Game and movie advertisement title generation and optimization method and system based on AI

Abstract

The invention discloses an AI-based game and movie advertisement title generation and optimization method and system, and the method specifically comprises the steps: collecting multi-source heterogeneous data, carrying out the processing through a user interest migration model, and generating a time sequence user interest vector; constructing a knowledge graph based on the product data, and fusing the time-sequential user interest vector to perform node weighting to generate a deep product feature vector; inputting the deep product feature vector into a hybrid generation architecture comprising a Transform model and a generative adversarial network, and generating candidate titles; performing evaluation and iterative optimization on the candidate titles through a multi-target reward model agent to obtain target titles; and executing self-adaptive intelligent putting on the target title. According to the method, automatic generation, precise customization and continuous optimization of the advertisement title are realized, and the creation efficiency, precision and individuation degree of the advertisement title are improved, so that the click rate and conversion rate of the advertisement in games and the movie industry are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an AI-based method and system for generating and optimizing game and short drama advertisement titles. Background Technology

[0002] In today's era of booming digital marketing, the gaming and short drama industries, as an important part of the entertainment industry, rely heavily on the accuracy and effectiveness of their advertising for product promotion and market share enhancement. Ad headlines, as the primary element for attracting user attention and stimulating interest, directly impact click-through rates and conversion rates. However, the current creation of game and short drama ad headlines primarily relies on manual labor, a method with several significant problems.

[0003] First, manually writing headlines is extremely inefficient. Given the rapidly evolving nature of advertising demands, manually creating headlines requires a significant investment of time and effort in brainstorming, writing, and revising. This makes it difficult to generate a large number of compliant headlines in a short period, failing to meet the market's timeliness requirements for advertising and potentially missing optimal promotional opportunities.

[0004] Secondly, manually created headlines lack solid data support. Headline creation often relies on the creator's personal experience and subjective feelings, lacking in-depth analysis of large amounts of user data, product data, and market trend data. This lack of scientific basis makes it difficult for the generated headlines to accurately match the interests and needs of the target users, and to effectively highlight the product's core selling points, resulting in insufficient appeal and persuasiveness of the advertisement.

[0005] Furthermore, the trends in the gaming and short drama industries change rapidly, with new gameplay, short drama themes, and popular culture elements constantly emerging. Human-created content struggles to capture these latest trends in a timely manner and cleverly integrate them into headlines, making the headlines appear outdated and unappealing to users seeking novelty, thus reducing the effectiveness of advertising.

[0006] Furthermore, manually created content suffers from significant limitations in terms of personalization. Different user groups possess varying interests, consumption habits, and aesthetic standards. Simultaneously, different advertising channels (such as social media platforms, video platforms, and search engines) have their own unique formatting guidelines and user group characteristics. User activity and focus also differ across time periods. Manually created content struggles to accurately personalize headlines to address these diverse factors, resulting in poor targeting and adaptability of ad headlines, failing to fully leverage the effectiveness of ad placement.

[0007] While existing technologies attempt to address the challenges of creating ad headlines to some extent, they still face numerous limitations. Firstly, automation is low; the ad headline generation process lacks efficient automated tools and methods, still requiring significant manual intervention and hindering large-scale, rapid headline generation. Secondly, accuracy is insufficient; the generated headlines do not accurately match the target audience or product characteristics, failing to convey the product's core information and failing to resonate with the target audience, resulting in low click-through rates and conversion rates. Furthermore, existing technologies lack effective data feedback and optimization mechanisms, making it impossible to continuously optimize and improve headlines based on actual ad performance, thus limiting the overall effectiveness of ad campaigns. Summary of the Invention

[0008] The purpose of this invention is to provide an AI-based method and system for generating and optimizing game and short drama advertisement titles. This method and system enables automated generation, precise customization, and continuous optimization of advertisement titles, improving the efficiency, accuracy, and personalization of advertisement title creation. This, in turn, increases the click-through rate and conversion rate of advertisements in the game and short drama industry, thereby solving at least one of the aforementioned problems in the prior art.

[0009] In a first aspect, the present invention provides an AI-based method for generating and optimizing game and short drama advertisement titles, the method specifically comprising: Collect multi-source heterogeneous data and process it using a user interest transfer model to generate time-series user interest vectors; A knowledge graph is built based on product data, and node weighting is performed by integrating time-series user interest vectors to generate deep product feature vectors. The deep product feature vector is input into a hybrid generative architecture that includes a Transformer model and a generative adversarial network to generate candidate titles, and a constraint generation mechanism is introduced during the generation process to ensure that the titles comply with regulations. The candidate titles are evaluated and iteratively optimized using a multi-objective reward model agent to obtain the target title. Based on the channel characteristic knowledge base, the output of the timing prediction model, and real-time user profiles, adaptive intelligent targeting is performed on the target titles.

[0010] Secondly, this invention provides an AI-based system for generating and optimizing game and short drama advertisement titles, the system specifically comprising: The data acquisition module is used to collect multi-source heterogeneous data and process it using a user interest transfer model to generate time-series user interest vectors. The vector weighting module is used to build a knowledge graph based on product data and integrate time-series user interest vectors to weight nodes and generate deep product feature vectors. The title generation module is used to input deep product feature vectors into a hybrid generation architecture that includes a Transformer model and a generative adversarial network to generate candidate titles, and introduces a constraint generation mechanism during the generation process to ensure title compliance. The title optimization module is used to evaluate and iteratively optimize candidate titles through a multi-objective reward model agent to obtain the target title; The title delivery module is used to perform adaptive intelligent delivery of target titles based on the channel characteristic knowledge base, the output of the delivery timing prediction model, and real-time user profiles.

[0011] Thirdly, the present invention provides a computer device, including: a memory and a processor, and a computer program stored in the memory, wherein when the computer program is executed on the processor, it implements the AI-based method for generating and optimizing game and short drama advertisement titles as described in any of the above methods.

[0012] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the AI-based method for generating and optimizing game and short drama advertisement titles as described in any of the above methods.

[0013] Compared with the prior art, the present invention has at least one of the following technical effects: 1. This invention enables automated generation, precise customization, and continuous optimization of advertising titles, improving the efficiency, accuracy, and personalization of advertising title creation, thereby increasing the click-through rate and conversion rate of advertisements in the game and short drama industries.

[0014] 2. This invention collects multi-source heterogeneous data and uses a user interest transfer model to generate time-series user interest vectors, which can comprehensively and accurately capture changes in user interests over time, providing a solid foundation for the subsequent generation of personalized advertising titles.

[0015] 3. This invention effectively improves the accuracy and generalization of user interest representation by cleaning and fusing multi-source heterogeneous data, and by generating time-series user interest vectors using a temporal attention mechanism and an adversarial domain adaptive training strategy.

[0016] 4. This invention constructs a knowledge graph based on product data and integrates time-seriesd user interest vectors to generate deep product feature vectors, achieving a deep integration of objective product characteristics and subjective user interests, and enhancing the comprehensive characterization of product features.

[0017] 5. This invention inputs deep product feature vectors into a hybrid generation architecture and introduces a constraint generation mechanism, which can generate diverse and creative candidate titles by leveraging the advantages of both models while ensuring title compliance.

[0018] 6. By generating an initial title sequence in the Transformer model, this invention can make full use of deep product feature vectors to gradually generate initial titles that accurately represent the core information of the product, providing a good foundation for subsequent optimization.

[0019] 7. This invention utilizes generative adversarial networks to stylize and enhance the creativity of the initial title sequence. Through a multi-task discriminator and adversarial training mechanism, it effectively improves the attractiveness and quality of the titles.

[0020] 8. This invention calls the constraint generation module to guide and filter the generation process in real time based on a multi-level compliance rule library, which can effectively reduce the risk of title compliance and ensure that the generated initial title sequence meets the specification requirements.

[0021] 9. This invention uses a multi-objective reward model agent to evaluate and iteratively optimize candidate titles. Based on multi-dimensional reward vectors and comprehensive reward values, it can continuously improve the quality of the output titles of the hybrid generation architecture using reinforcement learning algorithms, thereby obtaining better target titles.

[0022] 10. This invention adaptively and intelligently targets the target title based on channel characteristics, timing of delivery, and real-time user profiles, accurately matching the target advertising channel, user, and timing, thereby improving the accuracy and effectiveness of advertising delivery. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating an AI-based method for generating and optimizing game and short drama advertisement titles, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of an AI-based game and short drama advertisement title generation and optimization system provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0025] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0026] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0027] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0028] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0029] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0030] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0031] In this application embodiment, the entity executing the process includes a terminal device. This terminal device includes, but is not limited to, devices capable of executing the methods disclosed in this application, such as servers, computers, smartphones, and tablets. Figure 1 A flowchart illustrating an AI-based method for generating and optimizing game and short drama advertisement titles, as disclosed in an embodiment of the present invention, is shown below in detail: S101 collects multi-source heterogeneous data and processes it using a user interest transfer model to generate time-series user interest vectors.

