Automatic generation method and system for AI digital person recommendation strategy

Abstract

The invention relates to the technical field of data processing, and discloses an AI digital person recommendation strategy automatic generation method and system. The method comprises the steps of collecting target market data, extracting a behavior feature vector and constructing a strategy vector, predicting a promotion effect through a deep neural network, generating an optimal strategy vector based on a multi-agent game framework and a genetic algorithm, generating digital human expression content, action and scene configuration according to a strategy coefficient, and cross-market strategy reuse is realized by using transfer learning. According to the method, the culture adaptability and the generation efficiency of the AI digital person recommendation strategy are improved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and system for automatically generating AI digital human recommendation strategies. Background Technology

[0002] As more and more brands utilize AI digital human technology for cross-regional promotion, existing AI digital human recommendation strategies mainly adapt to different markets through simple language translation and content localization. They use preset recommendation strategy templates and standardized digital human images for unified promotion in various target markets. Basic data such as language habits and consumption levels in the target markets are collected manually, and then the promotional content is translated and converted. The digital human image and the content expressed by the digital human are basically consistent, with adjustments only made at the language level. This approach can quickly complete content deployment in multiple markets, but in practical applications, many shortcomings have gradually been exposed.

[0003] The main shortcomings of existing technologies lie in the lack of understanding of the deep cultural characteristics of different markets and the lack of intelligent generation capabilities for promotional strategies. Simple language translation cannot capture the attention preferences, aesthetic preferences, and consumption habits of consumers in the target market, resulting in recommended content that conflicts with local culture or fails to resonate with consumers. Uniform recommendation strategy templates ignore the differences in pricing sensitivity, promotion response, and channel preferences among different markets, leading to unreasonable allocation of promotional resources and low return on investment. Standardized digital avatars and expressions cannot adapt to the diverse needs of consumers in different cultural backgrounds for persona types and content styles. In addition, existing technologies lack dynamic game analysis of market competition and consumer decision-making behavior during the strategy generation process, resulting in a lack of adaptability and optimal guarantee when facing countermeasures from competitors.

[0004] Further analysis reveals that even if behavioral characteristics of the target market can be extracted and promotion strategy vectors can be constructed, establishing a precise mapping relationship between market characteristics and promotion effects remains a key challenge. This is because the impact mechanisms of different behavioral characteristics on conversion rates, ROI, and brand awareness are complex and non-linear, and simple linear models cannot accurately predict promotion effects. Furthermore, given the ability to predict promotion effects, automatically searching for the optimal strategy combination in a multi-party game environment that considers consumer response, strategy costs, and competitive risks requires solving the technical problems of multi-agent game modeling and efficient optimization algorithms. In addition, after obtaining the optimal strategy vector, transforming the abstract strategy coefficients into specific digital human execution plans, including content generation, image type selection, action sequence design, and scene element configuration, requires establishing mapping rules between strategy parameters and execution details. Finally, for multiple markets with similar characteristics, how to quickly reuse the successful experience of existing markets through transfer learning to reduce the strategy optimization costs and time of new markets involves technical challenges such as feature similarity calculation, market clustering, and cross-market knowledge transfer. Summary of the Invention

[0005] This application provides an automatic generation method and system for AI digital human recommendation strategies. It is used to construct a multi-agent game framework and use a genetic algorithm to solve the game objective function that includes consumer response payoffs, strategy costs, and competitive risks. This solves the problem that existing technologies lack intelligent decision-making and dynamic game capabilities in promotion strategy generation. By establishing a deep neural network to map the nonlinear relationship between cultural characteristics and promotion effects and combining it with transfer learning to achieve cross-market strategy reuse, the cultural adaptability and generation efficiency of AI digital human recommendation strategies are improved.

[0006] Firstly, this application provides a method for automatically generating AI digital human recommendation strategies, the method comprising: Step S1: Collect social media data and e-commerce review data from the target market, extract cultural feature vectors, construct a strategy vector containing pricing strategy coefficients, promotion strategy coefficients, content strategy coefficients, channel strategy coefficients, and digital human image strategy coefficients, and concatenate the cultural feature vectors with the strategy vectors to form an input feature matrix; Step S2: Input the input feature matrix into a deep neural network, train the mapping relationship between the cultural feature vector and the promotion effect, and output the promotion effect score vector of expected conversion rate, expected return on investment and expected brand awareness; Step S3: Construct a multi-agent game framework based on the promotion effect score vector, set up consumer agent, competitor agent and strategy agent, solve the game equilibrium of the strategy vector through genetic algorithm, and generate the optimal strategy vector; Step S4: Generate a hierarchical recommendation strategy decision tree based on the pricing strategy coefficient, promotion strategy coefficient, content strategy coefficient, channel strategy coefficient, and digital human image strategy coefficient in the optimal strategy vector; and generate digital human expression content, facial expression and action sequence, and scene element configuration for each consumer segment. Step S5: Calculate the cultural similarity matrix between cultural feature vectors of different target markets, divide similar markets into cultural clusters, and use transfer learning to transfer the digital human expression content, facial expression and action sequence and scene element configuration of the source market to the target market to generate a recommendation strategy.

[0007] Secondly, this application provides an AI digital human recommendation strategy automatic generation system, the AI ​​digital human recommendation strategy automatic generation system comprising: The extraction module is used to collect social media data and e-commerce review data of the target market, extract cultural feature vectors, construct a strategy vector containing pricing strategy coefficients, promotion strategy coefficients, content strategy coefficients, channel strategy coefficients and digital human image strategy coefficients, and concatenate the cultural feature vectors with the strategy vectors to form an input feature matrix; The training module is used to input the input feature matrix into a deep neural network, train the mapping relationship between the cultural feature vector and the promotion effect, and output the promotion effect score vector of expected conversion rate, expected return on investment and expected brand awareness. The generation module is used to construct a multi-agent game framework based on the promotion effect score vector, set up a consumer agent, a competitor agent, and a strategy agent, and solve the game equilibrium of the strategy vector through a genetic algorithm to generate the optimal strategy vector. The configuration module is used to generate a hierarchical recommendation strategy decision tree based on the pricing strategy coefficient, promotion strategy coefficient, content strategy coefficient, channel strategy coefficient and digital human image strategy coefficient in the optimal strategy vector, and to generate digital human expression content, facial expression and action sequence and scene element configuration for different consumer segments. The transfer module is used to calculate the cultural similarity matrix between cultural feature vectors of different target markets, divide similar markets into cultural clusters, and use transfer learning to transfer the digital human expression content, facial expression and action sequence and scene element configuration of the source market to the target market to generate a recommendation strategy.

[0008] Thirdly, an AI digital human recommendation strategy automatic generation device is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the AI ​​digital human recommendation strategy automatic generation device to execute the aforementioned AI digital human recommendation strategy automatic generation method.

[0009] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to perform the aforementioned method for automatically generating AI digital human recommendation strategies.

[0010] The technical solution provided in this application extracts cultural feature vectors by collecting social media data and e-commerce review data from the target market, and constructs a strategy vector containing pricing strategy coefficients, promotion strategy coefficients, content strategy coefficients, channel strategy coefficients, and digital human image strategy coefficients. The cultural feature vectors are concatenated with the strategy vectors to form an input feature matrix. This solves the problem in existing technologies where promotion strategy generation relies solely on surface-level language translation while neglecting deeper cultural connotations. It achieves a systematic extraction and quantitative expression of multi-dimensional cultural characteristics such as target market attention preferences, consumption habits, and aesthetic preferences. Furthermore, by structurally combining cultural features with promotion strategies, a complete data foundation is established for subsequent intelligent strategy generation. The input feature matrix is ​​then input into a deep neural network to train the mapping relationship between cultural feature vectors and promotion effects, outputting promotion effect score vectors for expected conversion rate, expected return on investment, and expected brand awareness. This overcomes the limitations of existing technologies. To address the shortcomings of existing technologies in accurately predicting promotional effects under different cultural backgrounds, this paper utilizes multi-layer nonlinear transformations of deep neural networks to capture the complex correlation patterns between cultural characteristics and promotional effects, avoiding the prediction errors of simple linear models. A multi-agent game framework is constructed based on promotional effect score vectors, setting up consumer agents, competitor agents, and strategy agents. A genetic algorithm is used to solve the game equilibrium of strategy vectors to generate the optimal strategy vector. This solves the problem of existing technologies lacking dynamic game analysis and intelligent decision-making capabilities in promotional strategy generation. The multi-agent game framework simulates the interaction between consumer decision-making behavior and competitor countermeasures in a real market environment. The genetic algorithm efficiently searches for the optimal solution in a complex strategy space through population evolution, avoiding the subjectivity and limitations of human experience-based decision-making. This ensures that the generated promotional strategy achieves optimal configuration under multi-objective constraints that comprehensively consider consumer response, cost control, and competitive risk.

