A traffic bill splitting method

Abstract

The application discloses a traffic bill cutting method and belongs to the technical field of operator communication. In view of the problems of the change of bill processing demand of traffic generated by users in the current communication operation system and the large increment of bills, the application first introduces various features of the user, outputs the key label of the user through an attention mechanism algorithm, secondly calculates the online category traffic weight value of the user as the traffic feature of the user, thirdly introduces the host performance as the host parameter information, and finally takes the various features of the user as the output of the recurrent neural network model, dynamically calculates the traffic step of the user at a certain moment, that is, the traffic authorization value. Finally, the classification of various feature users is realized, and the corresponding traffic step is granted, the growth of the monthly bill amount is controlled from 300% to 40% increase, and the system operation risk is effectively reduced. On the one hand, the accurate cutting of the bill can be realized, the bill amount is reduced, and the host storage expansion frequency is reduced, and on the other hand, the user use perception is improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of operator communication, and particularly relates to a traffic bill splitting method. BACKGROUND

[0002] In an operator system, a user uses a 5G network to surf the Internet by applying for traffic authorization to a business support system through a network core device, the core device controls through the feedback authorization traffic, and the support system realizes the use metering of the Internet traffic through the splitting of the authorized amount. The existing bill splitting method splits by fixed traffic or fixed time length. With the increase in the number of users and the increase in the Internet surfing time of users, the monthly bill amount increases by 300% through the fixed splitting of traffic bills, and the server resources need to be regularly expanded to meet the demand for bill processing, increasing the operating cost and operating risk. SUMMARY

[0003] In view of the problems of the change in the bill processing demand of the user traffic generated in the current communication operation system and the large bill increment, the present application provides a traffic bill splitting method.

[0004] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0005] K-means clustering is a clustering algorithm based on sample set division. K-means clustering divides the sample set into k subsets to form k classes, and divides n samples into k classes, each sample has the minimum distance to the center of its belonging class, and each sample belongs to only one class, which is k-means clustering. At the same time, according to the fact that a sample belongs to only one class, it also indicates that k-means clustering is a hard clustering algorithm.

[0006] A traffic bill splitting method, comprising the following steps:

[0007] Step 1, calculating the business category (RG) weight value of the user; RG represents the business category of the user surfing the Internet;

[0008] Step 1.1, first, the traffic of each business category of the user is taken as a sample for clustering, and each business category of the user is labeled;

[0009] Step 1.2, selecting the traffic weight value of the user;

[0010] First, the coefficient of variation v i of each business traffic flow i after labeling is calculated, which is expressed as: Where v i represents the coefficient of variation, σ i represents the standard deviation, represents the mean, and i represents the number of users;

[0011] Then, the total traffic of each service category of the user and the weight w are calculated i , is expressed as: Wherein, w i represents the traffic weight of the user, represents the sum of the coefficients of variation of each flow i .

[0012] The total weight of the traffic of the user service category is composed of the set W = {W1, W2, W3}, which is the input of the user traffic dynamic step judgment.

[0013] Step 2, extract the key information of the user based on the attention mechanism;

[0014] Step 2.1, extract the user information and simplify and number it to obtain the user information label T, and the user information label T is composed of the set J, expressed as J = {T1, T2,...,T n};

[0015] Step 2.2, according to the user information label, the joint structure feature fusion is carried out, and the regional traffic feature, the time period traffic feature and the fused habit identification traffic feature are obtained;

[0016] Step 2.3, the weight distribution calculation of the attention mechanism is carried out based on the encoder-decoder model of the attention mechanism, and the attention distribution probability distribution value a ij of each feature in the input sequence to the current output signal is obtained, expressed as: a ij = F(h i , H j ), wherein h j represents the hidden state of the input data X j in the encoder, H j represents the hidden state of the output data Y j in the decoder, and F represents the probability value that H j and h j are consistent with the user traffic usage habit;

[0017] The attention probability distribution value of each element in the input information to the output signal is calculated in the decoder module, that is, the corresponding intermediate semantic encoding is calculated for each output, so that different input elements are assigned different weight values, highlighting the data with large weight values and weakening the data with small weight values, thereby improving the classification effect.