[0032] In this embodiment, multi-source heterogeneous data is collected. This multi-source heterogeneous data encompasses multiple different types of data sources. User behavior data is obtained from game and short drama related platforms, including official game platforms, short drama streaming platforms, and social media platforms. On official game platforms, user behavior data for different games is collected, such as login duration, number of game level challenges, and in-game item purchase records. On short drama streaming platforms, user viewing behavior for various short dramas is recorded, such as viewing duration, number of episodes watched, whether a short drama was watched completely, and likes, comments, and shares. On social media platforms, user discussions on game and short drama related topics and related updates are collected. Simultaneously, basic user information data, such as age, gender, region, and occupation, is also collected. This information can be obtained from the information filled in by the user when registering on the platform or from legally authorized third-party data. In addition, market trend data is also an important source of information, including the latest developments in the gaming and short drama industries, popular gameplay, emerging short drama themes, and popular cultural elements. This data can be obtained through industry reports, news articles, and professional forums. By collecting data from these multiple channels, we ensure that the acquired data is comprehensive and diverse, providing a rich data foundation for subsequent processing.

[0033] The collected multi-source heterogeneous data is preprocessed, and then a user interest migration model is constructed. This model is based on deep learning technology and adopts a multi-layer neural network structure. The model input is the preprocessed multi-source heterogeneous data, including user behavior data, basic information data, and market trend data. In the model construction process, feature extraction is first performed on the input data. Convolutional neural networks (CNNs) are used to extract features from unstructured data such as images and text in user behavior data, such as extracting keywords and sentiment characteristics from user comments on short dramas. For structured data, such as user basic information and some behavioral data, fully connected neural networks are used for feature extraction. The extracted features of different types are fused to form a comprehensive feature vector. Then, recurrent neural networks (RNNs) and their variants, long short-term memory networks (LSTMs) or gated recurrent units (GRUs), are used to capture the changes in user interests over time. These network structures can process sequential data, model the user's interest state at different points in time, and predict the migration trend of user interests by learning the time series information in the user's historical behavior data. During model training, a large amount of historical data is used as the training set, and user interest annotation information is used as the supervision signal. The model parameters are continuously adjusted through the backpropagation algorithm, so that the model can accurately predict the migration of user interests.

[0034] Preprocessed data is processed using a trained user interest transfer model to generate time-series user interest vectors. The preprocessed data is then input into the trained model, which, based on user history, basic information, and market trends, and incorporating learned user interest transfer patterns, evaluates and quantifies user interests at different points in time. The model outputs a multi-dimensional vector, where each dimension represents the degree of user interest across different dimensions. The values ​​in the vector dynamically adjust over time, reflecting the shift in user interests. For example, a dimension might represent a user's interest in adventure games; as the user plays adventure games more frequently, the value of this dimension increases; conversely, if the user's interest in adventure games decreases, the value decreases. In this way, the generated time-series user interest vectors accurately and comprehensively reflect the dynamic changes in user interests, providing strong data support for the subsequent generation of game and short drama ad titles.

[0035] S102 constructs a knowledge graph based on product data and integrates time-series user interest vectors to perform node weighting, generating a deep product feature vector.

[0036] In this embodiment, product data is collected. For game products, data is collected from multiple dimensions, including game type (e.g., role-playing, strategy, casual), game visual style (realistic, cartoon, anime), gameplay features (single-player adventure, multiplayer competition, cooperative dungeons), game plot setting (fantasy, historical, science fiction), and in-game purchase item types (equipment, skins, items). For short drama products, data is collected including drama genre (romance, suspense, period drama), drama length (episode length, total length), drama plot highlights (plot twists, distinctive characters), drama production team information (director, screenwriter, lead actors, etc.), and data related to the drama streaming platform (e.g., number of views, likes, comments, etc.).

[0037] After collecting the raw product data, preprocessing is performed. Based on the preprocessed product data, a product knowledge graph is constructed. The entity types in the knowledge graph are determined. For game products, entities include the game itself, game type, visual style, gameplay, plot, in-app purchases, etc.; for short drama products, entities include the short drama, theme, duration, plot highlights, production team, and streaming platform, etc. The relationships between entities are clarified. For example, the relationship between a game and a game type is "belongs to," the relationship between a game and a visual style is "possesses," the relationship between a short drama and a theme is "belongs to," and the relationship between a short drama and a production team is "made by," etc.

[0038] Using graph database tools such as Neo4j, the defined entities and relationships are visualized and modeled. Each entity is treated as a node in the graph, and the relationships between entities are treated as edges between nodes, thus constructing a structured product knowledge network.

[0039] Temporally sequenced user interest vectors reflect users' varying degrees of interest in different aspects of games and short dramas at different times. For example, a user might have a higher interest in adventure games at one time and a higher interest in suspenseful short dramas at another. These vectors contain multi-dimensional information, each corresponding to a specific product feature or user interest point. These temporally sequenced user interest vectors are then incorporated into a constructed product knowledge graph. For each node in the knowledge graph, based on the correlation between the product feature it represents and the dimensions of the user interest vector, the corresponding value from the user interest vector is assigned to that node. For example, if a node in the knowledge graph represents "adventure game gameplay," and a dimension in the temporally sequenced user interest vector corresponds to "degree of interest in adventure game gameplay," then the value of that dimension is assigned to the "adventure game gameplay" node. In this way, each node in the knowledge graph carries user interest information from different times, achieving an initial fusion of user interests and product features.

[0040] The nodes in the knowledge graph are weighted based on temporal user interest vectors. Considering the time-sensitivity of user interests, recent user interests have a greater impact on product features; therefore, the values ​​in the temporal user interest vectors are adjusted by decay according to their time proximity. For example, user interest values ​​closer to the current time have higher weights, while user interest values ​​further back in time have lower weights.

[0041] For each node in the knowledge graph, a weighted value is calculated based on the adjusted user interest score and the node's importance within the product knowledge graph (e.g., some core product feature nodes may be more important than peripheral feature nodes). The weighted value can be calculated using a simple weighted summation method: multiply the user interest score by the node's importance coefficient and sum the results to obtain the final weighted value for each node. For example, a node representing a popular game mechanic, due to its inherent importance in the product knowledge graph and recent high user interest in that mechanic, will have a relatively large weighted value after weighted calculation.

[0042] After weighting the nodes, a deep product feature vector is generated based on the weighted product knowledge graph. All nodes in the knowledge graph are traversed, and the weighted value of each node and the product feature information it represents are extracted. This information is then integrated and encoded according to certain rules, such as classifying and combining different types of product feature information, and then sorting and quantifying them based on the weighted values.

[0043] The integrated and encoded information is converted into a multi-dimensional vector, which is the deep product feature vector. Each dimension of the vector corresponds to a specific product feature or combination of product features, and the value of the dimension reflects the importance of that feature after considering user interests. For example, one dimension of the deep product feature vector might represent "the product feature of combining popular gameplay that users have recently been paying attention to with suspenseful short dramas," and its value indicates the importance of this feature to users in the current market environment. The deep product feature vector generated in this way not only contains detailed information about the product itself but also fully considers the dynamic changes in user interests, enabling it to more accurately reflect the product's attractiveness and competitiveness in the market, and providing strong data support for subsequent ad headline generation.

[0044] S103 takes the deep product feature vector as input to a hybrid generative architecture that includes a Transformer model and a generative adversarial network to generate candidate titles, and introduces a constraint generation mechanism during the generation process to ensure title compliance.

[0045] In this embodiment, a hybrid generative architecture incorporating a Transformer model and a Generative Adversarial Network (GAN) is constructed. The Transformer model possesses powerful sequence modeling capabilities, enabling it to capture long-range dependencies and contextual information in the input data, giving it a significant advantage in processing data containing rich product information, such as deep product feature vectors. The GAN consists of a generator and a discriminator. The generator is responsible for generating titles, while the discriminator determines whether the generated titles are real or generated. Through adversarial training between the two, the quality and realism of the generated titles can be continuously improved.

[0046] In the architecture development process, the Transformer model is used as part of the Generative Adversarial Network (GAN) generator. Specifically, deep product feature vectors are used as input to the Transformer model. The Transformer model performs multi-head attention and feedforward neural network processing on the input vectors to deeply mine and integrate the product feature information in the vectors, outputting a feature representation with rich semantic information. This feature representation will serve as the initial input to the GAN generator, providing the foundation for subsequent title generation.

[0047] Simultaneously, a discriminator based on a generative adversarial network (GAN) is constructed. The discriminator employs a multi-layer neural network structure, taking as input generated titles (initially randomly generated, gradually becoming generator-generated as training progresses) and real advertisement title samples. The discriminator extracts and analyzes features from the input titles to determine whether a title is a real or generated sample, outputting a probability value representing the judgment. In this way, the generator and discriminator form a mutually adversarial and mutually reinforcing training system, continuously improving the quality of generated titles.

[0048] The deep product feature vector is input into the constructed hybrid generation architecture. The deep product feature vector contains detailed product information, such as game type, gameplay features, short drama themes, plot highlights, etc., and also incorporates time-series user interest information, reflecting users' interest preferences for the product at different times.

[0049] In the Transformer model, multi-head attention is performed on the input deep product feature vector. Multi-head attention mechanisms can focus on and extract information from the vector from different perspectives, capturing the complex relationships between product features. For example, for a role-playing game, a multi-head attention mechanism can simultaneously focus on the relationships between features such as the game's class system, equipment system, and storyline. After multi-head attention computation, the result is input into a feedforward neural network for further processing, resulting in a deeply processed feature representation.

