An AI-based method and system for automatic design of art advertisements

Abstract

This invention discloses an AI-based automatic design method and system for art advertisements, relating to the field of advertising design technology. First, visual and textual features are extracted from historical advertising materials to construct a creative gene vector, and a creative feature map is established based on this vector. Second, user interaction data is used to calculate creative popularity parameters, and real-time feedback information is used to weight and optimize the advertising creative. Then, the target product features are semantically matched with the creative gene vector, and a portability score is calculated by correcting semantic relevance and historical validity data. Finally, the most suitable creative gene is selected for content fusion and rendering to generate the final advertising content. Based on dynamic optimization using creative popularity and user feedback, the real-time performance and accuracy of advertising are improved, effectively solving the cold start problem for new product advertising creatives, and enabling cross-platform and cross-domain creative migration and adaptation.

Description

Technical Field

[0001] This invention relates to the field of advertising design technology, specifically to an AI-based automatic design method and system for art advertisements. Background Technology

[0002] With the rapid development of digital marketing, advertising creative design has become a core element of brand promotion and marketing. Traditional advertising design often relies on the experience of creative personnel and market research data. However, because advertising creative needs to constantly adapt to changing market trends, and the demand is high with short cycles, traditional design methods are no longer sufficient to meet the need for quickly generating personalized, high-quality advertising creatives. Against this backdrop, utilizing artificial intelligence (AI) technology to automatically generate advertising creatives has gradually become a hot research and application area.

[0003] However, existing AI ad generation methods still have some limitations. In the early stages of promoting a new brand or product, the lack of sufficient historical advertising data causes AI systems to face a cold start problem when generating creative content, making it difficult to generate high-quality and highly relevant ad content. Existing systems often rely on fixed historical advertising data and creative templates, which makes the generated ad creatives lack flexibility and adaptability, and unable to accurately match the unique characteristics of new products or brands. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an AI-based automatic design method and system for art advertisements.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] In a first aspect, the present invention discloses an AI-based automatic design method for art advertisements, comprising the following steps:

[0007] Obtain historical advertising material data, and extract the visual and textual features of the historical advertising material data to construct a creative gene vector;

[0008] Construct a creative feature map, which includes a networked data structure with creative gene vectors as nodes and the correlation strength between creative gene vectors as edges;

[0009] Acquire user interaction data on creative gene vectors within a preset time window, and calculate creative popularity parameters based on the interaction data;

[0010] Obtain the description information of the target product, process the description information into a target product feature vector, and calculate the initial semantic relevance between the target product feature vector and the creative gene vector in the creative feature map;

[0011] Using the aforementioned creative popularity parameter as a weight, the initial semantic relevance is weighted and calculated to obtain the corrected semantic relevance.

[0012] Based on historical validity data of corrected semantic relevance and creative gene vectors, a portability score is calculated.

[0013] Select creative gene vectors whose portability scores reach a preset score threshold as portable creative gene vectors.

[0014] Based on the portable creative gene vector and the target product feature vector, content fusion and rendering are performed to generate target advertising content.

[0015] Secondly, this invention discloses an AI-based automatic art advertising design system, comprising:

[0016] The data acquisition and processing module is used to acquire historical advertising material data and extract the visual and textual features of the historical advertising material data to form a creative gene vector;

[0017] The feature map construction module is used to construct a creative feature map, which includes a networked data structure with creative gene vectors as nodes and the correlation strength between creative gene vectors as edges.

[0018] The popularity parameter calculation module is used to acquire user interaction behavior data on creative gene vectors within a preset time window, and calculate creative popularity parameters based on the interaction behavior data.

[0019] The relevance calculation module is used to obtain the description information of the target product, process the description information into the target product feature vector, and calculate the initial semantic relevance between the target product feature vector and the creative gene vector in the creative feature map;

[0020] The relevance correction module is used to perform a weighted calculation on the initial semantic relevance using the creative popularity parameter as a weight, to obtain the corrected semantic relevance.

[0021] The scoring calculation module is used to calculate the portability score based on historical validity data of corrected semantic relevance and creative gene vectors;

[0022] The creative selection module is used to select creative gene vectors whose portability scores reach a preset score threshold as portable creative gene vectors.

[0023] The ad generation module is used to perform content fusion and rendering based on the portable creative gene vector and the target product feature vector to generate target ad content.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0025] 1. By semantic matching based on creative feature maps and creative gene vectors, combined with creative popularity calculation based on real-time user behavior data, we can ensure that even in the early stages when data is scarce, high-quality advertising creatives can be generated through intelligent migration mechanisms, thus solving the cold start problem.

[0026] 2. By calculating the semantic relevance between the target product feature vector and historical ad creatives, and weighting it based on the creative popularity parameter, a high degree of matching between ad creative content and target products can be achieved. The weighted semantic relevance further optimizes the quality of the creatives, making the generated ads not only consistent with the characteristics of the target products, but also market attractive, thus improving the accuracy and relevance of the ads. Attached Figure Description

[0027] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0028] Figure 1 This is a flowchart of the method of the present invention;

[0029] Figure 2 This is a data flow diagram of the present invention;

[0030] Figure 3 This is a system architecture diagram of the present invention. Detailed Implementation

[0031] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0032] Traditional advertising creative generation systems, relying on static matching mechanisms based on fixed historical data and templates, struggle to address the data scarcity present in the early stages of new product launches. This results in insufficient semantic relevance between generated ad content and the target product, and an inability to dynamically capture the impact of user behavior feedback on the popularity of creative elements. For example, when a new brand launches a smart wearable device, the system needs to generate creative content tailored to the product's characteristics based on zero historical advertising data. However, existing methods can only access basic materials from a general template library, failing to construct a networked structure reflecting the relationships between creative elements. This limits the matching calculation between the target product's feature vector and the creative's gene vector to shallow semantic similarity, ignoring the dynamic correction effect of real-time user interaction data on the effectiveness of creative elements. Ultimately, the generated content suffers from issues such as chaotic layout logic and discrepancies between visual style and product positioning. Without addressing these problems, user click-through rates and conversion rates after ad campaigns will consistently fall below industry benchmarks. Decreased reuse efficiency of creative elements leads to wasted system computing resources, while the new product launch cycle is forced to extend, impacting the brand's market penetration speed and the efficiency of building user awareness.