[0011] Based on the pricing strategy coefficient, promotion strategy coefficient, content strategy coefficient, channel strategy coefficient, and digital human image strategy coefficient in the optimal strategy vector, a hierarchical recommendation strategy decision tree is generated. Digital human expressive content, facial expression and action sequences, and scene element configurations are generated separately for different consumer segments. This solves the problem that the standardized digital human promotion content in existing technologies cannot adapt to the diverse needs of different cultural backgrounds and consumer groups. By transforming abstract strategy coefficients into concrete and executable digital human promotion schemes, an automated process from planning-level positioning to strategy-level content generation is achieved. Differentiated expressive content, action sequences, and scene configurations are generated for different consumer segments, avoiding the compatibility issues of a uniform promotion scheme across different groups. To address the shortcomings, this application calculates the cultural similarity matrix between cultural feature vectors of different target markets, divides similar markets into cultural clusters, and uses transfer learning to transfer the digital human's expressive content, facial expression sequences, and scene element configurations from the source market to the target market to generate recommendation strategies. By calculating cultural similarity and identifying market groups with similar cultural characteristics through market clustering, the application utilizes transfer learning technology to quickly reuse successful experiences from existing markets to new markets. Only a small amount of target market data is needed for fine-tuning to complete strategy adaptation. Overall, this application establishes a complete technical chain from cultural feature extraction, promotion effect prediction, game strategy optimization to specific solution generation and cross-market transfer, realizing the intelligent and automatic generation of AI digital human recommendation strategies. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of an embodiment of the AI ​​digital human recommendation strategy automatic generation method in this application. Figure 2 This is a schematic diagram of one embodiment of the AI ​​digital human recommendation strategy automatic generation system in this application. Figure 3 This is a schematic block diagram of the structure of the AI ​​digital human recommendation strategy automatic generation device in an embodiment of the present invention. Detailed Implementation

[0014] This application provides an AI digital human recommendation strategy automatic generation method and system. The terms first, second, third, fourth, etc. (if present) in the specification, claims, and accompanying drawings 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 described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms include or have, and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes 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 devices.

[0015] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the AI ​​digital human recommendation strategy automatic generation method in this application includes: Step S1: Collect social media data and e-commerce review data from the target market, extract cultural feature vectors, construct a strategy vector that includes pricing strategy coefficients, promotion strategy coefficients, content strategy coefficients, channel strategy coefficients, and digital human image strategy coefficients, and concatenate the cultural feature vectors with the strategy vectors to form an input feature matrix; Step S2: Input the input feature matrix into the deep neural network, train the mapping relationship between cultural feature vectors and promotion effects, and output the promotion effect score vectors of expected conversion rate, expected return on investment and expected brand awareness. Step S3: Construct a multi-agent game framework based on the promotion effect score vector, set up consumer agent, competitor agent and strategy agent, solve the game equilibrium of strategy vector through genetic algorithm, and generate the optimal strategy vector; Step S4: Generate a hierarchical recommendation strategy decision tree based on the pricing strategy coefficient, promotion strategy coefficient, content strategy coefficient, channel strategy coefficient, and digital human image strategy coefficient in the optimal strategy vector; and generate digital human expression content, facial expression and action sequences, and scene element configurations for different consumer segments. Step S5: Calculate the cultural similarity matrix between cultural feature vectors of different target markets, divide similar markets into cultural clusters, and use transfer learning to transfer the digital human expression content, facial expression and action sequences and scene element configurations from the source market to the target market to generate recommendation strategies.

[0016] It is understood that the executing entity of this application can be an AI digital human recommendation strategy automatic generation system, or it can be a terminal or a server; the specific implementation is not limited here. This application's embodiment uses a server as an example for illustration.

[0017] Specifically, raw text data is obtained by collecting social media data, e-commerce review data, and competitor promotion data from the target market through web crawling. This raw text data is then input into a multilingual BERT model for vectorization. The BERT model uses a self-attention mechanism to perform deep semantic encoding on the text, extracting parameters such as attention preference, consumption habits, aesthetic preferences, constraints, and festival customs to generate cultural feature vectors. When constructing strategy vectors, pricing strategy coefficients are set, including high-end positioning coefficients, mid-range positioning coefficients, and low-end positioning coefficients with values ​​ranging from 0 to 1. Promotion strategy coefficients include discount strength coefficients, gift type codes, and limited-time activity periods. Content strategy coefficients include emotional weighting coefficients and rational weighting coefficients. The gender weighting coefficient and channel strategy coefficient include the social media platform weight vector and KOL cooperation level coefficient. The digital human image strategy coefficient includes the professional image coefficient, the affinity feature coefficient, and the fashion feature coefficient. The cultural feature vector and the strategy vector are concatenated column by column to form the input feature matrix. For example, when a skin care brand expands its market, the BERT model identifies the consumer's attention preference parameters for natural ingredients, consumption habit parameters for online shopping, and aesthetic preference parameters for fresh packaging from the text to generate a cultural feature vector. At the same time, a strategy vector is constructed with the high-end positioning coefficient set to 0.7, the emotional weighting coefficient set to 0.6, and the professional image coefficient set to 0.5. The two vectors are concatenated to form the input feature matrix.

[0018] A five-layer fully connected deep neural network is constructed. The number of neurons in the input layer equals the sum of the dimensions of the cultural feature vector and the policy vector. The number of neurons in the three hidden layers are 256, 128, and 64, respectively. The number of neurons in the output layer is 3, corresponding to the expected conversion rate, expected return on investment, and expected brand awareness, respectively. After the input feature matrix is ​​input into the network, it is processed by linear transformation and ReLU activation function in each hidden layer to gradually extract high-level semantic features. The ReLU activation function maps negative values ​​to 0 while keeping positive values ​​unchanged, introducing non-linearity. The three neurons in the output layer output numerical predictions of the promotion effect. During the training process, a multi-objective weighted loss function is calculated, which includes three errors: mean squared error of expected conversion rate, mean squared error of expected return on investment, and mean squared error of expected brand awareness. The Adam optimizer is used for backpropagation training to update the network weight parameters. Training is terminated when the validation set loss no longer decreases after 20 consecutive training epochs. The trained network can receive the combination of cultural feature vector and policy vector to predict the corresponding promotion effect score vector. For example, after inputting the cultural feature vector and policy vector, the network outputs the expected conversion rate of 0.08, the expected return on investment of 2.5, and the expected brand awareness of 0.15, which constitute the promotion effect score vector.

[0019] The number of consumer agents is set equal to the number of consumer segments. Each consumer agent is configured with a consumer attribute vector including cultural preference parameters, price sensitivity, brand loyalty, and social media activity. The strategy vector and consumer attribute vector are input into the Sigmoid decision function to calculate the purchase probability. The Sigmoid function maps the dot product of the strategy vector and the consumer attribute vector to the interval between 0 and 1. In the dot product calculation, the pricing strategy coefficient is multiplied by price sensitivity, the promotion strategy coefficient is multiplied by brand loyalty, the content strategy coefficient is multiplied by cultural preference parameters, and the channel strategy coefficient is multiplied by social media activity, and then summed. The consumer response benefit is calculated based on the purchase probability and the promotion effect score vector. A competitor agent is set up to predict the competitor's countermeasure strategy based on the strategy vector and market state to calculate the competitive risk. The game objective function is constructed to include three terms: consumer response benefit, strategy cost, and competitive risk. A genetic algorithm is used with a population size of 100, a crossover probability of 0.8, and a mutation probability of 0.1. The game objective function is used as the fitness function for 500 generations of evolution. Each generation of evolution updates the population through selection, crossover, and mutation operations. The strategy vector corresponding to the individual with the highest fitness in the final population is the optimal strategy vector.

[0020] The optimal strategy vector is compared by comparing the values ​​of the high-end, mid-range, and low-end positioning coefficients. The brand's core positioning is determined based on the maximum coefficient, generating a brand slogan template and constructing a planning layer. For each consumer segment, the proportion of expressive content is allocated based on the values ​​of the emotional and rational weighting coefficients, generating digital human expressive content. The values ​​of the professional image coefficient, approachable feature coefficient, and fashionable feature coefficient are compared, and the corresponding digital human action library is selected based on the maximum coefficient to generate facial expression sequences. Background color schemes, decorative elements, and music styles are selected based on the aesthetic preference parameters and festival custom parameters in the cultural feature vector to generate scene element configurations. For example, in the optimal strategy vector, a high-end positioning coefficient of 0.6 (maximum) determines a high-end quality positioning; emotional weighting coefficients of 0.65 and rational weighting coefficients of 0.35 allocate the proportion of expressive content; and a professional image coefficient of 0.55 (maximum) selects a professional action library containing stable posture and appropriate expressions.

[0021] Calculate the cosine similarity between any two target market cultural feature vectors. The cosine similarity is equal to the dot product of the two vectors divided by the product of their norms, with a value ranging from 0 to 1 indicating the degree of similarity. Calculate the cosine similarity pairwise for all markets to generate a cultural similarity matrix. Based on the cultural similarity matrix, use a hierarchical clustering algorithm to segment the markets. Set a clustering distance threshold of 0.7 to group markets with a similarity greater than 0.7 into the same cultural cluster. Identify the source markets with completed strategy deployments and sufficient performance data, and the target markets to be optimized within each cultural cluster. Use the Actor network parameters trained in the source markets as the initialization parameters for the Actor network in the target markets. Fine-tune the Actor network parameters using a small amount of data from the target markets. Transfer the digital human expression content, facial expression sequences, and scene element configurations from the source markets to the target markets to generate recommendation strategies.

