An AI-based intelligent advertisement promotion method

Abstract

The application discloses an AI-based intelligent advertisement promotion method, relates to the technical field of advertisement information management, and comprises the following steps: performing visual analysis and text analysis on a promotion product to generate a product visual slice set and a product keyword set; constructing a scene data set based on scene data, comparing the product keyword set with a scene keyword set, and calculating a scene matching degree score of the promotion product and each scene; calculating the complementarity score of each associated product based on the associated product set of the e-commerce platform data and the promotion product, and sorting the associated products; generating a candidate scene graph based on the promotion product, the target scene, the top five associated products in the sorting, and the visual feature label; obtaining a semantic importance weight sequence by analyzing the promotion product script, obtaining a visual importance weight sequence by identifying each candidate scene graph, calculating a difference evaluation value of the two weight sequences, and selecting the picture with the minimum scene difference evaluation value as the optimal scene graph.

Description

Technical Field

[0001] This invention relates to the field of advertising information management technology, and in particular to an AI-based intelligent advertising promotion method. Background Technology

[0002] With the rapid development of the internet and e-commerce, advertising has become a key means for businesses to acquire users, enhance brand awareness, and drive sales conversion. However, traditional advertising methods mainly rely on manual creation and static content delivery, which is not only time-consuming and labor-intensive but also difficult to achieve large-scale personalized delivery, resulting in limited advertising effectiveness. Although artificial intelligence technology has been gradually applied to areas such as ad targeting and copywriting generation, existing solutions still have significant shortcomings: ad content and scenario matching are mostly based on simple keywords or categories, lacking multimodal fusion analysis of product visual and textual features, resulting in weak relevance between ads and user scenarios and insufficient appeal; related product recommendations mostly rely on historical behavioral data, lacking comprehensive consideration of functional complementarity, visual style consistency, and marketing strategy synergy, and the recommendation results are often limited to similar products, making it difficult to achieve intelligent matching across product categories and scenarios. Ad creative generation still heavily relies on manual design, which is costly and time-consuming, and cannot quickly respond to the real-time delivery needs of multiple scenarios and products; existing advertising systems generally lack dynamic optimization mechanisms based on user attention trajectories and psychological stages, and ad content cannot evolve personalizedally with the exposure process, resulting in a decline in advertising effectiveness as the frequency of exposure increases, making it difficult to continuously guide user interest and drive conversion. Summary of the Invention

[0003] This application provides an AI-based intelligent advertising promotion method that addresses the problems of existing technologies, such as reliance on manual creative work, crude matching, limited related recommendations, inefficient material generation, and lack of dynamic optimization. Through multimodal AI fusion analysis, generative AI-generated scene graphs with optimal selection, and dynamic tracking of user attention to adjust content, it achieves precise matching of advertisements to scenes, intelligent cross-category pairing, efficient automated generation, and continuous enhancement of attractiveness, thereby improving the relevance, conversion efficiency, and personalization of advertisements.

[0004] This application provides an AI-based intelligent advertising promotion method, including: S1: Perform visual and textual analysis on the promoted products, extract the visual feature tags and key selling points of the products, and generate a set of product visual slices and a set of product keywords. S2: Construct a scenario dataset based on scenario data, which includes a set of scenario keywords for each scenario; compare the product keyword set with the scenario keyword set, and calculate the scenario matching score between the promoted product and each scenario; S3: Select the target scenario based on the scenario matching score, and calculate the complementarity score of each related product based on the e-commerce platform data and the related product set of the promoted product, and sort the related products. S4: Based on the promoted product, target scenario, top 5 related products, and visual feature tags, generate several candidate scenario images; parse the promotional product copy to obtain a semantic importance weight sequence, identify each candidate scenario image to obtain a visual importance weight sequence, calculate the difference evaluation value between the two weight sequences, and select the image with the smallest scenario difference evaluation value as the optimal scenario image.

[0005] Furthermore, the visual analysis includes: using image segmentation technology, cutting different functional and appearance parts of the product image into independent visual slices, with each slice corresponding to several visual feature labels, forming a product visual slice set; The text analysis includes: performing natural language processing on the descriptive text of the promoted products to extract key nouns, verbs and adjectives that describe the product's functions, attributes and usage scenarios, forming a set of product keywords.

