Model training method and device, and operator package recommendation method

Abstract

The application discloses a model training method and device and a recommendation method of an operator package. The method comprises the following steps: obtaining historical operator package data and historical voice data of a target object; generating a first row vector corresponding to the historical operator package data, performing sentiment analysis on the historical voice data, and obtaining a second row vector; merging the first row vector and the second row vector to obtain a target matrix; inputting the target matrix into a generative adversarial network model; determining a sentiment tendency reward according to the first row vector and the second row vector; determining a target probability as an optimization parameter, updating the optimization parameter, and obtaining a trained generative adversarial network model if a loss function related to the target probability meets a preset convergence condition. The application solves the technical problem of low recommendation accuracy of an operator package caused by the fact that a related operator package recommendation method is limited by a static model and does not fully consider the dynamic change of user demand.

Description

Technical Field

[0001] This application relates to the field of computer technology, and more specifically, to a model training method, apparatus, and a method for recommending operator packages. Background Technology

[0002] In today's telecommunications service market, a wide variety of packages are available, covering different data allowances, call durations, and SMS messages, aiming to meet the needs of diverse user groups. However, current telecom operator package recommendation systems, relying on static models and simple rule systems, struggle to provide accurate and personalized services. While these methods can make initial recommendations based on basic user information (such as age, gender, and region) and package usage history, they fail to fully consider the diversity and dynamic changes in individual user needs, and ignore the complexity of user behavior patterns. This results in limited recommendation accuracy and user satisfaction, as static models struggle to capture users' real-time needs and emotional fluctuations, while simple statistical methods lack personalized and intelligent decision-making mechanisms.

[0003] There is currently no effective solution to the above problems. Summary of the Invention

[0004] This application provides a model training method, apparatus, and method for recommending operator packages, in order to at least solve the technical problem that the accuracy of operator package recommendations is low because the relevant operator package recommendation methods are limited by static models and do not fully consider the dynamic changes in user needs.

[0005] According to one aspect of this application, a model training method is provided, comprising: acquiring historical carrier package data and historical voice data of a target object, wherein the historical carrier package data is multi-dimensional information including attribute information and historical package usage information of the target object, and the historical voice data includes usage demand information for the carrier package data; generating a first row vector corresponding to the historical carrier package data; performing sentiment analysis on the usage demand information in the historical voice data to obtain a second row vector, wherein the components of the first row vector are different types of recommended packages for the target object, and the components of the second row vector are the target object's sentiment tendency information for different types of recommended packages; merging the first row vector and the second row vector to obtain a target matrix, wherein the rows of the target matrix are used to represent the package category of the recommended packages, and the columns are used to represent... The features corresponding to the package categories are defined as follows: The target matrix is ​​input into the Generative Adversarial Network (GAN) model to obtain target data with a similar distribution to the target matrix output by the generator in the GAN model, and the target probability that the target data output by the discriminator in the GAN model is real data. Based on the first and second row vectors, the sentiment reward is determined, and based on the sentiment reward, the reward function is determined. The target matrix and target data are defined as state vectors, and the recommended package for the target object is defined as action vectors. Based on the target probability state vector, action vector, and reward function, the objective function is determined. The target probability is defined as the optimization parameter to maximize the objective function. The optimization parameter is updated. In the discriminator, if the loss function with respect to the target probability satisfies the preset convergence condition, the trained GAN model is obtained.

[0006] Optionally, the objective function is determined based on the target probability state vector, action vector, and reward function, including: determining several pairs of state vectors and action vectors under the current optimization parameters as observations to obtain an observation matrix including multiple observations; determining the negative log-marginal likelihood function based on the target probability, observation matrix, and reward function, and using the negative log-marginal likelihood function as the objective function; updating the optimization parameters, including: obtaining the model parameters in the multi-output student T process, wherein the model parameters include: degrees of freedom, mean function, covariance function, parameter matrix, and optimization parameters; calculating the derivative of the negative log-marginal likelihood with respect to the model parameters in the multi-output student T process, and updating the model parameters in the opposite direction of the derivative.

[0007] Optionally, the sentiment incentive reward is determined based on the first row vector and the second row vector, including: determining the first weight based on the activity index of the historical operator package used by the target object; determining the second weight based on the package value index of the historical operator package used by the target object; and determining the sentiment incentive reward based on the first weight, the second weight, the first row vector, and the second row vector.

[0008] Optionally, a reward function is determined based on sentiment-based rewards, including: obtaining the target audience's satisfaction index with the historical operator packages used by the target audience, the operator's revenue index with the historical operator packages used by the target audience, and the market adaptability index of the historical operator packages used by the target audience; and weighting and summing the sentiment-based rewards, satisfaction index, revenue index, and market adaptability index to obtain the reward function.

[0009] Optionally, updating the optimization parameters includes: obtaining several state vector and action vector pairs under the current optimization parameters; calculating the reward function in each state vector and action vector, and updating the optimization parameters according to the calculation results, until the mathematical expectation of the state vector and action vector pair is greater than a preset threshold.

[0010] Optionally, generating the first row vector corresponding to historical operator package data includes: cleaning the historical operator package data and interpolating and padding missing data in the cleaned historical operator package data to obtain initial data with multiple dimensions, including: target object identification information, age information, gender information, region information, package subsidy information, package usage information, call information outside the package, SMS information outside the package, and data traffic information outside the package; extracting features from the initial data with multiple dimensions to obtain multiple initial features; normalizing the multiple initial features to obtain multiple first features; and generating different types of recommended packages for the target object based on the multiple first features to obtain the first row vector.

[0011] Optionally, sentiment analysis is performed on the usage demand information in the historical voice data to obtain the second row vector, including: converting the historical voice data into text data, performing feature extraction processing on the text data to obtain a text feature vector, wherein the feature extraction processing includes at least one of the following: word frequency statistics, syntactic analysis, and topic modeling; using a pre-trained sentiment analysis model to perform sentiment tendency analysis on the text feature vector to obtain a quantitative score of the target object's sentiment tendency towards different types of recommended packages; and converting the quantitative score into vector form to obtain the second row vector.

[0012] According to another aspect of this application, a method for recommending operator packages is also provided, comprising: acquiring operator package data and voice data of a target object, wherein the operator package data is multi-dimensional information including attribute information and package usage information of the target object, and the voice data includes usage demand information for the operator package data; inputting the vectorized operator package data and voice data into a trained generative adversarial network model to obtain a recommended package for the target object output by the trained generative adversarial network model, wherein the generative adversarial network model is obtained by training using the above model training method.

[0013] According to another aspect of this application, a model training apparatus is also provided, comprising: an acquisition module, configured to acquire historical carrier package data and historical voice data of a target object, wherein the historical carrier package data is multi-dimensional information including attribute information and historical package usage information of the target object, and the historical voice data includes usage demand information for the carrier package data; a generation module, configured to generate a first row vector corresponding to the historical carrier package data, and perform sentiment analysis on the usage demand information in the historical voice data to obtain a second row vector, wherein the components of the first row vector are different types of recommended packages for the target object, and the components of the second row vector are the target object's sentiment tendency information for different types of recommended packages; and a merging module, configured to merge the first row vector and the second row vector to obtain a target matrix, wherein the rows of the target matrix are used to represent the package categories of the recommended packages, and the columns are used to represent the table. The model is divided into four modules: a target matrix and a target data set. The target matrix is ​​input into the generative adversarial network (GAN) model to obtain target data with a similar distribution to the target matrix from the generator, and the target probability from the discriminator. The target data is then identified as real data. A determination module is used to determine the sentiment reward based on the first and second row vectors, and to determine the reward function based on the sentiment reward. The target matrix and target data are defined as state vectors, and the recommended package for the target object is defined as action vectors. The target function is determined based on the target probability state vector, action vector, and reward function. An update module is used to determine the target probability as optimization parameters to maximize the target function and update the optimization parameters. In the discriminator, if the loss function with respect to the target probability satisfies a preset convergence condition, the trained GAN model is obtained.

[0014] According to another aspect of this application, a non-volatile storage medium is also provided, the storage medium including a stored program, wherein the program, when running, controls the device where the storage medium is located to execute the above model training method.

