Cross-border communication package adaptive decision system and method based on multi-source data fusion

Abstract

The present invention belongs to the field of data processing technology, and discloses a cross-border communication package adaptive decision-making system and method based on multi-source data fusion; the method comprises: collecting user traffic data, monetary analysis data and historical payment data; the present invention collects multi-source data such as user traffic data, monetary analysis data and historical payment data, and can more comprehensively understand the needs of users in cross-border scenarios, and provide sufficiently accurate data support for adaptive decision-making of cross-border communication packages; further feature extraction is performed on user traffic data, and the user's daily traffic usage and usage intensity can be analyzed to obtain a basic traffic value, and the package data is collected and analyzed, and the economic background of the user's location and the user's payment ability are analyzed in combination with the monetary analysis data and the historical payment data, and the basic traffic value is selected in combination with the user's personalized needs, so as to accurately screen out a cross-border communication package that meets the user's expectations.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and more specifically, to a cross-border communication package adaptive decision-making system and method based on multi-source data fusion. Background Art

[0002] The Chinese patent application with publication number CN116415206A discloses a method, system, electronic device and computer storage medium for fusion of operator multi-data, including establishing a first-stage rule engine, configuring the data range to be extracted in the first-stage rule engine to form a data pool; the data stream engine extracts data in the data pool, converts the extracted data into a structured data stream, and generates a multi-source data set to be fused by the operator based on the structured data stream; embeds an adaptive streaming engine algorithm model in the first-stage rule engine to form a second-stage rule engine; the business system calls the second-stage rule engine according to business needs, passes the data to be matched to the second-stage rule engine, and the second-stage rule engine matches and processes the incoming data set according to the preset algorithm rules, and returns the matching results. The invention uses different data sources for different tasks, and uses different data fusion algorithms for different data sources, which greatly improves the accuracy and efficiency of data fusion.

[0003] Although the above method can meet most scenarios, research and practical application of the above method and existing technology have found that the above method and existing technology have at least the following defects:

[0004] (1) Operator data mainly comes from internal operator data such as base station signaling data sources, user terminal signaling data sources, and network signaling data sources. The data sources are relatively narrow and do not take into account data such as payment differences in different countries and regions. It is difficult to directly provide sufficiently accurate data support for adaptive decision-making of cross-border communication packages.

[0005] (2) It fails to deeply consider the personalized needs of users (such as their daily data usage and usage intensity), resulting in a large deviation between the recommended packages and users’ expectations.

[0006] In view of this, the present invention proposes a cross-border communication package adaptive decision system and method based on multi-source data fusion to solve the above problems. Summary of the invention

[0007] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned purpose, the present invention provides the following technical solution: a cross-border communication package adaptive decision method based on multi-source data fusion, comprising:

[0008] Collect user traffic data, currency analysis data and historical payment data;

[0009] Perform feature extraction on user traffic data to obtain user traffic features; perform feature extraction on currency analysis data to obtain currency analysis features; perform feature extraction on historical payment data to obtain historical payment features;

[0010] Use user traffic characteristics as input to the basic analysis model to obtain basic traffic values;

[0011] Obtain package service data and corresponding preferential policies of all cross-border communication packages of mobile communication cross-border service providers;

[0012] Obtain the package service data and preferential policies of the L cross-border communication packages corresponding to the basic traffic value, encode the L cross-border communication packages, mark them as package code values, concatenate the package code values ​​with the package service data and preferential policies as package data, use the monetary analysis features, historical payment features and package data as inputs to the personalized analysis model, and obtain the package code values ​​of the cross-border communication packages.

[0013] Furthermore, the training method of the personalized analysis model includes:

[0014] Pre-collect personalized training data, the personalized training data including personalized feature data and package code values ​​corresponding to the personalized feature data; the personalized feature data including currency analysis features, historical payment features and package data;

[0015] Each group of personalized feature data is used as the input of the personalized analysis model, and the personalized analysis model takes the package code value corresponding to each group of personalized feature data as the output, and takes the actual package code value corresponding to each group of personalized feature data as the prediction target; wherein, the sum of the minimized loss function values ​​of all package code values ​​is minimized as the training target, and the personalized analysis model is trained based on the stochastic gradient descent algorithm until the sum of the prediction errors reaches convergence and the training is stopped; the personalized analysis model is a deep neural network model.

[0016] Furthermore, the loss function value of the personalized analysis model is as follows:

[0017] ;

[0018] In the formula, is the loss function value; is the number of training data; is the prediction function of the personalized analysis model, based on the parameters For Training data Produce prediction output; For the The actual output corresponding to the training data, that is, the actual package encoding value, For the training data; The parameters of the personalized analysis model, namely the weights and biases of monetary analysis features, historical payment features, and package data, are optimized through stochastic gradient descent algorithm training.

