traffic and downlink traffic.

[0115] This matrix form of representation helps the user App usage traffic sequence synthesis module to learn the temporal correlation in the user App usage traffic sequence and the relationship between traffic types.

[0116] S3.3 Utilize the user attribute vector A synthesized in S2 generated , and use the CT-GAN of the user App usage traffic sequence synthesis module to synthesize the user App usage traffic sequence.

[0117] Specifically, the temporal generator G in CT-GAN t combines the TCN block and the Transformer encoding block to learn the multi-scale temporal patterns in the time series data. The input of G t is the concatenated vector of the temporal noise vector and A generated , where R represents the real number field, and n t represents the dimension of the temporal noise vector.

[0118] S3.3.1 For short-term temporal pattern learning, G t adopts the TCN block to capture short-term (e.g., daily and weekly) temporal patterns. TCN is a one-dimensional convolutional network using random convolution and dilated convolution. Specifically, let X t ∈R T represent the concatenated feature vector at time step t, where T represents the total number of time steps in the time series. First, represent the time X = {X1, X2,..., X t} is input into the causal convolution to predict the corresponding output at each time step. Then, the extended convolutional component uses a deep stack of extended convolutions to capture long-range temporal patterns. In the design of G t , two extended causal convolutional layers are used, and a normalization layer is included as one of the hidden layers in the TCN block. Subsequently, Leaky ReLU and Dropout are applied to prevent overfitting. The TCN part in G t can be expressed as:

[0119] h TCN = LeakyReLU(Dropout(TCN([T nosie , A generated )))

[0120] S3.3.2 For long-term temporal pattern learning, G t employs a Transformer encoding block to capture long-term (e.g., monthly granularity) dependencies in the user App usage traffic sequence. The Transformer encoding block uses the multi-head self-attention mechanism to dynamically weigh the importance of different time steps in the input sequence, enabling the model to capture long-term temporal patterns. Specifically, let represent the vectors output from the TCN block, where T is the length of the input sequence, and these vectors are then processed by the Transformer encoding block to produce the time series vector where represents the hidden state of the Transformer encoding block at time step t. Finally, a normalization function is used. The Transformer part in G t can be expressed as:

[0121] T generated = Norm(TransEncoder(h TCN ))

[0122] S3.4 The temporal discriminator D in CT-GAN t takes the time series vector T generated and the three-dimensional matrix T real as inputs and outputs the discrimination result, i.e., a binary label y ∈ {0, 1}, indicating whether T generated is real (y = 1) or synthetic (fake) (y = 0), defined as D t (R) and D t (F) respectively. By combining G t and D t , CT-GAN effectively captures the inherent temporal dependencies in the input user App usage traffic sequence.

[0123] S3.5 Optimize the timing generator G through continuous iteration t and the timing discriminator D t so that CT-GAN can learn the distribution characteristics of the App usage traffic sequence and synthesize the time series vector T that conforms to this distribution generated .

[0124] S4 synthesizes the user's spatial base station connection sequence through A generated using the spatial base station connection sequence synthesis module

[0125] S4.1 First, preprocess the user's spatial base station connection sequence in the real data, represent the user's spatial base station connection sequence in the real data as a two-dimensional list, each list in the two-dimensional list represents the spatial base station connection sequence of each user, and according to each list in the two-dimensional list, perform an embedding operation on the user's spatial base station connection sequence in the real data to represent the discrete base station identifier as a continuous and high-dimensional user's spatial base station connection sequence embedding vector The superscript i represents the i-th user, i = 1, 2,..., n.

