Adjustment method and device for putting cross-platform advertisement materials, electronic equipment and medium

Abstract

The invention relates to the technical field of adjustment of cross-platform advertisement material putting, and discloses an adjustment method and device for cross-platform advertisement material putting, electronic equipment and a medium, and the method comprises the steps: generating creative variants in batches by using a multi-modal large language model, shortening the material creation time, greatly improving the production efficiency and diversity, and reducing the production cost. Cultural compliance filtering is automatically completed through a preset rule base, the over-auditing rate is greatly improved, the compliance risk is reduced, an innovative cross-platform adaptation mechanism can intelligently convert core creativity into a form conforming to different platform preferences, efficient reuse of materials is achieved, a large amount of repeated manufacturing cost is saved, and finally, the method is suitable for popularization and application. And performing dynamic strategy adjustment on the real-time feedback data based on the reinforcement learning model. The method has the beneficial effects that the continuous autonomous optimization of the putting effect is realized, the advertisement click rate, the conversion rate and the overall return on investment are remarkably improved, and the end-to-end intelligent advertisement operation capability is formed.

Description

Technical Field

[0001] This invention relates to the field of adjustment technology for cross-platform advertising creatives, and more particularly to a method, apparatus, electronic device, and medium for adjusting cross-platform advertising creatives. Background Technology

[0002] Digital advertising faces the complex challenges of operating across multiple platforms. Different platforms exhibit significant differences in content formats, user interaction patterns, and algorithmic recommendation logic, making it difficult to directly reuse the same ad creative across platforms. To reach audiences on multiple platforms, advertisers often need to conduct manual creative design and a series of controlled experiments for each platform. This process is time-consuming, costly, and difficult to scale. Manual review is also prone to oversights, leading to operational risks. Furthermore, traditional optimization methods rely on periodic reviews and manual adjustments, resulting in delayed responses and an inability to capture dynamic market feedback in real time to maximize advertising effectiveness. Summary of the Invention

[0003] Based on this, it is necessary to address the existing issue of adjusting cross-platform advertising creatives, and propose a method, device, electronic device, and medium for adjusting cross-platform advertising creatives.

[0004] A method for adjusting cross-platform advertising creatives, the method comprising: Based on the specified advertising campaign objectives, obtain the corresponding historical campaign data; The historical delivery data is input into a preset multimodal large language model to generate multiple initial advertising materials containing text, images and video elements, thus obtaining an initial set of advertising variants; Based on a preset regional cultural rule base, the initial set of advertising variants is filtered for compliance to obtain a set of compliant advertising variants; Obtain multiple target advertising platforms to be advertised, and perform cross-platform adaptation of the initial advertising creatives in the set of compliant advertising variants to obtain the target advertising creatives corresponding to each initial advertising creative; The target advertising creatives are deployed on the target advertising platforms, and real-time deployment feedback data is collected. Based on the real-time delivery feedback data, the delivery strategy for each target advertising creative is dynamically adjusted through a reinforcement learning model.

[0005] Furthermore, the step of inputting the historical delivery data into a preset multimodal large language model to generate multiple initial advertising materials includes: Extract multimodal feature combinations that are associated with the preset positive delivery effect from the historical delivery data; The multimodal feature combination is input into a preset multimodal large language model as a constraint condition for generating initial advertising materials; The multimodal large language model generates multiple initial advertising materials based on the constraints.

[0006] Furthermore, the step of performing compliance filtering on the initial set of ad variants based on a preset regional cultural rule base to obtain a set of compliant ad variants includes: Determine the target region associated with the specified advertising delivery objective; Obtain the regional restriction rules and data compliance rules corresponding to the target region from the preset regional culture rule base; Based on the region limitation rules and the data compliance rules, conflicting visual elements or text descriptions in the initial set of ad variants are identified and filtered to obtain a set of compliant ad variants.

[0007] Further, the steps of acquiring multiple target advertising platforms to be deployed, and cross-platform adapting the initial ad creatives in the compliant ad variant set to obtain the target ad creatives corresponding to each initial ad creative, include: Obtain multiple target advertising platforms to be deployed; Obtain the platform feature vector for each of the target advertising platforms; wherein the platform feature vector is used to characterize its preference weights for text, static images, dynamic videos, and audio elements; Based on the platform feature vector, the platform's general semantic features and platform-specific style features in the initial advertising material are separated through a preset feature decoupling network; Based on the platform feature vector of the target platform, the platform-specific style features are remapped and combined with the platform's general semantic features to reconstruct and generate the target advertising material.

[0008] Furthermore, the step of dynamically adjusting the delivery strategy for each target advertising creative based on the real-time delivery feedback data using a reinforcement learning model includes: Construct a multi-objective reward function; The real-time delivery feedback data is input into the target reward function to obtain a comprehensive reward value; wherein, the real-time delivery feedback data includes at least two of the following: click-through rate, conversion rate, user dwell time, and interaction rate; Using the comprehensive reward value as the optimization objective, a near-end strategy optimization algorithm is used to dynamically adjust the allocation of the advertising budget and display frequency of each target advertising creative on different target advertising platforms.

[0009] Furthermore, after the steps of delivering each of the target advertising creatives on each of the target advertising platforms and collecting real-time delivery feedback data, the method further includes: Determine whether the cumulative amount of real-time delivery feedback data collected has reached the preset amount; If the amount of real-time delivery feedback data collected is greater than the preset amount, then the policy network parameters of the reinforcement learning model are incrementally updated. After the incremental update is completed, the number of accumulated real-time delivery feedback data will be reset to zero.

[0010] Furthermore, after the steps of delivering each of the target advertising creatives on each of the target advertising platforms and collecting real-time delivery feedback data, the method further includes: The real-time delivery feedback data is used as new historical delivery data and updated in the database used to obtain historical delivery data.

[0011] An adjustment device for delivering cross-platform advertising creatives, the device comprising: The acquisition module is used to obtain the corresponding historical advertising data based on the specified advertising target. The generation module is used to input the historical delivery data into a preset multimodal large language model to generate multiple initial advertising materials containing text, images and video elements, thus obtaining an initial set of advertising variants; The filtering module is used to perform compliance filtering on the initial set of ad variants based on a preset regional cultural rule base to obtain a set of compliant ad variants; The adaptation module is used to obtain multiple target advertising platforms to be deployed, and to perform cross-platform adaptation of the initial advertising materials in the set of compliant advertising variants to obtain the target advertising materials corresponding to each initial advertising material; The delivery module is used to deliver each of the target advertising materials to each of the target advertising platforms and collect real-time delivery feedback data. The adjustment module is used to dynamically adjust the delivery strategy of each target advertising creative based on the real-time delivery feedback data through a reinforcement learning model.

[0012] An electronic device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps: Based on the specified advertising campaign objectives, obtain the corresponding historical campaign data; The historical delivery data is input into a preset multimodal large language model to generate multiple initial advertising materials containing text, images and video elements, thus obtaining an initial set of advertising variants; Based on a preset regional cultural rule base, the initial set of advertising variants is filtered for compliance to obtain a set of compliant advertising variants; Obtain multiple target advertising platforms to be advertised, and perform cross-platform adaptation of the initial advertising creatives in the set of compliant advertising variants to obtain the target advertising creatives corresponding to each initial advertising creative; The target advertising creatives are deployed on the target advertising platforms, and real-time deployment feedback data is collected. Based on the real-time delivery feedback data, the delivery strategy for each target advertising creative is dynamically adjusted through a reinforcement learning model.