[0050] The feature representation output by the Transformer model is used as input to a Generative Adversarial Network (GAN) generator, which then begins generating candidate titles. Based on the input feature representation, and combining previously learned language patterns and title generation rules, the generator progressively generates words or phrases, ultimately assembling them into a complete candidate title. During the generation process, the generator continuously references feedback from the discriminator to adjust its generation strategy, improving the quality and realism of the generated titles. For example, if the discriminator determines that the generated title differs significantly from the real title, the generator will adjust its internal parameters and change its generation method to make the generated title more closely resemble the style and content of a real title.

[0051] To ensure the compliance of generated candidate titles, a constraint generation mechanism is introduced during the generation process. This mechanism mainly includes two aspects: content compliance constraints and format specification constraints.

[0052] Regarding content compliance, a database of prohibited terms and sensitive topics has been established. The prohibited terms database contains various words that violate laws, regulations, ethical standards, and platform rules. The sensitive topics database covers topics that may cause controversy or have adverse effects. During the generation of candidate titles, each generated word or phrase is checked in real time. If it is found to match content in the prohibited terms database or sensitive topics database, the generation of titles in the current direction is immediately stopped, and the generation strategy is adjusted to avoid generating non-compliant content.

[0053] Regarding formatting guidelines, corresponding title formatting guidelines are established based on different advertising channels and requirements. For example, ad titles on social media platforms may have character limits and punctuation usage guidelines; ad titles on video platforms may need to highlight the key features of the video and use specific wording. When generating candidate titles, these formatting guidelines are strictly followed. For instance, if the limit for ad titles on social media platforms is 20 characters, the title generator will count the number of characters in real time during title generation. If the limit is exceeded, the title content will be automatically adjusted to ensure compliance with the formatting requirements.

[0054] This embodiment utilizes a hybrid generative architecture that includes a Transformer model and a generative adversarial network, combined with a constraint generation mechanism, to generate a large number of qualified, high-quality, and compliant candidate titles, providing a wealth of options for subsequent title evaluation and optimization.

[0055] S104 uses a multi-objective reward model agent to evaluate and iteratively optimize candidate titles to obtain the target title.

[0056] In this embodiment, a multi-objective reward model agent is constructed, which integrates multiple objective dimensions related to ad headline evaluation. These objective dimensions cover key indicators in the ad delivery process and aim to comprehensively and holistically evaluate the quality of candidate headlines.

[0057] Specifically, the following key objective dimensions are set: First, click-through rate (CTR) is a crucial indicator of an ad title's ability to attract user attention. A higher CTR means the title successfully entices users to click and view ad details, making it a key objective for evaluating title appeal. Second, conversion rate reflects the effectiveness of the ad title in guiding users to make actual purchases, downloads, or other conversions. It is a core indicator for evaluating the commercial value of a title, and improving conversion rates directly enhances the product's marketing effectiveness. Third, relevance is assessed by examining the degree to which the title aligns with the product's core features and the interests of the target users, ensuring the title accurately conveys product information and attracts the target audience. Fourth, compliance is ensured by complying with laws, regulations, ethical standards, and platform rules, avoiding prohibited words, sensitive topics, and other non-compliant content to guarantee the legality and security of ad placement.

[0058] To achieve the evaluation of these target dimensions, the multi-objective reward model agent employs deep learning algorithms to construct the evaluation model. This model is trained using a large amount of historical advertising data and user feedback data, learning the complex relationships between different target dimensions and title features. For example, the model can learn which word combinations and expressions are more likely to improve click-through rates and conversion rates, and how to ensure that titles are highly relevant to the product and users and comply with regulations. During training, the model parameters are continuously adjusted to improve the accuracy of the model's evaluation of each target dimension.

[0059] The generated candidate titles are input into the constructed multi-objective reward model agent, which performs an initial evaluation of each candidate title. During the evaluation process, the agent scores the candidate titles from multiple objective dimensions, including click-through rate, conversion rate, relevance, and compliance.

[0060] For click-through rate (CTR) evaluation, the agent analyzes factors such as word choice, expression, and sentiment in the title, and combines this with historical data on the CTR performance of similar titles to predict the potential CTR of the current candidate title. For example, if the title uses highly attractive words such as "exclusive" or "limited-time free," and the expression is concise and clear, the agent may award a higher CTR score.

[0061] In terms of conversion rate evaluation, the agent considers the relevance of the title to the product's core selling points and its ability to guide users to convert. If the title accurately highlights the product's unique advantages, such as innovative gameplay or engaging storylines in short dramas, and uses guiding language such as "Experience Now" or "Click to Download," the agent may predict that the title will have a high conversion rate.

[0062] Relevance assessment primarily examines the degree to which the title matches the interests of the target users and the characteristics of the product. The AI ​​agent analyzes user profile data and product knowledge graphs to determine whether the title accurately targets the user group and conveys the product's key information. For example, if the title targets young users, uses vocabulary and expressions that align with their interests, and matches the product's youthful positioning, the AI ​​agent will award a high relevance score.

[0063] The compliance assessment involves a comprehensive check of the title to ensure it does not contain prohibited words, sensitive topics, or content that violates platform rules. The AI ​​agent compares the title with a database of prohibited words and sensitive topics to determine its compliance and assign a corresponding score.

[0064] Based on the initial evaluation results, the multi-objective reward model agent determines the score of each candidate title across all objective dimensions. For titles with lower scores, the agent analyzes their shortcomings across each objective dimension and formulates corresponding optimization strategies. For example, if a candidate title scores low in click-through rate, the agent might suggest adjusting the title's vocabulary and wording, adding more appealing elements such as using interrogative sentences, exclamations, or creating suspense to improve its attractiveness. If the title performs poorly in conversion rate, the agent might suggest further highlighting the product's core selling points and using more guiding language to encourage user conversion.

[0065] During the optimization process, the agent iteratively improves the candidate titles. After each optimization, the updated title is input again into the multi-objective reward model agent for evaluation, and the optimization strategy is adjusted based on the new evaluation results. Through multiple iterations, the scores of the candidate titles on each objective dimension are gradually improved, thus continuously enhancing their overall quality.

[0066] After multiple rounds of iterative optimization, when the candidate title's score on each target dimension reaches the preset threshold or the quality improvement is no longer significant after a certain number of iterations, the title is considered to have been optimized to a better state and is determined as the target title.

[0067] The target title not only boasts high expected performance in click-through rate and conversion rate, attracting target users and guiding them to convert, but it is also highly relevant to the product and users, accurately conveying product information, while complying with laws, regulations, and platform rules to ensure the legality and security of ad placement. In this way, the multi-objective reward model agent can filter out the optimal target title from a large number of candidate titles, providing high-quality title support for subsequent ad placement.

[0068] S105, based on the channel characteristic knowledge base, the output of the delivery timing prediction model, and real-time user profiles, performs adaptive intelligent delivery to the target title.

[0069] In this embodiment, a channel characteristic knowledge base is constructed. This knowledge base covers various common advertising channels, such as social media platforms, video platforms, search engines, and other related advertising channels. For each channel, its unique formatting specifications are recorded in detail, such as the character limit for titles, whether special characters are supported, and layout requirements; the characteristics of the user group of this channel are analyzed, including age distribution, gender ratio, interests, and spending power; and successful experiences and common problems in advertising on this channel are summarized.

[0070] When placing an ad, the system quickly retrieves relevant information about the corresponding channel from the channel characteristic knowledge base based on the preset channel selection instructions. For example, if the ad is placed on the Douyin platform, the system will obtain information such as the character limit for Douyin titles, supported emoji types, suitable title styles (e.g., lively, fun, creative), and common interests of users on the platform regarding games and short dramas, such as popular game genres and popular short drama themes.

[0071] The ad placement timing prediction model is based on multi-source data, including historical ad placement data, user behavior data, and market trend data. It is trained and optimized using machine learning algorithms to accurately predict the effectiveness of ad placements at different times. The model outputs a series of suggestions regarding ad placement timing, including predictions for ad performance at different times of the day, for different dates within a week, and for special holidays and events.

[0072] After obtaining the output of the ad placement timing prediction model, the system performs in-depth analysis. For example, the analysis reveals that users are more active on social media and video platforms between 8 and 10 pm on weekends, and have a stronger interest in and willingness to spend on games and short dramas. Ad placement during these times is likely to achieve higher click-through rates and conversion rates. Conversely, during weekday lunch breaks and after-get off work hours, users search for game and short drama-related information more frequently on search engines, making search engine platforms suitable for ad placement. Based on these analytical results, the optimal ad placement times for different channels are determined.

[0073] The real-time user profile generation system collects real-time user behavior data, such as browsing history, search keywords, likes and comments, and dwell time, and combines this with basic user information (age, gender, region, etc.). Using data mining and machine learning techniques, it generates detailed user profiles in real time. These profiles accurately reflect a user's current interests, consumption habits, and activity levels. For example, the system might detect that a user has recently been frequently browsing adventure game content and has liked and commented extensively on adventure short dramas on social media. Furthermore, this user is more active on weekday evenings and weekend afternoons. Based on this information, the system can determine that the user has a high interest in adventure games and short dramas and is more likely to pay attention to advertising content during these time periods.

[0074] By integrating information from the knowledge base of integrated channel characteristics, the output of the timing prediction model, and real-time user profiles, the system formulates an adaptive intelligent delivery strategy.