[0022] In one specific embodiment, step S1 includes: Raw text data is obtained by collecting social media data, e-commerce review data, and competitor promotion data from the target market using web crawling technology. The original text data is input into the multilingual BERT model for vectorization processing, and attention preference parameters, consumption habit parameters, aesthetic preference parameters, constraint condition parameters, and festival custom parameters are extracted to generate cultural feature vectors. Set the value ranges for pricing strategy coefficient, promotion strategy coefficient, content strategy coefficient, channel strategy coefficient, and digital human image strategy coefficient, and generate a strategy vector; The cultural feature vector and the policy vector are concatenated column by column to form the input feature matrix.

[0023] Specifically, when collecting social media data from a target market, web crawler technology automatically captures user comments, posts, and likes by setting API interfaces or webpage parsing rules for the target platform. E-commerce review data is obtained by accessing product detail pages on e-commerce platforms to capture user evaluations, ratings, purchase times, and other information. Competitor promotional data is obtained by monitoring advertising copy, promotional activities, and pricing information published by competitors on various channels. The crawler program periodically accesses target webpages at preset time intervals and extracts text content from HTML tags, storing it as raw text data. This raw text data contains a large amount of unstructured text, such as user comments like "I feel reassured by the product's natural ingredients," "The packaging design is beautiful," and "The price is a bit high, but it's worth it." When raw text data is input into a multilingual BERT model for vectorization, the BERT model first segments the text into word sequences. Each word is mapped to a fixed-dimensional word vector through an embedding layer. These word vectors are then input into a Transformer encoder, which uses a multi-layer self-attention mechanism to calculate a context-dependent semantic representation. The self-attention mechanism captures semantic relationships between words by calculating attention weights. The vector representation output by the encoder contains deep semantic information of the text. When extracting attention preference parameters from the vector output by the BERT model, the frequency and co-occurrence patterns of words expressing features such as "natural," "environmentally friendly," and "quality" are identified through keyword matching and semantic clustering to calculate attention preferences. The parameter values ​​are as follows: Consumer habit parameters are extracted by analyzing descriptions of purchase channels, purchase frequency, and decision-making factors in the text; aesthetic preference parameters are extracted by identifying evaluations of visual elements such as color, packaging, and design style in the text; constraint parameters are extracted by identifying content in the text that explicitly expresses exclusion intentions (such as ingredient labels, product attribute tags, and visual element markers); and seasonal preference parameters are extracted by analyzing the names of time nodes mentioned in the text, the concentration of purchase periods, and seasonal activities. These five types of parameters are combined to form a cultural feature vector. When setting pricing strategy coefficients, the values ​​of the high-end positioning coefficient, mid-range positioning coefficient, and low-end positioning coefficient are set to a range of 0 to 1, indicating the strength of the strategy tendency for that pricing level. Promotion strategy... The discount coefficient ranges from 0 to 1, representing discount percentages from 0% to no discount. The gift type code ranges from 1 to 10, with each integer corresponding to a gift type (e.g., 1 for trial packs, 2 for coupons, 3 for physical gifts, etc.). The limited-time event period ranges from 1 to 30 days, indicating the number of days the promotion lasts. The content strategy coefficients, including the emotional and rational weighting coefficients, range from 0 to 1, representing the ratio of emotional expression to rational explanation in the promotional content. The channel strategy coefficients, including the social media platform weight vector, are multi-dimensional vectors, with each dimension corresponding to a social media platform. The value of this dimension represents the resource allocation ratio for promotional content on that platform. The KOL collaboration level coefficient ranges from 0 to 1, representing the influence level of the collaborating KOLs from low to high.In the digital human image strategy coefficients, the professional image coefficient, the approachable feature coefficient, and the fashionable feature coefficient take values ​​from 0 to 1, indicating the degree to which the digital human image leans towards that persona type. All these coefficients are arranged and combined in the order of pricing, promotion, content, channel, and image to form a strategy vector. When concatenating the cultural feature vector and the strategy vector column-wise, assuming the cultural feature vector has a dimension of 50 and the strategy vector has a dimension of 15, the concatenation results in a 65-dimensional input feature matrix (50 + 15). Each row represents a sample, i.e., the cultural characteristics of a target market and the corresponding strategy configuration. For example, when a beauty brand expands into an emerging market, a web crawler might scrape 100,000 user reviews from social media. The original text data was constructed by scraping 50,000 product reviews from e-commerce platforms and 200 promotional activity messages from competitor websites. The BERT model processed comments like "I like products with natural ingredients that don't harm my skin," identifying keywords such as "natural" and "doesn't harm." Semantic analysis determined that the user's preference was for natural and safe products. The frequency of similar expressions in all comments was statistically analyzed to calculate the "natural and safe" dimension of the preference parameters. Additionally, comments like "I buy skincare products online every week" extracted parameters related to high online purchase frequency, comments like "Fresh and attractive packaging colors" extracted parameters related to a preference for fresh color schemes, and comments like "I absolutely won't buy products containing alcohol" extracted parameters related to a preference for natural and safe products. The study extracts alcohol content as a limiting parameter and extracts festival custom parameters related to concentrated consumption during shopping festivals from comments on hoarding and other related issues. These five parameter values ​​are combined to form a 50-dimensional cultural feature vector. When constructing the strategy vector, the high-end positioning coefficient is set to 0.7, the mid-range positioning coefficient to 0.2, and the low-end positioning coefficient to 0.1 to indicate a preference for high-end positioning; the discount coefficient is set to 0.3 to indicate a 30% discount; the gift type code is set to 3 to indicate the gift of physical samples; the limited-time activity period is set to 7 days; the emotional weight coefficient is set to 0.6, and the rational weight coefficient is set to 0.4 to indicate that the promotional content is mainly based on emotional narratives; the social media platform weight vector is set to the distribution ratio of each platform. For example, a coefficient of 0.4 on one platform, 0.35 on another, and 0.25 on yet another. A KOL collaboration level coefficient of 0.6 indicates a mid-tier KOL in the collaboration. A professional image coefficient of 0.5, an approachable characteristic coefficient of 0.4, and a fashionable characteristic coefficient of 0.1 indicate a digital avatar with a more professional image. These coefficient values ​​are used to form a 15-dimensional strategy vector. The 50-dimensional cultural feature vector and the 15-dimensional strategy vector are then concatenated column-wise to form a 65-dimensional input feature matrix. The first 50 columns of the matrix represent the values ​​of each dimension of the cultural feature vector, and the last 15 columns represent the coefficient values ​​of the strategy vector. Each element in the input feature matrix is ​​a real number representing the quantified value of the corresponding feature or strategy.

[0024] In one specific embodiment, the value ranges of pricing strategy coefficient, promotion strategy coefficient, content strategy coefficient, channel strategy coefficient, and digital human image strategy coefficient are set to generate a strategy vector, including: The pricing strategy coefficients include high-end positioning coefficient, mid-range positioning coefficient and low-end positioning coefficient, with values ​​ranging from 0 to 1; The promotion strategy coefficient includes the discount strength coefficient, the gift type code, and the limited-time event period. The discount strength coefficient ranges from 0 to 1, the gift type code ranges from 1 to 10 (integers), and the limited-time event period ranges from 1 to 30 days. The content strategy coefficients include the emotional weight coefficient, the rational weight coefficient, the independent scene coefficient, and the shared scene coefficient, with values ​​ranging from 0 to 1. The channel strategy coefficients include the social media platform weight vector and the KOL cooperation level coefficient. The digital human image strategy coefficients include the professional image coefficient, the approachable feature coefficient, and the fashionable feature coefficient, with values ​​ranging from 0 to 1. All coefficients are combined to form a strategy vector.