[0013] A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the following steps: Based on the specified advertising campaign objectives, obtain the corresponding historical campaign data; The historical delivery data is input into a preset multimodal large language model to generate multiple initial advertising materials containing text, images and video elements, thus obtaining an initial set of advertising variants; Based on a preset regional cultural rule base, the initial set of advertising variants is filtered for compliance to obtain a set of compliant advertising variants; Obtain multiple target advertising platforms to be advertised, and perform cross-platform adaptation of the initial advertising creatives in the set of compliant advertising variants to obtain the target advertising creatives corresponding to each initial advertising creative; The target advertising creatives are deployed on the target advertising platforms, and real-time deployment feedback data is collected. Based on the real-time delivery feedback data, the delivery strategy for each target advertising creative is dynamically adjusted through a reinforcement learning model.

[0014] The beneficial effects of this invention are as follows: By using a multimodal large language model to generate creative variations in batches, the creation time of materials is shortened, greatly improving production efficiency and diversity. By automatically completing cultural compliance filtering through a preset rule base, the approval rate is significantly improved and compliance risks are reduced. The innovative cross-platform adaptation mechanism can intelligently transform core creative ideas into forms that conform to the preferences of different platforms, realizing efficient reuse of materials and saving a lot of repetitive production costs. Finally, by relying on reinforcement learning models to dynamically adjust strategies based on real-time feedback data, continuous autonomous optimization of the campaign effect is achieved, significantly improving ad click-through rate, conversion rate and overall return on investment, forming an end-to-end intelligent advertising operation capability. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] in: Figure 1 This is an application environment diagram of a method for adjusting cross-platform advertising creatives in one embodiment; Figure 2 This is a flowchart of a method for adjusting cross-platform advertising creatives in one embodiment; Figure 3 A structural block diagram of an adjustment device for delivering cross-platform advertising creatives in one embodiment; Figure 4 This is a structural block diagram of an electronic device in one embodiment. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Figure 1 This is a diagram illustrating the adjusted application environment for delivering cross-platform advertising creatives in one embodiment. (Refer to...) Figure 1 This method for adjusting cross-platform advertising creatives is applied to a system for adjusting cross-platform advertising creatives. This system includes a terminal 110 and a server 120. The terminal 110 and server 120 are connected via a network. The terminal 110 can be a desktop terminal or a mobile terminal; a mobile terminal can be at least one of a mobile phone, tablet, or laptop. The server 120 can be a standalone server or a server cluster consisting of multiple servers. The terminal 110 is used to obtain corresponding historical delivery data, and the server 120 is used to adjust the delivery strategies for each target advertising creative.

[0019] like Figure 2 As shown, in one embodiment, a method for adjusting cross-platform advertising creatives is provided. This method can be applied to both terminals and servers; this embodiment illustrates its application to terminals. The method for adjusting cross-platform advertising creatives specifically includes the following steps: S1: Based on the specified advertising delivery target, obtain the corresponding historical delivery data; S2: Input the historical delivery data into a preset multimodal large language model to generate multiple initial advertising materials containing text, images and video elements, and obtain an initial set of advertising variants; S3: Based on a preset regional cultural rule base, perform compliance filtering on the initial set of advertising variants to obtain a set of compliant advertising variants; S4: Obtain multiple target advertising platforms to be deployed, and perform cross-platform adaptation of the initial advertising materials in the set of compliant advertising variants to obtain the target advertising materials corresponding to each initial advertising material; S5: Deploy each of the target advertising creatives on each of the target advertising platforms and collect real-time deployment feedback data; S6: Based on the real-time delivery feedback data, dynamically adjust the delivery strategy for each of the target advertising materials using a reinforcement learning model.

[0020] As described in step S1 above, based on the specified advertising campaign objective, the corresponding historical campaign data is obtained. The specified advertising campaign objective can be brand exposure, click-through rate (CTR), conversion rate (CVR), sales revenue (GMV), or interaction rate of a specific audience. During execution, it is necessary to access the advertising campaign database, analysis platform, or third-party data interface on the local terminal or server to extract historical campaign records related to the objective, including: campaign time, campaign platform, audience segmentation information, creative identifiers (text, image, video metadata), exposure, clicks, conversion events, dwell time, interaction events (likes, shares, comments), and attribution chain information. Data preprocessing is a key sub-process, which needs to complete deduplication, time alignment, missing value imputation, event deduplication and privacy anonymization (such as PII deletion or hashing), indicator calculation (such as CTR, CVR), and feature engineering (extracting multimodal features such as text keywords, image tags, video duration, and main subject). At the same time, metadata should be stored hierarchically by platform and geographic region to facilitate subsequent feature selection and model optimization by region or platform.

[0021] As described in step S2 above, the historical delivery data is input into a preset multimodal large language model to generate multiple initial advertising materials containing text, images, and video elements, resulting in an initial set of advertising variants. The multimodal large language model may include models with text, image, and video understanding / generation capabilities (such as visual-text joint models, conditional generation networks, or fusion-type LLMs). The input consists of preprocessed historical features and constraints, including high-potential feature combinations, target markets, audience profiles, and brand tone constraints. The generation process specifically involves: constructing conditional prompts or a trained conditional generator, setting diversity parameters (temperature, top-k sampling), specifying the modal output format (short text length, image resolution, video duration and frame rate), and employing parallel generation strategies to produce a large number of candidates (e.g., 200+ variants). The generated content should be initially automatically scored and screened. The scoring dimensions include semantic relevance, innovation, visualization quality, and consistency with brand style. Technically, classifiers / regressors or heuristic rules can be used to score the generated results and remove obviously low-quality items. In addition, the generation metadata (generation strategy, random seed, model version, context input) of each variant should be saved for subsequent analysis and traceability.

[0022] As described in step S3 above, the initial set of advertising variants is filtered for compliance based on a pre-set regional cultural rule base to obtain a set of compliant advertising variants. The regional cultural rule base should pre-store entries for multiple regions, including local data privacy and advertising compliance clauses (such as advertising ingredient statements). The compliance filtering process includes two parts: automated detection and optional manual review. Automated detection uses text review (sensitive word library, named entity recognition), image recognition (object detection, portrait pose detection, skin color / clothing rules), video frame sampling detection, and rule matching. It also needs to perform legal compliance checks (such as whether it contains personal data without consent). For detected problems, the system can implement correction strategies, such as automatically replacing sensitive elements (changing colors, cropping composition, replacing gestures), adjusting the copy, or marking it as requiring manual review. The compliance judgment should record the review log and correction history to ensure that the review path can be traced in case of disputes. At the same time, the rule base should support online updates and regional version control to respond to the latest limiting rules.