[0015] According to another aspect of this application, an electronic device is also provided, comprising: a memory and a processor, the processor being configured to run a program stored in the memory, wherein the program executes the above-described model training method during runtime.

[0016] According to another aspect of this application, a computer program is also provided, wherein the computer program, when executed by a processor, implements the above-described model training method.

[0017] According to another aspect of this application, a computer program product is also provided, the computer program product including a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores a computer program, which implements the above-described model training method when executed by a processor.

[0018] In this application, historical carrier package data and historical voice data of the target object are obtained. The historical carrier package data includes multi-dimensional information such as attribute information and historical package usage information of the target object, while the historical voice data includes usage demand information related to carrier package data. A first row vector corresponding to the historical carrier package data is generated. Sentiment analysis is performed on the usage demand information in the historical voice data to obtain a second row vector. The components of the first row vector represent different types of recommended packages for the target object, and the components of the second row vector represent the target object's sentiment towards different types of recommended packages. The first and second row vectors are merged to obtain a target matrix. The rows of the target matrix represent the package categories of the recommended packages, and the columns represent the features corresponding to the package categories. The target matrix is ​​input into a generative adversarial network (GAN) model to obtain target data with a similar distribution to the target matrix, output by the generator in the GAN model, and the generated... In the adversarial network model, the discriminator outputs the target data as the target probability of the real data. Based on the first and second row vectors, the sentiment reward is determined, and a reward function is established accordingly. The target matrix and target data are defined as state vectors, and the recommended package for the target object is defined as action vectors. Based on the target probability state vector, action vector, and reward function, the target function is determined. The target probability is defined as the optimization parameter to maximize the target function. The optimization parameter is updated. In the discriminator, if the loss function regarding the target probability satisfies the preset convergence condition, the trained generative adversarial network model is obtained. This approach fully considers the dynamic changes in user needs, thereby improving the accuracy of operator package recommendations. It also solves the technical problem of low recommendation accuracy caused by static models that fail to fully consider the dynamic changes in user needs. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0020] Figure 1 This is a flowchart of a model training method according to an embodiment of this application;

[0021] Figure 2This is a schematic diagram of a model training system according to an embodiment of this application;

[0022] Figure 3 This is a training and prediction graph for a multi-output student T process according to an embodiment of this application;

[0023] Figure 4 This is a schematic diagram of a method for recommending operator packages according to an embodiment of this application;

[0024] Figure 5 This is a structural diagram of a model training device according to an embodiment of this application;

[0025] Figure 6 This is a hardware structure block diagram of a computer terminal for a model training method according to an embodiment of this application. Detailed Implementation

[0026] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0027] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0028] In related technologies, traditional carrier package recommendations are often limited by static models or simple rule-based recommendation systems. Therefore, they fail to fully consider the dynamic changes in user needs, resulting in low accuracy in carrier package recommendations. To address this issue, this application provides a related solution, which is detailed below.

[0029] According to an embodiment of this application, a method embodiment for model training is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0030] Figure 1 This is a flowchart of a model training method according to an embodiment of this application, such as... Figure 1 As shown, the method includes the following steps:

[0031] Step S101: Obtain the target object's historical carrier package data and historical voice data. The historical carrier package data includes multi-dimensional information such as attribute information and historical package usage information of the target object, and the historical voice data includes information on the user's demand for carrier package data.

[0032] It should be noted that operator package data includes the service combination subscribed to by the target user by that operator, such as the user's monthly spending, subsidies, package usage, internet browsing behavior, call habits, and SMS usage. Historical voice data includes the user's package usage needs, feedback, and opinions expressed in communications with customer service.

[0033] Step S102: Generate a first row vector corresponding to historical operator package data, perform sentiment analysis on the usage demand information in the historical voice data to obtain a second row vector. The components in the first row vector are different types of recommended packages for the target audience, and the components in the second row vector are the target audience's sentiment preference information for different types of recommended packages.

[0034] In step S102 above, the original package data is cleaned and features are extracted. Missing data is interpolated, padded with zeros, and normalized to obtain the first row vector M = [m] corresponding to the package data based on the user's ID, age, gender, region, monthly consumption, subsidy status, package usage, internet behavior, call habits, and SMS usage. j1 ,m j2 ,..,m jn ], where m j This refers to the number of different package categories recommended to user j, with each category having a value of m. jj This represents the attribute value of the recommended package in category j.

[0035] Sentiment analysis techniques are used to process historical voice data, extracting users' emotional inclinations towards different package types, such as positive, neutral, or negative emotions. These sentiment inclinations are then converted into numerical components, forming a vector (the second row vector), where each component corresponds to a package type and represents the target user's emotional inclination towards that package. Specifically, speech recognition is performed on the voice data from communication with users, converting it into text. Feature extraction and sentiment analysis are then performed on the text, outputting a sentiment inclination matrix E = n*k, where n is the number of users and k is the number of package types. For user i, the sentiment inclination vector can be represented as E... i =[e i1 ,e i2 ,…,e ik ], each e ij This represents the intensity of user i's emotional preference for package category j.

[0036] Step S103: Merge the first row vector and the second row vector to obtain the target matrix, where the rows of the target matrix are used to represent the package category of the recommended package, and the columns are used to represent the features corresponding to the package category.

[0037] Specifically, the extracted first row vector (package data matrix) and second row vector (sentiment feature matrix) are merged into a single matrix called the target matrix. The rows of the matrix represent the package categories of the recommended packages, and the columns represent the features corresponding to that package category, including sentiment tendency, usage frequency, and consumption habits.

[0038] Step S104: Input the target matrix into the generative adversarial network model to obtain the target data output by the generator in the generative adversarial network model that has a similar distribution to the target matrix, and the target probability output by the discriminator in the generative adversarial network model that the target data is real data.

[0039] In step S104, the target matrix obtained in the previous steps is fed into the generator part of the Generative Adversarial Network (GAN) model as input. The generator generates new target data with a similar distribution to the target matrix based on the input data. This generated data aims to simulate potential users' preferences and usage patterns for packages, thereby enriching the model's training dataset. Simultaneously, the discriminator part of the GAN model is used to evaluate the generated data, determining the probability that the generated data is real data. The discriminator's output will be used for subsequent model optimization to improve the realism of the generated data.

[0040] Step S105: Determine the sentiment reward based on the first row vector and the second row vector, and determine the reward function based on the sentiment reward. Determine the target matrix and target data as state vectors, and determine the recommended package for the target object as action vectors. Determine the target function based on the target probability state vector, action vector, and reward function.

[0041] In step S105, the sentiment tendency R is added to the reward function. sentiment (t,a) is used to reward the model based on the user's emotional preference for the package. The stronger the preference, the higher the reward score, and R... sentiment (t,a) can be represented as:

[0042]

[0043] Among them, w i The weight of user i is determined based on the activity level of the subscription plan, w j It is the weight of package category j, which is determined based on the package value.

[0044] Therefore, the reward function R(t, a) is designed as follows:

[0045]

[0046] in, and λ sentiment R is a weighting coefficient used to measure the importance of different factors; user (t, a), R operator (t, a), R market (t, a) represent user satisfaction reward, operator revenue reward, and market adaptability reward, respectively.

[0047] Using the target matrix and target data as the state vector, and the recommended meal plan as the action vector, the target function is constructed using the reward function determined by sentiment-based reward. The target function takes the following form:

[0048] f(θ) = E[R(t,a)] = ∑ t,a P(t,a|θ)*R(t,a)

[0049] Where R(t,a) is the reward function and P(t,a|θ) is the probability that the model will take action a in state t given parameter θ.

[0050] Step S106: Determine the target probability as the optimization parameter to maximize the objective function. Update the optimization parameter. In the discriminator, if the loss function with respect to the target probability satisfies the preset convergence condition, the trained generative adversarial network model is obtained.

[0051] In step S106, the target probability of the target data's authenticity in the discriminator is determined as the optimization parameter. Adjusting the optimization parameter aims to maximize the objective function, thereby improving the effectiveness and personalization of the recommendation strategy. Gradient descent or other optimization algorithms, such as particle swarm optimization, are used to update the optimization parameter until the loss function regarding the target probability satisfies a preset convergence condition, ensuring the model can stably learn and optimize the recommendation strategy. Once the optimization parameter update is complete and the loss function converges, the trained generative adversarial network model is obtained.