[0126] Specifically, use the pre-trained embedding layer to convert the base station identifier of the user's spatial base station connection sequence in the real data into the corresponding embedding vector. This embedding layer is trained based on the user's spatial base station connection sequence dataset in the real data and learns the similarity and semantic information between the base stations

[0127] For each list in the two-dimensional list, traverse each base station identifier in it and convert it into the corresponding embedding vector through the embedding layer. In this way, a new two-dimensional list is obtained, where each list consists of the embedding vectors of the base station identifiers

[0128] For example, assume that each list in the two-dimensional list represents the spatial base station connection sequence of a user, where the base station identifier is an integer value. If the output dimension of the embedding layer is d, then each base station identifier can be input into the embedding layer to obtain an embedding vector with dimension d. After traversing the entire spatial base station connection sequence, a new two-dimensional list is obtained, where each list consists of the embedding vectors of the base station identifiers, with the shape of (n, (t, d)), where n represents the number of users, t represents the length of the spatial base station connection sequence of each user, and d represents the dimension of the embedding vector

[0129] S4.2 Utilize the user attribute vector A synthesized by S2 generated, the user - space base - station connection sequence synthesis module is used to synthesize the user - space base - station connection sequence. Considering the uniqueness of each user, it is necessary to capture the unique characteristics associated with the base stations of each user. To achieve this goal, the user - space base - station connection sequence synthesis module focuses on synthesizing user - specific space base - station connection sequences to reflect the inherent uniqueness of the user's movement pattern. For this purpose, in the CS - GAN of the space base - station connection sequence synthesis module, for each user u i is assigned a dedicated space generator and space discriminator, denoted as and respectively, and they have the same model structure. The subscript i represents the i - th user, where i = 1, 2, …, n.

[0130] S4.2.1 In , the input is a connection vector, which consists of two key elements: the space noise vector and the user attribute vector A a synthesized from the attribute generator G generated , where R represents the real - number field, and n s represents the dimension of the space noise vector. First, a LSTM network is used to capture the correlations in the user - space base - station connection sequence, and outputs a feature vector containing the correlations of the user - space base - station connection sequence denoted as follows:

[0131]

[0132] S4.2.2 Then, a personalized Linear layer is used to customize the output synthesized for each user u i . Specifically, to ensure that the synthesized base - station IDs come from the space base - station connection sequence of each user u i , a user - specific Linear layer is designed in . In this Linear layer, for each , a different output dimension is assigned to the Softmax layer, and this dimension corresponds to the number of unique base - station IDs associated with the user u i . The output of this Linear layer is a probability matrix representing the synthesized user - space base - station connection sequence where R represents the real - number field, n s represents the dimension of the space noise vector, BS represents the base - station IDs that each user has ever connected to, and P i is denoted as follows:

[0133]

[0134] Among them, and are the weight matrix and bias vector of the Linear layer of user u i respectively.

[0135] S4.3 Finally, the probability matrix P i and the embedding vector are input into the corresponding spatial discriminator . A Linear layer and a Leaky ReLU layer are used, and its output generates a label y ∈ {0, 1}, indicating whether the synthesized user space base station connection sequence is true (y = 1) or false (y = 0). These operations are defined as and

[0136] S4.4 By continuously iteratively optimizing the spatial generator and the spatial discriminator CS-GAN can learn the feature distribution of the user space base station connection sequence and synthesize the probability matrix P i of the user space base station connection sequence that conforms to this distribution. The synthesized user space base station connection sequence is determined by the probability matrix P i .

[0137] S5 Training and Loss Function Design

[0138] The entire model training process consists of two stages. First, train the attribute generator g a and the attribute discriminator D a to synthesize the user attribute vector A generated . Subsequently, train the temporal generator G t and the temporal discriminator D t to synthesize the time series vector T generated . In addition, train the spatial generator G s and the spatial discriminator D s to synthesize the user space base station connection sequence. For each stage, the Wasserstein distance with gradient penalty is used. This method ensures the robustness and stability of the model training process. In each stage, the training objective of the generator is to minimize the loss, while the training objective of the discriminator is to maximize the loss. The loss function is defined as follows:

[0139]

[0140] where γ1, γ2, γ3 are hyperparameters that balance the importance of the two losses, is the generated sample, represents G a 、G t 、Gs Generator distribution

[0141] In addition, in this embodiment, a computer system is disclosed, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the above method are implemented.