[0023] As described in step S4 above, multiple target advertising platforms to be deployed are obtained, and the initial advertising materials in the set of compliant advertising variants are cross-platform adapted to obtain the target advertising materials corresponding to each initial advertising material. First, the list of target platforms and their specifications and preferences are obtained: format restrictions (image / video resolution, duration, file size), text length and word count requirements, cover / thumbnail rules, interactive elements (CTA buttons, cards, challenge tags), user behavior preferences (long text or short sentences, vertical screen preference, etc.), and the platform's recommendation mechanism preference weights, etc. Technical implementation includes constructing a platform feature vector space to quantify the platform's weights in text density, visual dynamism, audio importance, interactive design, etc.; then, a feature decoupling network is used to separate the semantic core and presentation of the compliant variants, and the presentation is remapped based on the platform vectors (e.g., translating a text-based story into a vertical short video script, splitting long text into multi-frame subtitles, and replacing background music to match the target cultural rhythm). An attention-based cross-platform converter can be used to maintain consistency in marketing intent while adjusting elements to meet platform specifications. After adaptation, the target ad creatives for each platform are output, along with corresponding delivery metadata (format, version, platform constraint descriptions) for subsequent delivery and monitoring. During adaptation, technical compliance (encoding format, playback compatibility) should also be verified, and previews should be generated for manual confirmation. Feature decoupling networks refer to a neural network structure and training method used to decompose the representation of multimodal ad creatives into "platform-wide semantic features" (semantic / content latent) and "platform-specific style features" (style / domain latent). Its goal is to preserve the semantic core of the ad (such as product information and marketing intent) without changing with the platform, while representing the presentation (composition, color, rhythm, copywriting rhetoric, etc.) with controllable style vectors, facilitating subsequent style remapping and creative reconstruction based on the target platform's feature vectors. Through an encoding-fusion-branching-adversarial / constraint-reconstruction system, the semantics and style of the creatives are separated, allowing for controllable style remapping based on the target platform vectors while maintaining the marketing intent, thereby achieving efficient and traceable cross-platform creative adaptation.

[0024] As described in step S5 above, each of the target advertising creatives is deployed on its respective target advertising platform, and real-time deployment feedback data is collected. Implementation requires uploading creatives via platform API or advertising system, setting deployment parameters (budget, audience targeting, deployment time period, bidding strategy), and executing traffic allocation. Deployment requires configuration of tracking and event tracking, including exposure events, click events, conversion events, interaction events, and user session metrics. For videos, completion rate and average viewing time must be recorded. Real-time feedback data collection requires low latency and high reliability, typically achieved through message queues, stream processing platforms, or edge nodes to collect logs and perform preliminary aggregation. Data consistency (attribution issues, cross-device attribution, delayed arrival events) and noise (abnormal traffic, bot traffic) also need to be addressed. During deployment, data preservation and privacy compliance measures (log anonymization, minimizing personal data collection, data storage period strategies) should be clearly defined. Collected data should be updated in real-time to the model input pipeline according to a predefined indicator system to provide timeliness for subsequent reinforcement learning decisions. Simultaneously, platform review results and related indicators need to be monitored, and abnormal creatives should be paused promptly.

[0025] As described in step S6 above, based on the real-time delivery feedback data, the delivery strategy for each target advertising creative is dynamically adjusted through a reinforcement learning model. First, the elements of the reinforcement learning problem are defined: the state typically consists of current creative characteristics, platform characteristics, real-time metric vectors (CTR, CVR, ROAS, watch time), and budget remaining capacity; actions include adjusting the delivery budget allocation, display frequency, audience targeting parameters, or deactivating / activating creatives; the reward is a multi-objective function, which can combine click-through rate, conversion revenue, brand exposure weight, and budget efficiency, employing a weighted or multi-objective optimization scheme. Commonly used algorithms include Proximal Policy Optimization (PPO) or other policy gradient methods that balance stability and sample efficiency. These methods combine offline pre-training with online incremental updates to balance safety and adaptability. Implementation details include using short-cycle (e.g., every 15 minutes or at a preset cycle) batch updates of policy network parameters, introducing exploratory policies to avoid getting trapped in local optima, setting protection thresholds for abnormal distribution behavior, and maintaining interpretability (recording the reasons for policy changes). To improve sample utilization, causal inference or value function baselines can be used to correct offline data biases, and a safety rollback mechanism can be retained to revert to historically stable policies when the online policy shows a decline in effectiveness.

[0026] In one embodiment, step S2, which involves inputting the historical delivery data into a preset multimodal large language model to generate multiple initial advertising materials, includes: S201: Extract multimodal feature combinations associated with preset positive delivery effects from the historical delivery data; S202: Input the multimodal feature combination into a preset multimodal large language model as a constraint condition for generating initial advertising materials; S203: The multimodal large language model generates multiple initial advertising materials based on the constraints.

[0027] As described in step S201 above, multimodal feature combinations associated with preset positive campaign effects are extracted from the historical campaign data. Features and their cross-modal combinations significantly related to positive campaign effects (e.g., high click-through rate, high conversion rate, longer dwell time, or better ROAS) are identified from historical campaign records. A threshold can be set; values ​​exceeding this threshold are considered positive campaign effects. The specific threshold value can be set by relevant personnel based on actual conditions, so that subsequent creative generation can use these high-potential features as constraints. The input is preprocessed and anonymized historical campaign data, including text (copy, title), image metadata (tags, color histograms, composition features), video metadata (duration, shot switching frequency, main subject actions), user profiles, and platform context. Common methods include: statistical analysis (correlation coefficient, chi-square test), regression or tree models (such as XGBoost) for feature importance ranking, and causal inference or uplift modeling to distinguish features that truly bring incremental effects. For multimodal features, each modality can be mapped to a vector first (BERT class embedding for text, CNN / visual Transformer features for images, and temporal features or video embedding for videos), and then multimodal joint analysis (such as multimodal alignment, cross-modal correlation matrix, clustering or contrastive learning) can be used to identify high-frequency co-occurrence or complementary feature combinations. Interpretable tools such as SHAP and LIME can also be used to explain the contribution of the black-box model, outputting candidate combinations such as "short copy + product close-up + blue background + 9-second vertical video" or "title + interactive tags + soft-light portrait image." During implementation, attention should be paid to data stratification (grouping by platform, region, and audience) to avoid confusion between variables and historical strategy biases, and feature extraction rules and versions should be recorded to ensure traceability.

[0028] As described in step S202 above, the multimodal feature combination is input into a preset multimodal large language model as a constraint condition for generating initial advertising materials. The identified high-potential feature combinations are passed to the multimodal large language model in a machine-usable form as generation control conditions. Specifically, this involves constructing a structured conditional input (JSON or key-value pairs) containing modal attributes (e.g., {"copywriting style":"concise","image subject":"product close-up","color":"blue","video duration":"9s","presentation format":"vertical screen"}), which is read by the model's conditional encoder); then, natural language prompts are constructed, transforming the feature combinations into constraint-rich prompt text and combining it with examples (few-shots), such as "Generate a 9-second vertical short video script for female users in a certain region, with a concise and direct copywriting style, emphasizing discounts, and using blue as the main color"; finally, constraints are implanted within the model through soft prompts or control tokens / prefix tuning to guide the generation process in the latent space. It can also simultaneously inject brand guidelines, legal compliance summaries, and target platform format constraints to ensure that the generation does not violate prior rules. Technically, it is necessary to ensure that the condition representation is aligned across modalities (text labels and visual attributes can be correctly understood by the model), and to encode the priority and optionality of the conditions (required items / preferred items) to avoid conflicts and ambiguities during generation. During implementation, the prompts, control parameters, and model versions used should be recorded to facilitate reproduction and effect comparison, and differentiated constraint sets can be adopted for different platforms or audiences.