[0052] Preferably, a Bayesian optimization algorithm can be used to update the optimization parameters, specifically including the following steps: 1. Sampling: Sample a series of state-action pairs (t, a) based on the current parameter θ. 2. Evaluation: Evaluate the reward function R(t, a) for each pair (t, a). 3. Update: Update the parameter θ based on the evaluation result in order to maximize the objective function f(θ).

[0053] Based on the above steps, the following steps are taken: Historical carrier package data and historical voice data of the target object are obtained. The historical carrier package data includes multi-dimensional information such as attribute information and historical package usage information of the target object, while the historical voice data includes usage demand information related to carrier packages. A first row vector corresponding to the historical carrier package data is generated. Sentiment analysis is performed on the usage demand information in the historical voice data to obtain a second row vector. The components of the first row vector represent different types of recommended packages for the target object, and the components of the second row vector represent the target object's sentiment towards different types of recommended packages. The first and second row vectors are merged to obtain a target matrix. The rows of the target matrix represent the package categories of the recommended packages, and the columns represent the features corresponding to the package categories. The target matrix is ​​then input into a generative adversarial network (GAN) model to obtain a generative adversarial network (GAN) model. The generator in the network model outputs target data with a similar distribution to the target matrix, and the discriminator in the generative adversarial network model outputs the target data as the probability that the target data is real data. Based on the first and second row vectors, a sentiment reward is determined, and a reward function is determined accordingly. The target matrix and target data are defined as state vectors, and the recommended package for the target object is defined as action vectors. Based on the target probability state vector, action vector, and reward function, the objective function is determined. The target probability is defined as an optimization parameter to maximize the objective function. The optimization parameter is updated. In the discriminator, if the loss function with respect to the target probability satisfies the preset convergence condition, the trained generative adversarial network model is obtained. This approach fully considers the dynamic changes in user needs, thereby improving the technical effect of recommending operator packages.

[0054] The following are Figure 1The steps shown are illustrated and explained by way of example.

[0055] According to some optional embodiments of this application, the objective function can be determined based on the target probability state vector, action vector, and reward function, and can be achieved by the following method: determining several pairs of state vectors and action vectors under the current optimization parameters as observations to obtain an observation matrix including multiple observations; determining the negative log-marginal likelihood function based on the target probability, observation matrix, and reward function, and using the negative log-marginal likelihood function as the objective function.

[0056] In addition, the optimization parameters can be updated as follows: obtain the model parameters in the multi-output student T process, where the model parameters include: degrees of freedom, mean function, covariance function, parameter matrix, and optimization parameters; calculate the derivative of the negative log-marginal likelihood with respect to the model parameters in the multi-output student T process, and update the model parameters in the opposite direction of the derivative.

[0057] Through the above steps, based on users' historical package data, on-site voice data, and personalized sentiment tendencies, the package recommendation strategy can be continuously iterated and optimized through Bayesian optimization and MOSTP models. Ultimately, this aims to improve recommendation accuracy and user satisfaction, thereby achieving a more intelligent and personalized package recommendation service.

[0058] According to some alternative embodiments of this application, the sentiment reward is determined based on the first row vector and the second row vector, which can be achieved by: determining a first weight based on the activity index of the historical operator package used by the target object; determining a second weight based on the package value index of the historical operator package used by the target object; and determining the sentiment reward based on the first weight, the second weight, the first row vector, and the second row vector.

[0059] In the above embodiments, firstly, historical data on the service plans used by the target object are obtained from the operator's database, including but not limited to plan type, monthly spending, and frequency of plan usage. For each historical plan, its activity index is calculated. Activity can be comprehensively evaluated based on factors such as the frequency and duration of user use of the plan, and data consumption. For example, activity index = number of times the user uses the plan / subscription duration.

[0060] Then, based on the activity index of each historical plan, normalization techniques (such as minimum-maximum scaling) are used to map the index values ​​to the [0,1] range to ensure the comparability of weights. The normalized activity index is used as the user's plan preference weight, representing the degree of user preference for high-activity plans, which is the first weight. The weight can be further fine-tuned based on user feedback, plan satisfaction, and other information to reflect actual user preferences.

[0061] Next, a second weight is determined based on the historical carrier package value metrics used by the target user. Package value can be based on the ratio of services provided to costs, or the actual benefits the user receives after using the package (such as savings, additional services, etc.). For each historical package, its value metric is calculated. For example, package value metric = actual user benefits / package cost. Similar to activity metrics, the package value metric is normalized to ensure the reasonableness and comparability of the weights. The normalized package value metric is used as another dimension of package preference weight, reflecting the user's preference for high-value packages; this is the second weight. The second weight can also be dynamically adjusted based on user feedback and market changes.

[0062] The affective reward R is determined using the following formula. sentiment :

[0063]

[0064] Among them, w i It is the weight of user i, w j It is the weight of package category j, e ij It represents the intensity of user i's sentiment towards package category j (elements in the second row vector), m jj This represents the attribute value of the recommended package in category j (the element in the first row of the vector).

[0065] By following the steps above, we can determine the user's preference weights for different plans based on their historical plan usage and sentiment analysis results, and then calculate the sentiment-based reward. This process not only considers objective indicators of user behavior (such as activity level and plan value) but also integrates the user's sentiment, providing the model with more comprehensive and personalized feedback signals, which helps improve the accuracy of plan recommendations and user satisfaction.

[0066] In some optional embodiments of this application, the reward function is determined based on sentiment-based rewards, which can be achieved by: obtaining the target object's satisfaction index with the historical operator packages used by the target object, the operator's revenue index with the historical operator packages used by the target object, and the market adaptability index of the historical operator packages used by the target object; and performing a weighted summation of the sentiment-based rewards, satisfaction index, revenue index, and market adaptability index to obtain the reward function.

[0067] In the above embodiments, firstly, historical data on the target user's (user's) package usage is collected from the operator's user behavior database, including package type, usage patterns (such as data usage, call duration, and SMS volume), monthly spending, and subsidy status. Simultaneously, user feedback after each package change is recorded, such as whether to continue using the same package, whether to upgrade or downgrade, and the frequency of use after the package change. This data will be used to calculate satisfaction indicators. Based on the collected user feedback data, quantitative indicators of satisfaction are defined. For example, satisfaction can be calculated using the ratio of positive reviews in questionnaires or user feedback to the total number of feedback responses, or by analyzing the frequency of use after the package change compared to before the change to assess user acceptance of the new package. Then, the relationship between historical package data and user consumption behavior is analyzed to calculate the revenue the operator obtains from the target user's historical packages. This can include direct revenue (such as monthly package fees) and indirect revenue (such as additional service fees within the package). Simultaneously, market trends in the target user's region, competitors' package strategies, and changes in the target user's package preferences are considered to assess the market adaptability of the package. Secondly, sentiment analysis is performed on communication records with the target user to extract the user's positive or negative sentiment towards different package categories.

[0068] Finally, the reward function is obtained by weighting and summing the affective reward, satisfaction index, profit index, and market adaptability index. The reward function R(t, a) is designed as follows:

[0069]

[0070] in, and λ sentiment R is a weighting coefficient used to measure the importance of different factors; user (t, a), R operator (t, a), R market (t, a), R sentiment These represent user satisfaction rewards, revenue metrics, market adaptability metrics, and sentiment rewards, respectively.

[0071] By following these steps, it is possible to accurately capture user needs, operator revenue targets, market dynamics, and user sentiment, thereby constructing more personalized and effective package recommendation strategies. This not only improves user satisfaction but also enhances the operator's market competitiveness.

[0072] As some optional embodiments of this application, updating the optimization parameters can be achieved by the following method: obtaining several state vector and action vector pairs under the current optimization parameters; calculating the reward function in each state vector and action vector, and updating the optimization parameters according to the calculation results, until the mathematical expectation of the state vector and action vector pair is greater than a preset threshold.