[0142] To better illustrate the embodiments of the present invention, the following experiments are conducted in this embodiment:

[0143] The embodiments of the present invention are based on two mobile cellular network datasets of a large network operator, covering 9,933 mobile users and 28,524 cellular base stations. The datasets contain two different time intervals: from December 1, 2020 to December 31, 2020 (referred to as "Dataset-201212"), and from February 1, 2021 to February 28, 2021 (referred to as "Dataset-202102").

[0144] The embodiments of the present invention run on a machine equipped with 2 NVIDIA Tian V GPUs, each GPU equipped with 12GB of memory and 60GB of CPU memory. The experiments use the Python3.8 and PyTorch1.10 frameworks. To ensure training efficiency, the batch size is set to 25, the maximum number of training epochs is set to 200, and the spatio-temporal sequence length is set to 24. The initial learning rates of AT-GAN, CT-GAN, and CS-GAN are set to 3e-6, 3e-6, and 8e-7 respectively. These hyperparameters are determined through multiple experiments to achieve the best performance of their respective models. The specific experimental process is as follows:

[0145] I. Performance Metrics

[0146] The present invention uses different evaluation metrics to evaluate data fidelity, usability, and privacy, as follows.

[0147] · In terms of the evaluation of data fidelity, to evaluate the fidelity of the synthetic data, the Earth Mover's Distance (EMD) (also known as the Wasserstein-1 distance) is used to evaluate the distribution of the sequence synthetic data of the uplink traffic, downlink traffic, and base station ID with the real data, and the Jensen-Shannon divergence (JSD) is used for the distribution of the synthetic data of the user ID (uid), age, gender, and App category with the real data. Since the EMD of different data fields has different scales, the EMD of each field is normalized to [0.1, 0.9]. Specifically, given two probability distributions P and Q of a set X, the EMD is calculated as:

[0148] EMD(P,Q) = inf∑ x,yγ(x,y)·d(x,y)

[0149] where γ is a parameter representing the mass transferred from each element x in P to each element y in Q, and d(x,y) is the distance between elements x and y in the metric space. The JSD is calculated as follows:

[0150]

[0151] where x represents an element in set X, and P(x) and Q(x) are the probabilities of x in distributions P and Q respectively.

[0152] · In terms of the evaluation of data utility, the utility of data is evaluated through two typical downstream applications, namely, App usage traffic prediction and the prediction of the next connected base station. For App traffic usage prediction, two commonly used metrics are used: Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), which are calculated as follows:

[0153]

[0154]

[0155] where y i and are the true value and the predicted value of the i-th prediction respectively, and n is the total number of predictions. For the prediction evaluation of the next connected base station, Acc@K is used, which measures the ratio of the number of correctly predicted samples to the total number of predicted samples, and is calculated as follows:

[0156]

[0157] where N is the total number of users. If all the connected base stations appear in the set of predicted base stations in Top-K, Acc@K = 1; otherwise, Acc@K = 0. In the embodiments of the present invention, K = 3 is selected to calculate the prediction result.

[0158] II. Comparative Experiments

[0159] The present invention uses CTGAN, STAN, DoppelGANger, TabDDPM, NetShare to conduct a comparative evaluation of data fidelity and data utility with the technical solutions of the embodiments of the present invention, and the results are as follows:

[0160] (1)Regarding data fidelity, first compare the data fidelity of the CellSyn method with other baseline models. The average value is calculated for every 500 users out of 9933 users for each sample. Figures 3(a), 3(b), and 3(c) respectively depict the CDF plots of the EMD values of the uplink traffic, downlink traffic, and base station IDs under different models. The main observations are as follows: First, the performance of CellSyn is better than all baseline models. For example, at the 80th percentile, the EMD values of CTGAN, STAN, DGAN, TabDDPM, and NetShare for the uplink traffic are approximately 0.45, 0.46, 0.82, 0.82, and 0.33 respectively. In contrast, the EMD value of the CellSyn scheme is 0.28, and the EMD value of CellSyn for the downlink traffic is only 0.18. Through analysis, the performance advantage of CellSyn is attributed to its comprehensive consideration of the dynamics and multi-scale time patterns in the time series, which other baseline models do not consider carefully. Second, for the base station ID, the EMD value achieved by CellSyn is only 0.47, far better than other baseline models, which highlights CellSyn's ability to effectively model the spatial base station association pattern.