[0029] As described in step S203 above, the multimodal large language model generates multiple initial advertising materials based on the constraints. The generated structured constraints or prompts are input, and the output is a set of multimodal initial materials (including text, visual descriptions, image sketches, or video scripts, etc.). The generation process may employ a multimodal generation pipeline: the text generator (LLM) first produces the text and video script, then calls the visual generation module (such as a diffusion model or conditional GAN) to generate images or keyframes based on text and visual constraints, and then the video synthesis module combines the keyframes, transitions and audio into short video samples. During generation, diversity control parameters (temperature, top-k, top-p) and decoding strategies need to be set to balance quality and diversity. Simultaneously, batch generation can be performed in parallel to obtain a large number of candidates (such as 200+). To improve the generation quality, post-processing and automatic evaluation are often combined: text semantic consistency detection, image quality assessment (FID / CLIP score), brand consistency scoring and compliance screening. Each generated material should be accompanied by metadata (constraints used, random seed, model version, generation time), and duplicate detection and deduplication should be performed to retain a diverse and high-potential candidate set for the next step of compliance and adaptation processing. In addition, security strategies (such as sensitive content filtering, personal identification information recognition) and early warning mechanisms should be added to the generation pipeline to trigger manual review or rollback when inappropriate content is generated.

[0030] In one embodiment, step S3, which involves filtering the initial set of ad variants for compliance based on a preset regional cultural rule base to obtain a set of compliant ad variants, includes: S301: Determine the target area associated with the specified advertising delivery target; S302: Obtain the regional restriction rules and data compliance rules corresponding to the target region from the preset regional culture rule base; S303: Based on the region limitation rules and the data compliance rules, identify and filter conflicting visual elements or text descriptions in the initial set of ad variants to obtain a set of compliant ad variants.

[0031] As described in step S301 above, the target region associated with the specified advertising delivery target is determined. The purpose is to clarify the geographical and audience context in which each initial ad variant will be delivered, thereby providing a positioning basis for subsequent invocation of corresponding cultural and legal rules. The inputs typically include the geographic parameters of the advertising delivery target, the target audience profile (language, age range, cultural orientation, interest tags), and the delivery time window (whether it is during a specific holiday or sensitive period). In practice, this information can be obtained by parsing the delivery plan or business rules: reading the geographic ID from the delivery configuration file, reading the audience tags from the audience targeting settings, or translating the market code into an index of the rule set according to the business strategy mapping.

[0032] As described in step S302 above, regional restriction rules and data compliance rules corresponding to the target region are obtained from a preset regional cultural rule base. The goal is to accurately retrieve and load a set of cultural and legal compliance rules applicable to the identified target region. The regional cultural rule base should be hierarchical, containing multi-level entries: relevant restriction requirements, regional restriction content (colors, gestures, symbols, clothing, image sensitivity), special restrictions for holidays / anniversaries, platform-specific review preferences, and a sensitive word library for local language terminology. The acquisition process includes: executing a rule retrieval algorithm based on the regional ID and audience attributes obtained in step S301, supporting precise matching and inheritance. The technical implementation can store rules in a structured format (e.g., JSON / YAML) and provide meta-attributes such as rule priority, effective time, and applicable audience tags. For data compliance, the rule base should include PII processing specifications, personalized recommendation restrictions, etc., and also handle rule conflicts and version control: when multiple rules are applicable but contradictory, they should be merged or selected according to a predefined priority strategy, and the conflict resolution strategy and the selected rule version should be recorded. The rule base should also support online updates and emergency patching mechanisms to quickly respond to new restriction requirements.

[0033] As described in step S301 above, conflicting visual elements or text descriptions in the initial set of ad variants are identified and filtered based on the regional limitation rules and the data compliance rules, thereby obtaining a set of compliant ad variants. For text, sensitive word retrieval, named entity recognition (NER), sentiment / semantic classification, and contextual analysis (to determine puns or regionally specific meanings) are applied. For visuals, image recognition / object detection, portrait recognition, and posture analysis (e.g., detecting left-hand display, exposed specific clothing), color analysis (detecting whether prohibited colors are present), and lighting and composition detection are used. For video, keyframe sampling, subtitle / OCR recognition, and motion sequence analysis are used to detect elements that do not meet the limitations. The detection results are scored according to confidence level and compared with thresholds defined in the rule base. For elements with high confidence levels that do not meet the limitations, the system can directly execute preset handling strategies: automatic replacement (color, image cropping, material replacement), obscuring or blurring sensitive areas, adjusting text, or marking as unsuitable for placement. For cases with moderate confidence or semantic ambiguity, the system should trigger a manual review workflow, providing contextual information and variant history (generation constraints, model versions, etc.) for manual judgment and consultation with rule experts. Throughout the process, a complete review log must be maintained: detection results, triggered rule ID, handling operation, reviewer, and timestamp, ensuring traceability. Implementation should balance accuracy and recall, setting appropriate false positive tolerance strategies and periodically evaluating the performance of the detection model (false positive rate, false negative rate, review pass rate). Continuous optimization should be achieved through online learning or rule updates. Furthermore, if the number of filtered variants is insufficient or core information is corrupted, a regeneration or variant correction process can be automatically triggered to ensure both compliance and achievement of deployment goals.

[0034] In one embodiment, step S4, which involves acquiring multiple target advertising platforms to be advertised and cross-platform adapting the initial ad creatives in the compliant ad variant set to obtain the target ad creatives corresponding to each initial ad creative, includes: S401: Obtain multiple target advertising platforms to be deployed; S402: Obtain the platform feature vector of each of the target advertising platforms; wherein the platform feature vector is used to characterize its preference weights for text, static images, dynamic videos, and audio elements; S403: Based on the platform feature vector, the platform-wide semantic features and platform-unique style features in the initial advertising material are separated by a preset feature decoupling network; S404: Based on the platform feature vector of the target platform, remap the platform-specific style features and combine them with the platform's general semantic features to reconstruct and generate the target advertising material.

[0035] As described in step S401 above, multiple target advertising platforms to be deployed are obtained to determine the specific set of platforms that need to be adapted and deployed, providing a target basis for subsequent format conversion, strategy matching, and monitoring. The input is usually the deployment plan, marketing strategy, or business configuration (such as region, target audience, budget allocation, etc.), and the output is a structured list of platforms (platform ID, platform type, deployment entry, API credentials, priority, and deployment window). In practice, this can be achieved by reading the deployment configuration file, the management console, or by selection by the user / business system. At the same time, the platform registration / integration module needs to be called to obtain the platform access information and capability description (supported media types, interface versions, rate limits, billing modes, etc.). During integration, permission and credential management (OAuth, API Key), sandbox environment testing and online environment switching, and platform availability testing (network connectivity, API response health checks) need to be handled. In the scenario of multi-platform parallel deployment, platform grouping, priority setting, and canary deployment strategy should be supported, and platform metadata (geographic coverage, main user groups, historical performance summary) should be recorded for subsequent platform-specific customization adaptation. This step should also support dynamic expansion (adding / removing platforms) and trigger the corresponding adaptation re-evaluation process.

[0036] As described in step S402 above, the platform feature vector of each target advertising platform is obtained; wherein, the platform feature vector is used to characterize its preference weights for text, static images, dynamic videos, and audio elements. The purpose is to quantify the multidimensional preferences and constraints of each target platform into a computable vector representation, which facilitates comparison and mapping in subsequent algorithms. The dimensions of the platform feature vector may include, but are not limited to: text length weight, visual static weight, dynamic visual weight, audio weight, preferred aspect ratio and resolution, sensitivity of the recommendation algorithm to interactive signals, tolerance for push frequency, user preference distribution (age / gender / interest), creative rhythm preference (fast / slow rhythm), interactive element preference (significant / implicit CTA), and compliance strictness, etc. The acquisition methods can be combined: first, parsing the platform's public documents and SDK specifications; second, statistical learning based on historical delivery data (calculating the average CTR / viewing time of each type of material on the platform and normalizing it); third, model learning (training a platform encoder to encode a large number of successful platform materials into vectors and extract feature weights); and fourth, industry experience and manual calibration. During implementation, data from different sources needs to be weighted, fused, and normalized, and vector versions need to be managed (version number, update time, source description) to ensure the consistency and traceability of vectors when comparing and mapping across platforms.