[0073] In the above embodiment, firstly, the parameters in the reinforcement learning model and the Bayesian optimization model are initialized, including the parameter sets of the generator G and the discriminator S, as well as the hyperparameters of the Bayesian optimization model. The state space T and action space A of the reinforcement learning model are defined, where T is the set of state vectors containing user features and preference data, and A is the set of recommended package combinations. Then, based on the current optimization parameters θ, a series of combinations of state vectors and action vectors (t, a) are sampled. This can be done using a stochastic strategy or a model-based strategy, such as an ε-greedy strategy. It is ensured that the state vector t contains the user feature matrix and sentiment matrix obtained from the data preprocessing module. Simultaneously, the action vector a should contain the recommended package combinations under the reinforcement learning strategy. For each sampled state-action pair (t, a), the reward function R(t, a) is calculated, which includes factors such as user satisfaction, operator revenue, and market trends, and specifically incorporates the sentiment reward function Rt. sentiment .

[0074] Finally, the Bayesian optimization algorithm is used to update the parameter θ based on the calculated reward function value, aiming to maximize the expected reward E[R(t,a)]. The above steps are repeated, continuously sampling new state-action pairs and calculating rewards, until the expected value of the state vector and action vector pair exceeds a preset threshold or a predetermined number of iterations is reached. In other words, a preset threshold for the expected value is set; when the expected value E[R(t,a)] of the state-action pair calculated by the reinforcement learning model or the Bayesian optimization algorithm is greater than this threshold, the optimization process stops.

[0075] After the optimization process, the optimal package recommendation strategy is output. This strategy can provide the best package combination based on user characteristics and sentiment preferences to improve user satisfaction and operator revenue. For example, in one iteration, the state vector t represents a user's historical package usage and sentiment data, and the action vector a represents the recommended package combination. Under the current optimization parameter θ, the sampled state-action pair (t1, a1) is calculated, and the reward function R(t1, a1) is 80 points, which is lower than the preset threshold of 85 points. Therefore, the Bayesian optimization algorithm is used to update θ, the next state-action pair (t2, a2) is sampled, and the reward function is calculated again. This process is iterated until E[R(t, a)] exceeds the threshold, indicating that the model has found a recommendation strategy that can meet user satisfaction.

[0076] Through the above implementation steps, combined with Bayesian optimization algorithms based on reinforcement learning and generative adversarial networks, the package recommendation strategy can be dynamically adjusted to improve the personalization level and efficiency of recommendations, ultimately achieving accurate satisfaction of user needs and increased operator revenue.

[0077] In some optional embodiments of this application, generating the first row vector corresponding to historical operator package data can be achieved by the following method: cleaning the historical operator package data and interpolating and padding missing data in the cleaned historical operator package data to obtain initial data with multiple dimensions, wherein the multiple dimensions include: target object identification information, age information, gender information, region information, package subsidy information, package usage information, call information outside the package, SMS information outside the package, and data traffic information outside the package; extracting features from the initial data with multiple dimensions to obtain multiple initial features; normalizing the multiple initial features to obtain multiple first features; and generating different types of recommended packages for the target object based on the multiple first features to obtain the first row vector.

[0078] In the above embodiments, firstly, the collected historical operator package data is subjected to quality checks, removing or correcting records with obvious errors or anomalies. This includes checking whether the data format is correct, whether there are logical errors (such as negative age), and whether the data is complete. Next, missing values ​​in the data are identified. For example, in telecommunications package data, some users have not provided complete personal information, or there are unrecorded consumption behaviors in the usage data. For missing data, an interpolation zero-padding strategy is used. Specifically, if a feature (such as call time outside the package) is missing in some user records, then this part of the data will be filled with zeros. It should be noted that zero-padding is reasonable in telecommunications data because missing data can be interpreted as unused or zero consumption of a specific service, which is actually a usage pattern. For example, if user data lacks out-of-package data usage, then it can be assumed that the user did not use out-of-package data during the statistical period, so the value of that field is set to zero. Starting from the cleaned data, an initial dataset containing multiple dimensions is constructed. These dimensions include, but are not limited to: target object identification information (such as user ID), age information, gender information, region information, package subsidy information, package usage information, call information outside the package, SMS information outside the package, and data usage information outside the package. For each dimension, meaningful features are extracted from the data. For example, age information can be converted into age groups, region information can be encoded into region categories, and package subsidy and usage information can be converted into consumption behavior features. The purpose of feature extraction is to transform the raw data into a form that the model can understand and process, while also reducing the dimensionality of the data and reducing computational complexity. After data cleaning, missing data handling, feature extraction, and normalization, several first features are obtained. These features will be organized into a vector, namely the first row vector. This vector will be used as input to the GAN network model to generate personalized package recommendations.

[0079] By following the steps above, structured, complete, and standardized user feature vectors can be extracted from the original historical operator package data, laying a solid foundation for subsequent model training and personalized package recommendations. This data preprocessing method not only ensures data quality but also supports the training of complex models, thereby achieving more accurate and personalized package recommendation services.

[0080] As some alternative embodiments of this application, sentiment analysis of usage demand information in historical voice data to obtain a second row vector can be achieved by the following method: converting historical voice data into text data, performing feature extraction processing on the text data to obtain a text feature vector, wherein the feature extraction processing includes at least one of the following: word frequency statistics, syntactic analysis, and topic modeling; using a pre-trained sentiment analysis model to perform sentiment tendency analysis on the text feature vector to obtain a quantitative score of the target object's sentiment tendency towards different types of recommended packages; converting the quantitative score into vector form to obtain the second row vector.

[0081] In the above embodiments, speech recognition technology, such as Automatic Speech Recognition (ASR) models in deep learning, is used to convert the collected historical speech data of the target object into text data. The audio signal is converted into a computer-processable digital signal, and the ASR model is used to recognize the language content and output the corresponding text description. Feature extraction is performed on the converted text data to extract key information related to the package recommendation.

[0082] Feature extraction can include at least one of the following processing methods: Statistically analyzing the frequency of keywords related to package usage, demand expression, and preference description in the text. Word frequency information helps understand users' attention to specific package features. Utilizing Natural Language Processing (NLP) techniques to analyze the text's grammatical rules and sentence structure helps identify users' evaluative tone, affirmative or negative expressions regarding the package. Syntactic analysis helps understand users' sentiment tendencies and attitudes. Applying topic models, such as Latent Dirichlet Allocation (LDA) or Nonnegative Matrix Factorization (NMF), allows for topic analysis of the text, identifying common topics discussed, such as package value, service experience, and network quality. Topic modeling helps summarize broad text content into a few key topics, facilitating subsequent sentiment analysis.

[0083] Then, based on the results of word frequency statistics, syntactic analysis, and topic modeling, a text feature vector is constructed. Each component in this vector represents a feature (such as keyword frequency, topic weight, etc.), and its value indicates the importance of that feature in the text data. The vector can be constructed using the bag-of-words model, TF-IDF (term frequency-inverse document frequency), or other advanced representation methods (such as word2vec, BERT, etc.). A pre-trained sentiment analysis model, such as a deep learning-based model (such as LSTM, GRU, Transformer, etc.), is used to perform sentiment analysis on the text feature vector. The pre-trained model should be trained on a large amount of sentiment-annotated text data and be able to identify and quantify the sentiment tendency in the text, including positive, negative, or neutral emotions. The sentiment analysis model will output a quantitative score of the target audience's sentiment tendency towards different types of recommended packages. The score range can be from -1 to 1, where -1 represents strong negative emotion, 0 represents neutral, and 1 represents strong positive emotion.

[0084] Finally, the quantitative score of sentiment preference is transformed into a vector form, namely the second-row vector. The number of rows in this vector equals the number of target objects, and the number of columns equals the number of package types. Each element in the vector represents the intensity of a specific user's sentiment preference for a particular package type.

[0085] Through the above steps, historical voice data can be converted into text, key features can be extracted using text processing techniques, and then a pre-trained sentiment analysis model can be used to quantify users' sentiment towards different types of recommended packages, ultimately constructing a sentiment vector. This process not only considers users' actual usage data but also incorporates information about users' emotions and attitudes, providing a more comprehensive data foundation for subsequent personalized recommendation services.