[0018] For each different output element y j of the output data Y j , there is a different intermediate semantic encoding E j , expressed as wherein S represents the regional traffic feature, representing the time-division flow characteristics, C ij representing the fusion habit identification flow characteristics, T represents the total number of input data;

[0019] The decoder output feature is: X t = D(C j , x1, x2, x3,..., x j-1 ), wherein D represents the user flow usage feature, x1, x2, x3,..., x j-1 is the output element;

[0020] The joint user's service category clustering classification flow total weight set forms the user flow usage feature:

[0021] P = {W, x t} = (E j , x1, x2, x3,..., x j-1 );

[0022] Step 3, flow prediction based on recurrent neural network (GRU);

[0023] Using a recurrent neural network model, for an input sequence {x1, x2, x3,..., x n} of length n, the feature information in the sequence is learned by mapping the input sequence to the hidden layer state, and then higher level information is obtained by stacking multiple hidden layer feature information;

[0024] The decoder output feature X t is used as the input feature of the recurrent neural network model, and the forward propagation of the recurrent neural network model is represented as:

[0025] Z t = σ(W z *X t +U z *h t-1 +b z )

[0026] R t = σ(W r *X t +U r *h t-1 +b r )

[0027]

[0028]

[0029] Wherein, R t represents the reset gate at time t; Z t represents the update gate at time t; denotes the candidate hidden state at time t; h t denotes the current cell state at time t; h t-1 denotes the cell state at the previous time; denotes the Hadamard product; denotes the Sigmoid activation function, which converts data to a value of [0, 1]; W r denotes the weight matrix of the reset gate; W z denotes the weight matrix of the update gate; W h denotes the weight matrix of the candidate hidden state; W* and U* are convolution kernels; b* is a bias term, which is a parameter of the model in the training phase;

[0030] In order to generate a user traffic usage value corresponding to the current input feature at the output layer, a network including two convolution layers is established, and then a Softmax function is used to calculate the weight value,

[0031] S t = U s *tanh(W xs *X t +W hs *h t-1 +b s )+b us

[0032] X t is the feature value of the user at time t of the day, h t-1 is the cell state at the previous time, U s , W xs , and W hs all denote convolution kernels, and the size is 1x1;

[0033] S t obtains the traffic weight value corresponding to the current input X t , which is denoted as: wherein, denotes the attention weight of the input feature, and the size is between 0 and 1, and the sum of the attention weights of all input features = 1;

[0034] According to the feature vector at each time, the obtained traffic weight coefficient is multiplied by the input feature to obtain the dynamic traffic authorization value of the user at the current time, which is denoted as:

[0035] X t ' = A t ⊙x t .

[0036] For the dynamic traffic authorization value, the operator system generates a bill according to the usage of the traffic authorization value. When a dynamic traffic authorization value is used up, the bill is generated, and the splitting of the fixed value becomes the splitting of the dynamic value, reducing the number of bills.

[0037] The attention mechanism is applied to feature extraction in the present application, and the extracted user feature information is fused with joint structure features to extract more significant important features of the user. The attention on ambiguous classification features is reduced, and the calculation efficiency of the entire traffic prediction model is improved.

[0038] Further, the service categories of the user are divided into three categories: video category: flow1={RG 11 , RG 12 ,..., RG 1n}; webpage category: flow2={RG 21 , RG 22 ,..., RG 2n}; chat category: flow3={RG 31 , RG 32 ,..., RG 3n}.

[0039] Further, the user information labels in step 2.1 include: 1-device type, 2-gender, 3-age, 4-customer type, 5-online time, 6-professional type, 7-whether to exit the account, 8-whether to be extremely low usage, 9-user average monthly online traffic in the past three months, 10-whether to be fixed migration fusion package, 11-social influencer, 12-influencer, 13-current credit, 14-traffic usage, 15-value score, 16-5G traffic, 17-payment method preference, 18-stability, 19-whether high traffic user, 20-VIP level, 21-stability score, 22-total cumulative usage of the month, 23-cpu, 24-base station id.