[0037] As described in step S403 above, based on the platform feature vector, a preset feature decoupling network separates the platform-wide semantic features and platform-specific style features from the initial advertising material. The aim is to decouple the material's "semantic kernel" (such as marketing intent, product attributes, target information) from its "expressive style" (such as composition, color, camera rhythm, and copywriting rhetoric), preparing for style remapping for different platforms. This can be implemented using a multimodal encoder and a decoupling network: first, a text encoder (such as BERT-like algorithms) extracts the text semantic vector, and a visual encoder (such as CNN / ViT) extracts image / video features; then, the features from each modality are fed into a joint representation layer. Subsequently, a decoupling network (such as a variational autoencoder, conditional autoencoder, or a separation module with contrast loss) applies constraints, making the semantic vector insensitive to expressive style, while the style vector retains visual / auditory expressive characteristics. Contrastive learning or domain adversarial training can be introduced to ensure that the general semantics remain consistent across different platform domains (domain-invariant), while the style vector carries platform-specific information (domain-specific). Implementation details include defining a decoupling loss function, setting capacity boundaries for style / semantics, employing an attention mechanism to weight multimodal information, and using platform-labeled data for supervision during training. The output is a pair of vectors for each initial creative: a semantic vector and a style vector, recording decoupling confidence and interpretability metrics. Platform-wide semantic features express the core content and intent of the advertising creative and should remain consistent across different advertising platforms, regardless of the presentation format. These include product attributes (model, color, usage), intent (discount, limited-time offer, buy-one-get-one-free), target audience (e.g., "young women"), brand proposition, and key information points (price, selling points, CTA semantics). Platform-specific style features describe the presentation style or format of the creative and change with platform preferences and presentation media. This includes visual style (composition, color tone, filters, screen layout), temporal style (editing rhythm, shot switching frequency, video length), copywriting style (sentence length, tone, use of emoticons), interactive elements (challenge tags, CTA layout, bullet screen style), and audio style (background music type, volume curve).

[0038] As described in step S404 above, the platform-specific style features are remapped based on the platform feature vector of the target platform, and combined with the platform's general semantic features to reconstruct and generate the target advertising material. The decoupled semantic kernel is fused with the style features remapped for the target platform to generate final materials that meet the platform's preferences and specifications. This can be achieved using a cross-platform converter (based on attention mechanisms or conditional generation networks): the input consists of the semantic vector of the source material, the source style vector, and the target platform feature vector. The converter first calculates the style mapping matrix or transformation function (learnable through a small neural network) to project the source style into the target style space. Then, the projected style and semantic vector are input into a multimodal generation module using a fusion strategy (concatenate, cross-attention, or conditional layer) to generate text, static images, or video samples in the corresponding format. During generation, format constraints (resolution, duration, word count), platform interaction components (tags, challenge logos, CTA formats), and compliance annotations must be considered. Post-processing (subtitle segmentation, cover generation, encoding conversion) is performed, and the generated results undergo automatic quality checks (visual clarity, text coherence, brand consistency) and platform compatibility verification, triggering fine-tuning or manual confirmation as necessary. The reconstruction process must also retain metadata (mapping rules, version number, conversion confidence) to support A / B experiments, effect attribution, and future iterative optimization.

[0039] In one embodiment, step S6, which dynamically adjusts the delivery strategy for each target advertising creative based on the real-time delivery feedback data using a reinforcement learning model, includes: S601: Construct a multi-objective reward function; S602: Input the real-time delivery feedback data into the target reward function to obtain a comprehensive reward value; wherein, the real-time delivery feedback data includes at least two of click-through rate, conversion rate, user dwell time, and interaction rate; S603: Using the comprehensive reward value as the optimization target, and employing a near-end strategy optimization algorithm, dynamically adjust the allocation of the advertising budget and display frequency of each target advertising creative on different target advertising platforms.

[0040] As described in step S601 above, the purpose of constructing a multi-objective reward function is to synthesize multiple business metrics (such as click-through rate, conversion rate, user dwell time, interaction rate, etc.) in the campaign objectives into one or more reward signals that can be used for reinforcement learning optimization. The design of the multi-objective reward function needs to consider the differences in the scale of the metrics, the importance weight, delayed feedback and attribution issues. Common practices include normalizing or standardizing each individual metric (e.g., using historical mean-variance normalization or quantile mapping), and then assigning weights according to business priority to construct linear weighted sums or non-linear combinations (such as weighted product, threshold penalty, tiered segmented rewards). To handle delayed conversions, discount factors can be introduced or conversions within the future window can be allocated (attribution model). The expected revenue is then subtracted from the baseline or value function to stabilize the training signal. If simultaneously optimizing conflicting objectives, a multi-objective optimization strategy can be adopted: first, synthesizing multiple indicators into a single comprehensive reward to simplify the problem; second, using constrained rewards (maximizing the primary objective and using secondary objectives as constraints); and third, multi-task / vectorized rewards combined with Pareto frontier search. The design should also include budget consumption penalties, compliance penalties, and upper and lower bound constraints to prevent the strategy from deviating from business or compliance boundaries. Finally, offline simulation and sensitivity analysis of the reward function are performed to ensure that the reward signal reflects business value while possessing good training stability. In a specific embodiment, a specific formula for the multi-objective function is as follows: R = w1 * CTR + w2 * CVR + w3 * Normalized value of user dwell time; where w1, w2, and w3 are preset weight coefficients set according to business objectives, CTR is the click-through rate, and CVR is the conversion rate.

[0041] As described in step S602 above, the real-time delivery feedback data is input into the target reward function to obtain a comprehensive reward value; wherein, the real-time delivery feedback data includes at least two of the following: click-through rate, conversion rate, user dwell time, and interaction rate. Multiple real-time delivery metrics are mapped to scalar rewards that can be learned using pre-built rules or models. First, the real-time data undergoes preprocessing: noise reduction, latency correction (handling event arrival delays), and attribution allocation (attributing conversions to specific creatives / events based on impressions / clicks). Then, the data is aggregated at a granularity matching the reward function (e.g., per creative / per platform / per time window). Subsequently, each metric is converted to the same dimension using a defined standardization method (e.g., normalization by historical mean standard deviation or percentile scaling). Weights and penalties are applied to obtain an immediate comprehensive reward value. To enhance stability, instantaneous rewards are typically smoothed (e.g., exponential moving average) or baseline subtraction is used to calculate advantages, which is particularly important for subsequent policy gradient estimation. The sparsity or delayed reward problem must also be considered: this can be mitigated through reward shaping, introducing intermediate rewards (e.g., view completion rate as a short-term reward), or using models to predict future rewards. Low latency should be ensured (to achieve near real-time optimization), and complete time series and metadata should be preserved for auditing and offline training.