[0161] Figures 3(d) and 3(e) are the CDFs of the JSD values obtained by different models for gender and age respectively. From these figures, the superior performance of the proposed CellSyn scheme can be observed. CellSyn still maintains an overall advantage over other baseline models in terms of the ability to accurately capture the distribution of these categorical features.

[0162] Figures 4(a), 4(b), and 4(c) show the average performance and error bars of the two metrics on Dataset-202012. It can be observed that CellSyn is better than several baseline models in all metrics. In addition, when observing the results of the error bars, it is found that CellSyn shows a lower bias among all baseline models, indicating the robustness of its performance. Although the average EMD of the base station ID obtained by NetShare is slightly lower than that of CellSyn, it is observed that the generated base station ID sequence is not user-specific, which may have a negative impact on the data utility of downstream applications (e.g., the prediction of the next connected base station).

[0163] (2) Regarding data utility, the data utility of CellSyn is evaluated through two typical downstream applications: App traffic prediction and prediction of the next connected base station. The synthetic data (denoted as B) generated by the CellSyn method proposed in the present invention based on real data (denoted as A), both the real data and the synthetic data are sorted in chronological order based on timestamps. Then the data is divided into a 70% training set and a 30% test set. The training set consists of early data, while the test set consists of later data. The prediction model is trained on the training set, and its performance is evaluated on the test set. Finally, the performance metrics of two scenarios are compared: the real data training set (A) and the real data test set (A′) with the synthetic data training set (B) and the real data test set (A′), so as to be able to evaluate the generalization ability of training the model on synthetic data.

[0164] In the experiment of App traffic prediction, an important application of mobile cellular network data is to design an App traffic prediction algorithm. In this application, several specific fields in the dataset are focused on, including user ID, uplink traffic, downlink traffic, and timestamp. The goal is to predict the App usage traffic of each user at the next time step, which is a key test for the availability of synthetic App usage traffic. To evaluate the performance, three commonly used time series prediction models are adopted for evaluation: LSTM, ConvLSTM, and Transformer.

[0165] The results show that the real data reaches the lowest MAE and RMSE values, that is, the best prediction effect. At the same time, it is also observed that CellSyn is better than all other baseline models in all prediction metrics. For example, when using the LSTM model to predict uplink traffic on Dataset - 202012, the MAE values of the synthetic data from CTGAN, DGAN, STAN, TabDDPM, and NetShare are 615.32, 1009.92, 4547.32, 91.76, and 683.6 respectively, while the MAE of the synthetic data from CellSyn is only 75.21. Similar results can also be observed in downlink traffic prediction. For example, when using the ConvLSTM model on Dataset - 202102, the RMSE values of the synthetic data from CTGAN, DGAN, STAN, TabDDPM, and NetShare are 61.5, 272.1, 1737.1, 15.3, and 48.60 respectively, while the RMSE of the CellSyn synthetic data is closest to the real data, which is 10.8. The above observation results indicate that CellSyn can effectively learn the time patterns of App usage traffic.

[0166] In the experiment of predicting the next connected base station, the prediction of the next connected base station is another key application that relies on mobile cellular network data. The main goal of this task is to predict the next base station that a user may connect to based on past movement patterns and historical data. To evaluate the performance of synthetic data, four prediction models are adopted: LSTM, CNN, Attention, and GRU+Attention.

[0167] Figure 5 The Acc@3 of each prediction algorithm on the Dataset-202012 dataset is presented. It should be noted that since the prediction results of the synthetic data generated by CTGAN and DGAN are less than 0.02, their prediction results are not shown here. It can be observed that the real data achieves the best metric value. However, it is found that the synthetic data generated by CellSyn is superior to all other baseline models in all prediction metrics. For example, when using the Attention network on Dataset-202012, the Acc@3 scores of STAN, TabDDPM, and NetShare are 0.2, 0.21, and 0.3 respectively, while CellSyn reaches 0.62, which is very close to the real data. This benefits from the customized design of the CellSyn in the spatial base station sequence generation module, which can learn the correlation of the user base station sequence.