[0086] In some optional embodiments, a method for recommending operator packages is also provided. This method includes the following steps: acquiring operator package data and voice data of a target object, wherein the operator package data is multi-dimensional information including attribute information and package usage information of the target object, and the voice data includes usage demand information for the operator package data; inputting the vectorized operator package data and voice data into a trained generative adversarial network (GAN) model to obtain a recommended package for the target object output by the trained GAN model, wherein the GAN model is obtained by training using the above-described model training method.

[0087] In the above embodiments, firstly, detailed usage records of the target user's service plan are extracted from the telecom operator's database, including but not limited to user ID, age, gender, region, historical plan type, monthly spending limit, data usage, voice call minutes, and number of SMS messages. Recordings of conversations between the target user and the operator's customer service are then acquired; these recordings contain the user's specific needs for the service plan, feedback on their user experience, and suggestions for improving the existing plan. The collected plan data and voice data are preprocessed, including noise removal, missing value imputation, and outlier detection, to ensure data quality.

[0088] Then, speech recognition technology is used to convert the speech data into text format, and features are extracted from the textualized user demand information, including word frequency statistics, grammatical analysis, topic modeling, etc., to identify key points of concern to users, such as data traffic needs and price sensitivity of packages.

[0089] The collected operator package data is then converted into numerical vectors, where each dimension represents a specific attribute, such as age or spending amount. Appropriate encoding methods, such as one-hot encoding, label encoding, or normalization, are used to ensure data format compatibility with the model. A pre-trained sentiment analysis model is used to perform sentiment analysis on the converted text data, outputting users' positive or negative sentiment ratings for the package services. The sentiment analysis results are then converted into vector format, with each user corresponding to a sentiment vector, where each dimension represents the intensity of sentiment for a specific package type.

[0090] The vector data obtained in the above steps are integrated to form a comprehensive vector containing user attribute information, package usage information, and sentiment bias, which serves as the input to the Generative Adversarial Network (GAN) model. The trained GAN model is then invoked, and the integrated vector is fed into its generator. The generator outputs one or more recommended package sequences for the target audience, including recommendations for different package combinations. These generated sequences are validated by the GAN model's discriminator to ensure their consistency with real user data, thereby improving the accuracy and authenticity of personalized recommendations. The final personalized recommended packages are then presented to the target audience in a clear and easy-to-understand format, such as via SMS, email, or app push notifications, ensuring that users can easily comprehend the recommended content.

[0091] Through the specific implementation steps described above, this application can combine the target user's operator package data and voice demand information, and use a generative adversarial network model to output highly personalized recommended packages, effectively improving user satisfaction and the market competitiveness of telecom operators. This process fully considers the diversity of user needs and incorporates sentiment analysis, making the recommendations more humanized and accurate.

[0092] Figure 2 This is a schematic diagram of a model training system according to an embodiment of this application, such as... Figure 2 As shown, the system includes the following modules:

[0093] The data preprocessing module is used to collect user package usage data and on-site voice data, perform cleaning, feature extraction and sentiment analysis, and construct user feature matrix and sentiment tendency matrix.

[0094] A GAN network model consists of a generator G and a discriminator S. The generator G receives a random vector z as input and outputs simulated user data; the discriminator S receives real user data and the generator's output data and outputs the true probability of the data.

[0095] Reinforcement learning and Bayesian models are used, including the design of the state space, action space, and reward function. A reinforcement learning-based Bayesian optimization model is used to maximize the objective function constructed using the reward function based on user voice data and generated preference data, thereby obtaining the optimal package recommendation strategy. The reward function design considers the correlation between user satisfaction, operator revenue, and package availability, and also incorporates the user's sentiment tendency function into the reward function. A reward function R is designed to guide the model's training and optimization.

[0096] The model optimization module adjusts the parameters of the generator G and discriminator D, enabling the generator G to generate simulated data that matches user preferences, while the discriminator D accurately distinguishes between real and simulated data. The reinforcement learning model learns the reward function R to optimize the package recommendation strategy.

[0097] The following is combined Figure 2 The model training method proposed in this application is described in detail.

[0098] Step S201 involves collecting and preprocessing the model input data. The input data includes user data on their data plan usage and on-site voice data.

[0099] Specifically, the original package data is cleaned and its features are extracted. Missing data is interpolated, padded with zeros, and normalized to obtain a package data matrix M = [m...]. j1 ,m j2 ,..,m jn ], where m j This refers to the number of different package categories recommended to user j, with each category having a value of m. jjThis represents the attribute value of the recommended package in category j. Speech recognition is performed on the on-site communication voice data with users, which is then converted into text. Feature extraction and sentiment analysis are performed on the text, outputting a sentiment tendency matrix E = n*k, where n is the number of users and k is the number of package categories. For user i, the sentiment tendency vector can be represented as E... j =[e i1 ,e i2 ,…,e ik ], each e jj This represents the intensity of user i's emotional preference for package category j.

[0100] The extracted package data matrix and sentiment feature matrix are combined to obtain input data D. The rows in D represent the user's package data, and the columns represent a certain feature in the package data. The final input data D is obtained.

[0101] Step S202: Construct and train the GAN network model.

[0102] The user preference model is modeled by constructing a Generative Adversarial Network (GAN). The GAN network consists of two parts: a generator G and a discriminator S. The generator G is responsible for generating virtual user data that is similar to real user data, while the discriminator S is responsible for judging and evaluating the similarity between the generated data and the real data.

[0103] Specifically, the generator in the network structure that constructs the generator G can be represented as G = G(Z; θ G ), where z represents a random vector z and data x drawn from real data that follows a Gaussian distribution, θ G Let Z be the parameter set of the generator. Using set Z as the input to the generator, in this scheme, the generator employs three fully connected layers and one output layer, i.e., G(z) = tanh(W3*ReLU(W2*ReLU(W1*z))), where W1, W2, and W3 are weight matrices, and θ... G The generator G consists of a weight matrix, with ReLU and tanh as activation functions to introduce nonlinear characteristics. The output of the generator G is a vector, and after passing through the tanh activation function, its value is restricted to the range [-1, 1]. The loss function of the generator can optionally be LG = Et. x [MSE(x,x G )]+E_z[MSE(D(G(z)),1)], where MSE(x,x G ) is the mean squared error loss function, used to measure the loss of generated data x. G The difference between the generated data and the real data x; MSE(D(G(z)),1) is the mean squared error loss function, which measures the discriminator D's performance on the generated data x. GThe difference between the evaluation and the evaluation of real data x. Output virtual user data G(x) G )=x G x G Same as the D dimension. After inputting the training data, the generator's parameters θ are optimized by minimizing the loss function LG. G Optional loss function minimization algorithms include gradient descent, particle swarm optimization, and other swarm intelligence optimization algorithms.

[0104] In constructing the network structure of the discriminator S, the real data D and the generator's output data G(x) are used. G Given input data X, the discriminator outputs the probability P that the input data is genuine. This probability can be expressed as S(X) = 1 / (1 + exp(-s(W4*x))), where W4 is the weight matrix and X is the output x of the generator G. G Or, take the real data D. By designing the loss function LS of S, LS = -E[log(S(S)) + log(1-S(G(z)))], we can perform a minimization optimization on the loss function to obtain the output data of S.

[0105] Training a generative adversarial network (GAN) model involves alternately training the generator G and the discriminator S, minimizing the loss function LG of the generator and the loss function LS of the discriminator. After inputting training data, the model parameters are optimized by minimizing the loss function LG. Optional algorithms for minimizing the loss function include gradient descent, particle swarm optimization, and other swarm intelligence optimization algorithms.

[0106] Step S203, design of reinforcement learning and Bayesian optimization model.