[0025] Specifically, the pricing strategy coefficient settings include three dimensions: high-end positioning coefficient, mid-range positioning coefficient, and low-end positioning coefficient. The value range is 0 to 1, representing the brand's positioning tendency at different price levels. The sum of the three coefficients equals 1 to ensure the integrity of the positioning strategy. A high-end positioning coefficient close to 1 indicates that the brand mainly targets the high price range and pursues quality premium. A mid-range positioning coefficient close to 1 indicates that the brand mainly targets the mid-range price range and balances cost performance. A low-end positioning coefficient close to 1 indicates that the brand mainly targets the low price range and focuses on market penetration. The promotion strategy coefficient settings include three dimensions: discount strength coefficient, gift type code, and limited-time event period. The discount strength coefficient ranges from 0 to 1, where 0 indicates no discount (original price) and 0 indicates no discount (original price). 1 represents a 10% discount, 0.5 represents a 50% discount, and so on; the larger the value, the greater the discount. Gift type codes range from 1 to 10, distinguishing different gift types, such as 1 for trial packs, 2 for coupons, 3 for physical gifts, and 4 for points rewards, defining 10 common gift types. Limited-time event periods range from 1 to 30 days, indicating the number of days from start to end of the promotion, such as a 3-day flash sale or a 7-day holiday promotion. Content strategy coefficients include four dimensions: emotional weighting, rational weighting, independent scenario weighting, and shared scenario weighting. Emotional and rational weighting coefficients range from 0 to 1, representing the balance between emotional expression and rational explanation in the promotional content. The proportions are as follows: a high emotional weighting coefficient indicates that the promotional copy focuses on emotional resonance, brand stories, and user experiences; a high rational weighting coefficient indicates that the promotional copy focuses on rational content such as product parameters, efficacy data, and technological advantages. The sum of the two coefficients equals 1 to ensure complete coverage of the content strategy. The independent scenario coefficient and the shared scenario coefficient, ranging from 0 to 1, indicate whether the promotional content emphasizes individual or collective value. A high independent scenario coefficient indicates that the copy emphasizes individual appeals such as personalization, uniqueness, and self-actualization; a high shared scenario coefficient indicates that the copy emphasizes collective appeals such as belonging, common interests, and social recognition. The sum of the two coefficients equals 1 to ensure clarity of cultural orientation. The channel strategy coefficient setting includes the weighting of social media platforms. The system comprises two dimensions: a vector and a KOL collaboration level coefficient. The social media platform weight vector is a multi-dimensional vector, with each dimension corresponding to a social media platform. For example, the first dimension corresponds to short video platforms, the second to text and image platforms, and the third to live streaming platforms. The value of each dimension represents the budget allocation ratio for promotional resources on that platform. The sum of all dimension values ​​equals 1 to ensure the completeness of resource allocation. The fan scale coefficient ranges from 0 to 1, indicating the level of the number of followers of the collaborating account. A coefficient in the low range (0 to 0.3) corresponds to 1,000 to 10,000 followers, a coefficient in the mid-range (0.3 to 0.7) corresponds to 10,000 to 1 million followers, and a coefficient in the high range (0.7 to 1) corresponds to...0) For a user base exceeding 1 million, the digital avatar image strategy coefficients are set across three dimensions: professional image coefficient, approachable characteristic coefficient, and fashionable characteristic coefficient. A high professional image coefficient portrays the digital avatar as an industry expert or knowledge disseminator; a high approachable characteristic coefficient portrays the digital avatar as a friendly neighbor to connect with users; and a high fashionable characteristic coefficient portrays the digital avatar as a fashionista or trendsetter to attract a younger audience. The sum of the three coefficients equals 1 to ensure the uniqueness of the avatar's positioning. The high-end, mid-range, and low-end positioning coefficients are arranged in order to form a pricing strategy sub-vector. Discount strength coefficients, gift type codes, and other factors are also considered. The limited-time event cycle is arranged sequentially to form a promotion strategy sub-vector. The emotional weight coefficient, rational weight coefficient, independent scenario coefficient, and shared scenario coefficient are arranged sequentially to form a content strategy sub-vector. The dimensions of the social media platform weight vector and the KOL collaboration level coefficient are arranged sequentially to form a channel strategy sub-vector. The professional image coefficient, approachable characteristic coefficient, and fashionable characteristic coefficient are arranged sequentially to form a digital persona strategy sub-vector. The pricing strategy sub-vector, promotion strategy sub-vector, content strategy sub-vector, channel strategy sub-vector, and digital persona strategy sub-vector are then concatenated sequentially to form a complete strategy vector. The total dimension of the strategy vector equals the sum of the dimensions of each sub-vector.

[0026] In one specific embodiment, step S2 includes: A five-layer fully connected deep neural network was constructed. The number of neurons in the input layer was set to the sum of the dimensions of the cultural feature vector and the policy vector. The number of neurons in the three hidden layers were set to 256, 128 and 64 respectively. The number of neurons in the output layer was set to 3. The input feature matrix is ​​fed into the input layer of the deep neural network, and forward propagation is performed through the ReLU activation function to obtain the hidden layer feature representation; The expected conversion rate, expected return on investment, and expected brand awareness are output separately through the output layer neurons; Calculate the multi-objective weighted loss function, use the Adam optimizer for backpropagation training, update the weight parameters of the deep neural network, terminate training when the validation set loss no longer decreases for 20 consecutive training epochs, and generate a promotion effect score vector.

[0027] Specifically, when constructing a five-layer fully connected deep neural network, the first layer, the input layer, has the number of neurons set to the sum of the dimensions of the cultural feature vector and the strategy vector. The cultural feature vector includes 50 dimensions: attention preference parameters, consumption habit parameters, aesthetic preference parameters, constraint condition parameters, and festival custom parameters. The strategy vector includes 15 dimensions: pricing strategy coefficient, promotion strategy coefficient, content strategy coefficient, channel strategy coefficient, and digital human image strategy coefficient. Therefore, the number of neurons in the input layer is 50 plus 15 equals 65 neurons. The second layer, the first hidden layer, has 256 neurons, the third layer, the second hidden layer has 128 neurons, and the fourth layer, the third hidden layer has 64 neurons. The fifth layer, the output layer, has 3 neurons, corresponding to the expected conversion rate, expected return on investment, and expected brand awareness, respectively. Network layers are connected via weight matrices and bias vectors. The weight matrix from the input layer to the first hidden layer has a dimension of 65 x 256, indicating that each neuron in the input layer has a connection weight with each neuron in the first hidden layer. The weight matrix from the first hidden layer to the second hidden layer has a dimension of 256 x 128, the weight matrix from the second hidden layer to the third hidden layer has a dimension of 128 x 64, and the weight matrix from the third hidden layer to the output layer has a dimension of 64 x 3. When the input feature matrix is ​​input into the input layer of the deep neural network, each row of the input feature matrix represents a sample containing 65 feature values. After receiving the feature values, the input layer performs matrix multiplication with the weight matrix of the first hidden layer. The result of the matrix multiplication is a 65-dimensional input vector multiplied by a 65x256 weight matrix, yielding a 256-dimensional output vector. This 256-dimensional output vector is then added to a 256-dimensional bias vector to obtain the linear combination result of the first hidden layer. This linear combination result is then input into a ReLU activation function for non-linear transformation. The ReLU activation function operates on the rule that the output value equals the input value when the input value is greater than 0, and the output value equals 0 when the input value is less than or equal to 0. The ReLU activation function sets all negative values ​​in the 256-dimensional linear combination result to 0, while positive values ​​remain unchanged. The activated 256-dimensional vector is then used as the output of the first hidden layer and input into the second hidden layer. The first hidden layer repeats the same linear transformation and ReLU activation process, transforming the 256-dimensional input into a 128-dimensional output through a 128x256 weight matrix. The third hidden layer transforms the 128-dimensional input into a 64-dimensional output through a 64x128 weight matrix. These three hidden layers progressively extract high-level abstract representations of the input features. These hidden layer feature representations contain the complex non-linear relationship between cultural features and strategy configuration. The 64-dimensional feature vector output from the third hidden layer is input to the output layer. The output layer transforms this 64-dimensional feature vector into a 3-dimensional output vector through a 3x64 weight matrix. The first element of the 3-dimensional output vector corresponds to the expected conversion rate, representing the percentage of browsing users expected to complete a purchase under the current cultural context and strategy configuration.The second element, corresponding to the expected return on investment (ROI), indicates how many units of revenue are expected to be generated for every unit of promotion cost invested under the current strategy configuration. The third element, corresponding to the expected brand awareness, indicates how much brand awareness is expected to increase in the target market after the current promotion strategy is implemented. When training a deep neural network, a multi-objective weighted loss function is calculated. This function includes three sub-terms corresponding to three output objectives: the first sub-term is the mean squared error of the expected conversion rate (calculated as the square of the difference between the network's output expected conversion rate and the actual conversion rate of the training samples); the second sub-term is the mean squared error of the expected ROI (calculated as the square of the difference between the network's output expected ROI and the actual ROI of the training samples); and the third sub-term is the expected brand awareness. The mean squared error (MSE) is calculated as the square of the difference between the expected brand awareness of the network output and the actual brand awareness of the training samples. The total loss is obtained by multiplying each of the three sub-items by its corresponding weight coefficient and summing the results. The weight coefficients reflect the importance of different promotional objectives. When using the Adam optimizer for backpropagation training, the Adam optimizer adaptively adjusts the learning rate of each parameter based on the gradient information of the loss function with respect to the network weight parameters. Gradient calculation starts from the output layer and propagates back to the input layer. The output layer gradient equals the derivative of the loss function with respect to the output layer output value multiplied by the derivative of the output layer activation function. The gradient of the third hidden layer equals the output layer gradient multiplied by the transpose of the output layer weight matrix and then multiplied by the derivative of the ReLU activation function. The gradients of the second and first hidden layers follow the same pattern. Working backwards, the Adam optimizer calculates parameter updates based on the gradient value of each weight parameter, the first-moment estimate of the gradient, and the second-moment estimate. The first-moment estimate is an exponentially weighted moving average of historical gradients, reflecting the trend direction of the gradient. The second-moment estimate is an exponentially weighted moving average of the squares of historical gradients, reflecting the fluctuation range of the gradient. The parameter update amount equals the learning rate multiplied by the square root of the first-moment estimate divided by the second-moment estimate, plus a small constant. This parameter update amount is added to the original parameter values ​​to complete one parameter update. Repeating the process of forward propagation to calculate the loss, backpropagation to calculate the gradient, and updating the parameters constitutes one training epoch. During training, a validation set is used to evaluate the network's generalization performance. The validation set contains samples that did not participate in training. After each training epoch, the validation set is updated. The validation set samples are input into the network to calculate the validation set loss. The validation set loss reflects the network's prediction error on unseen data. When the validation set loss no longer decreases after 20 consecutive training epochs, it is determined that the network has sufficiently learned the patterns in the training data and will not continue to overfit. At this point, training is terminated. The trained network can accept any combination of cultural feature vectors and policy vectors and calculates the corresponding expected conversion rate, expected return on investment, and expected brand awareness through forward propagation. These three output values ​​form the promotion effect score vector. For example, when a cosmetics brand trains a neural network for an emerging market, the input feature matrix includes the cultural feature vector of that market and various combinations of policy vectors. The attention preference parameter in the cultural feature vector shows that consumers value the natural ingredients of the product.The strategy vectors feature high-end positioning coefficients, high emotional weighting coefficients, and high professional image coefficients. The network uses multi-layer nonlinear transformations to map these input features to expected conversion rate, expected return on investment, and expected brand awareness. Training data is derived from the brand's historical promotional campaign performance records in culturally similar markets.