[0019] Furthermore, the parameters of the personalized analysis model are optimized through stochastic gradient descent algorithm training. The methods include:

[0020] Each time, a personalized feature data is randomly extracted from the training data, and the parameters are updated according to the gradient of the loss function with respect to the personalized feature data;

[0021] In the Updated:

[0022] ;

[0023] ;

[0024] ;

[0025] in, The number of iterations of the personalized analysis model; is the momentum coefficient, the momentum coefficient Modification based on model complexity; For the The momentum variable for the iteration; No. The momentum variable for the iteration; is the momentum variable for the first iteration; The loss function value of the personalized analysis model is about the parameter The gradient of is the loss function value of the personalized analysis model The parameters of the iterations; is the loss function value of the personalized analysis model The parameters of the iteration include the weights and biases of the personalized analysis model.

[0026] Furthermore, the momentum coefficient Methods for correction based on model complexity include:

[0027] Step 1: Evaluate the model complexity, traverse each layer of the personalized analysis model, and obtain parameter data , the number of layers from the input layer to the output layer of the personalized analysis model and nonlinear operations on data ;

[0028] Calculating model complexity :

[0029]

[0030] in, , and They are , and The corresponding weight coefficient;

[0031] Step 2: Correct the momentum coefficient and calculate the attenuation factor of the momentum coefficient :

[0032]

[0033] in, is a mathematical constant; is the decay speed control parameter; To preset the model complexity threshold; calculate the momentum coefficient: ;in, is the preset original momentum coefficient;

[0034] Repeat the above training process until the loss function converges; the package code value output when the loss function converges is used as the final output result.

[0035] Furthermore, the user traffic data includes traffic usage data, Mileage data and corresponding traffic consumption data during the period, and Driving data and corresponding traffic consumption data for each driving route in h driving routes within the time period;

[0036] The driving data of each driving route includes the mileage and driving time of each driving route; the traffic usage data includes the total amount of remaining traffic within a preset time, the total amount of traffic used and the duration of traffic use; the currency analysis data includes the currency exchange rate fluctuation data between the user's area and the mobile communication service provider's area; the historical payment data includes the payment amounts of nearly k vehicle traffic packages;

[0037] The package service data includes package price, package quota and roaming duration; the preferential policies include the quantitative value of traffic discount and the amount of free traffic.

[0038] Furthermore, the user flow characteristics include flow usage characteristics, driving flow scores and speed flow scores; and the method for obtaining the user flow characteristics includes:

[0039] Feature extraction is performed on the traffic usage data to obtain traffic usage features; feature extraction is performed on the mileage data and the corresponding traffic consumption data to obtain the driving traffic score; feature extraction is performed on the mileage and driving time corresponding to each driving route in the h-segment driving route, as well as the corresponding traffic consumption data to obtain the speed traffic score.

[0040] Furthermore, the method for obtaining the traffic usage characteristics includes:

[0041] Statistics are collected on the monthly average traffic usage, monthly average usage rate and rate variance of users within the preset time; the monthly average traffic usage is: the ratio of the total traffic usage to the preset time; the monthly average usage rate is: the average of the ratios of the total traffic usage of each month within the preset time to the traffic usage duration of each month; the rate variance is: the variance of the monthly average usage rate of each month within the preset time; the monthly average traffic usage, monthly average usage rate and rate variance are concatenated as traffic usage features.

[0042] Furthermore, the method of obtaining the driving flow score includes:

[0043] The average flow consumption of each preset mileage is counted, and the average consumption is the ratio of the flow consumption data corresponding to the mileage data to the mileage data, and then multiplied by the preset mileage; the average consumption sequence of the total preset mileage is obtained as the input of the driving analysis model to obtain the driving flow score; the training method of the driving analysis model includes:

[0044] Collecting a group of traffic training data A in advance, the traffic training data includes an average consumption sequence and a driving traffic score corresponding to the average consumption sequence;

[0045] Each group of average consumption sequences is used as the input of the driving analysis model, and the driving analysis model takes the driving flow score corresponding to each group of average consumption sequences as the output, and the actual driving flow score corresponding to each group of average consumption sequences as the prediction target; minimizing the sum of the prediction errors of all driving flow scores is used as the training target; the driving analysis model is trained until the sum of the prediction errors reaches convergence and the training is stopped; the driving analysis model is a deep neural network model.

[0046] Furthermore, the method for obtaining the speed flow score includes:

[0047] The average driving speed and corresponding traffic consumption data of each driving route in the h driving routes in the GPS system are counted; the average driving speed of each driving route is the ratio of the driving mileage and driving time corresponding to each driving route. The average driving speed and traffic consumption data are used as speed relationship data and input into the speed relationship model to obtain the speed and traffic score.