[0107] First, define the state space T (containing the processed data D and the generated data x). G T can be represented as T = {t1, t2, ..., t} n}, where t i Let be the i-th state vector, and n be the total number of states; the action space A (recommended package combination) can be represented as A = {a1, a2, ..., a...}. m}, where a j This is the j-th action, and m is the total number of actions;

[0108] The design of the reward function R(t, a) can affect the model's output, and thus the probability of user acceptance after the package recommendation. In this embodiment, the reward function includes, but is not limited to, user satisfaction, operator revenue, package combinations, market trends, and package survival time; specifically, a sentiment-based reward function R is incorporated into the reward function. sentiment (t,a) is used to reward the model based on the user's emotional preference for the package. The stronger the preference, the higher the reward score, and R... sentiment(t,a) can be represented as:

[0109]

[0110] Among them, w i The weight of user i is determined based on the activity level of the subscription plan, w j It is the weight of package category j, which is determined based on the package value.

[0111] Therefore, the reward function R(t, a) is designed as follows:

[0112]

[0113] in, and λ sentiment R is a weighting coefficient used to measure the importance of different factors; user (t, a), R operator (t, a), R market (t, a) represent user satisfaction reward, operator revenue reward, and market adaptability reward, respectively.

[0114] Step S204: Model optimization and parameter adjustment.

[0115] Further construct a Bayesian optimization algorithm model, combining the real data D and the generated data G(x) G The reward function matrix R is used as input data, where the real data D and the generated data G(x) are used as input data. G Let f be the training input, and the reward function matrix R be the training output. The design space T' is the randomized state space T. The Bayesian optimization model can effectively find the global optimum of the black-box objective function f within an invariant design space. By updating prior knowledge of the objective function and obtaining corresponding observations, it updates the posterior distribution to be closer to f, thus quickly finding the optimal solution. This can be represented as:

[0116] In this Bayesian optimization model, the optimal solution is found by maximizing the objective function f, which is designed as follows:

[0117] f(θ) = E[R(t,a)] = ∑ t,a P(t,a|θ)*R(t,a)

[0118] Where E[] represents the expectation operation, R(t,a) is the reward function, and P(t,a|θ) is the probability of taking action a in state t given the parameter θ, which is the hyperparameter of the Bayesian optimization algorithm.

[0119] The Bayesian optimization algorithm performs the following steps: 1. Sampling: Sample a series of state-action pairs (t, a) based on the current parameter θ. 2. Evaluation: Evaluate the reward function R(t, a) for each pair (t, a). 3. Update: Update the parameter θ based on the evaluation results to maximize the objective function f(θ).

[0120] Under an optional Bayesian optimization model, a multi-output student T-process model is used to approximate the objective function value.

[0121] The Multi-Output Student's T-Process (MOSTP) is an extension of the Single-Output Student's T-Process. By definition, if f is a multivariate Student's T-Process on X with parameter degrees of freedom v > 2, and the vector-valued mean function u satisfies... Covariance function (also known as kernel function) k: satisfies and the semi-definite parameter matrix Ω∈R d×d Therefore, we can define that any finite set of vector-valued variables has a student T-distribution with a joint matrix variable:

[0122]

[0123] Where f,u∈R d Let be a row vector, whose components are functions. M∈R n×d M ij =μ j (x i ), Σ∈R n×n , where Σ ij =k(x i ,x j ).

[0124] MOSTP can be represented as: f ~ MOSTP(v,u,k,Ω). Based on the definition of a multivariate student-T process, the MOSTP model is as follows:

[0125] Given n pairs of observations x i ∈R p ,y i ∈R 1×d Assume:

[0126] f~MOSTP(v,u,k′,Ω),v>2

[0127] y i =f(x) i ), i = 1, ..., n

[0128] Where v is the degree of freedom of MOSTP, its predicted distribution can be obtained by the following formula:

[0129]

[0130] in:

[0131]

[0132]

[0133] Based on the expected value and covariance of the student T-distribution matrix variable, the predicted mean and covariance are given by the following formula:

[0134]

[0135]

[0136] In the MOSTP model, the observations follow a Student's T-distribution of matrix variables. The negative log-marginal likelihood function is:

[0137]

[0138] Therefore, the MOSTP parameters include an additional parameter—the degrees of freedom ν. The negative logarithmic marginal likelihood is relative to the parameter ν. θ i ,φ ij , The derivative is as follows:

[0139]

[0140] Among them, U=K′+YΩ -1 Y T τ=v+n-1, ψn() is a function lnΓ n (·) The derivative with respect to ν.

[0141] Given training samples, MOSTP is trained. Then, given arbitrary test samples, MOSTP outputs the predicted mean and variance of the multiple outputs corresponding to the test samples. The input samples are matrix variables, and the output values ​​are matrix variables. Finally, the trained model is optimized by adjusting the reward function parameters to output the optimal result.

[0142] Figure 3 This is a training and prediction graph for a multi-output student T process according to an embodiment of this application, such as... Figure 3As shown, the process includes two stages: training and prediction. Specifically, features extracted from user historical package data and on-site voice data are preprocessed and represented as matrices or vectors. These are then fed into the MOSTP model for processing. This model can capture and learn the complex relationships between input and output data, especially when dealing with high-dimensional and multi-output scenarios. Through model training, its internal parameters (such as parameter degrees of freedom ν, covariance matrix, etc.) are adjusted to optimize the accuracy and reliability of predictions. After processing by the MOSTP model, it not only outputs predicted values ​​but also provides an estimate of the uncertainty of the prediction results, i.e., the prediction variance. The prediction results are also output in matrix or vector form and used in subsequent reinforcement learning and Bayesian optimization algorithms to build a more accurate package recommendation strategy.

[0143] Figure 4 This is a schematic diagram of a method for recommending operator packages according to an embodiment of this application, such as... Figure 4 As shown, the process first extracts historical user plan data for the target audience from the telecom operator's database. This data includes, but is not limited to, user ID, age, gender, region, historical plan type, monthly spending limit, data usage, voice call minutes, and SMS message count. It also includes on-site voice data from users, including but not limited to specific service needs, user experience feedback, and suggestions for improving existing plans. The collected plan data and voice data undergo preprocessing, including noise removal, missing value imputation, and outlier detection, to ensure data quality.

[0144] Then, speech recognition technology is used to convert the speech data into text format. Feature extraction is performed on the textualized user needs information, including word frequency statistics, syntactic analysis, and topic modeling, to identify key user concerns, such as data usage needs and price sensitivity. The collected operator package data is then converted into numerical vectors, where each dimension represents a specific attribute, such as age or spending amount. Appropriate encoding methods, such as one-hot encoding, label encoding, or normalization, are used to ensure data format compatibility with the model. A pre-trained sentiment analysis model is used to perform sentiment analysis on the converted text data, outputting positive or negative sentiment ratings for user service packages. The sentiment analysis results are then converted into vector format, with each user corresponding to a sentiment vector, where each dimension represents the intensity of sentiment for a specific package type.

[0145] The vector data obtained in the above steps is input into the operator's package recommendation system for integration, forming a comprehensive vector containing user attribute information, package usage information, and sentiment bias. This comprehensive vector serves as the input to the generative adversarial network (GAN) model. The trained GAN model is then invoked, and the integrated vector is fed into its generator. The generator outputs one or more recommended package sequences for the target user, including recommendations for different package combinations. The generated package sequences are validated by the GAN model's discriminator to ensure their consistency with real user data, thereby improving the accuracy and authenticity of personalized recommendations.

[0146] The final personalized recommendation package will be presented to the target audience in a clear and easy-to-understand format, including the recommended package, the user's likely acceptance level, and the recommendation script template, to ensure that the user can easily understand the recommended content.

[0147] Figure 5 This is a structural diagram of a model training device according to an embodiment of this application, such as... Figure 5 As shown, the device includes:

[0148] The acquisition module 51 is used to acquire the target object's historical operator package data and historical voice data. The historical operator package data includes multi-dimensional information such as attribute information and historical package usage information of the target object, and the historical voice data includes information on the user's demand for operator package data.

[0149] The generation module 52 is used to generate a first row vector corresponding to historical operator package data, perform sentiment analysis on the usage demand information in the historical voice data, and obtain a second row vector. The components in the first row vector are different types of recommended packages for the target object, and the components in the second row vector are the target object's sentiment tendency information for different types of recommended packages.

[0150] The merging module 53 is used to merge the first row vector and the second row vector to obtain the target matrix. The rows of the target matrix are used to represent the package category of the recommended package, and the columns are used to represent the features corresponding to the package category.