[0075] For different channels, the target titles should be tailored to their formatting guidelines and user demographics. For example, when advertising on social media platforms, the titles should be adjusted to be more lively, engaging, and interactive, reflecting the style of the platform's users, while ensuring the title length remains within the specified limits. When advertising on search engine platforms, the titles should highlight keywords and concisely convey the core product information to improve search ranking and click-through rates.

[0076] Based on the timing prediction results, the adjusted target titles will be accurately delivered to the corresponding channels during the optimal delivery time period.

[0077] By combining real-time user profiles, personalized advertising can be implemented for user groups with different interests, consumption habits, and activity levels. For example, for users interested in adventure games and short dramas, targeted headlines related to adventure can be displayed during their real-time active periods; for users with higher spending power, headlines for high-end, premium games and short dramas can be displayed to improve the targeting and conversion rate of the ads.

[0078] During the campaign, the system continuously monitors ad performance metrics such as click-through rate (CTR), conversion rate, and impressions. Based on the monitoring results, the system adjusts the campaign strategy in real time, optimizing or replacing underperforming titles and increasing the frequency of ads for high-performing titles to maximize campaign effectiveness. Through the formulation and execution of this adaptive intelligent campaign strategy, it ensures that target titles are precisely delivered to different user groups across different channels and time periods, effectively improving click-through and conversion rates and enhancing product marketing results.

[0079] In some embodiments, step S101 above, which involves collecting multi-source heterogeneous data and processing it using a user interest migration model to generate a time-series user interest vector, specifically includes: Multi-source heterogeneous data containing text, visual data, and user behavior sequences are collected from multiple internet platforms. The multi-source heterogeneous data is then cleaned and multimodal fusion is performed with a user-centric approach to form a time-series sequence of user behavior. The user behavior time series is input into the user interest transfer model for training. The key behaviors in the user behavior time series are weighted and encoded through the time series attention mechanism to extract preliminary user interest features. During model training, an adversarial domain adaptive training strategy is used to enable the user interest transfer model to extract a generalized user interest representation that is independent of the specific data source platform from the initial user interest features, and output a time-series user interest vector.

[0080] In this embodiment, multi-source heterogeneous data, including text, visual data, and user behavior sequences, is collected extensively from multiple internet platforms. The collected heterogeneous data is cleaned to remove duplicates, errors, and invalid data. Multimodal fusion is performed with the user at the center, integrating the cleaned text, visual, and user behavior sequence data. For each user, relevant data from different platforms is correlated to form a complete user dataset. For example, a user's comments on a game on social media platforms are correlated with their behavior sequence of watching game commentary videos on video streaming platforms, combined with text data from reading game reviews on game information platforms, to form a multimodal dataset about the user's interest in the game. By processing all user data in this way, a user behavior time-series sequence is formed, which comprehensively and dynamically reflects changes in users' interest in games and short dramas at different points in time.

[0081] User behavior time-series sequences are input into a user interest transfer model for training. This model employs a temporal attention mechanism, which automatically identifies key behaviors within the user behavior time-series. During training, the model evaluates each behavior in each user behavior time-series, assigning different weights based on the behavior's contribution to the user's interest. For example, repeatedly watching a game's commentary video multiple times within a short period is given a higher weight, as it better reflects the user's strong interest in the game; while occasionally browsing information about a game is given a lower weight.

[0082] By using a temporal attention mechanism to weight and encode key behaviors, the model can extract preliminary user interest features. These preliminary user interest features include users' interest tendencies towards different games and short dramas at various points in time, but they are still influenced by specific data source platforms. For example, users' interest performance on social media platforms may differ from their interest performance on video streaming platforms because the user groups and usage scenarios of different platforms are different.

[0083] During model training, an adversarial domain-adaptive training strategy is employed. The aim of this strategy is to enable the user interest transfer model to extract a generalized user interest representation that is independent of the specific data source platform from the initial user interest features.

[0084] Specifically, a domain discriminator is set up to determine which data source platform the initial user interest features originate from. The user interest transfer model acts as a generator, aiming to generate generalized user interest representations that can deceive the domain discriminator, making it unable to accurately determine the platform from which these features originate. During training, the generator and the domain discriminator engage in adversarial training. The generator continuously optimizes the generated generalized user interest representations, making them even more difficult for the domain discriminator to identify the data source platform; the domain discriminator, on the other hand, continuously improves its discriminative ability, attempting to accurately determine the feature source.

[0085] Through this adversarial training, the user interest transfer model can ultimately output a temporal user interest vector. This vector not only contains information about changes in users' interests over time, but also has generalizability, is not limited by specific data source platforms, and can more accurately and comprehensively reflect users' true interests in games and short dramas, providing strong support for generating accurate ad titles in the future.

[0086] In some embodiments, step S102 above, which involves constructing a knowledge graph based on product data and integrating time-series user interest vectors for node weighting to generate a deep product feature vector, specifically includes: Based on the multimodal description data of the target product, natural language processing technology is used to automatically identify and extract product entities and their relationships, and construct a fine-grained product knowledge graph. The entity nodes in the fine-grained product knowledge graph are mapped to a semantic vector space to obtain a static vector representation of each entity node; Calculate the semantic correlation between the temporal user interest vector and the static vector representation of each entity node, and use the semantic correlation as the dynamic importance weight of each entity node corresponding to the target user. Dynamic importance weights are injected into the graph neural network, which then performs weighted information propagation and global aggregation on the fine-grained product knowledge graph, outputting a deep product feature vector that includes both objective product characteristics and subjective user interests.

[0087] In this embodiment, multimodal description data of the target product is collected, including the product introduction page on the product's official website, which contains detailed product function descriptions, features, and other textual information; product promotional videos, in which the visuals and narration can intuitively demonstrate the product's appearance, gameplay, storyline, etc.; and user reviews and feedback, including user comments and ratings on the product from major app stores and social media platforms.

[0088] Natural Language Processing (NLP) techniques are used to process these multimodal descriptive data. For text data, Named Entity Recognition (NAME) technology is employed to automatically identify product entities, such as character names, item names, and game modes in games, and main characters and key plot elements in short dramas. Simultaneously, relation extraction techniques are used to determine the relationships between these entities, such as skill associations between game characters and emotional relationships between characters in a storyline. For video data, key visual information is extracted using video content understanding technology, and speech recognition technology is combined to convert narration into text, further identifying entities and relationships within it. Through these processes, a fine-grained product knowledge graph is constructed. This graph uses nodes to represent product entities and edges to represent relationships between entities, comprehensively and meticulously presenting various aspects of product information.

[0089] The entity nodes in the constructed fine-grained product knowledge graph are mapped to a semantic vector space. This process is achieved using a pre-trained language model, which is trained on large-scale text data and can learn rich semantic information. The text description corresponding to each entity node is input into the pre-trained language model, and the model outputs a fixed-dimensional vector, which is the static vector representation of that entity node. This static vector representation can capture the semantic features of entity nodes, such as the personality traits of a game character or the theme and style of a short drama plot, and remains relatively stable without being affected by external factors.

[0090] This study calculates the semantic relevance between the temporal user interest vector and the static vector representations of each entity node. The temporal user interest vector reflects changes in user interest in games and short dramas over time, including the user's attention to and preference for different types and features of products. By calculating semantic relevance, the importance of each entity node to the target user can be measured. Specifically, methods such as cosine similarity are used to evaluate the similarity between two vectors. The higher the similarity, the more closely the entity node matches the user's interests, and the higher its importance in the user's mind. The calculated semantic relevance is used as the dynamic importance weight of each entity node corresponding to the target user. This weight is dynamically adjusted as user interests change, thus reflecting changes in the user's attention to different aspects of the product in a timely manner.

[0091] The calculated dynamic importance weights are then injected into the graph neural network. A graph neural network is a neural network model capable of processing graph-structured data, performing information propagation and aggregation operations on nodes and edges in the graph. After injecting the dynamic importance weights, the graph neural network weights the information in the fine-grained product knowledge graph according to the magnitude of the weights.

[0092] During information dissemination, each node receives information from connected nodes and weights this information based on dynamic importance weights. For example, if an entity node has a higher dynamic importance weight, it will have a greater impact on other nodes during information dissemination. Through multiple rounds of information dissemination, nodes can obtain more comprehensive information that better aligns with user interests.

[0093] Finally, the graph neural network globally aggregates the weighted information, integrating the information from each node into a unified vector—a deep product feature vector that contains both the product's objective characteristics and the user's subjective interests. This vector includes both the product's inherent attributes, such as the game type and difficulty level of a game, and the plot type and production quality of a short drama, as well as the user's personalized interests and preferences. This provides accurate and comprehensive product feature information for subsequent ad headline generation, making the generated headlines more aligned with the needs and interests of the target users.

[0094] In some embodiments, step S103 above, which involves inputting the deep product feature vector into a hybrid generation architecture comprising a Transformer model and a generative adversarial network to generate candidate titles, and introducing a constraint generation mechanism during the generation process to ensure title compliance, specifically includes: Input the deep product feature vector into the Transformer model to generate an initial title sequence that represents the core information of the product. The initial title sequence is input into the generator of the generative adversarial network for stylized rendering and creative enhancement to obtain an optimized title sequence; The constraint generation module is invoked to provide real-time guidance and filtering for word selection and sequence structure during the generation process based on preset compliance rules.