[0048] Furthermore, the currency analysis feature includes a historical fluctuation feature and a fluctuation amplitude feature; and the method for obtaining the currency analysis feature includes:

[0049] The mean and standard deviation of currency exchange rate fluctuation data within a preset time period are counted and concatenated to obtain a fluctuation vector, which is used as a historical fluctuation feature; the coefficient of variation of exchange rate fluctuation data within a preset time period, that is, the ratio of the standard deviation to the mean, is counted, and the coefficient of variation is used as a fluctuation amplitude feature;

[0050] The method for obtaining the historical payment characteristics includes:

[0051] The payment amounts of nearly k vehicle traffic packages are concatenated as historical payment features.

[0052] The cross-border communication package adaptive decision system based on multi-source data fusion is used to implement the cross-border communication package adaptive decision method based on multi-source data fusion, including:

[0053] Data collection module: used to collect user traffic data, currency analysis data and historical payment data;

[0054] Data analysis module: used to extract features from user traffic data to obtain user traffic features; extract features from currency analysis data to obtain currency analysis features; extract features from historical payment data to obtain historical payment features;

[0055] Basic analysis module: used to take user traffic characteristics as input of the basic analysis model to obtain basic traffic values;

[0056] The second acquisition module is used to obtain package service data and corresponding preferential policies of all cross-border communication packages of mobile communication cross-border service providers;

[0057] Personalized risk module: used to obtain the package service data and preferential policies of L cross-border communication packages corresponding to the basic traffic value, encode the L cross-border communication packages, mark them as package code values, concatenate the package code values ​​with the package service data and preferential policies as package data, and use the currency analysis features, historical payment features and package data as inputs to the personalized analysis model to obtain the package code values ​​of the cross-border communication packages.

[0058] The technical effects and advantages of the cross-border communication package adaptive decision-making system and method based on multi-source data fusion of the present invention are as follows:

[0059] The present invention collects multi-source data such as user traffic data, monetary analysis data and historical payment data, so as to have a more comprehensive understanding of user needs in cross-border scenarios and provide sufficiently accurate data support for adaptive decision-making of cross-border communication packages; further feature extraction is performed on user traffic data, so as to analyze the user's daily traffic usage and usage intensity to obtain the basic traffic value, collect package data, analyze the package data, combine the monetary analysis data and historical payment data to analyze the economic background of the user's location and the user's payment ability, and select the basic traffic value in combination with the user's personalized needs, so as to accurately screen out the cross-border communication packages that meet the user's expectations. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 This is a flow chart of the cross-border communication package adaptive decision-making method based on multi-source data fusion mentioned in Embodiment 1 of the present invention;

[0061] Figure 2 This is a schematic diagram of extracting user traffic features mentioned in Example 1 of the present invention;

[0062] Figure 3 This is a schematic diagram of the personalized analysis model training process mentioned in Example 1 of the present invention;

[0063] Figure 4 This is a block diagram of the cross-border communication package adaptive decision-making system based on multi-source data fusion mentioned in Example 2 of the present invention. DETAILED DESCRIPTION

[0064] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0065] Example 1

[0066] See also Figure 1 As shown, this embodiment provides a cross-border communication package adaptive decision method based on multi-source data fusion, including:

[0067] Collect user traffic data, currency analysis data and historical payment data; user traffic data includes traffic usage data, Mileage data and corresponding traffic consumption data during the period, and Driving data and corresponding traffic consumption data for each driving route in the h driving routes within the time period.

[0068] Currency analysis data includes currency exchange rate fluctuation data between the user's region and the mobile communication service provider's region, which is obtained by connecting to the international financial market. The currency exchange rate levels in different regions are different. The currency exchange rate fluctuation data between the user's region and the mobile communication service provider's region can reflect the economic development level and price level of the user's region to a certain extent, which is helpful for analyzing the user's economic background; for users in regions with low currency exchange rates and relatively backward economic development, more affordable and cost-effective cross-border communication packages can be recommended to suit their consumption capacity.

[0069] The historical payment data is obtained through the user account information, including the payment amount of nearly k in-vehicle data packages. By collecting historical payment data through user account information, we can analyze the user's consumption level and recommend suitable data packages for users; for example, for high-spending users, we can recommend high-end cross-border communication packages that include more value-added services, such as unlimited data cross-border communication packages.

[0070] Traffic usage data includes the remaining total traffic, total traffic usage and traffic usage duration within the preset time; it is obtained through online statistics; combining the remaining total traffic, total traffic usage and traffic usage duration can help understand the user's overall traffic management. If the user's total traffic usage is high, the remaining total is often low and the traffic usage duration is long, it means that the user has a large demand for traffic, and a basic traffic value with a large traffic quota can be generated to meet the user's high traffic demand, such as an unlimited traffic package or a large-capacity traffic package.