[0042] As described in step S603 above, with the comprehensive reward value as the optimization target, the near-end strategy optimization algorithm is used to dynamically adjust the allocation of the advertising budget and the display frequency of each target advertising material on different target advertising platforms. First, the state, actions, and policy output of reinforcement learning need to be defined: the state can include material features, platform features, current real-time metric vectors, and budget margin; actions can be defined as continuously assigning weights to each material (budget proportion, display frequency, bid multiplier, or whether to enable it); the output is usually processed by softmax or normalization to meet budget and frequency constraints, and stable policy gradient algorithms such as Proximal Policy Optimization (PPO) are used. The process includes: collecting an interaction trajectory online with the current policy (a state-action-reward sequence over several time windows), using GAE to estimate the advantage, constructing a loss function and updating it in multiple rounds on mini-batches, while using clipping or KL penalty to limit the policy update magnitude to ensure training stability. To balance sample efficiency and safety, offline pre-training, simulated environment playback, and online incremental updates can be combined; to avoid business risks caused by overexploration, an entropy regularization term is added to control the degree of exploration and business protection thresholds are set (such as minimum ROI, budget cap, and campaign cooldown period). In addition, a monitoring and rollback mechanism needs to be implemented: real-time monitoring of key business metrics (CTR / CVR / cost, etc.), and automatic rollback to historical strategies or manual intervention when metrics are abnormal. In terms of deployment, micro-batch updates and layered deployment (offline training on the central server and incremental updates on edge nodes) are considered to achieve low latency response and scalability. Finally, the training hyperparameters (learning rate, clip range, update steps, batch size) are continuously tuned, and the strategy is regularly evaluated offline on historical traffic to ensure long-term robustness.

[0043] In one embodiment, after step S5 of delivering each of the target advertising creatives to each of the target advertising platforms and collecting real-time delivery feedback data, the method further includes: S611: Determine whether the cumulative amount of real-time delivery feedback data collected has reached the preset amount; S612: If the number of real-time delivery feedback data collected is greater than the preset number, then the policy network parameters of the reinforcement learning model are incrementally updated. S613: After the incremental update is completed, the number of accumulated real-time delivery feedback data will be cleared to zero.

[0044] As described in step S611 above, it is determined whether the number of accumulated real-time delivery feedback data has reached the preset quantity. This aims to control the amount and quality of data used for incremental model updates, avoiding model instability due to insufficient samples or data skew. In implementation, the definition of "the number of real-time delivery feedback data" must first be clarified: is it based on the number of samples per individual creative, the number of samples aggregated on the platform, or the total number of events (exposure / clicks / conversions)? The "preset quantity" can be a fixed threshold, a minimum sample size based on a confidence interval, or a dynamic threshold (adjusted according to sample variance, target significance requirements, or budget window). The specific implementation process includes: reading the current count from the data accumulation module and comparing it with the threshold configured in the rule engine; simultaneously verifying data integrity (checking temporal continuity, missing events, and duplicate records) and data quality (removing identified robot traffic or abnormal peaks); if the event has not yet arrived due to platform delays, compensation for the waiting time should be based on the delay distribution prediction. This judgment step should record the judgment basis and time point, and support granular judgment by material, platform or geographic region. If necessary, it can trigger partial updates (only executed on a subset that has reached the threshold) or continue to accumulate to achieve a more robust sample size. In addition, a maximum waiting time limit should be set to avoid waiting indefinitely for unreachable samples. If the time limit is exceeded, a decision should be made according to business rules (such as performing a weak update with existing data or skipping this update).

[0045] As described in step S612 above, if the amount of real-time delivery feedback data collected exceeds a preset amount, incremental updates are performed based on the policy network parameters of the reinforcement learning model. First, accumulated data needs to be retrieved from the data warehouse and preprocessed uniformly: data cleaning, event attribution, latency correction, feature standardization, and necessary privacy desensitization. Incremental update methods can include online learning, micro-batch training, or batch updates based on an experience replay buffer. Common methods include using a small learning rate, gradient updates with a limited number of steps, multi-round optimization of PPO in small batches, or fine-tuning based on the trust domain, to adapt to the new distribution while preserving the performance of the original policy. To ensure security, several protection mechanisms are required: a) Use pruning / KL penalties to limit the differences between the old and new policies; b) Perform offline backtesting or use small-volume canary experiments to verify changes in key metrics of the new policy before updating; c) Retain a snapshot and version number of the model before updating for rollback; d) Monitor training loss, reward curves, and business metrics during the update process, and automatically interrupt and roll back if an anomaly occurs. The technical implementation also includes concurrency control (to avoid conflicts from multiple parallel updates), update interpretability logging (update reason, training sample summary, hyperparameters, runtime), and compliant audit logs. After the update is completed, a gradual rollout should be implemented to progressively increase the traffic of the new strategy and continuously monitor its online performance.

[0046] As described in step S613 above, after the incremental update is completed, the number of accumulated real-time delivery feedback data is cleared to zero. The purpose is to reset the data accumulation state, start a new sample collection window for the next round of incremental updates, and ensure the idempotency and traceability of data use. This must be implemented following the principle of "confirmation before clearing": the clearing operation is only performed after the incremental update is successful and passes the preset verification and monitoring checks; if the update fails or anomalies are found in subsequent verifications, the data should be retained and marked as "to be retried" or stored in backup for diagnostic purposes. The clearing operation involves updating counters / indexes, moving or archiving used data (packaging and archiving the original events and training set snapshots and recording the storage location), and writing to the audit log (including the time interval of the data used, the amount of data, the updated model version, and the verification summary). In distributed deployment scenarios, distributed locks or transaction mechanisms need to be implemented to ensure the atomicity of the zeroing operation and avoid concurrent race conditions. At the same time, data retention strategies and compliance requirements (such as anonymization obligations) should be considered. Used data should be deleted, de-identified, or permanently encrypted and archived according to the strategy. Finally, after zeroing, the configuration of the next collection cycle (such as resetting the counting threshold and updating the time window) should be triggered and the monitoring system should be notified so that the next round of incremental updates can be carried out continuously and in a standardized manner.

[0047] In one embodiment, after step S5 of delivering each of the target advertising creatives to each of the target advertising platforms and collecting real-time delivery feedback data, the method further includes: S621: The real-time delivery feedback data is used as new historical delivery data and updated in the database used to obtain historical delivery data.

[0048] As described in step S621 above, the real-time feedback data generated during the online campaign phase is systematically and reliably integrated into the historical campaign database for subsequent analysis and generation, thus forming a closed-loop data flow to support subsequent feature extraction, model training, and constraint updates. The specific implementation includes the following key steps and points: Data reception and preprocessing: First, raw event streams are received from real-time acquisition pipelines (such as message queues, stream processing platforms, or edge node synchronization), including exposure, clicks, conversions, viewing duration, interaction events, and attribution information. The received data is cleaned and standardized, including timestamp unification (time zone, format), event deduplication, field completion, outlier detection and marking (such as abnormal high-frequency clicks, robot traffic identification), and delay correction logic (compensating for data loss or sequence errors caused by platform feedback delays) and preliminary attribution allocation (allocating conversion events to corresponding materials / channels according to a predetermined attribution model).

[0049] In one specific implementation, the following checks can be performed before writing to the history database: completeness (whether necessary fields are missing), consistency (whether field types and enumeration values ​​conform to specifications), relevance (whether material ID and platform ID exist in the metadata table), and credibility (determining whether it is abnormal or fraudulent traffic based on a statistical model). Records that fail the verification are placed in the "abnormal / pending review" queue according to the policy and trigger manual or automated review; qualified records proceed to the next writing stage.