[0095] In this embodiment, the deep product feature vector is input into the Transformer model. The Transformer model has powerful sequence generation capabilities, enabling it to perform deep analysis on the input deep product feature vector. Its internal multi-layer self-attention mechanism and feedforward neural network structure can automatically capture the correlations and importance between different dimensions of information in the product feature vector.

[0096] During processing, the Transformer model progressively transforms the deep product feature vector into an initial title sequence representing the product's core information. For example, for a role-playing game, the deep product feature vector contains information such as the game's fantasy world view, the main characters' skill characteristics, and unique gameplay. The Transformer model analyzes and processes this information to generate an initial title sequence such as "Fantasy world, dazzling character skills, unique gameplay awaits your challenge," which initially summarizes the product's key selling points and core information.

[0097] The generated initial title sequence is input into the generator of a Generative Adversarial Network (GAN). The GAN consists of a generator and a discriminator; the generator's role is to stylize and creatively enhance the initial title sequence. Internally, the generator employs a complex neural network structure, enabling it to learn numerous stylistic characteristics of excellent advertising headlines, such as humor, conciseness, and emotional impact. Upon receiving the initial title sequence, the generator reorganizes and refines it based on the learned style patterns. For example, for the game's initial title mentioned above, the generator might optimize it to "Step into a fantastical world, control cool characters, unlock unique gameplay, and embark on a thrilling adventure," by adding more evocative words and vivid descriptions to make the title more attention-grabbing and enhance its creativity and appeal.

[0098] Throughout the title generation process, the constraint generation module is invoked. This module provides real-time guidance and filtering for word selection and sequence structure during the generation process based on preset compliance rules.

[0099] The pre-defined compliance rules cover multiple aspects, including legal and regulatory requirements, social ethics, and platform regulations. For example, in terms of word choice, the use of words containing inappropriate content is prohibited; in terms of sequence structure, the rules ensure that the title sentences are fluent, clear in meaning, and free from logical confusion or ambiguity.

[0100] The constraint generation module monitors and intervenes in real time at every step of the title generation process. When the generator selects a word that violates compliance rules, the constraint generation module immediately blocks its use and guides the generator to choose a more appropriate word. For example, if the generator originally intended to use a discriminatory term to describe a game character, the constraint generation module will recognize this and force the generator to replace it with a positive, ethically sound word. Furthermore, regarding sequence structure, if the generated title has grammatical errors or unclear logic, the constraint generation module will adjust it promptly to ensure that the final candidate titles are both attractive and creative, while fully complying with all compliance requirements. In this way, a large number of high-quality and compliant candidate titles can be generated efficiently, providing a rich selection for subsequent advertising campaigns.

[0101] Furthermore, the step of inputting the deep product feature vector into the Transformer model to generate an initial title sequence for representing the core information of the product specifically includes: The projection layer of the Transformer model maps the deep product feature vectors into conditional embedding vectors of the same dimension as the word vectors. At each time step, the generated partial title word sequence is converted into word vectors, and the word vectors are combined with the conditional embedding vectors and input into the encoder of the Transformer model. The encoder performs context encoding on the combined word vector and conditional embedding vector, and outputs the probability distribution of the next candidate word on the entire vocabulary. The generated word for the current time step is obtained by sampling according to the probability distribution, and the generated word is appended to a part of the title word sequence to update the input; Iteratively generate words for each time step until the sequence termination condition is met, and output the initial title sequence.

[0102] In this embodiment, the acquired deep product feature vectors are input into the projection layer of the Transformer model. The projection layer is a key part of the Transformer model for dimensionality transformation, and its function is to map the deep product feature vectors into conditional embedding vectors of the same dimension as the word vectors. Deep product feature vectors contain rich and complex information about the product, such as the game type, visual style, and core gameplay of a game, and the plot type, main character traits, and production highlights of a short drama. However, the dimension of these feature vectors may not be consistent with the dimension of the word vectors required for subsequent processing. The projection layer uses a specific linear transformation to convert these deep product feature vectors of different dimensions into conditional embedding vectors with the same dimension as the pre-trained word vectors. For example, if the pre-trained word vectors have a dimension of 512, the projection layer will also convert the deep product feature vectors into 512-dimensional conditional embedding vectors, enabling these vectors to be effectively combined with the word vectors in subsequent model processing, providing product-related conditional information for title generation.

[0103] At each time step in title generation, the already generated partial title word sequence is first transformed. Using a pre-trained word vector table, each word in the partial title word sequence is converted into a corresponding word vector. The word vector table is trained on large-scale text data; each word corresponds to a fixed-dimensional vector that captures the semantic information of the word.

[0104] Next, the transformed word vectors are combined with the conditional embedding vectors obtained from the previous mapping. This combination is not a simple concatenation, but rather a specific fusion method, such as weighted summation of the word vectors and conditional embedding vectors or further fusion through a fully connected layer after concatenation, allowing the word vectors to incorporate product-related conditional information. The fused vectors are then fed into the encoder of the Transformer model. The encoder is one of the core components of the Transformer model, enabling in-depth processing and analysis of the input vectors.

[0105] The Transformer model's encoder receives the combined word vectors and conditional embeddings and performs contextual encoding on them. Internally, the encoder consists of multiple identical layers stacked together, each containing a multi-head self-attention mechanism and a feedforward neural network.

[0106] Multi-head self-attention mechanisms enable models to focus on information at different positions in the input sequence, capturing long-distance dependencies between words. For example, when generating game ad titles, the model can simultaneously focus on the correlation between game type and core gameplay information from two different positions. Feedforward neural networks further perform non-linear transformations on the output of the self-attention mechanism, enhancing the model's expressive power.

[0107] Through the processing of the encoder, the model can fully understand the context information of the input sequence and output the probability distribution of the next candidate word over the entire vocabulary. This probability distribution reflects the likelihood of each word being the next generated word under the current context conditions. For example, when generating a title for a role-playing game, given the current context "a fantasy-style", the model may assign relatively high probabilities to words such as "character", "adventure", "magic", etc.

[0108] According to the probability distribution output by the encoder, a sampling strategy is adopted to obtain the generated word at the current time step. Common sampling strategies include greedy sampling, beam search sampling, etc. Greedy sampling selects the word with the highest probability as the generated word at each step. This method is simple and direct, but may fall into local optimality. Beam search sampling, on the other hand, retains a certain number of words with the highest probabilities at each step, continues to generate the subsequent sequence, and finally selects the optimal sequence from them. This method can generate higher-quality titles, but the computational cost is relatively large.

[0109] After obtaining the generated word at the current time step, append this generated word to the partial title word sequence, thereby updating the input sequence. For example, the initial partial title word sequence is empty. At the first time step, the word "a" is generated, then the updated partial title word sequence becomes "a"; at the second time step, based on "a", continue to generate the next word. Suppose "fantasy" is generated, then the partial title word sequence is updated to "a fantasy", and so on.

[0110] Iterate continuously according to the above steps. At each time step, generate a word and update the partial title word sequence until the sequence termination condition is met. The sequence termination condition can be the generation of a specific end symbol or reaching the preset maximum title length.

[0111] When the termination condition is satisfied, the current partial title word sequence becomes the complete initial title sequence. This initial title sequence can accurately represent the core information of the product. For example, for a strategy short drama, the generated initial title sequence may be "Strategy short drama, brain-burning plot, exciting duels waiting for you to watch". This sequence summarizes the type of the short drama, the plot characteristics, and the attraction, providing a basis for subsequent title optimization and advertising placement.

[0112] Furthermore, inputting the initial title sequence into the generator of the generative adversarial network for stylistic rendering and creative enhancement to obtain an optimized title sequence specifically includes: Input the initial title sequence into the generative adversarial network. The generator for sequence-to-sequence style rendering encodes the initial title sequence and fuses the deep product feature vector to retain the core information of the product; Extract predefined style vectors from a pre-defined high-performance ad title library, and use the predefined style vectors to stylize and decode the encoded initial title sequence to obtain a stylized initial title sequence. The stylized initial headline sequence is fed into a multi-task discriminator along with real high-performing ad headlines from the high-performing ad headline library. By performing multiple tasks, including authenticity discrimination and attractiveness assessment, corresponding discrimination signals are generated. The generator parameters are updated using the discriminative signal through an adversarial training mechanism, and the title sequence is optimized based on the updated generator output.

[0113] In this embodiment, the initial title sequence is input into the generator of a generative adversarial network. This generator has the capability for sequence-to-sequence style rendering, and its internal structure includes an encoder module. The encoder performs word-by-word analysis on the initial title sequence, converting each word into a corresponding word vector. It then uses a multi-layer neural network structure to capture the semantic relationships between words and the contextual information of the sequence, thereby encoding the initial title sequence into a fixed-dimensional vector representation.

[0114] Meanwhile, to ensure that the core information of the product is not lost during stylized rendering, a deep product feature vector is introduced into the generator. This deep product feature vector contains key information about the mobile game, such as game type (role-playing, strategy, etc.), visual style (realistic, cartoon, etc.), and core gameplay (competitive, simulation, etc.). Through a specific fusion mechanism, such as concatenating the encoded initial title sequence vector with the deep product feature vector and then performing a non-linear transformation through a fully connected layer, the two are fully integrated. This ensures that the title is stylized while accurately conveying the product's core selling points during subsequent generation. For example, for a cartoon-style strategy simulation mobile game with an initial title sequence of "A fun game," after encoding and fusing the deep product feature vector, the model can understand that it is a cartoon-style, strategy simulation mobile game, laying the foundation for subsequent stylized rendering.