[0071] Mileage data and corresponding traffic consumption data during the period, and The driving data and corresponding traffic consumption data of each driving route within the time period are obtained by connecting to the GPS system. The driving data of each driving route is the mileage and driving time of each driving route. The mileage data can reflect the user's travel range, and the corresponding traffic consumption data can reflect the user's traffic demand. If the user is If the mileage is long and the traffic consumption is high during the time period, a cross-border communication package with a larger traffic value can be selected for the user; the driving data of each route can reflect the user's driving habits, and the corresponding traffic consumption data can reflect the user's traffic demand during driving; if the user's driving data is small but the traffic consumption data is high, this indicates that the user's traffic demand is high, and the basic traffic value can be appropriately increased.

[0072] Perform feature extraction on user traffic data to obtain user traffic features; perform feature extraction on monetary analysis data to obtain monetary analysis features; perform feature extraction on historical payment data to obtain historical payment features.

[0073] User traffic characteristics include traffic usage characteristics, driving traffic score and speed traffic score; see Figure 2 ,The methods for obtaining user traffic characteristics include:

[0074] Feature extraction is performed on the traffic usage data to obtain traffic usage features; feature extraction is performed on the mileage data and the corresponding traffic consumption data to obtain the driving traffic score; feature extraction is performed on the mileage and driving time corresponding to each driving route in the h-segment driving route, as well as the corresponding traffic consumption data to obtain the speed traffic score.

[0075] Methods for obtaining traffic usage characteristics include:

[0076] The monthly average traffic usage, monthly average usage rate and rate variance of users within the preset time are counted; the monthly average traffic usage is: the ratio of the total traffic usage to the preset time; the monthly average usage rate is: the average of the ratio of the total traffic usage in each month within the preset time to the traffic usage duration in each month; the rate variance is: the variance of the monthly average usage rate in each month within the preset time; the monthly average traffic usage, monthly average usage rate and rate variance are spliced ​​as traffic usage characteristics. For example, if the monthly average traffic usage is 13.62GB / month; the monthly average usage rate is 125MB / hour; and the rate variance is 0.000027GB² / hour², the traffic usage characteristics (13.62, 125, 0.000027) are obtained. Through traffic usage characteristics, operators can customize basic traffic values ​​for users. For example, for users with higher monthly average traffic usage, high-traffic cross-border communication packages can be provided, while for users with lower monthly average traffic usage, low-traffic cross-border communication packages can be provided. This personalized service can improve user satisfaction and reduce churn.

[0077] Ways to earn traffic scores include:

[0078] The average traffic consumption per preset mileage is calculated (e.g., the average traffic consumption per 100 km). The average consumption is the ratio of the traffic consumption data corresponding to the mileage data to the mileage data, multiplied by the preset mileage. It is calculated according to the following formula:

[0079] ;

[0080] Among them, B is the average consumption; C is the traffic consumption data corresponding to the mileage data; D is the preset mileage; and E is the mileage data.

[0081] The average consumption sequence of the total preset mileage (for example, if the total preset mileage is 1000 km, there are 10 average consumption sequence data) is obtained as the input of the driving analysis model to obtain the driving flow score.

[0082] The training methods of the driving analysis model include:

[0083] A group of traffic training data is collected in advance, and the traffic training data includes an average consumption sequence and a driving traffic score corresponding to the average consumption sequence.

[0084] Each group of average consumption sequences is used as the input of the driving analysis model, and the driving analysis model takes the driving flow score corresponding to each group of average consumption sequences as the output, and the actual driving flow score corresponding to each group of average consumption sequences as the prediction target; minimizing the sum of the prediction errors of all driving flow scores is used as the training target; the driving analysis model is trained until the sum of the prediction errors reaches convergence and the training is stopped; the driving analysis model is a deep neural network model.

[0085] The loss function value of the driving analysis model is the mean square error.

[0086] Prediction error formula: , a is the group number of the average consumption sequence; A is the number of groups in the average consumption sequence; is the predicted driving flow score corresponding to the average consumption sequence of group a; is the actual driving flow score corresponding to the average consumption sequence of group a.

[0087] The driving traffic score can understand the user's daily travel habits and the intensity of traffic usage during daily travel, and use this to make more accurate adjustments to the basic traffic value. For example, if a user's driving traffic score is high, it means that he consumes a lot of traffic during driving. The operator can consider providing the user with a higher basic traffic value to meet his needs.

[0088] Methods for obtaining speed flow scores include:

[0089] The average driving speed and corresponding traffic consumption data of each driving route in the h driving routes in the GPS system (for example, a navigation path is a driving route) are counted; the average driving speed of each driving route is the ratio of the driving mileage and driving time corresponding to each driving route. The average driving speed and traffic consumption data are used as speed relationship data and input into the speed relationship model to obtain the speed-traffic score.