[0040] Further, the regional traffic feature set is represented as S={U6, U9, U 11 , U 12 , U 16 , U 17 , U 19 , U 24}.

[0041] Further, the time period traffic feature is divided into three periods according to the time feature θ: θ1: 8 o'clock-12 o'clock, 14 o'clock-18 o'clock; θ2: 12 o'clock-14 o'clock, 18 o'clock-24 o'clock; θ3: 0 o'clock-8 o'clock; the feature attribute of the user's traffic usage in different time is represented as wherein l represents different time in a day, the current time of the user is calculated and the 24-hour traffic of the user in the previous day is calculated The traffic usage similarity value is measured; the time weight parameter of θ1 is 1.5, the time weight parameter of θ2 is 1.7, and the time weight parameter of θ3 is 0.8.

[0042] Further, the fusion habit identification traffic feature is that the specific traffic behavior of the network user is different in different usage habits; the network page touch point click rate, the occupation type, and the network type of the user are distinguished.

[0043] The click amount C of the user to different network page touch points ij is represented as:

[0044]

[0045] Wherein, i represents the number of users, and j represents the network page touch point. i→j click represents the click view operation of the user i to the network page touch point j. i→j lable represents the collection operation of the user i to the network page touch point j.

[0046] Further, the encoder module and the decoder module are included in the attention mechanism-based encoder-decoder model in step 2.3.

[0047] Further, the weight distribution calculation of the attention mechanism in step 2.3 is a joint feature hidden layer state and input signal hidden state semantics comparison according to the regional traffic feature, the time period traffic feature, and the fusion habit identification traffic feature.

[0048] Further, the recurrent neural network model includes two gate units, which are update gate Z t and reset gate R t ; wherein the update gate is information that determines the memory information of the hidden state to be saved to the current time; the reset gate controls whether the new input information is combined with the memory information of the hidden state.

[0049] Compared with the prior art, the present application has the following advantages:

[0050] The present application first introduces multiple features of the user, outputs the key label of the user through the attention mechanism algorithm, calculates the online category traffic weight value of the user as the traffic feature of the user, introduces the host performance as the host parameter information, and finally takes the multiple features of the user as the output of the recurrent neural network model, dynamically calculates the traffic step length of the user at a certain moment, and the traffic authorization value. Through this method, the accurate division of the bill can be realized, the user experience can be improved, the bill amount can be reduced, and the host storage expansion frequency can be reduced.

[0051] The main idea behind this method is that when a user has sufficient data balance, the data allowance step size in the data sliding window is larger, thereby improving the real-time dynamic estimation of data allowance values. Traditional data call detail record (CDR) authorization values ​​are generated in a single way, based on the user's usage scenario. Such methods only consider user behavior information and ignore the impact of user behavior habits on CDR authorization values. This invention's method starts from multiple dimensions such as user social information and data usage habits, utilizing behavioral characteristics such as user gender, age, RG value, region, total monthly usage, data plan, whether the user is a high-data-traffic user, and recent monthly data usage trends. Ultimately, it achieves user classification based on various characteristics and assigns corresponding data allowance steps, controlling the monthly CDR increase from 300% to 40%, effectively reducing system operational risks. On the one hand, it enables precise CDR segmentation, reducing the number of CDRs and the frequency of host storage expansion; on the other hand, it improves the user experience. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the encoding / decoding model structure of the attention mechanism;

[0053] Figure 2 This is a schematic diagram of the GRU structure;

[0054] Figure 3 Dynamic traffic authorization process diagram;

[0055] Figure 4 A diagram illustrating the traffic of the three main business categories. Detailed Implementation

[0056] Example 1

[0057] The traffic call detail record (CDR) segmentation method of this invention is a dynamic traffic authorization optimization method based on an attention-mechanism recurrent neural network, such as... Figure 3 The dynamic traffic authorization process is illustrated in the diagram below;

[0058] Step 1: Calculate the user's service category (RG) weight value;

[0059] Step 1.1: First, cluster the traffic of each user's business category as a sample and label each user's business category; take the user's internet traffic data, classify it according to the business category using a clustering algorithm, and then define the business in each category into three categories: video, web, and chat.