[0033] To address the aforementioned challenges, this application first explores how to construct a dynamically linked network of creative elements to resolve the issue of insufficient semantic matching under data scarcity. Traditional methods rely on fixed templates, resulting in isolated storage of creative elements and failing to reflect potential relationships between them. To resolve this, this application considers encoding the visual and textual features of historical advertising materials into creative gene vectors, constructing a networked feature map through co-occurrence frequencies, and forming a reasonable creative element association structure. To address the issue that user interaction data cannot dynamically adjust the matching degree, this application introduces a time-window-based creative popularity parameter calculation mechanism. This mechanism adjusts the initial semantic relevance by capturing real-time user behavior feedback and simultaneously assesses the portability of creative elements using historical validity data, ensuring that creative gene vectors with both semantic adaptability and user appeal are selected during the cold start phase.

[0034] In this regard, such as Figure 1 As shown, this application proposes an AI-based automatic design method for art advertisements, comprising the following steps:

[0035] Historical advertising creative data is obtained, and its visual and textual features are extracted to form a creative gene vector. Historical advertising creative data refers to a collection of advertising content containing visual elements and textual descriptions. Specifically, image recognition technology and natural language processing technology can be used to extract visual and textual features. By fusing the two types of features into a vector form, the core attributes of the advertising creative can be represented, which helps to establish a quantifiable creative expression model. The creative gene vector is a vectorized expression form composed of visual and textual features. Specifically, it can be implemented by using deep neural networks to encode the features of advertising creatives. Subsequently, the semantic correlation between different advertising creatives can be reflected by the distance metric in the vector space.

[0036] A creative feature graph is constructed, which includes a networked data structure with creative gene vectors as nodes and the correlation strength between creative gene vectors as edges. The creative feature graph refers to the association network constructed with creative gene vectors as nodes. Specifically, the weight of the edges can be determined by calculating the cosine similarity or co-occurrence frequency between vectors, forming a knowledge graph that reflects the association relationship of creative elements. This provides a structured data foundation for subsequent intelligent recommendation of creative elements.

[0037] The system acquires user interaction data on creative gene vectors within a preset time window and calculates creative popularity parameters based on the interaction data. Creative popularity parameters are quantitative indicators that reflect the degree of user attention to specific creative elements. Specifically, they can be calculated by statistically analyzing the ratio of click-through rate, number of shares, and exposure, and are used to dynamically evaluate the popularity of creative elements in the current market environment.

[0038] Obtain the description information of the target product and process the description information into a target product feature vector. Calculate the initial semantic relevance between the target product feature vector and the creative gene vector in the creative feature map. The target product feature vector refers to the transformation of product description information into a numerical feature representation. Specifically, text embedding technology can be used to convert product attributes into high-dimensional vectors, enabling semantic similarity calculation with the creative gene vector.

[0039] Using the aforementioned creative popularity parameter as weight, the initial semantic relevance is weighted and calculated to obtain the corrected semantic relevance. The corrected semantic relevance refers to the comprehensive score that integrates product feature matching degree and creative popularity. Specifically, the initial similarity can be adjusted by a dynamic weighting algorithm, which can balance the creative relevance and timeliness requirements.

[0040] Based on historical validity data of corrected semantic relevance and creative gene vectors, a portability score is calculated. The portability score is a quantitative indicator that evaluates the ability of creative elements to adapt to new products. Specifically, it can be calculated in multiple dimensions by combining historical campaign performance data and current semantic matching degree to ensure that the selected creatives not only meet the product characteristics but also have market-validated effectiveness.

[0041] Select creative gene vectors whose portability scores reach a preset score threshold as portable creative gene vectors.

[0042] Based on the aforementioned portable creative gene vector and target product feature vector, content fusion and rendering are performed to generate target advertising content. Content fusion and rendering refers to the process of combining selected creative elements with product features to generate an advertisement. Specifically, generative adversarial networks can be used to achieve layout generation and style transfer, and an end-to-end model can be used to complete the automated synthesis from creative elements to a complete advertisement.

[0043] The core innovation of this application lies in constructing a dynamically evolving creative feature map, which integrates the popularity trend, semantic relevance, and historical effectiveness of historical creative elements in multiple dimensions. When historical data of new products is lacking, it can intelligently select and adapt the optimal creative elements based on networked feature associations and real-time user feedback, thus breaking through the limitations of traditional template-based advertising generation methods.

[0044] The working process and principle of this application are as follows: First, historical advertising material data is obtained, and visual and textual features are extracted to form creative gene vectors. Then, a creative feature map is constructed, with creative gene vectors as nodes and association strength as edges, forming a networked data structure. Next, user interaction behavior data on creative gene vectors within a preset time window is obtained, and creative popularity parameters are calculated. Target product description information is obtained and processed into target product feature vectors, and the initial semantic relevance with the creative gene vectors in the creative feature map is calculated. Using the creative popularity parameter as weight, the initial semantic relevance is weighted to obtain a corrected semantic relevance. Based on the corrected semantic relevance and historical validity data of creative gene vectors, a portability score is calculated. Creative gene vectors with portability scores reaching a preset score threshold are selected as portable creative gene vectors. Finally, based on the portable creative gene vectors and target product feature vectors, content fusion and rendering are performed to generate target advertising content.