[0090] The training method of the speed relationship model can refer to the training method of the driving analysis model.

[0091] The speed and traffic score can reflect the traffic usage habits of users during travel; different users have significantly different traffic usage habits during travel, and a higher score means that the vehicle's traffic consumption data is higher at a specific speed. By analyzing the speed and traffic score, operators can better understand the traffic consumption at different speeds and make more reasonable adjustments to the basic traffic value to meet user needs.

[0092] The currency analysis characteristics include historical fluctuation characteristics and fluctuation range characteristics; the methods for obtaining the currency analysis characteristics include:

[0093] The mean and standard deviation of currency exchange rate fluctuation data within a preset time period are counted and spliced ​​to obtain a fluctuation vector, which is used as a historical fluctuation feature; the coefficient of variation of exchange rate fluctuation data within a preset time period, that is, the ratio of the standard deviation to the mean, is counted, and the coefficient of variation is used as a fluctuation amplitude feature. Currency analysis features can reflect the user's economic background. Different currency exchange rate levels imply the economic conditions in the user's region; for example, if the user's region has high historical fluctuation characteristics and fluctuation amplitude characteristics, it may be more sensitive to prices; when choosing a cross-border communication package, it will tend to be more cost-effective.

[0094] Methods for obtaining historical payment characteristics include:

[0095] The payment amounts of nearly k vehicle traffic packages are spliced ​​together as historical payment features. Historical payment features can evaluate the user's spending power. For example, users with higher payment amounts may have stronger spending power and higher requirements for package quality and service content. High-end cross-border communication packages can be recommended to such users. On the contrary, users with lower payment amounts may be more sensitive to prices, so affordable cross-border communication packages can be recommended to users.

[0096] The user flow characteristics are used as the input of the basic analysis model to obtain the basic flow value; the training method of the basic analysis model can refer to the training method of the driving analysis model.

[0097] For example, the user's traffic usage characteristics are (18.62, 125, 0.000027); the driving traffic score is 69; the speed traffic score is 56; the dynamic portrait label obtained is high frequency, medium speed and high consumption; and the basic traffic value obtained is 20GB.

[0098] For example, the user's traffic usage characteristics are (8.95, 62, 0.000030); the driving traffic score is 75; the speed traffic score is 52; the dynamic portrait label obtained is medium frequency, medium speed and low consumption; and the basic traffic value obtained is 10GB.

[0099] For example, the user's traffic usage characteristics are (11.12, 96, 0.000025); the driving traffic score is 88; the speed traffic score is 76; the dynamic portrait label obtained is high-frequency high-speed consumption; and the basic traffic value obtained is 15GB.

[0100] Through traffic usage characteristics, operators can preliminarily determine users' traffic needs. Further combined with driving traffic scores and speed traffic scores, operators can more accurately understand users' driving habits and can make adaptive adjustments to users' traffic needs.

[0101] The training method of the basic analysis model can refer to the training method of the driving analysis model.

[0102] Obtain package service data and corresponding preferential policies of all cross-border communication packages of mobile communication cross-border service providers. Package service data includes package price, package quota and roaming duration; preferential policies include the quantitative value of traffic discount (e.g., a cross-border communication package with an original price of 100 yuan is discounted by 20% during the preferential activity, so the quantitative value of traffic discount is 80%) and the amount of free traffic; obtaining package service data and corresponding preferential policies can accurately obtain the content of various packages, and provide users with accurate package recommendation strategies in combination with user portraits;

[0103] Obtain the package service data and preferential policies of L cross-border communication packages corresponding to the basic traffic value, encode the L cross-border communication packages, mark them as package code values, concatenate the package code values ​​with the package service data and preferential policies as package data, use the monetary analysis features, historical payment features and package data as inputs of the personalized analysis model, and obtain the package code value of the selected personalized cross-border communication package; wherein, the value of L is large enough, when the number of cross-border communication packages corresponding to the basic traffic value is less than L, fill them with zeros to ensure that the number of cross-border communication packages reaches L.

[0104] The training methods for personalized analysis models include:

[0105] Pre-collect personalized training data, the personalized training data including personalized feature data and package code values ​​corresponding to the personalized feature data; the personalized feature data including currency analysis features, historical payment features and package data;

[0106] Each group of personalized feature data is used as the input of the personalized analysis model, and the personalized analysis model takes the package code value corresponding to each group of personalized feature data as the output, and takes the actual package code value corresponding to each group of personalized feature data as the prediction target; wherein, the sum of the minimized loss function values ​​of all package code values ​​is minimized as the training target, and the personalized analysis model is trained based on the stochastic gradient descent algorithm until the sum of the prediction errors reaches convergence and the training is stopped; the personalized analysis model is a deep neural network model.