[0115] Predefined style vectors are extracted from a pre-defined high-performing ad headline library. This library is carefully selected and organized, containing a large number of ad headlines that perform well in the market and possess unique styles. Through in-depth analysis of these headlines, style vectors representing different styles are extracted, such as humorous, grand, and concise.

[0116] The encoded initial title sequence is stylized and rewritten using extracted predefined style vectors. The decoder, another crucial module of the generator, receives both the encoded vectors and the predefined style vectors as input. During decoding, the decoder reorganizes and refines the initial title sequence based on the guidance of the predefined style vectors. For example, choosing a humorous style vector might rewrite "A fun game" as "This game is so fun it'll make you laugh out loud"; choosing a grand and majestic style vector might rewrite it as "A stunning mobile game, start your legendary journey." This method yields a stylized initial title sequence, making it more attractive and unique.

[0117] The stylized initial headline sequence is fed along with real high-performing ad headlines from a high-performing ad headline library into a multi-task discriminator. The multi-task discriminator is a key component of the generative adversarial network (GAN), capable of performing multiple tasks including truth / falsehood detection and attractiveness assessment.

[0118] In the authenticity discrimination task, the discriminator needs to determine whether the input headline is a genuine high-performing ad headline or a stylized headline generated by a generator. By learning the features of a large number of real and generated headlines, the discriminator can accurately identify the differences between the two and generate corresponding authenticity discrimination signals. For example, if the generated stylized headline differs significantly from the real headline in grammar, semantics, or style, the discriminator will give a lower authenticity score.

[0119] In the attractiveness assessment task, the discriminator evaluates the attractiveness of the headline. It analyzes factors such as vocabulary choice, expression, and emotional tone to determine whether the headline can attract the user's attention. For example, a headline with strong emotional connotations and vivid vocabulary may receive a higher attractiveness score. By performing these two tasks, the multi-task discriminator generates corresponding discriminative signals, which are used to guide the generator's parameter updates.

[0120] The generator's parameters are updated using the discriminant signal generated by the discriminator through an adversarial training mechanism. Adversarial training is the core idea of ​​generative adversarial networks (GANs), where the generator and discriminator continuously optimize their performance through adversarial competition. The generator aims to generate headlines that can deceive the discriminator, causing it to judge them as genuine, high-performing, and highly attractive ad headlines; while the discriminator aims to accurately distinguish between genuine and generated headlines and provide a reasonable attractiveness assessment.

[0121] During training, the generator adjusts its parameters based on feedback signals from the discriminator, such as adjusting the weights of the encoder and decoder, to improve the quality of the generated titles. Through multiple iterations of training, the generator's performance gradually improves, enabling it to generate more realistic and attractive titles. When the generator reaches a good performance state, an optimized title sequence is generated based on the updated generator output. These optimized title sequences retain the core information of the product while possessing a unique style and high attractiveness, effectively attracting user attention and improving ad click-through rates and conversion rates. For example, for the aforementioned cartoon-style strategy simulation mobile game, the final optimized title sequence might be "Cartoon Strategy Simulation Masterpiece, Start Your Joyful Adventure, Come and Play!" Such a title accurately conveys the game information while attracting users to click and download.

[0122] Furthermore, the constraint generation module, based on preset compliance rules, provides real-time guidance and filtering for word selection and sequence structure during the generation process, specifically including: Construct a multi-level compliance rule library that includes a lexical-level negative list, pattern-level regular expression rules, and a semantic-level compliance classifier; In each time step of word generation in the Transformer model, the constraint generation module is called to dynamically mask the probability distribution of candidate words output by the Transformer model based on the lexical negative list, and correct the illegal patterns in the title word sequence output by the Transformer model based on the pattern-level regularization rules. Input the currently generated title word sequence into the semantic-level compliance classifier to obtain a scoring signal used to characterize the sequence compliance risk; The scoring signal is used as a reward to update the parameters of the Transformer model through a policy gradient reinforcement learning algorithm, so as to guide the Transformer model to generate an initial header sequence with low compliance risk.

[0123] In this embodiment, a multi-level compliance rule library is constructed, comprising a lexical-level negative list, pattern-level regular expression rules, and a semantic-level compliance classifier. This is the foundation for ensuring title compliance. Specifically, a comprehensive lexical-level negative list is formed by collecting and organizing words explicitly prohibited by laws and regulations, industry-sensitive words, and platform-defined prohibited words. Common illegal title patterns are analyzed, such as exaggerated expressions and deceptive consumption patterns, and these patterns are defined using regular expressions to form pattern-level regular expression rules. A semantic-level compliance classifier is trained using a large-scale dataset of compliant and illegal titles. This classifier can understand the semantic information of titles and determine whether a title poses a violation risk. For example, for titles that subtly express illegal content, such as those using homophones or metaphors to spread harmful information, the semantic-level compliance classifier can accurately identify them by analyzing the semantic features of the title. During training, deep learning models, such as convolutional neural networks (CNNs) or recurrent neural networks (RNNs), are used to extract features and classify titles, continuously improving the accuracy and generalization ability of the classification.

[0124] At each time step of word generation in the Transformer model, the constraint generation module intervenes in the generation process in real time. At each time step, the Transformer model outputs a candidate word probability distribution, reflecting the likelihood of each word becoming the next generated word under the current context. The constraint generation module dynamically masks the candidate word probability distribution based on a lexical-level negative list. Specifically, it sets the probability values ​​of words in the negative list to a very small value (e.g., close to 0), ensuring that these words are almost never selected during sampling. As the title word sequence is gradually generated, the constraint generation module monitors the sequence in real time for any violations that conform to pattern-level regularization rules. Once a violation is detected, the title word sequence is immediately corrected.

[0125] The generated title word sequence is input into a semantic-level compliance classifier to obtain a scoring signal characterizing the sequence's compliance risk. The semantic-level compliance classifier performs a comprehensive semantic analysis of the input title word sequence, considering factors such as vocabulary, grammar, and context, to provide a compliance risk score. The score range can be set between 0 and 1; a lower score indicates a higher compliance risk, while a higher score indicates a more compliant title. For example, a title containing clearly non-compliant semantics might receive a lower score, such as 0.2, while a fully compliant title would receive a higher score, such as 0.9. This scoring signal will serve as an important basis for subsequent model parameter updates.

[0126] The scoring signal output by the semantic compliance classifier is used as a reward to update the parameters of the Transformer model through a policy gradient reinforcement learning algorithm, so as to guide the Transformer model to generate an initial title sequence with low compliance risk.

[0127] The basic idea of ​​policy gradient reinforcement learning is to adjust the model's parameters to increase the probability of generating high-reward (i.e., low-compliance-risk) headlines. In each training step, the Transformer model generates a sequence of headline words based on the current parameters and receives a corresponding compliance risk score as a reward. Then, based on the reward, the policy gradient algorithm is used to calculate the direction and magnitude of parameter updates, adjusting the model's parameters accordingly. For example, if a generated headline receives a high compliance score (reward), the model tends to retain the parameter settings used to generate that headline; if the headline has a low compliance score, the model adjusts the parameters to avoid generating similar high-risk headlines again in subsequent generation. Through continuous iterative training, the Transformer model can gradually learn the strategy for generating compliant headlines, thereby generating more initial headline sequences with low compliance risk.

[0128] In some embodiments, step S104 above, where the candidate title is evaluated and iteratively optimized by a multi-objective reward model agent to obtain the target title, specifically includes: Using historical advertising data, a multi-objective reward model agent is constructed and trained based on a neural network containing a shared encoding layer and a multi-task output head; The candidate titles to be evaluated are input into the trained multi-objective reward model agent to obtain the multi-dimensional reward vector corresponding to each candidate title, and the comprehensive reward value of each candidate title is calculated according to the preset business weights. Based on the comprehensive reward value, a reinforcement learning algorithm is used to calculate and update the model parameters of a hybrid generative architecture that includes a Transformer model and a generative adversarial network, so that the updated hybrid generative architecture can output target titles with higher comprehensive reward values.

[0129] In this embodiment, a large amount of historical advertising data is collected, covering various information on past advertising campaigns in the gaming and short drama industries. This includes key metrics such as click-through rates, conversion rates, and user dwell time for different types of ad titles across different channels, time periods, and target user groups, as well as corresponding user feedback information, such as user comments, likes, and shares. This rich historical data provides a solid foundation for building an accurate multi-objective reward model agent.

[0130] This paper constructs a multi-objective reward model agent based on a neural network architecture that includes a shared encoding layer and multi-task output heads. The shared encoding layer performs unified feature extraction and encoding on the input ad title and related data, converting the raw text information and other auxiliary information into vector representations that the computer can understand and process. The multi-task output heads output corresponding reward values ​​for different evaluation objectives, such as click-through rate prediction, conversion rate prediction, and user satisfaction evaluation. This architecture design can fully utilize the correlation between different tasks, improving the model's generalization ability and evaluation accuracy.