[0045] This solution constructs a creative feature map and establishes relationships between creative gene vectors, facilitating more accurate semantic matching even with limited data. Introducing a creative popularity parameter dynamically adjusts the initial semantic relevance, reflecting user preferences for creative elements in real time. By combining historical validity data to assess the portability of creative elements, the system ensures that the selected creative gene vectors are both suitable for the target product and attractive. This dynamic association and real-time adjustment mechanism enables the system to generate more personalized and high-quality advertising content in the early stages of new product promotion.

[0046] like Figure 2 The diagram shown is a data flow chart of this application;

[0047] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0048] First, historical ad images and copy are retrieved from the ad creative database. A convolutional neural network is used to extract visual features from the images, and a natural language processing model is used to extract textual features from the copy. These features are then vectorized to form creative gene vectors. Next, a creative feature map is constructed based on the co-occurrence frequency of the creative gene vectors, and the cosine similarity between vectors is calculated as the association strength. Then, user interaction data such as clicks, shares, and skips for each creative gene vector within the last 30 days are obtained, and a popularity index relative to the average level is calculated as the creative popularity parameter. For the target product, keywords in its descriptive text are extracted and quantified into target product feature vectors. The cosine similarity between the target product feature vector and each creative gene vector in the creative feature map is calculated as the initial semantic relevance. The creative popularity parameter is normalized and used as a weighting coefficient, multiplied by the initial semantic relevance to obtain the corrected semantic relevance. Combining historical validity data such as click-through rates of the creative gene vectors over the past 3 months, a weighted average method is used to calculate the portability score. Creative gene vectors with scores exceeding 0.7 are selected as portable creative gene vectors. Finally, a generative adversarial network is used, with a portable creative gene vector as a condition and the target product feature vector as input, to generate advertising images and copy that conform to the product characteristics.

[0049] By employing the above-described solution, this application effectively addresses the cold-start problem of advertising creative generation in the initial stages of new product promotion. By constructing a dynamically linked network of creative elements, the accuracy of semantic matching is improved. Introducing a creative popularity parameter based on real-time user interaction data allows the system to dynamically capture and reflect changes in user preferences for creative elements. Combining historical validity data to evaluate the portability of creative elements ensures the quality and relevance of the generated advertising content. This method significantly improves the matching degree between advertising creatives and new product characteristics, enhancing the attractiveness and effectiveness of the generated content. Consequently, user click-through rates and conversion rates are increased after ad placement, the reuse efficiency of creative elements is improved, the market promotion cycle of new products is shortened, and brand awareness is established more quickly.

[0050] This application further proposes a process for obtaining the corrected semantic relevance by weighted calculation of the initial semantic relevance, including:

[0051] Obtain the historical popularity sequence corresponding to the creative gene vector, wherein the historical popularity sequence contains creative popularity parameters for multiple historical time windows;

[0052] The multiple historical time windows are divided into a basic popularity window group and a trend analysis window group. The basic popularity window group contains a first preset number of consecutive time windows, and the trend analysis window group contains a second preset number of discretely distributed time windows. The time windows of the basic popularity window group are arranged continuously to capture the short-term stability of creative popularity. The time windows of the trend analysis window group are discretely distributed, and historical data is selected at intervals to capture long-term trend fluctuations.

[0053] The baseline value of the basic popularity window group and the intensity of the popularity trend of the trend analysis window group are calculated respectively. The baseline value is obtained by calculating the mean or median of the creative popularity parameters within the basic popularity window group, reflecting the historical baseline level. The intensity of the popularity trend is calculated by the average rate of change of the slope of the popularity parameters of adjacent windows within the trend analysis window group, reflecting the magnitude of the trend change.

[0054] Based on the current creative popularity parameters, the popularity benchmark value, and the popularity trend intensity, a dynamic weight value is generated through a weight synthesis algorithm;

[0055] The initial semantic relevance is adjusted by weighting the dynamic weight values ​​to obtain the corrected semantic relevance.

[0056] During the generation of dynamic weight values, the ratio of the current creative popularity parameter to the popularity benchmark value measures the relative popularity deviation. The popularity trend strength is mapped to a trend confidence factor through a preset interval. The two are multiplied to obtain a comprehensive popularity index, and the final weight is determined by combining the dynamic weight mapping table.

[0057] Specifically, the time window is divided using a combination of continuous and discrete methods. The baseline heat index window group selects heat index data from the most recent three consecutive months to calculate the baseline value, while the trend analysis window group selects heat index data from the last month of each quarter within the past year to calculate the heat index trend strength. The heat index trend strength is determined by fitting the linear slope of the quarterly data using the least squares method, and the absolute value is taken as the heat index trend strength value. When the heat index trend strength is below a preset lower limit, the confidence factor takes the highest value, indicating a stable trend; when the heat index trend strength is in the middle range, the confidence factor decreases linearly with the strength value; when the heat index trend strength exceeds the upper limit, the confidence factor takes the lowest value, indicating a drastic trend fluctuation.

[0058] The dynamic weighted mapping table generates weight values ​​using piecewise linear interpolation based on the combination of the comprehensive popularity index and the benchmark popularity value. For example, when the comprehensive popularity index is 1.2 and the benchmark value is 0.8, the mapping table outputs a weight value of 0.75. This weight value is used to dynamically adjust the initial semantic relevance, allowing creative gene vectors with high popularity and stable trends to obtain higher corrected relevance. Through this dynamic adjustment mechanism, the semantic matching degree of advertising creatives can reflect the fluctuation characteristics of market popularity in real time, effectively improving the market adaptability of target advertising content.

[0059] Through the above technical solution, this application achieves dynamic weighted adjustment of the initial semantic relevance. This takes into account the historical popularity trends of the creative gene vector, making the corrected semantic relevance more accurately reflect the actual correlation between the creative gene vector and the target product. Furthermore, by introducing a basic popularity window group and a trend analysis window group, both sensitivity to recent popularity data and the ability to capture long-term popularity trends are ensured, thereby improving the accuracy and stability of semantic relevance correction.