[0107] The loss function value of the personalized analysis model is as follows:

[0108] ;

[0109] In the formula, is the loss function value; is the number of training data; is the prediction function of the personalized analysis model, based on the parameters For Training data Produce prediction output; For the The actual output corresponding to the training data, that is, the actual package encoding value, For the training data; To personalize the analysis model’s parameters, namely the weights and biases of currency analysis features, historical payment features, and package data, they need to be optimized through training with the stochastic gradient descent algorithm.

[0110] Reference Figure 3 , optimize the parameters of the personalized analysis model through stochastic gradient descent algorithm training The methods include:

[0111] Initialize model parameters and momentum variables, select a batch of training data, perform forward propagation based on the personalized analysis model and calculate the output and loss function of the personalized analysis model, calculate the gradient of the loss function with respect to the output layer, backpropagate the gradient from the output layer to the hidden layer and update the momentum variables and parameters;

[0112] Each time, a personalized feature data is randomly extracted from the training data, and the parameters are updated according to the gradient of the loss function with respect to the personalized feature data;

[0113] In the Updated:

[0114] ;

[0115] ;

[0116] ;

[0117] in, The number of iterations of the personalized analysis model; is the momentum coefficient, the momentum coefficient Modification based on model complexity; For the The momentum variable for the iteration; No. The momentum variable for the iteration; is the momentum variable for the first iteration; The loss function value of the personalized analysis model is about the parameter The gradient of is the loss function value of the personalized analysis model The parameters of the iterations; is the loss function value of the personalized analysis model The parameters of the iteration include the weights and biases of the personalized analysis model.

[0118] Momentum coefficient Methods for correction based on model complexity include:

[0119] Step 1: Evaluate model complexity: traverse each layer of the personalized analysis model to obtain parameter data , the number of layers from the input layer to the output layer of the personalized analysis model and nonlinear operations on data ;

[0120] Calculating model complexity :

[0121]

[0122] in, , and They are , and The corresponding weight coefficient;

[0123] Step 2: Correct the momentum coefficient: Calculate the attenuation factor of the momentum coefficient :

[0124]

[0125] in, is a mathematical constant; is the decay speed control parameter; To preset the model complexity threshold; calculate the momentum coefficient: ;in, is the preset original momentum coefficient;

[0126] Determine whether all training data batches have been traversed. If not, continue to select training data in the training batch to repeat the above training process until all training data batches have been traversed. If training has been completed, determine whether the preset training rounds have been reached, such as 50 times. If not, the learning rate can be updated and the above training process can be repeated until the loss function converges or the preset training rounds are reached. If it has been reached, the training is completed to obtain the corresponding personalized analysis model that has been trained. After each round, the model will conduct a complete learning of the entire training data set. As the number of rounds increases, the loss value of the model will usually gradually decrease, and the personalized service strategy output by the model will become closer and closer to the true label. During the training process, the performance of the model can be evaluated on the validation set regularly (such as every 10 rounds) to observe whether the model is overfitting or underfitting, so as to adjust the training parameters (such as learning rate, batch size, etc.) in time. Optimize the parameters of the personalized analysis model through the stochastic gradient descent algorithm This method can speed up model training, effectively process large-scale data, and avoid falling into local optimality to a certain extent.

[0127] In the training of personalized analysis models, the stochastic gradient descent algorithm combined with the momentum method that is aware of model complexity can dynamically adjust the momentum coefficient according to the complexity of the model. When the model complexity is high, the momentum coefficient is reduced to reduce the impact of historical gradient accumulation, making parameter updates more cautious. This helps to avoid excessive parameter updates due to excessive momentum, making the model training process more stable, especially when faced with complex personalized data, and can better converge. The momentum method combined with model complexity awareness can also optimize the training process by dynamically adjusting the momentum. It can avoid invalid historical gradient accumulation when the model complexity increases, allowing the model to adapt to new gradient changes more quickly, thereby improving training efficiency, reducing training time, and achieving a better personalized model performance more quickly.

[0128] Example 2

[0129] See also Figure 4 This embodiment provides a cross-border communication package adaptive decision system based on multi-source data fusion, including:

[0130] Data collection module: used to collect user traffic data, currency analysis data and historical payment data.

[0131] Data analysis module: used to extract features from user traffic data to obtain user traffic features; extract features from currency analysis data to obtain currency analysis features; extract features from historical payment data to obtain historical payment features;

[0132] Basic analysis module: used to take user traffic characteristics as input of the basic analysis model to obtain basic traffic values;

[0133] The second acquisition module is used to obtain package service data and corresponding preferential policies of all cross-border communication packages of mobile communication cross-border service providers;

[0134] Personalized risk module: used to obtain the package service data and preferential policies of L cross-border communication packages corresponding to the basic traffic value, encode the L cross-border communication packages, mark them as package code values, concatenate the package code values ​​with the package service data and preferential policies as package data, and use the currency analysis features, historical payment features and package data as inputs to the personalized analysis model to obtain the package code values ​​of the cross-border communication packages.