[0168] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for synthesizing user mobile cellular network data based on a deep generative adversarial network, characterized in that It includes the following steps: Construct a user multi-dimensional attribute synthesis module, which is used to learn the attribute characteristics of real users with different attributes and synthesize user attributes according to the attribute characteristics of real users; the user multi-dimensional attribute synthesis module is an attribute generative adversarial network, including an attribute generator and an attribute discriminator; Construct a user usage traffic sequence synthesis module, which is used to learn the traffic time-series characteristics of real traffic sequences of real users with different attributes and synthesize the traffic time series of users with different attributes according to the traffic time-series characteristics; The input of the user usage traffic sequence synthesis module includes the output of the user multi-dimensional attribute synthesis module; The user usage traffic sequence synthesis module is a conditional time-series generative adversarial network, including: a time-series generator and a time-series discriminator; Construct a user spatial base station connection sequence synthesis module, which is used to learn the spatial connection characteristics of real spatial base station connection sequences of real users with different attributes and synthesize the spatial base station connection sequences of users with different attributes according to the spatial connection characteristics; the input of the user spatial base station connection sequence synthesis module includes the output of the user multi-dimensional attribute synthesis module; the user spatial base station connection sequence synthesis module is a conditional spatial generative adversarial network, including: a spatial generator and a spatial discriminator; Jointly train the user multi-dimensional attribute synthesis module, the user usage traffic sequence synthesis module, and the user spatial base station connection sequence synthesis module, and use the trained user multi-dimensional attribute synthesis module, the user usage traffic sequence synthesis module, and the user spatial base station connection sequence synthesis module to synthesize data; The loss function of the joint training is as follows: Among them, G a is an attribute generator, D a is an attribute discriminator, L(G a , D a ) is the loss between the attribute generator and the attribute discriminator, G t is a temporal generator, D t is a temporal discriminator, L(G t , D t ) is the loss between the temporal generator and the temporal discriminator, G s is a spatial generator, D s is a spatial discriminator, L(G s , D s ) is the loss between the spatial generator and the spatial discriminator, γ1, γ2, γ3 are hyperparameters for balancing the importance of losses, is the generated sample, represents the distribution of the G a , G t , G s generators, E represents the expectation function, represents the gradient descent function.

2. The method for synthesizing user mobile cellular network data based on a deep generative adversarial network according to claim 1, characterized in that, The attribute generator includes an attribute encoder and an attribute decoder. The attribute encoder generates an encoded feature h based on an input attribute noise vector e , and the attribute decoder is used to decode the encoded feature h e to obtain a synthesized attribute vector A generated ; The attribute discriminator is used to take the synthetic attribute vector A generated and compare it with the real attribute vector A real to determine the authenticity of the synthetic attribute vector A generated , and adjust the network parameters of the attribute generator according to the judgment result until the attribute discriminator determines that the synthetic attribute vector A generated is true.

3. The method for synthesizing user mobile cellular network data based on a deep generative adversarial network according to claim 2, wherein The user multi-dimensional attribute synthesis module encodes the input real user attributes using one-hot encoding to obtain a real encoding vector, and constructs a real attribute vector A based on the real encoding vectors of multiple different users real ; The attribute encoder includes a first Linear layer, a Leaky ReLU layer, and a BatchNorm layer connected in sequence, the attribute decoder includes a second Linear layer and a Softmax layer connected in sequence, and the output end of the BatchNorm layer is connected to the input end of the second Linear layer; The attribute discriminator includes a third Linear layer and a second Leaky ReLU layer connected in sequence; the input of the third Linear layer is the synthetic attribute vector A generated concatenated with the real attribute vector A real vector.

4. The method for synthesizing user mobile cellular network data based on a deep generative adversarial network according to any one of claims 1-3, characterized in that, The timing generator is used to synthesize the input synthetic attribute vector A generated and the time series vector T of the usage traffic of the corresponding user generated ; The timing discriminator is used to compare the time series vector T generated with the real usage traffic sequence T real and determine the authenticity of the time series vector T according to the comparison result generated , and adjust the network parameters of the timing generator according to the judgment result until the timing discriminator determines that the synthesized time series vector T generated is true.