[0050] An incremental write strategy is employed to avoid duplicate writes or overwriting of important historical data. This can be achieved by using idempotent writes based on unique event IDs or composite keys (material ID + event type + time window), or by first writing to a temporary table and then merging the snapshot into the main history database via transactions or batch processing. If using a distributed database or data lake (such as Kafka → Hudi / Delta / Iceberg → data warehouse), its transaction / versioning mechanisms must be utilized to ensure atomicity and immutability, and batch numbers and source identifiers must be added to write operations.

[0051] Each write operation should include metadata: data source, collection time window, processing version (ETL / validation / attribution version), and model or strategy version (e.g., algorithm version used for attribution or anomaly detection). Table / partition version management should be implemented for the historical database to support traceable rollback and auditing. When the upstream schema changes, a compatibility strategy (backward / forward compatible conversion) is required, and a change description must be recorded in the metadata. Privacy compliance and secure storage: Privacy protection measures must be implemented before writing to the historical database, such as desensitizing, hashing, or deleting Personally Identifiable Information (PII) according to a policy; data retention periods should be recorded, and automatic cleanup or anonymization processes should be implemented to comply with relevant requirements. The write channel and storage layer should enable transmission and static encryption, access control, and audit logs, limiting access to sensitive fields to only authorized services / personnel. Indexing, partitioning, and query optimization: To support efficient feature extraction and model training, the historical database should be reasonably partitioned according to commonly used query dimensions (time, platform, material ID, region, etc.) and necessary indexes or materialized views should be established. High-frequency access aggregation metrics can be pre-calculated and stored to reduce subsequent offline / online feature engineering costs. For scenarios with large amounts of data, a tiered storage approach is adopted (hot storage for near real-time training and cold storage for long-term archiving) to optimize cost and latency.

[0052] After the write operation is complete, the downstream workflow should be triggered: notify the feature engineering module to update incremental features, update the training / online training queue, trigger the incremental update judgment of the reinforcement learning model (such as in conjunction with S611 and S612), and push monitoring alarms to assess write quality and data distribution drift. Record write statistics (number of records, anomaly rate, latency) and incorporate them into data quality indicator monitoring, supporting automatic rollback or manual intervention.

[0053] Archiving, auditing, and backtracking capabilities: All write batches must retain searchable snapshots and audit records (including original event samples, processing logs, and decision-making basis) to facilitate tracking and reproduction in case of model performance anomalies or compliance reviews. A data retention and archiving strategy must be developed to balance auditing needs with storage costs.

[0054] Reference Figure 3 The present invention also provides an adjustment device for delivering cross-platform advertising creatives, the device comprising: The acquisition module 902 is used to acquire the corresponding historical advertising data based on the specified advertising target. The generation module 904 is used to input the historical delivery data into a preset multimodal large language model to generate multiple initial advertising materials containing text, images and video elements, and obtain an initial set of advertising variants; The filtering module 906 is used to perform compliance filtering on the initial set of advertising variants based on a preset regional cultural rule base to obtain a set of compliant advertising variants; The adaptation module 908 is used to obtain multiple target advertising platforms to be deployed, and to perform cross-platform adaptation of the initial advertising materials in the set of compliant advertising variants to obtain the target advertising materials corresponding to each initial advertising material; The delivery module 910 is used to deliver each of the target advertising materials to each of the target advertising platforms and collect real-time delivery feedback data. The adjustment module 912 is used to dynamically adjust the delivery strategy of each target advertising material based on the real-time delivery feedback data through a reinforcement learning model.

[0055] In one embodiment, the generation module 904 includes: An extraction submodule is used to extract multimodal feature combinations that are associated with preset positive delivery effects from the historical delivery data; The input submodule is used to input the multimodal feature combination into a preset multimodal large language model as a constraint condition for generating the initial advertising material; The initial advertising material generation submodule is used to generate multiple initial advertising materials by the multimodal large language model based on the constraints.

[0056] In one embodiment, the filtering module 906 includes: The target area determination submodule is used to determine the target area associated with the specified advertising delivery target; The data compliance rule acquisition submodule is used to acquire the regional restriction rules and data compliance rules corresponding to the target region from a preset regional culture rule library; The identification submodule is used to identify and filter conflicting visual elements or text descriptions in the initial set of ad variants based on the region limitation rules and the data compliance rules, thereby obtaining a set of compliant ad variants.

[0057] In one embodiment, the adapter module 908 includes: The target advertising platform acquisition submodule is used to acquire multiple target advertising platforms to be advertised. The platform feature vector acquisition submodule is used to acquire the platform feature vector of each of the target advertising platforms; wherein, the platform feature vector is used to characterize its preference weights for text, static images, dynamic videos, and audio elements; The separation submodule is used to separate the platform-wide semantic features and platform-unique style features in the initial advertising material based on the platform feature vector and through a preset feature decoupling network. The reconstruction submodule is used to remap the platform-specific style features based on the platform feature vector of the target platform, and reconstruct and generate the target advertising material by combining the platform's general semantic features.

[0058] In one embodiment, the adjustment module 912 includes: A submodule is built to construct a multi-objective reward function; The comprehensive reward value acquisition submodule is used to input the real-time delivery feedback data into the target reward function to obtain the comprehensive reward value; wherein, the real-time delivery feedback data includes at least two of click-through rate, conversion rate, user dwell time, and interaction rate; The adjustment submodule is used to dynamically adjust the allocation of the advertising budget and display frequency of each target advertising creative on different target advertising platforms, with the comprehensive reward value as the optimization target and the near-end strategy optimization algorithm.

[0059] In one embodiment, the device for adjusting the delivery of cross-platform advertising creatives further includes: The judgment module is used to determine whether the amount of accumulated real-time delivery feedback data has reached the preset amount; The incremental update module is used to incrementally update the policy network parameters of the reinforcement learning model if the number of collected real-time delivery feedback data is greater than a preset number. The feedback module is used to reset the number of accumulated real-time delivery feedback data to zero after the incremental update is completed.

[0060] In one embodiment, the device for adjusting the delivery of cross-platform advertising creatives further includes: The update module is used to update the database used to obtain historical delivery data by taking the real-time delivery feedback data as new historical delivery data.

[0061] Figure 4 An internal structural diagram of an electronic device in one embodiment is shown. This electronic device can specifically be a terminal or a server, and more specifically, a computer device. Figure 4 As shown, the electronic device includes a processor, a memory, and a network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement a method for adjusting the delivery of cross-platform advertising materials. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to implement a method for adjusting the delivery of cross-platform advertising materials. Those skilled in the art will understand that... Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0062] In one embodiment, an electronic device is provided, including a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps: Based on the specified advertising campaign objectives, obtain the corresponding historical campaign data; The historical delivery data is input into a preset multimodal large language model to generate multiple initial advertising materials containing text, images and video elements, thus obtaining an initial set of advertising variants; Based on a preset regional cultural rule base, the initial set of advertising variants is filtered for compliance to obtain a set of compliant advertising variants; Obtain multiple target advertising platforms to be advertised, and perform cross-platform adaptation of the initial advertising creatives in the set of compliant advertising variants to obtain the target advertising creatives corresponding to each initial advertising creative; The target advertising creatives are deployed on the target advertising platforms, and real-time deployment feedback data is collected. Based on the real-time delivery feedback data, the delivery strategy for each target advertising creative is dynamically adjusted through a reinforcement learning model.