[0060] This application further proposes the following process for generating dynamic weight values:

[0061] Calculate the current creative popularity parameter Compared with the aforementioned heat reference value relative heat ratio :

[0062] Relative heat ratio The ratio of the two reflects the degree to which the instantaneous temperature deviates from the baseline level.

[0063] Determine the intensity of the heat trend With respect to the preset trend strength range Relationship:

[0064] when At that time, the trend confidence factor takes the following values:

[0065] ;

[0066] when At that time, the trend confidence factor takes the following values:

[0067] ;

[0068] when At that time, the trend confidence factor takes the following values:

[0069] ;

[0070] in, , These are the preset trend strength thresholds. , These are the preset upper and lower limits of the confidence factor, and ;

[0071] relative heat ratio With trend confidence factor Multiply to obtain a comprehensive popularity index. ;

[0072] The comprehensive heat index Compared with the heat baseline value The inputs are fed into the dynamic weight mapping table, the dynamic weight mapping table is queried, and the final dynamic weight values ​​are output. .

[0073] Trend confidence factor Based on the intensity of the heat trend The interval in which it is located is determined by linear interpolation or boundary values;

[0074] For example: when the intensity of the heat trend Below When adopted The value is 0.9; the intensity of the heat trend. Higher than When adopted The value of is 0.3, and it decreases proportionally in the middle range.

[0075] The dynamic weight mapping table maps weight values ​​to preset combinations of indicators and benchmark values; for example, when the comprehensive popularity indicator... The value is 1.2 and the heat reference value. When the value is 0.8, the mapping weight is 0.75.

[0076] Specifically, in calculating the relative heat ratio Then, through the intensity of the heat trend The confidence factor is dynamically adjusted within the given interval.

[0077] For example: intensity of popularity trend At 0.4 and In between;

[0078] like and 3, then the trend confidence factor for:

[0079] .

[0080] relative heat ratio With trend confidence factor Multiply to obtain the comprehensive heat index Combined with the heat baseline value The value 0.6 is used to query the mapping table, and the dynamic weight value 0.95 is output.

[0081] This process quantifies the degree of deviation between the confidence level of trend strength and the instantaneous popularity, enabling the weight values ​​to be adaptively adjusted. This avoids weight bias caused by trend fluctuations or data dispersion, thereby improving the accuracy of correcting semantic relevance.

[0082] Through the above technical solution, this application achieves adaptive adjustment of dynamic weight values. Therefore, the system can flexibly adjust the semantic relevance weights based on real-time changes and trend strength of creative popularity, improving the timeliness and relevance of advertising creatives. Furthermore, by introducing a trend confidence factor, this solution effectively balances the impact of short-term fluctuations and long-term trends, enhancing the stability and reliability of weight calculation. Specifically, when the popularity trend is not obvious, the system relies more on current popularity data; while when the popularity trend is significant, the system appropriately reduces its reliance on current popularity, thereby avoiding excessive pursuit of short-term fluctuations. This dynamic adjustment mechanism makes the selection of advertising creatives more intelligent and precise, helping to improve the effectiveness of advertising placement.

[0083] This application further proposes the following calculation process for the portability score:

[0084] The historical validity data are associated with the basic popularity window group and the trend analysis window group according to their corresponding time windows;

[0085] Based on the historical validity data corresponding to the basic heat window group, the basic validity value is calculated. The basic validity value The median of historical validity data within this time window group; the basic validity value. By calculating the median and filtering out extreme fluctuations, it can reflect the central tendency of historical validity.

[0086] The trend validity value is calculated based on the historical validity data corresponding to the trend analysis window group. The trend validity value This is the average rate of change of historical validity data within the same time window group; trend validity value. The persistence of changes in effectiveness is captured by averaging the rates of change of adjacent windows;

[0087] Establish a dynamic normalization mapping layer to correct semantic relevance. Basic validity value Trend validity value Each value is mapped to a unified preset numerical range to obtain normalized semantic relevance. Normalized basic validity value and normalized trend validity value The dynamic normalization mapping layer maps indicators with different dimensions to the same interval, eliminating the difference in magnitude between indicators.

[0088] Portability scores are calculated using a composite scoring formula. The scoring formula is as follows:

[0089] ;

[0090] in, , , The preset weighting coefficients, and .

[0091] The scoring formula balances the contributions of semantic relevance, historical stability, and trend changes by pre-setting weighting coefficients.

[0092] Specifically, historical validity data is divided into a basic window group and a trend window group. The median calculation of the basic window group can exclude short-term abnormal fluctuations and retain long-term stable characteristics, while the average rate of change of the trend window group can identify a persistent trend of increasing or decreasing validity. A dynamic normalization mapping layer corrects semantic relevance. Basic validity value Trend validity value Standardize each indicator separately, for example, by linearly scaling each indicator to the 0-1 range to ensure the comparability of indicators across different dimensions.

[0093] In the scoring formula, , , These values ​​can be set to 0.5, 0.3, and 0.2 respectively, prioritizing semantic relevance while also considering the impact of historical stability and trend changes. Through weighted summation, the portability score integrates semantic matching, historical performance stability, and effectiveness trends, improving the accuracy of ad creatives in matching target products.

[0094] Through the above technical solution, this application achieves an accurate assessment of the portability of creative gene vectors. Therefore, the system can comprehensively consider the semantic relevance, historical validity, and changing trends of creative ideas, thereby selecting the most suitable creative gene vector for the target product. This assessment method improves the targeting and effectiveness of advertising creatives, reduces the possibility of unsuitable creatives being selected, and thus enhances the quality of automatically generated advertising content.

[0095] This application further proposes the following process for calculating the initial semantic relevance:

[0096] The neighborhood subgraph of the target creative gene vector is extracted from the creative feature map. The neighborhood subgraph contains the target creative gene vector and a preset number of neighboring creative gene vectors directly connected by edges. The extraction of the neighborhood subgraph is achieved by traversing the nodes directly connected to the target creative gene vector in the creative feature map. The preset number can be dynamically adjusted according to the map density.