[0028] In one specific embodiment, step S3 includes: The number of consumer agents is set to the number of consumer segments, and each consumer agent is configured with a consumer attribute vector consisting of cultural preference parameters, price sensitivity, brand loyalty, and social media activity. The strategy vector and consumer attribute vector are input into the Sigmoid decision function for calculation, and the purchase probability of each consumer agent is output. The consumer response revenue is calculated based on the purchase probability and the promotion effect score vector. Set up competitor agents, predict competitor countermeasures and calculate competitive risks based on the market state consisting of policy vectors, current market share, consumer response rate and price range distribution; A game objective function is constructed that includes consumer response benefits, strategy costs, and competitive risks. A genetic algorithm is used with a population size of 100, a crossover probability of 0.8, and a mutation probability of 0.1. The game objective function is used as the fitness function to perform 500 generations of evolutionary iterations to obtain the optimal strategy vector.

[0029] Specifically, the number of consumer agents is determined based on the consumer segmentation results of the target market. Consumer segmentation uses clustering algorithms to divide consumers in the market into different groups according to characteristics such as age, income level, and consumption preferences. Each group corresponds to one consumer agent. When configuring the consumer attribute vector for each consumer agent, the cultural preference parameter is extracted from the aforementioned cultural feature vector to reflect the group's degree of identification with specific cultural elements. Price sensitivity indicates the importance the group attaches to price factors in purchasing decisions, with a value of 0 to 1; the higher the value, the more sensitive the group is to price changes. Brand loyalty indicates the group's loyalty to existing brands, with a value of 0 to 1; the higher the value, the less likely the group is to be attracted by new brands. Social media activity indicates the group's loyalty to existing brands. The frequency of a group's use and engagement on social media platforms ranges from 0 to 1; higher values ​​indicate greater reach through social media. When inputting the strategy vector and consumer attribute vector into the Sigmoid decision function, the product of the pricing strategy coefficient in the strategy vector and the price sensitivity in the consumer attribute vector is first calculated. The high-end positioning coefficient in the pricing strategy coefficient is multiplied by the price sensitivity to obtain the weight of the price factor on the purchase decision. Consumers with high price sensitivity have a lower acceptance of the high-end positioning strategy, resulting in a smaller product value; conversely, consumers with low price sensitivity have a higher acceptance of the high-end positioning strategy, resulting in a larger product value. Simultaneously, the inverse relationship between the discount strength coefficient in the promotion strategy coefficient and brand loyalty is calculated. Consumers with high brand loyalty... Consumers with low brand loyalty show a higher responsiveness to promotional activities. The content strategy coefficient, multiplied by the emotional weighting coefficient and the cultural preference parameter, reflects the match between the promotional content and the consumer's cultural identity. Consumers with high cultural preference parameters are more likely to resonate with content that aligns with their cultural characteristics. The channel strategy coefficient, multiplied by the social media platform weight vector and social media activity, reflects the reach efficiency of channel placement. Consumers with high social media activity are more easily reached through social media channels. The digital avatar strategy coefficient, multiplied by the consumer's aesthetic preference parameter, reflects the acceptance of the digital avatar. The sum of these multiplication results yields the comprehensive impact factor, which is then input into the Sigmoid function. The policy function performs a nonlinear mapping. The Sigmoid function maps any real number to the interval between 0 and 1. During calculation, the comprehensive influence factor is first subtracted from a preset decision threshold to obtain an offset. This offset is multiplied by a sensitivity coefficient to control the steepness of the function curve. A larger sensitivity coefficient indicates that the consumer's decision is more sensitive to policy changes. The product is then multiplied by a negative exponent to obtain the negative power of the natural logarithm. The result of 1 divided by 1 plus this power is the Sigmoid function output value. The output value represents the purchase probability of the consumer agent, ranging from 0 to 1. A purchase probability close to 0 indicates that the group is unlikely to buy, a purchase probability close to 1 indicates that the group has a very high probability of buying, and a purchase probability close to 0.5 indicates that the group's purchase intention is at a moderate level.When calculating consumer response revenue based on purchase probability and promotion effect score vector, the purchase probability is multiplied by the expected conversion rate in the promotion effect score vector to obtain the actual conversion rate. The actual conversion rate is then multiplied by the market size of the consumer group to obtain the expected number of buyers. The expected number of buyers is multiplied by the single purchase amount to obtain the expected sales revenue. The expected sales revenue is multiplied by the expected return on investment to obtain the net revenue contributed by the group. The weighted sum of the net revenues of all consumer agents yields the total consumer response revenue. The weights are determined based on the market importance and contribution of each group. When setting up competitor agents, the competitor agents judge the aggressiveness of the promotion strategy based on the strategy vector. A higher low-end positioning coefficient in the pricing strategy coefficient of the strategy vector indicates the adoption of a price competition strategy. In the strategy coefficients, a higher discount rate indicates a stronger promotional effort. In the content strategy coefficients, a higher emotional weighting indicates the adoption of an emotional promotional strategy. After sensing these aggressive strategies, competitor agents will determine whether to retaliate based on market conditions. The current market share in the market condition section represents the brand's market share; a higher market share indicates a greater perceived threat from competitors, making them more likely to take countermeasures. The consumer response rate indicates the actual level of consumer response to the promotional strategy; a sudden increase in the consumer response rate indicates that competitors are losing customers and require further countermeasures. The price range distribution represents the distribution of products at different price points in the market; a price range distribution showing an increase in low-priced products indicates that competitors anticipate the start of a price war and need to take countermeasures. Following price reductions, the competitor's predicted countermeasures include following the price cut, increasing promotional efforts, boosting advertising, and launching new products. Each countermeasure has a different impact on the competitor's own promotional strategy. Following price reductions leads to a loss of price advantage; increasing promotional efforts diverts some price-sensitive consumers; boosting advertising increases brand exposure and competition; and launching new products distracts consumers. The competitive risk is calculated by summing the impact of each countermeasure multiplied by its probability of occurrence. The probability of a competitor adopting a countermeasure is determined based on historical competitive behavior data and the current market situation. The game objective function includes three parts: consumer response payoff, strategy cost, and competitive risk. The consumer response payoff is the total consumer response calculated above. The revenue and strategy cost components include the opportunity cost of the pricing strategy (i.e., the potential revenue from the low-price market forgone by adopting a high-end positioning), the direct expenses of the promotion strategy (i.e., the cost of discounts and free gifts), the content production cost (i.e., the cost of developing and producing promotional materials for the digital human), the channel placement cost (i.e., the advertising expenditure on various social media platforms), and the digital human maintenance cost (i.e., the cost of updating and maintaining the digital human's image and technology). The competition risk component is the expected revenue loss caused by competitors' countermeasures, as calculated above. The game objective function equals the consumer's response revenue minus the strategy cost minus the competition risk. When using a genetic algorithm to solve for the maximum value of the game objective function, the population size is set to 100, indicating that 100 candidate strategy vectors are maintained simultaneously. The initial population is generated by randomly generating the coefficient values ​​of the strategy vectors.Each strategy vector is called an individual. The fitness value of each individual is calculated by substituting its strategy vector into the game's objective function. A higher fitness value indicates that the strategy vector performs better after considering benefits, costs, and risks. The selection operation based on fitness values ​​uses a roulette wheel selection method. In roulette wheel selection, the probability of each individual being selected is proportional to its fitness value; individuals with higher fitness values ​​are more likely to be retained in the next generation, while individuals with lower fitness values ​​are more likely to be eliminated. After selection, a crossover operation is performed. The crossover probability is set to 0.8, meaning there is an 80% probability of crossing two individuals. During crossover, a random position in the strategy vector is selected as the crossover point, and the strategies of the two individuals at the crossover point are... The selection, crossover, and mutation operations are used to generate two new individuals by exchanging coefficient values. Crossover allows for the combination of different advantageous traits among superior individuals. Mutation randomly selects an individual with a probability of 0.1 and applies a small random perturbation to a coefficient value in that individual's policy vector, with the perturbation amplitude controlled within ±10% of the original value. Mutation increases population diversity and prevents the algorithm from getting trapped in local optima. Selection, crossover, and mutation constitute one generation of evolution. After the first generation, 100 new individuals are generated. The fitness values ​​of all new individuals are recalculated, and the next generation of selection, crossover, and mutation operations are performed. This evolutionary process is repeated for 500 generations. After this, the fitness values ​​of individuals in the population gradually converge, and the policy vector corresponding to the individual with the highest fitness in the population is the optimal policy vector. For example... When expanding its market, a skincare brand segmented its consumers into three consumer groups: young, middle-aged, and elderly. The young consumer attribute vector showed lower price sensitivity (because they prioritize product experience over price), lower brand loyalty (because they are more willing to try new brands), and higher social media activity (because they frequently use social media). The middle-aged group showed moderate price sensitivity, high brand loyalty, and moderate social media activity. The elderly group showed high price sensitivity, the highest brand loyalty, and low social media activity. In the strategy vector, a high-end positioning coefficient was set and multiplied by the young group's low price sensitivity to obtain a large product. A medium discount coefficient was set and multiplied by the elderly group's... Multiplying high price sensitivity results in a large product value. A high emotional weighting coefficient, multiplied by the high cultural preference parameter of the younger group, also yields a large product. The social media platform weight vector, with a high weight assigned to short video platforms, multiplied by the high social media activity of the younger group, results in a high reach efficiency. The sum of these multiplications is then input into a Sigmoid function. The younger group, due to its larger overall influence factor, outputs a higher purchase probability. The older group, due to the mismatch between price sensitivity and low discounts, outputs a lower purchase probability. Multiplying the purchase probability of the younger group by the market size and expected conversion rate of that group yields the consumer response revenue contributed by the younger group. The competitor's agent observes a high high-end positioning coefficient in the strategy vector, indicating a low threat of a price war.Observing the significant investment in social media platforms, it was determined that increased brand exposure necessitated a corresponding increase in advertising as a countermeasure. Competitors' countermeasures led to some younger consumers being diverted to competitor advertising, resulting in a calculated competitive risk value. The game objective function equals the sum of consumer response benefits for young, middle-aged, and elderly groups, minus strategy costs, and then minus competitive risk. In the initial population, an individual with an excessively high-end positioning coefficient resulted in a very low purchase probability among the elderly, leading to low overall fitness and elimination. Another individual with an excessively high promotion intensity coefficient caused a surge in strategy costs, resulting in low fitness and elimination. Individuals with moderate high-end positioning coefficients, moderate promotion intensity coefficients, and high emotional weight coefficients were retained due to their high fitness, balancing the responses of different groups and cost control. After 500 generations of evolution, the population converged to the individual with the highest fitness, and the coefficients in the optimal strategy vector for that individual reached their optimal configuration.