[0151] Input module 54 is used to input the target matrix into the generative adversarial network model to obtain the target data with a similar distribution to the target matrix output by the generator in the generative adversarial network model, and the target probability that the target data output by the discriminator in the generative adversarial network model is real data.

[0152] The determination module 55 is used to determine the sentiment reward based on the first row vector and the second row vector, and to determine the reward function based on the sentiment reward. The target matrix and target data are determined as state vectors, the recommended package for the target object is determined as action vectors, and the target function is determined based on the target probability state vector, action vector and reward function.

[0153] The update module 56 is used to determine the target probability as the optimization parameter to maximize the objective function and update the optimization parameter. In the discriminator, if the loss function with respect to the target probability satisfies the preset convergence condition, the trained generative adversarial network model is obtained.

[0154] Optionally, the determination module 55 is also used to perform the following steps: determining several pairs of state vectors and action vectors under the current optimization parameters as observations, and obtaining an observation matrix including multiple observations; determining the negative log-marginal likelihood function based on the target probability, the observation matrix and the reward function, and using the negative log-marginal likelihood function as the target function.

[0155] Optionally, the update module 56 is also configured to perform the following steps: obtain the model parameters in the multi-output student T process, wherein the model parameters include: degrees of freedom, mean function, covariance function, parameter matrix and optimization parameters; calculate the derivative of the negative log marginal likelihood with respect to the model parameters in the multi-output student T process, and update the model parameters in the opposite direction of the derivative.

[0156] Optionally, the generation module 52 is also used to perform the following steps: determining a first weight based on the activity index of the historical operator packages used by the target object; determining a second weight based on the package value index of the historical operator packages used by the target object; and determining a sentiment reward based on the first weight, the second weight, the first row vector, and the second row vector.

[0157] Optionally, the merging module 53 is also used to perform the following steps: obtaining the target object's satisfaction index with the historical operator package used by the target object, the operator's revenue index with the historical operator package used by the target object, and the market adaptability index of the historical operator package used by the target object; and performing a weighted summation of the sentiment reward, satisfaction index, revenue index, and market adaptability index to obtain a reward function.

[0158] Optionally, the update module 56 is also used to perform the following steps: obtain several state vector and action vector pairs under the current optimization parameters; calculate the reward function in each state vector and action vector, and update the optimization parameters according to the calculation results, until the mathematical expectation of the state vector and action vector pair is greater than a preset threshold.

[0159] Optionally, the generation module 52 is further configured to perform the following steps: clean the historical operator package data, and perform interpolation and zero-padding on the missing data in the cleaned historical operator package data to obtain initial data with multiple dimensions, wherein the multiple dimensions include: target object identification information, age information, gender information, region information, package subsidy information, package usage information, call information outside the package, SMS information outside the package, and data traffic information outside the package; extract features from the initial data with multiple dimensions to obtain multiple initial features, normalize the multiple initial features to obtain multiple first features; and generate different types of recommended packages for the target object based on the multiple first features to obtain a first row vector.

[0160] Optionally, the generation module 52 is further configured to perform the following steps: converting historical speech data into text data, performing feature extraction processing on the text data to obtain a text feature vector, wherein the feature extraction processing includes at least one of the following: word frequency statistics, syntactic analysis, and topic modeling; performing sentiment analysis on the text feature vector using a pre-trained sentiment analysis model to obtain a quantitative score of the target object's sentiment towards different types of recommended packages; and converting the quantitative score into a vector form to obtain a second row vector.

[0161] The model training device is also used to perform the following steps: acquiring the operator package data and voice data of the target object, wherein the operator package data includes multi-dimensional information including attribute information and package usage information of the target object, and the voice data includes usage demand information of the operator package data; inputting the vectorized operator package data and voice data into the trained generative adversarial network model to obtain the recommended package for the target object output by the trained generative adversarial network model, wherein the generative adversarial network model is obtained by training using the above model training method.

[0162] It should be noted that the above Figure 5 The modules in the above can be program modules (e.g., a set of program instructions that implement a specific function) or hardware modules. For the latter, they can be represented in the following forms, but are not limited to these: each of the above modules is represented by a processor, or the functions of each of the above modules are implemented by a processor.

[0163] It should be noted that, Figure 5 Preferred embodiments of the shown examples can be found in [reference needed]. Figure 1 The relevant descriptions of the embodiments shown will not be repeated here.

[0164] Figure 6 A hardware block diagram of a computer terminal for implementing a model training method is shown. Figure 6As shown, the computer terminal 60 may include one or more processors 602 (shown as 602a, 602b, ..., 602n in the figure) 602 (processor 602 may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 604 for storing data, and a transmission module 606 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of a BUS bus), a network interface, a power supply, and / or a camera. Those skilled in the art will understand that... Figure 6 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, the computer terminal 60 may also include... Figure 6 The more or fewer components shown, or having the same Figure 6 The different configurations shown.

[0165] It should be noted that the aforementioned one or more processors 602 and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element within the computer terminal 60. As involved in the embodiments of this application, the data processing circuits serve as a form of processor control (e.g., selection of a variable resistor termination path connected to an interface).

[0166] The memory 604 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the model training method in this embodiment. The processor 602 executes various functional applications and data processing by running the software programs and modules stored in the memory 604, thereby realizing the above-mentioned model training method. The memory 604 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 604 may further include memory remotely located relative to the processor 602, and these remote memories can be connected to the computer terminal 60 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0167] The transmission module 606 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the computer terminal 60. In one example, the transmission module 606 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission module 606 may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0168] The display may be, for example, a touchscreen liquid crystal display (LCD) that allows the user to interact with the user interface of the computer terminal 60.

[0169] It should be noted here that, in some optional embodiments, the above... Figure 6 The computer terminal shown may include hardware elements (including circuitry), software elements (including computer code stored on a computer-readable medium), or a combination of both hardware and software elements. It should be noted that... Figure 6 This is only one instance of a specific particular instance, and is intended to illustrate the types of components that may exist in the aforementioned computer terminal.

[0170] It should be noted that, Figure 6 The computer terminal shown is used to execute Figure 1 The model training method shown above is also applicable to this electronic device, and will not be repeated here.

[0171] This application also provides a non-volatile storage medium, which includes a stored program, wherein the program, when running, controls the device where the storage medium is located to execute the above model training method.

[0172] A program on a non-volatile storage medium performs the following functions: It acquires historical carrier package data and historical voice data of a target object. The historical carrier package data includes multi-dimensional information such as attribute information and historical package usage information about the target object, while the historical voice data includes usage demand information related to the carrier package data. It generates a first row vector corresponding to the historical carrier package data, performs sentiment analysis on the usage demand information in the historical voice data to obtain a second row vector. The components of the first row vector represent different types of recommended packages for the target object, and the components of the second row vector represent the target object's sentiment towards different types of recommended packages. It merges the first and second row vectors to obtain a target matrix, where the rows of the target matrix represent the package category of the recommended packages, and the columns represent the package categories. The corresponding features are as follows: The target matrix is ​​input into the Generative Adversarial Network (GAN) model to obtain target data with a similar distribution to the target matrix output by the generator in the GAN model, and the target probability that the target data output by the discriminator in the GAN model is real data. Based on the first and second row vectors, the sentiment reward is determined, and based on the sentiment reward, the reward function is determined. The target matrix and target data are determined as state vectors, and the recommended package for the target object is determined as action vectors. Based on the target probability state vector, action vector, and reward function, the objective function is determined. The target probability is determined as the optimization parameter to maximize the objective function. The optimization parameter is updated. In the discriminator, if the loss function with respect to the target probability satisfies the preset convergence condition, the trained GAN model is obtained.

[0173] This application also provides an electronic device, including a memory and a processor, wherein the processor is used to run a program stored in the memory, wherein the program executes the above-described model training method during runtime.