[0097] Calculate the direct similarity between the target product feature vector and the target creative gene vector;

[0098] Calculate the similarity between the target product feature vector and each of the adjacent creative gene vectors in the neighborhood subgraph to form a neighborhood similarity set;

[0099] Based on the association strength of each edge in the neighborhood subgraph, the neighborhood similarity set is weighted and aggregated to obtain neighborhood enhanced similarity. In the weighted aggregation process, the association strength of each edge is used as the weight coefficient of the corresponding adjacent creative gene vector, and neighborhood enhanced similarity is generated through linear weighting or nonlinear transformation.

[0100] The target direct similarity and the neighborhood enhanced similarity are fused to calculate the initial semantic relevance. An aggregation consistency index is introduced into the fusion process to dynamically adjust the ratio of direct similarity to neighborhood enhanced similarity. For example, when the standard deviation of the neighborhood similarity set is low, it indicates high neighborhood consistency, and the fusion weight of the neighborhood enhanced similarity is increased.

[0101] Specifically, when calculating the initial semantic relevance, the local network structure of the target creative gene vector is first obtained through a neighborhood subgraph, where the association strength between adjacent nodes reflects the co-occurrence or complementarity between creative elements. Target direct similarity measures the degree of matching between the target product and the core creative gene, while the neighborhood similarity set, calculated by traversing adjacent nodes, covers extended creative elements associated with the core creative gene. After weighted aggregation based on association strength, the neighborhood enhanced similarity characterizes the overall matching level between the target product and the local structure of the creative gene network. During the final fusion process, the aggregation consistency index dynamically adjusts the weights of the two types of similarity. This method effectively integrates the local network information of the creative gene, improving the comprehensiveness and accuracy of semantic relevance assessment.

[0102] Through the above technical solution, this application fully utilizes neighborhood information in the creative feature map, improving the accuracy and robustness of semantic relevance calculation. By considering the neighborhood information of the target creative gene vector, the bias that may be caused by single vector matching is reduced, making the relevance calculation results more comprehensive and reliable. At the same time, the graph-based calculation method also improves the system's adaptability to new products and new brands, effectively alleviating the cold start problem.

[0103] This application further proposes the following fusion process between the target direct similarity and the neighborhood enhanced similarity:

[0104] Obtain the neighborhood similarity set, calculate the arithmetic mean and standard deviation of the neighborhood similarity set, and use the ratio of the standard deviation to the arithmetic mean as the aggregation consistency index. Aggregate Consistency Index The ratio of the standard deviation to the arithmetic mean of the neighborhood similarity set reflects the dispersion of neighborhood similarity. The lower the ratio, the more concentrated the neighborhood similarity distribution.

[0105] Based on the convergence consistency index Calculate the neighborhood fusion coefficient :

[0106] ,in, This is a preset global adjustment factor, whose value is set by staff based on historical experience. The value range is usually set to 0.3 to 0.6, and it is used to control the global influence intensity of neighborhood similarity.

[0107] Based on neighborhood fusion coefficient Fusion target direct similarity Enhanced similarity with neighboring regions :

[0108] ,in This represents the initial semantic relevance.

[0109] Direct similarity of targets Enhanced similarity with neighboring regions The fusion adopts a linear weighting method, and the weight allocation is based on the neighborhood fusion coefficient. Dynamic adjustment.

[0110] Specifically, when the aggregation consistency index of the neighborhood similarity set A higher value indicates greater fluctuations in neighborhood similarity, at which point the neighborhood fusion coefficient... Reduce direct similarity between targets The weights are increased to reduce noise interference;

[0111] Conversely, when the aggregation consistency index When the value is low, the neighborhood fusion coefficient Improve neighborhood similarity The weights are increased to take advantage of the synergistic effect of neighborhood data.

[0112] For example, when the arithmetic mean of the neighborhood similarity set is 0.75 and the standard deviation is 0.15, the aggregation consistency index... It is 0.2;

[0113] If global adjustment factor If we take 0.5, then the neighborhood fusion coefficient for At this point, the initial semantic relevance Based on 60% direct similarity of the target Enhanced similarity with 40% of the neighborhood The decision is made jointly. This dynamic adjustment mechanism can optimize the fusion weights based on the distribution characteristics of neighborhood data, thereby improving the robustness of the initial semantic relevance.

[0114] Through the above technical solution, this application achieves dynamic fusion of direct target similarity and enhanced neighborhood similarity. Therefore, the distribution characteristics of neighborhood similarity are fully considered when calculating the initial semantic relevance, improving the accuracy and robustness of similarity calculation. Furthermore, by introducing an aggregation consistency index and a global adjustment factor, the fusion process can adaptively adjust according to the distribution of neighborhood similarity, thus obtaining more reasonable fusion results in different scenarios.

[0115] This application further proposes a method for calculating creative popularity parameters based on the aforementioned interactive behavior data, including:

[0116] The interactive behavior data is decomposed into positive interactive data and negative interactive data; wherein, the positive interactive data includes ad click-through rate and ad share count, and the negative interactive data includes ad skip rate and negative comment ratio;

[0117] Preset weighting coefficients are assigned to the positive and negative interaction data respectively, and a weighted net interaction value is calculated. The positive and negative interaction data are weighted using these preset weighting coefficients, which are dynamically adjusted based on the historical conversion data of the advertising platform. The net interaction value is calculated using a linear weighted model, with positive interaction data having a higher weight than negative interaction data.

[0118] The exposure data of the creative gene vector within a preset time window is obtained, and the ratio of the net interaction value to the exposure data is calculated as the basic interaction efficiency value. During the calculation of the basic interaction efficiency value, the exposure data is obtained in real time through the log data of the advertising system.

[0119] The basic interaction efficiency value is compared with the average interaction efficiency value of all creative gene vectors within the same time period to calculate the relative popularity index, and the relative popularity index is used as the creative popularity parameter.