[0030] In one specific embodiment, step S4 includes: Compare the values ​​of the high-end positioning coefficient, mid-end positioning coefficient, and low-end positioning coefficient in the optimal strategy vector, determine the core brand positioning based on the maximum coefficient, generate the corresponding brand slogan template, and construct the planning layer of the hierarchical recommendation strategy decision tree. For each consumer segment, digital human-generated content is generated based on a content strategy coefficient, and the proportion of emotional expression content and rational explanation content is allocated based on the values ​​of the emotional weight coefficient and the rational weight coefficient. Compare the values ​​of the professional image coefficient, the approachable feature coefficient, and the fashionable feature coefficient, and select the corresponding digital human action library based on the maximum coefficient to generate an expression and action sequence; Based on the aesthetic preference parameters and festival custom parameters in the cultural feature vector, a background color scheme, decorative elements, and music style are selected to generate a scene element configuration.

[0031] Specifically, by comparing the values ​​of the high-end, mid-range, and low-end positioning coefficients in the optimal strategy vector, three coefficient values ​​are extracted from the pricing strategy portion of the optimal strategy vector. By comparing each coefficient individually, the largest value is determined: the largest high-end positioning coefficient indicates the brand should adopt a high-quality approach; the largest mid-range positioning coefficient indicates the brand should adopt a cost-effective approach; and the largest low-end positioning coefficient indicates the brand should adopt a mass-market approach. After determining the core brand positioning, corresponding brand slogan templates are generated. The high-quality brand slogan template emphasizes premium terms such as excellence, luxury, and refinement; the cost-effective brand slogan template emphasizes balanced terms such as smart choice and exceptional value; and the mass-market brand slogan template emphasizes accessibility and affordable prices. The inclusive vocabulary and brand slogan templates serve as the planning layer of the hierarchical recommendation strategy decision tree, providing overall guidance for content generation in subsequent strategy and operational layers. When generating digital human-based content for specific consumer segments, the overall tone of the content is first determined based on the emotional and rational weighting coefficients within the content strategy coefficients. A sum of 1 for both ensures complete content coverage. The emotional weighting coefficient represents the proportion of emotional expression in the content, while the rational weighting coefficient represents the proportion of rational explanation. Emotional content includes brand storytelling (recounting the brand founder's initial aspirations and the brand's development history) and user experience sharing (showcasing real users' experiences and improvement effects). Emotional resonance is evoked through scenario-based descriptions to awaken consumers' emotional identification. Rational explanations include ingredient analysis (detailed listing of the product's active ingredients and their mechanisms of action), efficacy data presentation (providing clinical test results), and explanations of technological advantages (introducing the product's patented technologies or unique processes). When allocating time based on the emotional and rational weighting coefficients, the total duration of the digital avatar's content is divided according to the ratio of these two coefficients. When the emotional weighting coefficient is 0.6, 60% of the total content time is allocated to emotional expression, and the remaining 40% to rational explanation. The script for the content is written first for the emotional expression section, followed by the rational explanation section. The emotional expression text uses first-person or second-person narration to enhance intimacy. The text employs metaphors and personification to enhance its persuasiveness, while the rational explanatory copy uses objective statements and data / charts to strengthen its argument. The specific expression of the content is tailored to different consumer segments: for younger audiences, the emotional expression section uses more internet slang and youthful language; for middle-aged audiences, the rational explanation section emphasizes the product's practical effects and long-term value; and for older audiences, the rational explanation section emphasizes the product's safety and reliability. By comparing the values ​​of the professional image coefficient, the approachable characteristic coefficient, and the fashionable characteristic coefficient, the three coefficient values ​​of the digital persona strategy component in the optimal strategy vector are extracted, and the largest coefficient value is determined through individual comparison.The highest professional image coefficient indicates the digital avatar should be portrayed as an industry expert; the highest approachable characteristic coefficient indicates a friendly and approachable image; and the highest fashion characteristic coefficient indicates a trendsetter image. The corresponding digital avatar action library is selected based on the highest coefficient. This library is a pre-built database containing characteristic actions for different avatar types. The professional image action library includes a stable stance (feet parallel, body weight balanced), a serious but appropriate expression (relaxed facial muscles, natural brows, slightly upturned corners of the mouth, but not excessive), gestures controlled below the shoulders (arms naturally hanging down or slightly raised, without exaggerated movements), and a firm and focused gaze (direct eye contact with the camera or audience to create a professional impression). The approachable characteristic action library includes... A relaxed standing posture (slightly leaning forward to create a sense of closeness), a smiling expression (a noticeably upturned mouth revealing some teeth), an open gesture (palms outstretched to indicate welcome and acceptance), and a gentle, warm gaze (soft eyes conveying friendliness). The fashionable action library includes dynamic postures (body moving to the rhythm of music or changing positions), rich expressions (frequent facial expressions showcasing vitality), exaggerated gestures (large gestures used in conjunction with speech for emphasis), and lively, changeable eyes (rapidly shifting gaze to display youthful energy). Specific action data is extracted from the selected action library to generate an expression and action sequence. This sequence is arranged along a timeline, with each point in time corresponding to a specific action state of the digital human. Action states include facial expression parameters such as eyebrow position, eye opening / closing degree, and mouth corner curvature. The system considers numerical values ​​such as degrees, body movement parameters (head angle, arm position, body posture, etc.), and the duration of each movement (the number of seconds each movement is held). Facial expressions and movement sequences are synchronized with the content the digital human is conveying. When telling emotional stories, the digital human's expressions are softer and warmer, and its gestures more approachable. When presenting rational data, the digital human's expressions are more serious and focused, and its gestures more stable. When selecting scene elements based on aesthetic preference parameters and festival custom parameters from the cultural feature vector, the aesthetic preference parameters include color preference, visual style preference, and decorative element preference dimensions. The color preference dimension value indicates the target market consumers' preference for different color schemes. A high preference for warm colors leads to the selection of orange, red, and yellow as the main background color, while cool colors... When the preference for neutral colors is high, blue, green, and purple are chosen as the main background colors; when the preference for neutral colors is high, white, gray, and beige are chosen as the main background colors. The numerical value of the visual style preference dimension indicates the acceptance of different visual styles by consumers in the target market. When the preference for minimalist style is high, a background design with simple lines and minimal elements is chosen; when the preference for gorgeous style is high, a background design with intricate decorations and rich details is chosen; when the preference for natural style is high, a background design with natural elements such as plants and landscapes is chosen. The numerical value of the decorative element preference dimension indicates the recognition of different decorative types by consumers in the target market. When the preference for geometric patterns is high, geometric decorations such as circles, squares, and triangles are added to the background; when the preference for plant patterns is high, plant decorations such as leaves, flowers, and vines are added to the background.When there is a high preference for abstract patterns, artistic abstract lines and color blocks are added to the background. The festival custom parameters include the names of important festivals and festival cultural symbols. Important festival names identify shopping or cultural festivals valued by consumers in the target market. Festival cultural symbols represent visual elements related to the festival, such as lanterns, colored lights, gift boxes, and flowers. During festival promotional activities, these festival cultural symbols are integrated into the scene element configuration. Festival decorative elements are added to the background, festival representative colors are used in the color scheme, and festival-related music is selected. The music style is chosen based on the music preference dimension in the aesthetic preference parameters: upbeat music is suitable for creating a vibrant atmosphere for young people, soothing music is suitable for creating a relaxing atmosphere for middle-aged people, and traditional music is suitable for creating a familiar atmosphere for older people. The background color scheme, decorative elements, and music style are combined to form a complete scene element configuration. The scene element configuration is stored in a data file containing parameters such as color RGB values, decorative element image paths, music file paths, and element position coordinates. For example, when a beauty brand targets young women, the optimal strategy vector with the highest high-end positioning coefficient determines the brand's core positioning as a high-end quality route, generating a brand slogan template of "Luxury Skin Care". "Radiating Confidence" – The emotional weighting factor outweighs the rational weighting factor, therefore the emotional expression portion of the content occupies more time. The script first tells the emotional story of the brand founder's unwavering pursuit of natural skincare, then showcases real user experiences of smooth, delicate skin after using the product, and finally briefly explains the plant extracts and clinical test data. The professional image factor is maximized, so a professional image action library is selected. The digital human maintains a stable posture while telling the brand story, with a serious yet warm facial expression and moderate gestures. When showcasing user experience, the expression softens slightly, and the gesture points to the user's photo on the screen. When explaining ingredient data, the expression returns to professional seriousness, and the gesture points to the ingredient list. The aesthetic preference parameter in the cultural characteristic vector shows that this group prefers a fresh and natural visual style and plant-based decorations. The scene element configuration selects light green and white as the main background colors, adding green plant leaves and small flower decorations to create a natural and fresh atmosphere. Soft natural sound effects such as flowing water and birdsong enhance immersion. The festival custom parameter shows that this group values ​​shopping festivals; during shopping festivals, the scene configuration adds exclusive shopping festival logos and limited-time offer information to the background corners.