5. The method for synthesizing user mobile cellular network data based on a deep generative adversarial network according to claim 4, wherein The input of the timing generator includes a timing noise vector and a synthetic attribute vector A generated , and the timing generator includes: A splicing layer for splicing the temporal noise vector and the synthetic attribute vector A generated together; TCN block for capturing the synthetic attribute vector A based on the timing noise vector generated Use the short-term timing features of the traffic; Transformer encoding block, synthesizing a time series vector T based on the short-term temporal features generated ; The timing discriminator includes a fourth Linear layer and a third Leaky ReLU layer connected in sequence; the input of the fourth Linear layer is the time series vector T generated and the concatenated vector with the true usage traffic sequence T real .

6. The method for synthesizing user mobile cellular network data based on a deep generative adversarial network according to claim 5, wherein The TCN block satisfies: h TCN = Leaky ReLU(Dropout(TCN([T nosie ,A generated ))) where h TCN is the short-term temporal feature; TCN is the temporal convolutional network, Dropout is the overfitting function, and Leaky ReLU is the rectified linear activation function; The Transformer encoding block satisfies: T generated = Norm(TransEncoder(h TCN )) Wherein, TransEncoder is an autoencoder, and Norm is a normalization function; The actual usage traffic sequence T real is obtained in the following manner: Obtain the usage traffic sequences of multiple real users, perform normalization processing on the usage traffic sequences, and convert the multiple normalized usage traffic sequences into a three-dimensional matrix, and the dimensions of the three-dimensional matrix include user attributes, traffic time steps, and traffic types.

7. The method for synthesizing user mobile cellular network data based on a deep generative adversarial network according to any one of claims 1 to 3, characterized in that The spatial generator is used to synthesize the user spatial base station connection sequence according to the input spatial noise vector and the user attribute vector; the spatial discriminator is used to identify the authenticity of the synthesized user spatial base station connection sequence, and adjust the network parameters of the spatial generator according to the identification result until the spatial discriminator identifies the synthesized user spatial base station connection sequence as true.

8. The method for synthesizing user mobile cellular network data based on a deep generative adversarial network according to claim 7, characterized in that, The space generator includes a plurality of sub-generators, the space discriminator includes a plurality of sub-discriminators, and the plurality of sub-generators, the plurality of sub-discriminators, and the plurality of users correspond to each other one by one. Each sub-generator is used to capture the correlation feature of the spatial base station connection sequence of its corresponding user, and synthesize the user spatial base station connection sequence of the corresponding user according to the correlation feature of the spatial base station connection sequence. Each sub-generator is used to identify whether the synthesized user spatial base station connection sequence of its corresponding sub-generator is true, and adjust the network parameters of the corresponding sub-generator according to the identification result until the corresponding sub-generator identifies that the synthesized user spatial base station connection sequence is true.

9. The method for synthesizing user mobile cellular network data based on a deep generative adversarial network according to claim 8, wherein Each sub-generator includes an LSTM network, a Linear network, and a Softmax network connected in sequence; The LSTM network is used to capture the correlation feature of the spatial base station connection sequence of its corresponding user; The Linear network is used to calculate the probability matrix of the user spatial base station connection sequence according to the correlation feature of the spatial base station connection sequence; The probability matrix satisfies: Among them, P i represents the probability matrix of the i-th user u i . And are the weight matrix and bias vector of the Linear layer of the i-th user u i respectively; represents the eigenvector of the spatial base station connection sequence correlation of the i-th user u i ; represents the spatial noise vector of the i-th user u i , and A generated represents the composite attribute vector of the i-th user u i . Each sub-discriminator includes a fifth Linear layer and a fourth Leaky ReLU layer connected in sequence; the input of the fifth Linear layer is the concatenated vector of the synthesized user spatial base station connection sequence and the real user spatial base station connection sequence.

10. A computer system, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of any one of the above claims 1 to 9.

Images