[0060] Video type: flow1 = {RG 11 RG 12 , ..., RG 1n};

[0061] Web page type: flow2 = {RG 21 RG22 ..., RG 2n};

[0062] Chat: flow3 = {RG 31 , RG 32 ..., RG 3n}.

[0063] The video type corresponds to the service such as iQiyi video, Tencent video, etc., the webpage type corresponds to the service such as today's headlines, etc., and the chat type corresponds to the service such as mobile Baidu, WeChat, etc., as shown in Figure 4 .

[0064] Through the clustering algorithm, the service is clustered into three clusters, each cluster corresponds to a different traffic area, the large flow part is the video type, the medium flow part is the webpage type, and the small flow part is the chat type.

[0065] Step 1.2, selection of user traffic weight value;

[0066] First, calculate the coefficient of variation v i of each service flow i after marking, which is represented as: where v i represents the coefficient of variation, σ i represents the standard deviation, represents the mean, and i represents the number of users;

[0067] Then, calculate the total flow and weight w i of each service category of the user, which is represented as: where w i represents the user flow weight, represents the sum of the coefficients of variation of each flow i ;

[0068] Through the calculation of the total weight of the flow of the user service category, a set W = {0.8, 0.4, 0.3}

[0069] Step 2, extract the key information of the user based on the attention mechanism;

[0070] Step 2.1, 24 key label information is extracted from user portrait information through a big data platform, and the key features are classified. The user information labels include: 1-device type, 2-gender, 3-age, 4-customer type, 5-online time, 6-professional type, 7-whether to exit the account, 8-whether to be extremely low usage, 9-user's average online traffic in the past three months, 10-whether to be fixed migration fusion package, 11-social media influencer, 12-net red influencer, 13-current credit, 14-traffic usage, 15-value score, 16-5G traffic, 17-payment method preference, 18-stability, 19-whether high traffic tendency user, 20-VIP level, 21-stability score, 22-total cumulative usage total traffic this month, 23-cpu, 24-base station id.

[0071] User information is extracted, simplified and numbered to obtain user information label T. The user information label T is grouped into set J, which is represented as J = {T1, T2,..., T n};

[0072] Step 2.2, according to the user information label, the joint structure feature fusion is carried out to obtain: regional traffic feature, time period traffic feature, and fusion habit identification traffic feature.

[0073] Further, the regional traffic feature set is represented as: S = {U6, U9, U 11 ,U 12 ,U 16 ,U 17 ,U 19 ,U 24} business circle, school, village, building, airport, station, suburb, scenic spot.

[0074] The time period traffic feature is divided into three periods according to the time feature θ: θ1: 8 o'clock-12 o'clock, 14 o'clock-18 o'clock; θ2: 12 o'clock-14 o'clock, 18 o'clock-24 o'clock; θ3: 0 o'clock-8 o'clock; the feature attribute of user's traffic usage in different time is represented as wherein, l represents different time in a day, the traffic usage of the user in the current time is calculated and the similarity value of the user's traffic usage in the previous 24 hours is measured.

[0075] Table 1 Time feature and busy time corresponding relationship

[0076]

[0077] The time weight parameter of θ1 is 1.5, the time weight parameter of θ2 is 1.7, and the time weight parameter of θ3 is 0.8.

[0078] The fusion habit identification traffic feature: video class, student; webpage class, office worker; chat class, middle-aged. Commercial district, school, airport, station 5G network, rural area, building, suburb, scenic spot 4 / 5G network. The specific traffic behavior of the network user is different under different use habits; according to the user's access webpage touch point click rate, occupation type, network type, the user is distinguished;

[0079] The user's click volume C of different webpage touch points ij is represented as:

[0080]

[0081] Wherein, i represents the number of users, j represents the webpage touch point; J i→j click represents the click view operation of user i to webpage touch point j, J i→j lable represents the collection operation of user i to webpage touch point j;