[0135] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

[0136] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An adaptive decision-making method for cross-border communication packages based on multi-source data fusion, characterized in that: include: Collect user traffic data, currency analysis data and historical payment data; the user traffic data includes traffic usage data, Mileage data and corresponding traffic consumption data during the period, and The driving data and corresponding traffic consumption data of each driving route in the h driving routes within the time period; the currency analysis data includes the currency exchange rate fluctuation data between the user's area and the mobile communication service provider's area; the historical payment data includes the payment amounts of nearly k in-vehicle traffic packages; Extract features from user traffic data to obtain user traffic features; Extract features from currency analysis data to obtain currency analysis features; Extract features from historical payment data to obtain historical payment features; Use user traffic characteristics as input to the basic analysis model to obtain basic traffic values; Obtain package service data and corresponding preferential policies of all cross-border communication packages of mobile communication cross-border service providers; Obtain the package service data and preferential policies of the L cross-border communication packages corresponding to the basic traffic value, encode the L cross-border communication packages, mark them as package code values, concatenate the package code values ​​with the package service data and preferential policies as package data, use the monetary analysis features, historical payment features and package data as inputs to the personalized analysis model, and obtain the package code values ​​of the cross-border communication packages.

2. The cross-border communication package adaptive decision method based on multi-source data fusion according to claim 1 is characterized in that: The training method of the personalized analysis model includes: Pre-collect personalized training data, the personalized training data including personalized feature data and package code values ​​corresponding to the personalized feature data; the personalized feature data including currency analysis features, historical payment features and package data; Each group of personalized feature data is used as the input of the personalized analysis model, and the personalized analysis model takes the package code value corresponding to each group of personalized feature data as the output, and takes the actual package code value corresponding to each group of personalized feature data as the prediction target; wherein, the sum of the minimized loss function values ​​of all package code values ​​is minimized as the training target, and the personalized analysis model is trained based on the stochastic gradient descent algorithm until the sum of the prediction errors reaches convergence and the training is stopped; the personalized analysis model is a deep neural network model.

3. The cross-border communication package adaptive decision method based on multi-source data fusion according to claim 2 is characterized in that: The loss function value of the personalized analysis model is as follows: ； In the formula, is the loss function value; is the number of training data; is the prediction function of the personalized analysis model, based on the parameters For Training data Produce prediction output; For the The actual output corresponding to the training data, that is, the actual package encoding value, For the training data; The parameters of the personalized analysis model, namely the weights and biases of monetary analysis features, historical payment features, and package data, are optimized through stochastic gradient descent algorithm training.

4. The cross-border communication package adaptive decision method based on multi-source data fusion according to claim 3 is characterized in that: Optimize the parameters of the personalized analysis model through stochastic gradient descent training The methods include: Each time, a personalized feature data is randomly extracted from the training data, and the parameters are updated according to the gradient of the loss function with respect to the personalized feature data; In the Updated: ； ； ； in, The number of iterations of the personalized analysis model; is the momentum coefficient, the momentum coefficient Modification based on model complexity; For the The momentum variable for the iteration; No. The momentum variable for the iteration; is the momentum variable for the first iteration; The loss function value of the personalized analysis model is about the parameter The gradient of is the loss function value of the personalized analysis model The parameters of the iterations; is the loss function value of the personalized analysis model The parameters of the iteration include the weights and biases of the personalized analysis model.

5. The cross-border communication package adaptive decision method based on multi-source data fusion according to claim 4 is characterized in that: Momentum coefficient Methods for correction based on model complexity include: Step 1: Evaluate the model complexity, traverse each layer of the personalized analysis model, and obtain parameter data , the number of layers from the input layer to the output layer of the personalized analysis model and nonlinear operations on data ; Calculating model complexity : ； in, , and They are , and The corresponding weight coefficient; Step 2: Correct the momentum coefficient and calculate the attenuation factor of the momentum coefficient : ； in, is a mathematical constant; is the decay speed control parameter; To preset the model complexity threshold; calculate the momentum coefficient: ;in, is the preset original momentum coefficient; Repeat the above training process until the loss function converges; the package code value output when the loss function converges is used as the final output result.

6. The cross-border communication package adaptive decision method based on multi-source data fusion according to claim 1 is characterized in that: The driving data of each driving route is the driving mileage and driving time of each driving route; the traffic usage data includes the total remaining traffic volume, the total traffic usage volume and the traffic usage duration within a preset time; The package service data includes package price, package quota and roaming duration; the preferential policies include the quantitative value of traffic discount and the amount of free traffic.