[0120] The relative popularity index uses the standard deviation normalization method to map the basic interaction efficiency value to a standardized scoring system based on the industry average.

[0121] Specifically, in positive interaction data, ad click-through rate is statistically analyzed through click event tracking, and ad share counts are obtained through social platform interfaces. In negative interaction data, ad skip rate is recorded through video player event callbacks, and the proportion of negative comments is calculated after sentiment analysis of user comments using a natural language processing model. Preset weighting coefficients are set differently according to ad type; for example, the weighting coefficient for skip rate is higher for video ads than for banner ads. Exposure data is deduplicated to exclude duplicate exposure records from the same user. The calculation of the basic interaction efficiency value uses a smoothing method where the numerator is the net interaction value and the denominator is the exposure plus one, avoiding division by zero errors caused by zero exposures. The relative popularity index is calculated by subtracting the industry average from the basic interaction efficiency value of the current creative gene vector and then dividing by the industry standard deviation to generate a popularity parameter that follows a standard normal distribution.

[0122] Through the above technical solution, this application achieves precise quantification of creative popularity. By decomposing interactive behavior data into positive and negative categories and assigning different weights, it can more comprehensively reflect users' true reactions to advertising creatives. Introducing exposure volume as a benchmark to calculate the interaction efficiency value avoids the bias that may result from relying solely on absolute interaction data. By comparing with the average level to obtain a relative popularity index, the popularity of advertising creatives of different periods and scales can be compared horizontally. This method improves the accuracy and comparability of creative popularity assessment, providing reliable data support for subsequent creative selection and optimization.

[0123] This application further proposes the construction of creative feature maps, including:

[0124] The co-occurrence frequency (COF) among the creative gene vectors is calculated. The COF is determined based on the number of times each creative gene vector co-occurs in historical ad creative data. The COF is calculated by statistically analyzing the number of times different creative gene vectors appear simultaneously within the same ad creative data. A higher COF value corresponds to a higher COF, indicating a stronger correlation between the creative gene vectors.

[0125] The networked data structure is formed by constructing edges between the creative gene vectors based on the co-occurrence frequency. Edge construction involves mapping the co-occurrence frequency to an association strength value, for example, normalizing the co-occurrence frequency to the range of 0 to 1, which serves as the edge weight. Furthermore, the co-occurrence frequency statistics can be limited to specific advertising categories or time periods to enhance the domain-specificity of the association relationships.

[0126] Specifically, in historical ad creative data, each ad creative contains multiple creative gene vectors, such as visual layout features and copywriting style features within the same ad. By traversing all historical ad creatives, the frequency of each pair of creative gene vectors appearing in the same creative is recorded.

[0127] For example, if creative gene vectors A and B co-occur 30 times in 100 historical ads, their co-occurrence frequency is 30%. When generating edges based on this frequency, the 30% is converted into an association strength value, for example, directly used as a weight or adjusted through a piecewise function. Thus, the edges in the networked data structure can objectively reflect the combination patterns of creative elements in practical applications, avoiding the subjective bias of manually setting associations. Edges constructed using co-occurrence frequencies ensure that the creative feature map is closer to real-world advertising design scenarios, thereby improving the matching degree between target ads and product characteristics in the subsequent content generation stage.

[0128] Through the above technical solution, this application effectively solves the problem of inaccurate creative correlation construction caused by insufficient historical data in the prior art. By quantifying the intrinsic correlation between creative elements through co-occurrence frequency, the networked data structure can objectively reflect the actual combination rules of advertising creative elements, thereby improving the logical coherence and combination rationality of advertising content generation, and providing a reliable basis for subsequent semantic relevance calculation.

[0129] This application further proposes the following process for generating the target advertising content:

[0130] The portable creative gene vector is deconstructed to separate the structural gene vector and the style gene vector.

[0131] Among them, the structural gene vector is extracted through specific encoding channels of the decoupled network and is used to control the macro layout of the advertising content, including but not limited to: topological features such as the coordinates of the main body position, the boundary of the copy area, the proportion of white space, and the spatial relationship of visual elements.

[0132] Style gene vector: Extracted through another encoding channel of the decoupled network, used to control the visual and textual style of advertising content, including but not limited to: color distribution histogram, font type features, material texture description, lighting effect parameters and other performance features;

[0133] Through this deconstruction step, the system breaks down complex creative expressions into relatively independent structural and stylistic components, laying the foundation for subsequent precise control and integration.

[0134] The target product feature vector and the structural gene vector are fused in the first round, and the layout generation model is used to calculate and generate the structural skeleton of the target advertisement.

[0135] The layout generation model can be a conditional variational autoencoder or a Transformer-based layout generator;

[0136] The layout generation model is calculated based on the following logic:

[0137] Analyze the layout patterns (such as the golden ratio, symmetrical composition, etc.) contained in the structural gene vectors.

[0138] By combining the visual characteristics of the product itself (such as the product's aspect ratio and the number of core selling points) in the target product feature vector, the layout parameters are automatically adjusted and optimized to generate a composition skeleton that retains the essence of the original creative layout while adapting to the characteristics of the target product. The composition skeleton generated in this step is a structured layout scheme that clarifies the position, size, and spatial relationship of each element in the advertising canvas, providing a precise framework for subsequent content filling.

[0139] After the composition skeleton is determined, style content is generated in parallel; the target product feature vector and the style gene vector are fused in a second round, and visual elements and text content that match the composition skeleton are calculated and generated through the style rendering model.

[0140] Style rendering models can use style transfer networks or conditional GANs. The style rendering model is calculated based on the following logic:

[0141] Analyze the aesthetic features (such as color matching and font style) contained in the style gene vector, and combine them with the brand tone and product attributes in the target product feature vector to generate visual elements (such as background patterns and decorative elements) and copywriting content (such as advertising slogans and product descriptions) that match the composition skeleton. This step ensures that the generated visual and text content not only inherits the aesthetic characteristics of the source idea in style, but also maintains consistency with the brand image of the target product.