[0082] Step 2.3, the encoder-decoder model based on attention mechanism includes an encoder module and a decoder module; the structure of the encoder-decoder model based on attention mechanism is as shown in Figure 1 The weight distribution calculation of the attention mechanism based on the encoder-decoder model is the hidden layer state of the joint feature composed of the regional traffic feature, the time period traffic feature and the fusion habit identification traffic feature and the hidden state semantics of the input signal are compared with the user traffic use habit. Get the attention distribution probability distribution value a ij of each feature in the input sequence to the current output signal, represented as: a ij =F(h i ,H j ), wherein h j represents the hidden state of the input data X j in the encoder, H j represents the hidden state of the output data Y j in the decoder, and F represents the probability value of h j and h j complying with the user traffic use habit;

[0083] The attention probability distribution value of each element in the input information to the output signal is calculated in the decoder module, that is, the corresponding intermediate semantic code is calculated for each output, so that different input elements are assigned different weight values, highlighting the data with large weight values and weakening the data with small weight values, thereby improving the classification effect.

[0084] Each different output element y j corresponds to a different intermediate semantic code C j, is expressed as wherein S represents regional traffic characteristics, denotes the time period traffic characteristics, C ij denotes the fusion habit identification traffic characteristics, and T represents the total number of input data;

[0085] The decoder output feature is:

[0086] X t = D(C j , x1, x2, x3,..., x j-1 ), wherein D represents the user traffic usage characteristics;

[0087] The feature after clustering and classifying the user's service category log is F = {W, x t} = (C j , x1, x2, x3,..., x j-1 ), according to the user's usage of regional information, the user's usage of traffic is divided into 8 categories, and the corresponding traffic usage increases from low to high, as shown in Table 2:

[0088] Table 2

[0089]

[0090] Step 3, traffic prediction based on recurrent neural network (GRU);

[0091] The recurrent neural network model includes two gate units, namely update gate Z t and reset gate R t ; wherein the update gate is the information that decides to save the memory information of the hidden state to the current time; the reset gate controls whether to combine the new input information with the memory information of the hidden state. The structure of the recurrent neural network (GRU) is shown in Figure 2 .

[0092] Using the recurrent neural network model, for an input sequence {x1, x2, x3,..., x n} of length n, the feature information in the sequence is learned by mapping the input sequence to the hidden layer state, and then higher level information is obtained by stacking multiple hidden layer feature information;

[0093] The decoder output feature X t is used as the input feature of the recurrent neural network model, and the forward propagation of the recurrent neural network model is represented as:

[0094] Z t = σ(W z *X t +U z *h t-1 +b z)

[0095] R t = σ(W r *X t +U r *h t-1 +b r )

[0096]

[0097]

[0098] wherein R t represents the reset gate at time t; Z t represents the update gate at time t; represents the candidate hidden state at time t; h t represents the current cell state at time t; h t-1 represents the cell state at the previous time; represents Hadamard product; σ represents Sigmoid activation function, which converts data to [0, 1] values; W r represents the weight matrix of the reset gate; W z represents the weight matrix of the update gate; W h represents the weight matrix of the candidate hidden state; W* and U* are convolution operation convolution kernels; b* is a bias term, which is a parameter of the model in the training stage;

[0099] In order to generate the user traffic usage value corresponding to the current input feature in the output layer, a network comprising two convolutional layers is established, and then a Softmax function is used to calculate the weight value,

[0100] S t = U s *tanh(W xs *X t +W hs *h t-1 +b s )+b us

[0101] X t is the feature value of the user at the t time of the day, h t-1 is the cell state at the previous time, U s , W xs , and W hs all represent convolution kernels, and the size is 1x1;

[0102] S t obtains the traffic weight value corresponding to the current input X t , which is represented as: wherein, Attention weight representing input features, size between 0~1, the sum of all input features attention weight = 1;

[0103] According to the feature vector at each time, the obtained weight coefficient is multiplied by the input feature to obtain the dynamic traffic authorization value of the user at the current time; The formula is as follows:

[0104] X t '=A t ⊙x t .

[0105] Through the neural network, the traffic prediction of each period is carried out, and the prediction result is the corresponding traffic piece and time piece of the user under each service category as shown in the following table:

[0106] Table 3

[0107]

[0108] According to the prediction result, different dynamic traffic values are issued according to the user in different time periods. If the user uses the video service in the morning at 9 o'clock in the business district, it is judged that the traffic balance is sufficient, and the 5G traffic step is directly allocated.