[0131] During the model training phase, the collected historical advertising data is divided into training, validation, and test sets. The training set is used to train the constructed multi-objective reward model agent. By continuously adjusting the model's parameters, the prediction results on the training set gradually approach the actual labeled reward values. During training, the validation set is used to monitor and evaluate the model's performance in real time to prevent overfitting. Once the model's performance on the validation set reaches a stable and satisfactory state, the test set is used for final evaluation, ensuring that the model can accurately and reliably evaluate advertising titles in practical applications. For example, based on the historical advertising data of the aforementioned strategy mobile games and urban romance short dramas, after multiple rounds of training and optimization, the multi-objective reward model agent can accurately predict the performance of different titles in various scenarios.

[0132] Once the multi-objective reward model agent is trained, candidate titles to be evaluated are input into the model. These candidate titles are generated using a hybrid generative architecture that incorporates a Transformer model and a Generative Adversarial Network, covering titles of various styles and expressions. The shared encoding layer of the multi-objective reward model agent extracts and encodes features for each candidate title, converting it into a vector representation.

[0133] Then, the multi-task output head calculates the reward value for each candidate title based on different evaluation objectives, such as click-through rate, conversion rate, and user satisfaction, ultimately generating a multi-dimensional reward vector for each candidate title. For example, for a candidate title "This strategy mobile game will take you on a thrilling journey," the multi-objective reward model agent might output a vector containing multiple dimensions such as click-through rate reward value, conversion rate reward value, and user satisfaction reward value, such as [0.8, 0.7, 0.9], representing the expected performance of the title in terms of click-through rate, conversion rate, and user satisfaction, respectively.

[0134] Next, the multi-dimensional reward vector is calculated based on preset business weights to obtain the comprehensive reward value for each candidate title. The business weights are set according to the actual business needs and goals of the game and short drama industry. For example, if the current focus is on improving the click-through rate of ads, then the weight of the click-through rate reward value will be set higher; if the focus is on long-term user retention and conversion, then the weight of the conversion rate reward value will be increased accordingly.

[0135] Based on the calculated comprehensive reward value for each candidate title, a reinforcement learning algorithm is used to update the parameters of the hybrid generative architecture, which includes a Transformer model and a generative adversarial network. The basic idea of ​​reinforcement learning is to guide the model's learning and optimization through a reward mechanism, enabling the model to output titles that achieve higher comprehensive reward values ​​in subsequent generation processes.

[0136] In reinforcement learning, the hybrid generative architecture is viewed as an agent, where the generated candidate titles represent the agent's actions, and the overall reward value is the feedback reward for these actions. The agent continuously adjusts its policy, i.e., the model parameters, based on the reward signals to maximize future cumulative rewards. For example, if a candidate title receives a high overall reward value, the reinforcement learning algorithm will increase the probability of generating similar titles; conversely, if a title receives a low overall reward value, the algorithm will reduce the likelihood of generating that type of title.

[0137] Through multiple rounds of iterative optimization, the model parameters of the hybrid generation architecture are continuously updated and adjusted, enabling the updated architecture to output target titles that achieve higher overall reward values. These target titles not only perform exceptionally well in key metrics such as click-through rate and conversion rate, but also better meet user interests and needs, highlighting the product's core selling points. For example, after multiple iterations and optimizations, the final generated target title for the aforementioned strategy mobile game might be "Super-exciting strategy mobile game, thrilling battles, leading you to dominate the entire server!" This title achieved a high overall reward value in the evaluation of the multi-objective reward model agent, effectively attracting the attention of target users and improving the advertising performance.

[0138] In some embodiments, step S105 above, which involves performing adaptive intelligent delivery to the target title based on the channel characteristic knowledge base, the output of the delivery timing prediction model, and the real-time user profile, specifically includes: Based on the target advertising channel identifier, query the pre-built channel characteristic knowledge base to obtain channel characteristics, including the format specifications of the target advertising channel, the user profile of the target group, and the historical performance characteristics. Input the real-time user profile and historical behavior time series data of the target users into the preset delivery timing prediction model to obtain a timing score that represents the suitability of delivery at the current moment. The target title to be advertised, channel characteristics, timing score, and real-time user profile are all input into the dynamic decision engine. Through a multi-objective decision function, the expected utility value of the target title for the target advertising channel, target user, and current timing is calculated comprehensively. Based on the expected utility value, the target title is adaptively delivered to the target users through the target advertising channels.

[0139] In this embodiment, the target advertising channel is clearly identified. For example, if an advertisement is to be placed on a well-known social media platform, that platform is the target advertising channel. Then, a pre-built channel characteristic knowledge base is queried based on this identifier. The channel characteristic knowledge base is constructed through extensive data collection and analysis in the early stages, and it records detailed information about each advertising channel. For the aforementioned social media platform, its format specifications can be obtained from the channel characteristic knowledge base, such as the title being limited to 20-30 characters and requiring the inclusion of specific hashtags; user profiles, such as the platform's users being primarily young people with a high interest in fresh and exciting content, and a relatively large proportion of female users; and historical performance characteristics, such as the title style of similar game and short drama advertisements previously placed on the platform being lively and suspenseful, and the type of title with high click-through and conversion rates. These channel characteristics provide a foundation for subsequent precise targeting.

[0140] The real-time user profiles and historical behavioral time-series data of the target users are input into a pre-defined ad placement timing prediction model. The real-time user profile includes basic user information such as age, gender, and location, as well as current interests and preferences, such as frequent recent searches for role-playing games and suspense / mystery short dramas. Historical behavioral time-series data records user activity at different times over a period of time, such as higher activity levels between 8-10 PM, during which time click-through rates and conversion rates for ads are also relatively high. By analyzing and learning from this data, the ad placement timing prediction model can predict the suitability of ad placement at the current moment and output a timing score to characterize the suitability. For example, if the model calculates a timing score of 0.8 for the current moment (with a score range of 0-1, closer to 1 indicating higher suitability), it indicates that the current moment is a relatively suitable time to place ads.

[0141] The target title to be advertised, the acquired channel characteristics, the timing score, and the real-time user profile are all input into the dynamic decision engine. The dynamic decision engine is an intelligent system built on a multi-objective decision function that comprehensively considers the impact of multiple factors on ad performance. The multi-objective decision function combines information such as the format specifications of the target advertising channel, the user profile and historical performance characteristics, the current timing score, and the real-time interests and preferences of the target users to comprehensively calculate the expected utility value of the target title for the target advertising channel, target users, and current timing. For example, for the target title of the role-playing game mentioned above, "Thrilling role-playing, embark on a mysterious adventure," combined with the channel characteristics of the social media platform (mostly young and female users who like new and exciting content), a timing score of 0.8, and the real-time profile of the target users (young women interested in role-playing games), the dynamic decision engine calculates through the multi-objective decision function that the expected utility value of this title in the current situation is 0.75 (the utility value range is set from 0 to 1, with higher values ​​indicating better expected performance).

[0142] Based on the calculated expected utility value, the target title is adaptively delivered to the target users through the target advertising channels. If the expected utility value is high, it means that delivering the target title to the target users and on the target advertising channels at the current moment can achieve good advertising results, and the delivery operation will be executed immediately. For example, if the calculated expected utility value is 0.75, reaching the preset delivery threshold (assuming the delivery threshold is 0.7), then the target title "Thrilling role-playing, embark on a mysterious adventure" will be delivered to the target young female users through social media platforms. At the same time, the system will monitor the performance of the advertising in real time, such as click-through rate and conversion rate, so as to further optimize and adjust the delivery strategy based on the actual results. If the expected utility value does not reach the delivery threshold, the delivery will not be carried out for the time being, and the system will wait for a more suitable time or further optimize the target title before delivery. In this way, adaptive intelligent delivery of target titles is achieved, improving the accuracy and effectiveness of advertising.

[0143] Reference Figure 2 An embodiment of the present invention provides an AI-based system 2 for generating and optimizing game and short drama advertisement titles. The system 2 specifically includes: Data acquisition module 201 is used to collect multi-source heterogeneous data and process it using a user interest transfer model to generate time-series user interest vectors. The vector weighting module 202 is used to construct a knowledge graph based on product data and integrate time-series user interest vectors to perform node weighting and generate deep product feature vectors. The title generation module 203 is used to input the deep product feature vector into a hybrid generation architecture containing a Transformer model and a generative adversarial network to generate candidate titles, and introduces a constraint generation mechanism during the generation process to ensure title compliance. The title optimization module 204 is used to evaluate and iteratively optimize candidate titles through a multi-objective reward model agent to obtain the target title; The title delivery module 205 is used to perform adaptive intelligent delivery of target titles based on the channel characteristic knowledge base, the output of the delivery timing prediction model, and real-time user profiles.

[0144] It is understandable that, such as Figure 1 The content shown in the embodiments of the AI-based game and short drama ad title generation and optimization method is applicable to the embodiments of this AI-based game and short drama ad title generation and optimization system. The specific functions implemented by the embodiments of this AI-based game and short drama ad title generation and optimization system are as follows: Figure 1 The illustrated method for generating and optimizing game and short drama ad titles based on AI is the same as the one shown above, and achieves the same beneficial effects. Figure 1 The beneficial effects achieved by the AI-based game and short drama ad title generation and optimization method shown in the embodiment are the same.