[0032] In one specific embodiment, step S5 includes: Calculate the cosine similarity between the cultural feature vectors of any two target markets to generate a cultural similarity matrix; Market segmentation is performed using a hierarchical clustering algorithm based on a cultural similarity matrix. A clustering distance threshold of 0.7 is set, and target markets with a similarity greater than 0.7 are grouped into the same cultural cluster. Identify the source market and target market to be optimized within each cultural cluster that have completed strategy deployment and have sufficient effect data, and use the Actor network parameters trained in the source market as the initialization parameters for the Actor network in the target market; The parameters of the Actor network are fine-tuned using a small amount of data from the target market. The digital human expression content, facial expression and action sequences, and scene element configurations from the source market are transferred to the target market to generate a recommendation strategy.

[0033] Specifically, when calculating the cosine similarity between the cultural feature vectors of any two target markets, the cultural feature vectors of the two markets are extracted. These vectors include parameters related to attention preferences, consumption habits, aesthetic preferences, constraints, and festival customs. Cosine similarity is obtained by calculating the dot product of the two vectors and dividing it by the product of their norms. The dot product is calculated by multiplying the corresponding dimensions of the two vectors and then summing the results. The vector norm is calculated by taking the square root of the sum of the squares of the values ​​in each dimension of the vector. The cosine similarity value ranges from 0 to 1; a value closer to 1 indicates greater similarity in the cultural features of the two markets, while a value closer to 0 indicates greater cultural differences. Cosine similarity is calculated pairwise for all target markets. Cosine similarity is used to generate a cultural similarity matrix. This matrix is ​​a symmetric matrix where rows and columns correspond to different target markets. Matrix elements represent the similarity between corresponding markets. When using a hierarchical clustering algorithm to segment markets based on the cultural similarity matrix, the algorithm gradually merges markets into clusters through a bottom-up agglomerative approach. Initially, each target market forms a separate cluster. During the algorithm iteration, the two clusters with the highest cultural similarity are selected for merging at each iteration. The similarity between the merged cluster and other clusters is obtained by calculating the average similarity between all markets within the cluster and all markets within the other cluster. A clustering distance threshold of 0.7 is set to indicate that the similarity between two clusters is less than 0.The merging process stops at 7:00 AM. A higher clustering distance threshold results in more clusters and stronger cultural similarities within each cluster; conversely, a lower threshold results in fewer clusters and greater cultural differences within each cluster. After the clustering algorithm terminates, multiple cultural clusters are obtained. Each cultural cluster contains multiple target markets with similar cultural characteristics. When identifying the source and target markets within each cultural cluster, the source market is defined as the market where strategy deployment has been completed and sufficient performance data has been accumulated. Source market performance data includes the actual conversion rate, actual return on investment, actual brand awareness improvement, and consumer response under different strategy configurations from historical promotional activities. The target market is an emerging market requiring strategy optimization or a market with insufficient performance data. The Actor network represents the strategy in reinforcement learning. The network is responsible for outputting promotion strategy decisions based on market conditions. The input to the Actor network is the market culture feature vector and the current strategy vector, and the output is a strategy adjustment suggestion. The Actor network parameters of the source market are obtained by training with historical data from the source market, including the weight matrix and bias vector of each layer. When using the Actor network parameters trained in the source market as the initial parameters of the Actor network in the target market, all weights and bias values ​​of the source market Actor network are directly copied into the target market Actor network. When fine-tuning the Actor network parameters using a small amount of data from the target market, the fine-tuning training only uses a small sample of data from the target market. The fine-tuning process is carried out through the backpropagation algorithm based on the target market data. The prediction error updates the network parameters. Fine-tuning uses a small learning rate, making only minor adjustments to the network parameters. Fewer training epochs are used to avoid overfitting the limited data from the target market. After fine-tuning, the Actor network retains the general generalization rules learned from the source market while adapting to the specific cultural characteristics of the target market. When transferring the digital human's expressive content from the source market to the target market, the content is adjusted based on the differences in the cultural feature vectors of the two markets. Differences in the attention preference parameters within the cultural feature vectors reflect the differences in product preferences between the two markets. The value expression in the content is adjusted according to the attention preference parameters of the target market. Differences in the aesthetic preference parameters reflect the differences in visual aesthetics between the two markets. The value expression in the content is adjusted according to the aesthetic preferences of the target market. The algorithm adjusts the digital human's facial expressions and actions, as well as scene element configurations, based on preference parameters. Differences in holiday customs parameters reflect the differences between the two markets regarding important festivals and cultural symbols. Based on the target market's holiday customs parameters, it adjusts holiday decorations in the scene during specific periods. When migrating facial expression and action sequences from the source market to the target market, it maintains the overall structure and rhythm of the action sequences. It fine-tunes the amplitude and frequency of specific actions based on the target market's cultural characteristics. When migrating scene element configurations from the source market to the target market, it maintains the overall style of the scene. It adjusts the background color scheme and specific selection of decorative elements based on the target market's aesthetic preferences. When generating the recommendation strategy, it integrates the optimization strategies of all cultural clusters to form a complete promotion plan covering multiple target markets globally.

[0034] The above describes the method for automatically generating AI digital human recommendation strategies in the embodiments of this application. The following describes the system for automatically generating AI digital human recommendation strategies in the embodiments of this application. Please refer to [link / reference]. Figure 2 One embodiment of the AI ​​digital human recommendation strategy automatic generation system in this application includes: The extraction module is used to collect social media data and e-commerce review data of the target market, extract cultural feature vectors, construct a strategy vector containing pricing strategy coefficients, promotion strategy coefficients, content strategy coefficients, channel strategy coefficients and digital human image strategy coefficients, and concatenate the cultural feature vectors with the strategy vectors to form an input feature matrix; The training module is used to input the input feature matrix into a deep neural network, train the mapping relationship between the cultural feature vector and the promotion effect, and output the promotion effect score vector of expected conversion rate, expected return on investment and expected brand awareness. The generation module is used to construct a multi-agent game framework based on the promotion effect score vector, set up a consumer agent, a competitor agent, and a strategy agent, and solve the game equilibrium of the strategy vector through a genetic algorithm to generate the optimal strategy vector. The configuration module is used to generate a hierarchical recommendation strategy decision tree based on the pricing strategy coefficient, promotion strategy coefficient, content strategy coefficient, channel strategy coefficient and digital human image strategy coefficient in the optimal strategy vector, and to generate digital human expression content, facial expression and action sequence and scene element configuration for different consumer segments. The transfer module is used to calculate the cultural similarity matrix between cultural feature vectors of different target markets, divide similar markets into cultural clusters, and use transfer learning to transfer the digital human expression content, facial expression and action sequence and scene element configuration of the source market to the target market to generate a recommendation strategy.

[0035] above Figure 2 The AI ​​digital human recommendation strategy automatic generation system in this embodiment of the invention is described in detail from the perspective of modular functional entities. The AI ​​digital human recommendation strategy automatic generation device in this embodiment of the invention is described in detail from the perspective of hardware processing.

[0036] Reference Figure 3 This invention also provides an AI digital human recommendation strategy automatic generation device, which can be a server, and its internal structure can be as follows: Figure 3As shown, the AI ​​digital human recommendation strategy automatic generation device includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor, designed for computing and control, provides computational and control capabilities. The memory of the AI ​​digital human recommendation strategy automatic generation device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and the database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the AI ​​digital human recommendation strategy automatic generation device stores the data corresponding to this embodiment. The network interface of the AI ​​digital human recommendation strategy automatic generation device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the above-described method.