[0142] Based on the composition skeleton, the visual elements and text content are combined and rendered to generate the final target advertising content.

[0143] Precisely arrange visual elements and text content according to the spatial constraints of the composition framework:

[0144] Place the main product image in the designated position on the composition skeleton, fill the generated advertising copy in the preset text area, adjust the hierarchy and size ratio of each element according to the layout requirements, and use the graphics rendering engine to synthesize the layout elements into complete target advertising content (such as JPEG, PNG images or MP4 video files).

[0145] Through the above technical solution, this application achieves decoupled control of structural layout and style features in the process of advertising content generation, and effectively solves the creative adaptability problem caused by insufficient historical data when promoting new products through a phased fusion mechanism. Specifically, the independent processing of structural gene vectors and style gene vectors ensures the coordinated optimization of macro-layout and micro-visual elements of the advertisement, avoiding style conflicts or information redundancy problems caused by single feature fusion in traditional methods. Thus, even in the absence of historical data support, it can still generate advertising content with professional design quality and accurate matching of product characteristics.

[0146] like Figure 3 The diagram shown is a system architecture diagram of this application. This application further proposes an AI-based automatic art advertisement design system, including a data acquisition and processing module, a feature map construction module, a popularity parameter calculation module, a relevance calculation module, a relevance correction module, a scoring calculation module, a creative selection module, and an advertisement generation module.

[0147] The data acquisition and processing module constructs creative gene vectors by extracting visual and textual features from historical advertising materials;

[0148] The feature map construction module uses creative gene vectors as nodes and constructs a networked data structure based on co-occurrence frequency to build association strength.

[0149] The popularity parameter calculation module decomposes interactive behavior data into positive and negative indicators, calculates the ratio of net interaction value to exposure as the basic interaction efficiency value, and compares it with the average efficiency to obtain the relative popularity index.

[0150] The relevance calculation module extracts the neighboring nodes of the target creative gene vector through the neighborhood subgraph, calculates the direct similarity of the target and the neighborhood enhanced similarity, and obtains the initial semantic relevance after fusion;

[0151] The relevance correction module adjusts the weight values ​​according to the dynamic weight mapping table to perform weighted correction on the initial semantic relevance.

[0152] The scoring calculation module associates historical validity data with different time window groups, calculates basic validity values ​​and trend validity values, and generates a portable score through normalization mapping and scoring synthesis formula;

[0153] The creative selection module filters creative gene vectors that meet the scoring criteria.

[0154] The ad generation module separates the structural gene vector and the style gene vector, performs layout generation and style rendering separately, and finally combines them to generate the target ad content.

[0155] By automating the entire process from historical data analysis to new advertising content generation through data transfer and collaborative processing between modules, the cold start problem is solved and the adaptability of creative content is improved.

[0156] Through the above technical solutions, this application effectively solves the problem of low creative matching accuracy caused by insufficient data when generating new brand advertisements. By representing the correlation of advertising elements through a dynamic graph structure and adjusting feature weights in conjunction with real-time interactive data, the system can accurately capture potentially effective creative elements during the cold start phase. The semantic matching capability of low-exposure creative genes is enhanced based on a neighborhood similarity propagation mechanism, avoiding the over-reliance on a single similarity indicator in traditional methods. The two-level content generation architecture enables independent optimization of advertising layout and style elements, significantly improving the fit between generated content and target product characteristics.

[0157] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. An AI-based automatic design method for artistic advertising, characterized in that: The method comprises the following steps: obtaining historical advertising material data, and extracting visual features and text features of the historical advertising material data to form creative gene vectors; constructing a creative feature map, the creative feature map comprising a network data structure with the creative gene vectors as nodes and the correlation strength between the creative gene vectors as edges; obtaining user interaction behavior data of the creative gene vectors within a preset time window, and calculating a creative heat parameter based on the interaction behavior data; obtaining description information of a target product, and processing the description information into a target product feature vector, and calculating an initial semantic correlation degree between the target product feature vector and the creative gene vectors in the creative feature map; weighting the initial semantic correlation degree by using the creative heat parameter as a weight to obtain a modified semantic correlation degree; calculating a portability score based on the modified semantic correlation degree and historical effectiveness data of the creative gene vectors; selecting a creative gene vector with a portability score reaching a preset score threshold as a portable creative gene vector; based on the portable creative gene vector and the target product feature vector, performing content fusion and rendering to generate target advertising content.

2. The AI-based automatic design method of artistic advertisements according to claim 1, characterized in that: The process of weighting the initial semantic correlation degree to obtain the modified semantic correlation degree comprises the following steps: obtaining a historical heat sequence corresponding to the creative gene vector, the historical heat sequence comprising creative heat parameters of a plurality of historical time windows; dividing the plurality of historical time windows into a basic heat window group and a trend analysis window group; wherein the basic heat window group comprises a first preset number of continuous time windows, and the trend analysis window group comprises a second preset number of time windows distributed discretely; respectively calculating a heat benchmark value of the basic heat window group and a heat trend strength of the trend analysis window group; generating a dynamic weight value by a weight synthesis algorithm based on the current creative heat parameter, the heat benchmark value and the heat trend strength; weighting and adjusting the initial semantic correlation degree by using the dynamic weight value to obtain the modified semantic correlation degree.

3. The AI-based automatic design method of artistic advertisement according to claim 2, characterized in that: The process of generating the dynamic weight value comprises the following steps: calculating the current creative heat parameter relative heat ratio to the heat baseline value :​ ； determining the intensity of the trend of the heat in relation to a preset trend intensity interval ​ When the trend confidence factor takes the value: ； When the trend confidence factor takes the value: ； When the trend confidence factor takes the value: ； wherein, , are preset trend strength thresholds, respectively, , are preset confidence factor upper and lower limits, respectively, and ; The relative hotness ratio is multiplied by a trend confidence factor to obtain a composite hotness indicator ; The comprehensive heat index is inputted into a dynamic weight mapping table together with a heat reference value to query the dynamic weight mapping table and output a final dynamic weight value to query the dynamic weight mapping table and output a final dynamic weight value .