[0145] It should be noted that the information interaction and execution process between the above systems are based on the same concept as the method embodiments of the present invention. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.

[0146] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0147] Reference Figure 3The present invention also provides a computer device 3, including: a memory 302 and a processor 301, and a computer program 303 stored on the memory 302. When the computer program 303 is executed on the processor 301, it implements the AI-based game and short drama advertisement title generation and optimization method as described in any of the above methods.

[0148] The computer device 3 may be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device 3 may include, but is not limited to, a processor 301 and a memory 302. Those skilled in the art will understand that... Figure 3 The computer device 3 is merely an example and does not constitute a limitation on the computer device 3. It may include more or fewer components than shown in the figure, or combine certain components, or different components, such as input / output devices, network access devices, etc.

[0149] The processor 301 may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0150] In some embodiments, the memory 302 may be an internal storage unit of the computer device 3, such as a hard disk or memory of the computer device 3. In other embodiments, the memory 302 may be an external storage device of the computer device 3, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 3. Furthermore, the memory 302 may include both internal and external storage units of the computer device 3. The memory 302 is used to store the operating system, applications, boot loader, data, and other programs, such as the program code of the computer program. The memory 302 can also be used to temporarily store data that has been output or will be output.

[0151] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the AI-based method for generating and optimizing game and short drama advertisement titles as described in any of the above methods.

[0152] In this embodiment, 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, all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0153] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0154] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0155] In the embodiments disclosed in this application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling or direct coupling or communication connection may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0156] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

Claims

1. A method for generating and optimizing game and short drama ad titles based on AI, characterized in that, The method specifically includes: Collect multi-source heterogeneous data and process it using a user interest transfer model to generate time-series user interest vectors; A knowledge graph is built based on product data, and node weighting is performed by integrating time-series user interest vectors to generate deep product feature vectors. The deep product feature vector is input into a hybrid generative architecture that includes a Transformer model and a generative adversarial network to generate candidate titles, and a constraint generation mechanism is introduced during the generation process to ensure that the titles comply with regulations. The candidate titles are evaluated and iteratively optimized using a multi-objective reward model agent to obtain the target title. Based on the channel characteristic knowledge base, the output of the timing prediction model, and real-time user profiles, adaptive intelligent targeting is performed on the target titles.

2. The method according to claim 1, characterized in that, The process of collecting multi-source heterogeneous data and processing it using a user interest transfer model to generate time-series user interest vectors specifically includes: Multi-source heterogeneous data containing text, visual data, and user behavior sequences are collected from multiple internet platforms. The multi-source heterogeneous data is then cleaned and multimodal fusion is performed with a user-centric approach to form a time-series sequence of user behavior. The user behavior time series is input into the user interest transfer model for training. The key behaviors in the user behavior time series are weighted and encoded through the time series attention mechanism to extract preliminary user interest features. During model training, an adversarial domain adaptive training strategy is used to enable the user interest transfer model to extract a generalized user interest representation that is independent of the specific data source platform from the initial user interest features, and output a time-series user interest vector.

3. The method according to claim 1, characterized in that, The process of constructing a knowledge graph based on product data and integrating time-series user interest vectors for node weighting to generate a deep product feature vector specifically includes: Based on the multimodal description data of the target product, natural language processing technology is used to automatically identify and extract product entities and their relationships, and construct a fine-grained product knowledge graph. The entity nodes in the fine-grained product knowledge graph are mapped to a semantic vector space to obtain a static vector representation of each entity node; Calculate the semantic relevance between the temporal user interest vector and the static vector representation of each entity node, and use the semantic relevance as the dynamic importance weight of each entity node corresponding to the target user; Dynamic importance weights are injected into the graph neural network, which then performs weighted information propagation and global aggregation on the fine-grained product knowledge graph, outputting a deep product feature vector that includes both objective product characteristics and subjective user interests.

4. The method according to claim 1, characterized in that, The process involves inputting deep product feature vectors into a hybrid generative architecture that includes a Transformer model and a generative adversarial network to generate candidate titles. A constraint generation mechanism is introduced during the generation process to ensure title compliance. Specifically, this includes: Input the deep product feature vector into the Transformer model to generate an initial title sequence that represents the core information of the product. The initial title sequence is input into the generator of the generative adversarial network for stylized rendering and creative enhancement to obtain an optimized title sequence; The constraint generation module is invoked to provide real-time guidance and filtering for word selection and sequence structure during the generation process based on preset compliance rules.

5. The method according to claim 4, characterized in that, The step of inputting the deep product feature vector into the Transformer model to generate an initial title sequence for representing the core information of the product specifically includes: The projection layer of the Transformer model maps the deep product feature vectors into conditional embedding vectors of the same dimension as the word vectors. At each time step, the generated partial title word sequence is converted into word vectors, and the word vectors are combined with the conditional embedding vectors and input into the encoder of the Transformer model. The encoder performs context encoding on the combined word vector and conditional embedding vector, and outputs the probability distribution of the next candidate word on the entire vocabulary. The generated word for the current time step is obtained by sampling according to the probability distribution, and the generated word is appended to a part of the title word sequence to update the input; Iteratively generate words for each time step until the sequence termination condition is met, and output the initial title sequence.

6. The method according to claim 4, characterized in that, The process of inputting the initial title sequence into the generator of a generative adversarial network for stylized rendering and creative enhancement to obtain an optimized title sequence specifically includes: The initial title sequence is input into a generative adversarial network, encoded by a generator for sequence-to-sequence style rendering, and fused with deep product feature vectors to retain core product information. Extract predefined style vectors from a pre-defined high-performance ad title library, and use the predefined style vectors to stylize and decode the encoded initial title sequence to obtain a stylized initial title sequence. The stylized initial headline sequence is fed into a multi-task discriminator along with real high-performing ad headlines from the high-performing ad headline library. By performing multiple tasks, including authenticity discrimination and attractiveness assessment, corresponding discrimination signals are generated. The generator parameters are updated using the discriminative signal through an adversarial training mechanism, and the title sequence is optimized based on the updated generator output.

7. The method according to claim 5, characterized in that, The call constraint generation module provides real-time guidance and filtering for word selection and sequence structure during the generation process based on preset compliance rules, specifically including: Construct a multi-level compliance rule library that includes a lexical-level negative list, pattern-level regular expression rules, and a semantic-level compliance classifier; In each time step of word generation in the Transformer model, the constraint generation module is called to dynamically mask the probability distribution of candidate words output by the Transformer model based on the lexical negative list, and correct the illegal patterns in the title word sequence output by the Transformer model based on the pattern-level regularization rules. Input the currently generated title word sequence into the semantic-level compliance classifier to obtain a scoring signal used to characterize the sequence compliance risk; The scoring signal is used as a reward to update the parameters of the Transformer model through a policy gradient reinforcement learning algorithm, so as to guide the Transformer model to generate an initial header sequence with low compliance risk.

8. The method according to claim 1, characterized in that, The process of evaluating and iteratively optimizing candidate titles using a multi-objective reward model agent to obtain the target title specifically includes: Using historical advertising data, a multi-objective reward model agent is constructed and trained based on a neural network containing a shared encoding layer and a multi-task output head; The candidate titles to be evaluated are input into the trained multi-objective reward model agent to obtain the multi-dimensional reward vector corresponding to each candidate title, and the comprehensive reward value of each candidate title is calculated according to the preset business weights. Based on the comprehensive reward value, a reinforcement learning algorithm is used to calculate and update the model parameters of a hybrid generative architecture that includes a Transformer model and a generative adversarial network, so that the updated hybrid generative architecture can output target titles with higher comprehensive reward values.

9. The method according to claim 1, characterized in that, The process of adaptive intelligent targeting of target titles based on channel characteristic knowledge base, delivery timing prediction model output, and real-time user profiles specifically includes: Based on the target advertising channel identifier, query the pre-built channel characteristic knowledge base to obtain channel characteristics, including the format specifications of the target advertising channel, the user profile of the target group, and the historical performance characteristics. Input the real-time user profile and historical behavior time series data of the target users into the preset delivery timing prediction model to obtain a timing score that represents the suitability of delivery at the current moment. The target title to be advertised, channel characteristics, timing score, and real-time user profile are all input into the dynamic decision engine. Through a multi-objective decision function, the expected utility value of the target title for the target advertising channel, target user, and current timing is calculated comprehensively. Based on the expected utility value, the target title is adaptively delivered to the target users through the target advertising channels.

10. An AI-based system for generating and optimizing game and short drama advertisement titles, characterized in that, The system specifically includes: The data acquisition module is used to collect multi-source heterogeneous data and process it using a user interest transfer model to generate time-series user interest vectors. The vector weighting module is used to build a knowledge graph based on product data and integrate time-series user interest vectors to weight nodes and generate deep product feature vectors. The title generation module is used to input deep product feature vectors into a hybrid generation architecture that includes a Transformer model and a generative adversarial network to generate candidate titles, and introduces a constraint generation mechanism during the generation process to ensure title compliance. The title optimization module is used to evaluate and iteratively optimize candidate titles through a multi-objective reward model agent to obtain the target title; The title delivery module is used to perform adaptive intelligent delivery of target titles based on the channel characteristic knowledge base, the output of the delivery timing prediction model, and real-time user profiles.

Images