[0063] By leveraging multimodal large language models to generate creative variations in batches, the creation time of materials is shortened, greatly improving production efficiency and diversity. Through a preset rule base, cultural compliance filtering is automatically completed, significantly increasing the approval rate and reducing compliance risks. The innovative cross-platform adaptation mechanism can intelligently transform core creative ideas into forms that conform to the preferences of different platforms, enabling efficient reuse of materials and saving a lot of repetitive production costs. Finally, relying on reinforcement learning models to dynamically adjust strategies based on real-time feedback data, continuous autonomous optimization of campaign performance is achieved, significantly improving ad click-through rates, conversion rates, and overall return on investment, forming an end-to-end intelligent advertising operation capability.

[0064] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, causes the processor to perform the following steps: Based on the specified advertising campaign objectives, obtain the corresponding historical campaign data; The historical delivery data is input into a preset multimodal large language model to generate multiple initial advertising materials containing text, images and video elements, thus obtaining an initial set of advertising variants; Based on a preset regional cultural rule base, the initial set of advertising variants is filtered for compliance to obtain a set of compliant advertising variants; Obtain multiple target advertising platforms to be advertised, and perform cross-platform adaptation of the initial advertising creatives in the set of compliant advertising variants to obtain the target advertising creatives corresponding to each initial advertising creative; The target advertising creatives are deployed on the target advertising platforms, and real-time deployment feedback data is collected. Based on the real-time delivery feedback data, the delivery strategy for each target advertising creative is dynamically adjusted through a reinforcement learning model.

[0065] By leveraging multimodal large language models to generate creative variations in batches, the creation time of materials is shortened, greatly improving production efficiency and diversity. Through a preset rule base, cultural compliance filtering is automatically completed, significantly increasing the approval rate and reducing compliance risks. The innovative cross-platform adaptation mechanism can intelligently transform core creative ideas into forms that conform to the preferences of different platforms, enabling efficient reuse of materials and saving a lot of repetitive production costs. Finally, relying on reinforcement learning models to dynamically adjust strategies based on real-time feedback data, continuous autonomous optimization of campaign performance is achieved, significantly improving ad click-through rates, conversion rates, and overall return on investment, forming an end-to-end intelligent advertising operation capability.

[0066] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0067] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0068] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for adjusting cross-platform advertising creatives, characterized in that, The method includes: Based on the specified advertising campaign objectives, obtain the corresponding historical campaign data; The historical delivery data is input into a preset multimodal large language model to generate multiple initial advertising materials containing text, images and video elements, thus obtaining an initial set of advertising variants; Based on a preset regional cultural rule base, the initial set of advertising variants is filtered for compliance to obtain a set of compliant advertising variants; Obtain multiple target advertising platforms to be advertised, and perform cross-platform adaptation of the initial advertising creatives in the set of compliant advertising variants to obtain the target advertising creatives corresponding to each initial advertising creative; The target advertising creatives are deployed on the target advertising platforms, and real-time deployment feedback data is collected. Based on the real-time delivery feedback data, the delivery strategy for each target advertising creative is dynamically adjusted through a reinforcement learning model.

2. The method for adjusting cross-platform advertising materials according to claim 1, characterized in that, The step of inputting the historical delivery data into a preset multimodal large language model to generate multiple initial advertising materials includes: Extract multimodal feature combinations that are associated with the preset positive delivery effect from the historical delivery data; The multimodal feature combination is input into a preset multimodal large language model as a constraint condition for generating initial advertising materials; The multimodal large language model generates multiple initial advertising materials based on the constraints.

3. The method for adjusting cross-platform advertising materials according to claim 1, characterized in that, The step of performing compliance filtering on the initial set of ad variants based on a preset regional cultural rule base to obtain a set of compliant ad variants includes: Determine the target region associated with the specified advertising delivery objective; Obtain the regional restriction rules and data compliance rules corresponding to the target region from the preset regional culture rule base; Based on the region limitation rules and the data compliance rules, conflicting visual elements or text descriptions in the initial set of ad variants are identified and filtered to obtain a set of compliant ad variants.

4. The method for adjusting cross-platform advertising materials according to claim 1, characterized in that, The steps of acquiring multiple target advertising platforms to be advertised and cross-platform adapting the initial ad creatives in the compliant ad variant set to obtain the target ad creatives corresponding to each initial ad creative include: Obtain multiple target advertising platforms to be deployed; Obtain the platform feature vector for each of the target advertising platforms; wherein the platform feature vector is used to characterize its preference weights for text, static images, dynamic videos, and audio elements; Based on the platform feature vector, the platform's general semantic features and platform-specific style features in the initial advertising material are separated through a preset feature decoupling network; Based on the platform feature vector of the target platform, the platform-specific style features are remapped and combined with the platform's general semantic features to reconstruct and generate the target advertising material.

5. The method for adjusting cross-platform advertising materials according to claim 1, characterized in that, The step of dynamically adjusting the delivery strategy for each target advertising creative based on the real-time delivery feedback data using a reinforcement learning model includes: Construct a multi-objective reward function; The real-time delivery feedback data is input into the target reward function to obtain a comprehensive reward value; wherein, the real-time delivery feedback data includes at least two of the following: click-through rate, conversion rate, user dwell time, and interaction rate; Using the comprehensive reward value as the optimization objective, a near-end strategy optimization algorithm is used to dynamically adjust the allocation of the advertising budget and display frequency of each target advertising creative on different target advertising platforms.

6. The method for adjusting cross-platform advertising materials according to claim 1, characterized in that, After the steps of delivering each of the target advertising creatives to each of the target advertising platforms and collecting real-time delivery feedback data, the method further includes: Determine whether the cumulative amount of real-time delivery feedback data collected has reached the preset amount; If the amount of real-time delivery feedback data collected is greater than the preset amount, then the policy network parameters of the reinforcement learning model are incrementally updated. After the incremental update is completed, the number of accumulated real-time delivery feedback data will be reset to zero.

7. The method for adjusting cross-platform advertising materials according to claim 1, characterized in that, After the steps of delivering each of the target advertising creatives to each of the target advertising platforms and collecting real-time delivery feedback data, the method further includes: The real-time delivery feedback data is used as new historical delivery data and updated in the database used to obtain historical delivery data.

8. An adjustment device for delivering cross-platform advertising materials, characterized in that, The device includes: The acquisition module is used to obtain the corresponding historical advertising data based on the specified advertising target. The generation module is used to input the historical delivery data into a preset multimodal large language model to generate multiple initial advertising materials containing text, images and video elements, thus obtaining an initial set of advertising variants; The filtering module is used to perform compliance filtering on the initial set of ad variants based on a preset regional cultural rule base to obtain a set of compliant ad variants; The adaptation module is used to obtain multiple target advertising platforms to be deployed, and to perform cross-platform adaptation of the initial advertising materials in the set of compliant advertising variants to obtain the target advertising materials corresponding to each initial advertising material; The delivery module is used to deliver each of the target advertising materials to each of the target advertising platforms and collect real-time delivery feedback data. The adjustment module is used to dynamically adjust the delivery strategy of each target advertising creative based on the real-time delivery feedback data through a reinforcement learning model.

9. A computer-readable storage medium, characterized in that, The device stores a computer program that, when executed by a processor, causes the processor to perform the steps of the method for adjusting cross-platform advertising creatives as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, The device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method for adjusting cross-platform advertising creatives as described in any one of claims 1 to 7.

Images