4. The AI-based automatic design method of artistic advertisement according to claim 2, characterized in that: The process of calculating the portability score comprises the following steps: associating the historical effectiveness data to the basic heat window group and the trend analysis window group respectively according to the time windows corresponding to the historical effectiveness data. based on historical effectiveness data corresponding to the base heat window group, compute a base effectiveness value , the base effectiveness value is a median of the historical effectiveness data within the time window group; based on historical validity data corresponding to the set of trend analysis windows, calculating a trend validity value , the trend validity value is an average of rates of change of historical validity data for adjacent windows within the set of time windows A dynamic normalization mapping layer is established to map the corrected semantic relevance , the base effectiveness value , the trend effectiveness value to a unified preset numerical interval respectively to obtain normalized semantic relevance , the normalized base effectiveness value , the normalized trend effectiveness value ; The portability score is calculated by a score synthesis formula , the score synthesis formula being: ； wherein, , , is a preset weight coefficient, and .

5. The AI-based automatic design method of artistic advertisements according to claim 1, characterized in that: The process of calculating the initial semantic correlation degree comprises the following steps: extracting a neighborhood subgraph of a target creative gene vector from the creative feature map, the neighborhood subgraph comprising the target creative gene vector and a preset number of adjacent creative gene vectors directly connected to the target creative gene vector through edges; calculating a target direct similarity between the target product feature vector and the target creative gene vector; calculating a similarity between the target product feature vector and each of the adjacent creative gene vectors in the neighborhood subgraph to form a neighborhood similarity set; weighting and aggregating the neighborhood similarity set based on the correlation strength of each edge in the neighborhood subgraph to obtain a neighborhood enhanced similarity; fusing the target direct similarity and the neighborhood enhanced similarity to calculate the initial semantic correlation degree.

6. The AI-based automatic design method of artistic advertisement according to claim 5, characterized in that: The process of fusing the target direct similarity and the neighborhood enhanced similarity comprises the following steps: obtaining the set of neighborhood similarities and calculating an arithmetic mean and a standard deviation of the set of neighborhood similarities and using the ratio of the standard deviation to the arithmetic mean as an aggregated consistency indicator ; Based on a consensus index of aggregation Computing neighborhood fusion coefficients : wherein, is a preset global adjustment factor; Based on neighborhood fusion coefficients Fusion target direct similarity With neighborhood enhanced similarity : wherein is the initial semantic relevance.

7. The AI-based automatic design method of artistic advertisements according to claim 1, characterized in that: calculating a creativity heat parameter based on the interaction behavior data comprises: decomposing the interaction behavior data into positive interaction data and negative interaction data, wherein the positive interaction data comprises an advertisement click rate and an advertisement sharing number, and the negative interaction data comprises an advertisement skipping rate and a negative comment proportion; respectively assigning preset weight coefficients to the positive interaction data and the negative interaction data, and calculating a weighted net interaction value; obtaining exposure data of the creativity gene vector within a preset time window, and calculating a ratio of the net interaction value to the exposure data as a basic interaction efficiency value; comparing the basic interaction efficiency value with an average interaction efficiency value of all creativity gene vectors in the same time period, and calculating a relative heat index, and taking the relative heat index as the creativity heat parameter.

8. The AI-based automatic design method of artistic advertisement according to claim 1, characterized in that: the construction of the creativity feature map comprises: calculating a co-occurrence frequency between creativity gene vectors, the co-occurrence frequency being determined based on a common occurrence number of the creativity gene vectors in historical advertisement material data; constructing edges between the creativity gene vectors based on the co-occurrence frequency, and forming the networked data structure.

9. The AI-based automatic design method of artistic advertisement according to claim 1, characterized in that: the generation process of the target advertisement content is: deconstructing the portable creativity gene vector, separating a structure gene vector and a style gene vector, the structure gene vector being used to control the macro layout of the advertisement content, and the style gene vector being used to control the visual and text style of the advertisement content; performing a first round of fusion of the target product feature vector and the structure gene vector, and calculating a composition skeleton of the target advertisement through a layout generation model; performing a second round of fusion of the target product feature vector and the style gene vector, and calculating visual elements and copy contents matched with the composition skeleton through a style rendering model; combining and rendering the visual elements and the copy contents based on the composition skeleton, and generating the final target advertisement content.

10. An AI-based automatic design system for artistic advertisement, characterized in that: comprises: a data acquisition and processing module, configured to obtain historical advertisement material data, and extract visual features and text features of the historical advertisement material data to form creativity gene vectors; a feature map construction module, configured to construct a creativity feature map, the creativity feature map comprising a networked data structure with creativity gene vectors as nodes and an association strength between creativity gene vectors as edges; a heat parameter calculation module, configured to obtain interaction behavior data of users on creativity gene vectors within a preset time window, and calculate a creativity heat parameter based on the interaction behavior data; a relevance calculation module, configured to obtain description information of a target product, process the description information into a target product feature vector, and calculate an initial semantic relevance between the target product feature vector and creativity gene vectors in the creativity feature map; a relevance correction module, configured to take the creativity heat parameter as a weight, and perform weighted calculation on the initial semantic relevance to obtain a corrected semantic relevance; a score calculation module, configured to calculate a portability score based on the corrected semantic relevance and historical effectiveness data of the creativity gene vector; a creative selection module configured to select a creative gene vector whose portability score reaches a preset score threshold as a portable creative gene vector; an advertisement generation module configured to perform content fusion and rendering based on the portable creative gene vector and a target product feature vector to generate target advertisement content.

Images