[0109] For the dynamic traffic authorization value, the operator system generates a bill according to the use of the traffic authorization value. When a dynamic traffic authorization value is used up, a bill is generated, and the splitting of the fixed value becomes the splitting of the dynamic value, reducing the number of bills.

[0110] Example 2 cost analysis

[0111] Table 4 resource input

[0112] Way Investment cycle Number of resources No dynamic split Per year 10 servers Dynamic split Three years 5 servers

[0113] Table 4 shows the consumption of system resources before and after dynamic splitting. When dynamic splitting is not performed, because the number of traffic bills is multiplied, bill storage resource expansion is needed every year, and each expansion is 10 servers, with a price of about 800,000 yuan. After using dynamic splitting, because the growth is only 30%, only one expansion is needed every three years, and 5 servers are needed each time, with a cost of only 400,000 yuan. The overall operating cost is greatly reduced. (Such as operating expenses) and many other aspects to reflect the advantages of the method of the present application.

[0114] The details of the application not described herein are considered known to those skilled in the art. Although the foregoing specific embodiments of the application have been described in some detail for the purposes of clarity of understanding, it will be apparent that certain modifications can be practiced within the spirit and scope of the application, as defined by the appended claims.

Claims

1. A method for traffic call detail record (CDR) segmentation, characterized in that: The method comprises the following steps: Step 1, calculating the service category weight value of the user; Step 1.1, first clustering the traffic under each service category of the user as a sample, and marking each service category of the user; Step 1.2, selecting the traffic weight value of the user; First, the coefficient of variation v i of each business flow flow i after marking is calculated, expressed as: where v i represents the coefficient of variation, σ i represents the standard deviation, represents the mean, and i represents the number of users; Then, the total traffic and weight w of each service category of the user are calculated i , is expressed as: wherein w i represents the traffic weight of the user, represents the sum of the coefficients of variation of each flow i ; The total weight of the traffic under the service category of the user is grouped into a set W = {W1, W2, W3} as the input of the user traffic dynamic step judgment; Step 2, extracting the key information of the user based on the attention mechanism; Step 2.1: Extract user information, simplify and number it to obtain user information tags T. Assemble user information tags T into a set J, represented as J = {T1, T2, ..., T}. n }; Step 2.2, according to the user information label, the joint structure feature fusion is carried out, and the regional traffic feature, the time period traffic feature and the fusion habit identification traffic feature are obtained; Step 2.3, the attention mechanism weight distribution calculation based on the attention mechanism based encoder-decoder model, to obtain the attention distribution probability distribution value a of each feature in the input sequence to the current output signal ij , is expressed as: a ij = F(h i , H j ), wherein h j represents the hidden state of the input data X j in the encoder, H j represents the hidden state of the output data Y j in the decoder, and F represents the calculation of H j and h j comply with the probability value of the user traffic usage habit; For output data Y j Each different output element y j Corresponds to a different intermediate semantic encoding E j , denoted as Where S represents the regional traffic characteristics, Denotes the traffic characteristics by time period, C ij Denotes the fusion habit identification traffic characteristics, and T represents the total number of input data; The decoder outputs features: X t = D(C j , x1, x2, x3,..., x j-1 ), where D denotes user traffic usage features, x1, x2, x3,..., x j-1 are output elements; The total weight of the traffic after the clustering and classification of the service category of the user is grouped into a set, and the user traffic usage feature is formed as: P = {W, x t} = (E j , x1, x2, x3,..., x j-1 ); Step 3, traffic prediction based on recurrent neural network; Using a recurrent neural network model, for an input sequence of length n {x1, x2, x3, ..., x...} n The system learns features from the sequence by mapping the input sequence to the hidden layer state, and then obtains higher-level information by stacking the features from multiple hidden layers. The decoder output feature X t As the input feature of the recurrent neural network model, the forward propagation of the recurrent neural network model is represented as: Z t = σ(W z * X t + U z * h t-1 + b z ) R t = σ(W r * X t + U r * h t-1 + b r ) wherein R t denotes the reset gate at time t; Z t denotes the update gate at time t; denotes the candidate hidden state at time t; h t denotes the current cell state at time t; h t-1 denotes the cell state at the previous time; denotes the Hadamard product; denotes the Sigmoid activation function, which converts data to a value of [0, 1]; W r denotes the weight matrix of the reset gate; W z denotes the weight matrix of the update gate; W h denotes the weight matrix of the candidate hidden state; W* and U* are convolutional operation convolution kernels; b* is a bias term, which is a parameter of the model in the training phase; A network is built with two convolutional layers, and then a Softmax function is used to calculate the weight values, S t = U s * tanh(W xs * X t + W hs * h t-1 + b s + b us X t is the user's day t time feature value, h t-1 is the last time unit state, U s , W xs , W hs all represent convolution kernels; S t get the corresponding flow weight, denoted as: t wherein, denotes the attention weight of the input feature, the size is between 0 and 1, and the sum of the attention weights of all input features = 1;​ According to the feature vector at each time, the obtained traffic weight coefficient is multiplied by the input feature to obtain the dynamic traffic authorization value of the user at the current time; it is represented as: X t ' = A t ⊙ x t .