[0174] The processor runs a program that performs the following functions: It acquires historical carrier package data and historical voice data of the target object, wherein the historical carrier package data includes multi-dimensional information such as attribute information and historical package usage information of the target object, and the historical voice data includes usage demand information related to the carrier package data; it generates a first row vector corresponding to the historical carrier package data; it performs sentiment analysis on the usage demand information in the historical voice data to obtain a second row vector, wherein the components of the first row vector represent different types of recommended packages for the target object, and the components of the second row vector represent the target object's sentiment tendency information regarding different types of recommended packages; it merges the first and second row vectors to obtain a target matrix, wherein the rows of the target matrix represent the package category of the recommended packages, and the columns represent the package categories. The corresponding features are as follows: The target matrix is ​​input into the Generative Adversarial Network (GAN) model to obtain target data with a similar distribution to the target matrix output by the generator in the GAN model, and the target probability that the target data output by the discriminator in the GAN model is real data. Based on the first and second row vectors, the sentiment reward is determined, and based on the sentiment reward, the reward function is determined. The target matrix and target data are determined as state vectors, and the recommended package for the target object is determined as action vectors. Based on the target probability state vector, action vector, and reward function, the objective function is determined. The target probability is determined as the optimization parameter to maximize the objective function. The optimization parameter is updated. In the discriminator, if the loss function with respect to the target probability satisfies the preset convergence condition, the trained GAN model is obtained.

[0175] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0176] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0177] In the above embodiments of this application, the information collected is information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data all comply with relevant laws, regulations and standards, take necessary protective measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0178] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0179] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0180] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0181] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to related technologies, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0182] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A model training method, characterized in that, include: Obtain historical carrier package data and historical voice data of the target object, wherein the historical carrier package data is multi-dimensional information including attribute information and historical package usage information of the target object, and the historical voice data includes usage demand information of carrier package data; A first row vector corresponding to the historical operator package data is generated. Sentiment analysis is performed on the usage demand information in the historical voice data to obtain a second row vector. The components in the first row vector are different types of recommended packages for the target object, and the components in the second row vector are the target object's sentiment tendency information for different types of recommended packages. Merge the first row vector and the second row vector to obtain the target matrix, wherein the rows of the target matrix are used to represent the package category of the recommended package, and the columns are used to represent the features corresponding to the package category; The target matrix is ​​input into the generative adversarial network model to obtain target data with a similar distribution to the target matrix output by the generator in the generative adversarial network model, and the target probability that the target data is real data output by the discriminator in the generative adversarial network model. Based on the first row vector and the second row vector, a sentiment reward is determined, and based on the sentiment reward, a reward function is determined. The target matrix and the target data are determined as a state vector, and the recommended package for the target object is determined as an action vector. Based on the target probability, the state vector, the action vector, and the reward function, a target function is determined. The target probability is determined as the optimization parameter to maximize the objective function. The optimization parameter is updated. In the discriminator, if the loss function with respect to the target probability satisfies the preset convergence condition, the trained generative adversarial network model is obtained.

2. The method according to claim 1, characterized in that, Based on the target probability, the state vector, the action vector, and the reward function, the target function is determined, including: Several state vectors and action vectors under the current optimization parameters are determined as observations, resulting in an observation matrix containing multiple observations; based on the target probability, the observation matrix, and the reward function, a negative log-marginal likelihood function is determined, and the negative log-marginal likelihood function is used as the target function. Update the optimization parameters, including: The model parameters for obtaining the multi-output student T process include: degrees of freedom, mean function, covariance function, parameter matrix, and optimization parameters. Calculate the derivative of the negative logarithmic marginal likelihood with respect to the model parameters in the multi-output student T process, and update the model parameters in the opposite direction of the derivative.

3. The method according to claim 1, characterized in that, Based on the first row vector and the second row vector, determine the sentiment reward, including: The first weight is determined based on the activity index of the historical operator packages used by the target object; The second weight is determined based on the package value index of the historical operator packages used by the target object; The sentiment reward is determined based on the first weight, the second weight, the first row vector, and the second row vector.

4. The method according to claim 1 or 3, characterized in that, Based on the stated emotional tendency reward, determine the reward function, including: Obtain the target object's satisfaction index with the historical operator packages used by the target object, the operator's revenue index with the historical operator packages used by the target object, and the market adaptability index of the historical operator packages used by the target object; The reward function is obtained by weighting and summing the emotional tendency reward, the satisfaction index, the revenue index, and the market adaptability index.

5. The method according to claim 1, characterized in that, Update the optimization parameters, including: Obtain several state vector and action vector pairs under the current optimization parameters; calculate the reward function in each state vector and action vector, and update the optimization parameters according to the calculation results, until the mathematical expectation of the state vector and action vector pair is greater than a preset threshold.

6. The method according to claim 1, characterized in that, Generate the first row vector corresponding to the historical operator package data, including: The historical operator package data is cleaned, and missing data in the cleaned historical operator package data is interpolated and padded with zeros to obtain initial data in multiple dimensions. The multiple dimensions include: the target object's identification information, age information, gender information, region information, package subsidy information, package usage information, call information outside the package, SMS information outside the package, and data traffic information outside the package. Feature extraction is performed on the initial data from multiple dimensions to obtain multiple initial features. The multiple initial features are then normalized to obtain multiple first features. Based on the multiple first features, different types of recommendation packages are generated for the target object to obtain the first row vector.

7. The method according to claim 1, characterized in that, Sentiment analysis is performed on the usage demand information in the historical voice data to obtain a second row vector, including: The historical speech data is converted into text data, and the text data is subjected to feature extraction processing to obtain a text feature vector. The feature extraction processing includes at least one of the following: word frequency statistics, syntactic analysis, and topic modeling. A pre-trained sentiment analysis model is used to perform sentiment analysis on the text feature vectors to obtain a quantitative score of the target object's sentiment towards different types of recommended packages; the quantitative score is then converted into a vector form to obtain the second row vector.

8. A method for recommending operator packages, characterized in that, include: The operator package data and voice data of the target object are obtained. The operator package data includes multi-dimensional information such as attribute information and package usage information of the target object, and the voice data includes usage demand information of the operator package data. The vectorized operator package data and voice data are input into the trained generative adversarial network model to obtain the recommended package for the target object output by the trained generative adversarial network model, wherein the generative adversarial network model is trained by the model training method described in any one of claims 1 to 7.

9. A model training device, characterized in that, include: The acquisition module is used to acquire the historical carrier package data and historical voice data of the target object. The historical carrier package data is multi-dimensional information including attribute information and historical package usage information of the target object, and the historical voice data includes usage demand information for the carrier package data. The generation module is used to generate a first row vector corresponding to the historical operator package data, perform sentiment analysis on the usage demand information in the historical voice data to obtain a second row vector, wherein the components in the first row vector are different types of recommended packages for the target object, and the components in the second row vector are the target object's sentiment tendency information for different types of recommended packages; The merging module is used to merge the first row vector and the second row vector to obtain a target matrix, wherein the rows of the target matrix are used to represent the package category of the recommended package, and the columns are used to represent the features corresponding to the package category; The input module is used to input the target matrix into the generative adversarial network model to obtain target data with a similar distribution to the target matrix output by the generator in the generative adversarial network model, and the target probability that the target data is real data output by the discriminator in the generative adversarial network model. The determination module is used to determine the sentiment reward based on the first row vector and the second row vector, and to determine the reward function based on the sentiment reward. The target matrix and the target data are determined as state vectors, the recommended package for the target object is determined as action vectors, and the target function is determined based on the target probability, the state vector, the action vector, and the reward function. An update module is used to determine the target probability as an optimization parameter to maximize the objective function and update the optimization parameter. In the discriminator, if the loss function with respect to the target probability satisfies a preset convergence condition, a trained generative adversarial network model is obtained.

10. A non-volatile storage medium, characterized in that, The non-volatile storage medium includes a stored program, wherein, when the program is executed, it controls the device where the non-volatile storage medium is located to execute the model training method according to any one of claims 1 to 7 and the operator package recommendation method according to claim 8.

11. An electronic device, characterized in that, include: A memory and a processor, the processor being configured to run a program stored in the memory, wherein the program, when running, executes the model training method according to any one of claims 1 to 7 and the operator package recommendation method according to claim 8.

12. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the model training method according to any one of claims 1 to 7 and the operator package recommendation method according to claim 8.

Images