[0037] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the AI ​​digital human recommendation strategy automatic generation device to which the present invention is applied.

[0038] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the AI ​​digital human recommendation strategy automatic generation method.

[0039] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0040] 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 the present invention, in essence, or the part that contributes to the prior art, 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 an AI digital human recommendation strategy automatic generation 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 the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0041] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI digital human recommendation strategy automatic generation method, characterized in that, The method comprises: S1 step: collecting social media data and e-commerce comment data of target markets, extracting cultural feature vectors, constructing strategy vectors containing pricing strategy coefficients, promotion strategy coefficients, content strategy coefficients, channel strategy coefficients and digital human image strategy coefficients, and splicing the cultural feature vectors and the strategy vectors to form an input feature matrix; S2 step: inputting the input feature matrix into a deep neural network, training the mapping relationship between the cultural feature vectors and the promotion effect, and outputting a promotion effect score vector of expected conversion rate, expected investment return rate and expected brand awareness; S3 step: constructing a multi-agent game framework based on the promotion effect score vector, setting consumer agents, competitor agents and strategy agents, solving the game equilibrium of the strategy vector through a genetic algorithm, and generating an optimal strategy vector; S4 step: generating a hierarchical recommendation strategy decision tree according to the pricing strategy coefficients, promotion strategy coefficients, content strategy coefficients, channel strategy coefficients and digital human image strategy coefficients in the optimal strategy vector, and generating digital human expression content, expression action sequence and scene element configuration for consumer subgroups respectively; S5 step: calculating the cultural similarity matrix between cultural feature vectors of different target markets, dividing similar markets into cultural clusters, and using transfer learning to transfer the digital human expression content, expression action sequence and scene element configuration of the source market to the target market to generate a recommended strategy.

2. The AI digital human recommendation strategy automatic generation method according to claim 1, characterized in that, The S1 step comprises: Collecting social media data, e-commerce comment data and competitor promotion data of target markets through crawler technology to obtain raw text data; Inputting the raw text data into a multilingual BERT model for vectorization processing to extract attention preference parameters, consumption habit parameters, aesthetic preference parameters, restriction condition parameters and festival custom parameters, and generate the cultural feature vector; Setting the value range of the pricing strategy coefficient, the promotion strategy coefficient, the content strategy coefficient, the channel strategy coefficient and the digital human image strategy coefficient to generate the strategy vector; Splicing the cultural feature vector and the strategy vector by column to form the input feature matrix.

3. The AI digital human recommendation strategy automatic generation method according to claim 2, characterized in that, The setting of the value range of the pricing strategy coefficient, the promotion strategy coefficient, the content strategy coefficient, the channel strategy coefficient and the digital human image strategy coefficient to generate the strategy vector comprises: The pricing strategy coefficient includes a high-end positioning coefficient, a medium-end positioning coefficient and a low-end positioning coefficient, and the value range is 0 to 1; The promotion strategy coefficient includes a discount intensity coefficient, a gift type code and a time-limited activity period, wherein the discount intensity coefficient has a value range of 0 to 1, the gift type code has a value range of an integer from 1 to 10, and the time-limited activity period has a value range of 1 to 30 days; The content strategy coefficient includes an emotional proportion coefficient, a rational proportion coefficient, an independent scene coefficient and a shared scene coefficient, and the value range is 0 to 1; The channel strategy coefficient is set to include a social media platform weight vector and a KOL cooperation level coefficient, the digital human image strategy coefficient is set to include a professional image coefficient, an affinity feature coefficient and a fashion feature coefficient, and the values of all the coefficients range from 0 to 1, and the strategy vector is formed by combining all the coefficients.

4. The AI digital human recommendation strategy automatic generation method according to claim 3, characterized in that, The S2 step comprises: A five-layer fully connected deep neural network is constructed, the number of input layer neurons is set to be the sum of the dimensions of the cultural feature vector and the strategy vector, the number of neurons in three hidden layers is set to be 256, 128 and 64 respectively, and the number of output layer neurons is set to be 3; The input feature matrix is input into the input layer of the deep neural network, forward propagation calculation is performed through a ReLU activation function, and hidden layer feature representation is obtained; The expected conversion rate, the expected investment return rate and the expected brand awareness are output by the output layer neurons respectively; A multi-objective weighted loss function is calculated, an Adam optimizer is used for back propagation training, the weight parameters of the deep neural network are updated, the training is terminated when the validation set loss does not decrease for 20 consecutive training rounds, and the promotion effect score vector is generated.

5. The AI digital human recommendation strategy automatic generation method according to claim 4, characterized in that, The S3 step comprises: The number of consumer agents is set to be the number of consumer subgroups, and each consumer agent is configured with a consumer attribute vector composed of cultural preference parameters, price sensitivity, brand loyalty and social media activity; The strategy vector and the consumer attribute vector are input into a Sigmoid decision function for calculation, and the purchase probability of each consumer agent is output, and the consumer response revenue is calculated according to the purchase probability and the promotion effect score vector; The competitor agent is set based on the strategy vector and the market state composed of current market share, consumer response rate and price interval distribution, the competitor countermeasure is predicted, and the competition risk is calculated; A game objective function is constructed, which includes the consumer response revenue, the strategy cost and the competition risk, a genetic algorithm is used to set the population size to 100, the crossover probability to 0.8 and the mutation probability to 0.1, the game objective function is used as the fitness function for 500 generations of evolutionary iteration, and the optimal strategy vector is obtained.

6. The AI digital human recommendation strategy automatic generation method according to claim 5, characterized in that, The S4 step comprises: The numerical values of the high-end positioning coefficient, the medium-end positioning coefficient and the low-end positioning coefficient in the optimal strategy vector are compared, the brand core positioning is determined according to the maximum coefficient, the corresponding brand slogan template is generated, and the planning layer of the hierarchical recommendation strategy decision tree is constructed; For the consumer subgroups, the digital human expression content is generated according to the content strategy coefficient, and the proportion of emotional expression content and rational explanation content is allocated according to the numerical values of the emotional proportion coefficient and the rational proportion coefficient; The numerical values of the professional image coefficient, the affinity feature coefficient and the fashion feature coefficient are compared, and the corresponding digital human action library is selected according to the maximum coefficient to generate the expression action sequence; The background color scheme, decoration elements and music style are selected according to the aesthetic preference parameters and holiday custom parameters in the cultural feature vector to generate the scene element configuration.

7. The AI digital human recommendation strategy automatic generation method according to claim 1, characterized in that, The S5 step comprises: calculating the cosine similarity between the cultural feature vectors of any two target markets to generate the cultural similarity matrix; performing market division based on the cultural similarity matrix using a hierarchical clustering algorithm, setting the clustering distance threshold to 0.7, and grouping target markets with a similarity greater than 0.7 into the same cultural cluster; identifying source markets and target markets to be optimized in each cultural cluster that have completed strategy deployment and sufficient effect data, and using the Actor network parameters trained by the source markets as the initialization parameters of the Actor network of the target markets; fine-tuning the Actor network parameters using a small amount of data from the target market, and migrating the digital human expression content, expression action sequence, and scene element configuration from the source market to the target market to generate the recommended strategy.

8. An AI digital human recommendation strategy automatic generation system, characterized in that, An AI digital human recommendation strategy automatic generation system for implementing the AI digital human recommendation strategy automatic generation method of any one of claims 1-7, comprising: an extraction module configured to collect social media data and e-commerce review data of a target market, extract cultural feature vectors, construct a strategy vector comprising a pricing strategy coefficient, a promotion strategy coefficient, a content strategy coefficient, a channel strategy coefficient, and a digital human image strategy coefficient, and splice the cultural feature vectors and the strategy vector to form an input feature matrix; a training module configured to input the input feature matrix into a deep neural network, train the mapping relationship between the cultural feature vectors and the promotion effect, and output a promotion effect score vector of expected conversion rate, expected return on investment, and expected brand awareness; a generation module configured to construct a multi-agent game framework based on the promotion effect score vector, set a consumer agent, a competitor agent, and a strategy agent, solve the game equilibrium of the strategy vector through a genetic algorithm, and generate an optimal strategy vector; a configuration module configured to generate a hierarchical recommendation strategy decision tree according to the pricing strategy coefficient, the promotion strategy coefficient, the content strategy coefficient, the channel strategy coefficient, and the digital human image strategy coefficient in the optimal strategy vector, and generate digital human expression content, expression action sequence, and scene element configuration for consumer subgroups; a migration module configured to calculate the cultural similarity matrix between the cultural feature vectors of different target markets, divide similar markets into cultural clusters, and migrate the digital human expression content, expression action sequence, and scene element configuration from the source market to the target market using transfer learning to generate a recommended strategy.

9. An AI digital human recommendation strategy automatic generation device, characterized in that, A computer device comprising a memory and a processor, wherein the memory stores a computer program capable of running on the processor, and the processor implements the AI digital human recommendation strategy automatic generation method of any one of claims 1-7 when executing the computer program.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program causes the processor to execute the AI digital human recommendation strategy automatic generation method of any one of claims 1-7 when the computer program is run on the processor.

Images