7. The cross-border communication package adaptive decision method based on multi-source data fusion according to claim 6 is characterized in that: The user flow characteristics include flow usage characteristics, driving flow scores and speed flow scores; The method for obtaining the user traffic characteristics includes: Feature extraction is performed on the traffic usage data to obtain traffic usage features; feature extraction is performed on the mileage data and the corresponding traffic consumption data, and the corresponding feature extraction results are used as the input of the driving analysis model to obtain the driving traffic score; feature extraction is performed on the mileage and driving time corresponding to each driving route in the h-segment driving route, as well as the corresponding traffic consumption data, and the corresponding feature extraction results are used as the input of the speed relationship model to obtain the speed flow score.

8. The cross-border communication package adaptive decision method based on multi-source data fusion according to claim 7 is characterized in that: The method for obtaining the traffic usage characteristics includes: Statistics are collected on the monthly average traffic usage, monthly average usage rate and rate variance of users within the preset time; the monthly average traffic usage is: the ratio of the total traffic usage to the preset time; the monthly average usage rate is: the average of the ratios of the total traffic usage of each month within the preset time to the traffic usage duration of each month; the rate variance is: the variance of the monthly average usage rate of each month within the preset time; the monthly average traffic usage, monthly average usage rate and rate variance are concatenated as traffic usage features.

9. The cross-border communication package adaptive decision method based on multi-source data fusion according to claim 7 is characterized in that: Ways to earn traffic scores include: The average flow consumption of each preset mileage is counted, and the average consumption is the ratio of the flow consumption data corresponding to the mileage data to the mileage data, and then multiplied by the preset mileage; the average consumption sequence of the total preset mileage is obtained as the input of the driving analysis model to obtain the driving flow score; the training method of the driving analysis model includes: Collecting group A of traffic training data in advance, the traffic training data includes an average consumption sequence and a driving traffic score corresponding to the average consumption sequence; wherein the driving traffic score is used to measure the extent of mobile data traffic consumption by the user during driving of the mobile device, and the driving traffic score is proportional to the actual amount of mobile data traffic consumed by the user during driving; Each group of average consumption sequences is used as the input of the driving analysis model, and the driving analysis model takes the driving flow score corresponding to each group of average consumption sequences as the output, and the actual driving flow score corresponding to each group of average consumption sequences as the prediction target; minimizing the sum of the prediction errors of all driving flow scores is used as the training target; the driving analysis model is trained until the sum of the prediction errors reaches convergence and the training is stopped; the driving analysis model is a deep neural network model.

10. The cross-border communication package adaptive decision method based on multi-source data fusion according to claim 7 is characterized in that: Methods for obtaining the speed flow score include: The average driving speed and corresponding traffic consumption data of each driving route in the h driving routes in the GPS system are counted; the average driving speed of each driving route is the ratio of the driving mileage and driving time corresponding to each driving route, and the average driving speed and traffic consumption data are used as speed relationship data and input into the speed relationship model to obtain a speed-traffic score; the speed-traffic score is used to measure the user's mobile data traffic consumption intensity at different driving speeds, and the speed-traffic score is proportional to the traffic consumption per unit time in a specific speed range.

11. The cross-border communication package adaptive decision method based on multi-source data fusion according to claim 6 is characterized in that: The currency analysis characteristics include historical fluctuation characteristics and fluctuation amplitude characteristics; Methods for obtaining currency analysis features include: The mean and standard deviation of currency exchange rate fluctuation data within a preset time period are counted and concatenated to obtain a fluctuation vector, which is used as a historical fluctuation feature; the coefficient of variation of exchange rate fluctuation data within a preset time period, that is, the ratio of the standard deviation to the mean, is counted, and the coefficient of variation is used as a fluctuation amplitude feature; The method for obtaining the historical payment characteristics includes: The payment amounts of nearly k vehicle traffic packages are concatenated as historical payment features.

12. The cross-border communication package adaptive decision system based on multi-source data fusion is characterized by: The method for implementing the cross-border communication package adaptive decision-making method based on multi-source data fusion as described in any one of claims 1 to 11 comprises: Data collection module: used to collect user traffic data, currency analysis data and historical payment data; Data analysis module: used to extract features from user traffic data to obtain user traffic features; extract features from currency analysis data to obtain currency analysis features; extract features from historical payment data to obtain historical payment features; Basic analysis module: used to take user traffic characteristics as input of the basic analysis model to obtain basic traffic values; The second acquisition module is used to obtain package service data and corresponding preferential policies of all cross-border communication packages of mobile communication cross-border service providers; Personalized risk module: used to obtain the package service data and preferential policies of L cross-border communication packages corresponding to the basic traffic value, encode the L cross-border communication packages, mark them as package code values, concatenate the package code values ​​with the package service data and preferential policies as package data, and use the currency analysis features, historical payment features and package data as inputs to the personalized analysis model to obtain the package code values ​​of the cross-border communication packages.

Images