2. The method of claim 1, wherein: The service category of the user is divided into three categories: video category: flow1={RG 11 , RG 12 ,..., RG 1n}; webpage category: flow2={RG 21 , RG 22 ,..., RG 2n}; chat category: flow3={RG 31 , RG 32 ,..., RG 3n}.

3. The method of claim 1, wherein: The user information label in the step 2.1 includes: 1-device type, 2-gender, 3-age, 4-customer type, 5-online time, 6-professional type, 7-whether to exit the account, 8-whether to be extremely low usage, 9-user's average online traffic in the past three months, 10-whether to be fixedly fused package, 11-social influencer, 12-net red influencer, 13-current credit, 14-traffic usage, 15-value score, 16-5G traffic, 17-payment method preference, 18-stability, 19-whether high traffic tendency user, 20-VIP level, 21-stability score, 22-monthly cumulative total traffic, 23-cpu of main machine performance, 24-base station id.

4. The method of claim 1, wherein: The regional traffic feature set is represented as: S={U6, U9, U 11 , U 12 , U 16 , U 17 , U 19 , U 24}.

5. The method of claim 1, wherein: The time-division traffic feature is according to time feature θ, and the online time of the user is divided into three segments: θ1: 8 o'clock-12 o'clock, 14 o'clock-18 o'clock; θ2: 12 o'clock-14 o'clock, 18 o'clock-24 o'clock; θ3: 0 o'clock-8 o'clock; the feature attribute of the traffic use condition of the user in different time is represented as Wherein, l represents different time in a day, and the current time of the user is calculated The traffic use condition of the user in the previous day is similar to the traffic use condition of the user in the current day The similarity value is measured; the time weight parameter of θ1 is 1.5, the time weight parameter of θ2 is 1.7, and the time weight parameter of θ3 is 0.

8.

6. The method of claim 1, wherein: The fusion habit identification traffic feature: the specific traffic behavior of the online user is different under different usage habits; according to the user's access webpage touch point click rate, professional type and network type, the user is distinguished; The number of clicks C of the user on different web page touch points ij is represented as: wherein i represents the number of users, j represents the web touchpoint; J i→j click represents the click-through operation of the user i on the web touchpoint j, J i→j lable represents the collection operation of the user i on the web touchpoint j.

7. The method of claim 1, wherein: The encoder module and the decoder module are included in the step 2.3 based on the attention mechanism of the encoder-decoder model.

8. The method of claim 1, wherein: The weight distribution calculation of the attention mechanism in the step 2.3 is the hidden layer state of the joint feature composed of the regional traffic feature, the time period traffic feature and the fusion habit identification traffic feature and the hidden state semantics of the input signal are compared with the user traffic usage habit.

9. The method of claim 1, wherein: The recurrent neural network model includes two gating units, namely the update gate Z. t and reset door R t Among them, the update gate determines whether the memory information of the hidden state is saved to the current moment; the reset gate controls whether new input information is combined with the memory information of the hidden state.

Images