[0006] Furthermore, the scenario matching score includes a core selling point scenario fit score, a keyword co-occurrence frequency score, and a historical marketing performance score. The core selling point scenario fit score is obtained by matching and scoring the product's core selling points with the core needs of the scenario. The keyword co-occurrence frequency score is obtained by statistically analyzing the co-occurrence frequency of product keywords and scenario keywords in historical advertising copy and converting it into a score. The historical marketing performance score is obtained by calculating the average click-through rate of candidate scenarios in historical advertising campaigns and converting it into a score. The scenario matching score is then calculated through a weighted fusion. , in, For the scene With product promotion Scene matching score, The score is based on the relevance of the core selling points to the specific scenarios. The score is the co-occurrence frequency of keywords. Score the historical marketing effectiveness. , and For the corresponding weight coefficients, and .

[0007] Furthermore, the complementary scoring system comprises a functional complementary score, a visual style consistency score, and a marketing strategy bonus score. The functional complementary score is calculated by analyzing the functional relationship between the promoted product and related products in the usage process. The visual style consistency score is calculated by determining the degree of consistency between the related product and the promoted product and the dominant style of the scenario. The marketing strategy bonus score is calculated by considering the business strategy factors of the related product. The complementary score is obtained through a weighted fusion calculation. , in, To promote products Related products With scene Complementarity score, For functional complementarity scores, Score for visual style consistency. Add points to marketing strategies , and For the corresponding weight coefficients, and .

[0008] Furthermore, the semantic importance weight sequence is obtained by parsing promotional product copy using natural language processing. The parsed elements include core selling point keywords, scene atmosphere keywords, and implicit related demand keywords. Among them, core selling point keywords are words with a repetition rate greater than 10%, scene atmosphere keywords use adjectives describing the emotional aspects of the scene, and implicit related demand keywords infer potential user needs from the context, generating a standardized semantic importance weight sequence: , in, It is a sequence of semantic importance weights. In the promotional copy, the first The semantic importance weights of each key element. This represents the total number of key elements extracted from the product promotion copy. .

[0009] Furthermore, the visual importance weight sequence is obtained through visual analysis of each candidate scene image. Using an object detection model and image segmentation technology, the main elements in the image are identified, including promoted products, related products, and the scene background. Based on their visual salience in the image, a standardized visual importance weight sequence is generated. , in, For visual importance weight sequence, For the first The visual importance weight of each major element in the scene graph. An index is used to identify the first element in the set. There are 1 elements, with values ​​ranging from 1 to 1. , This represents the total number of main elements extracted from the candidate scene graph. ; The visual saliency is calculated based on area proportion, positional centrality, and color contrast: , in, As visual importance weight, For area percentage, For location centrality, For color contrast, For the corresponding weights, and .

[0010] Furthermore, the scene difference evaluation value includes: matching text elements in the semantic importance weight sequence with visual elements in the visual importance weight sequence to form matching pairs; comparing the consistency between the semantic importance weight sequence and the visual importance weight sequence through the matching pairs, and calculating a consistency score. , in, For consistency score, For the first Index of matching pairs The total number of matching pairs. For the first in the copy The semantic importance weight of each text element. For the scene diagram and the first The visual importance weight of visual elements that match each text element. Summation is performed on all matching text-image pairs; The scenario difference assessment value is calculated based on the consistency score; the higher the consistency score, the lower the scenario difference assessment value. , in, The scene difference evaluation value, ranging from [0,1], is selected from the generated candidate images. The smallest image is selected as the optimal scene image.

[0011] Furthermore, the method also includes: constructing a set of core elements of advertising content, including a sequence of feature word weights, scene type identifiers, and a list of associated products; calculating the attention stability coefficient and interaction depth coefficient based on user attention trajectory data, and dividing the advertising stage into the cognition stage, the consideration stage, and the decision-making stage; constructing a variation dimension library based on user historical interaction data, including narrative angle variation and composition layout variation; and selecting the corresponding variation dimension combination through a variation strategy function according to the current stage and user preferences to generate personalized advertising content.

[0012] Furthermore, the attention stability coefficient is the ratio of the longest duration for which a user gazes at the same area within a single exposure to the total exposure duration, and its value ranges from [0,1]. The interaction depth coefficient is calculated based on user-initiated interaction behavior and has a value range of [0,1]. The user-initiated interaction behavior includes click behavior and hover behavior. The advertising exposure stages are divided according to the attention stability coefficient and the interaction depth coefficient: if it is the first exposure or the attention stability coefficient is <0.3, it is judged as the cognition stage; if the number of exposures is ≥2 and 0.3≤attention stability coefficient<0.6, and the interaction depth coefficient>0.1, it is judged as the consideration stage; if the number of exposures is ≥3 and the attention stability coefficient is ≥0.6, and the interaction depth coefficient>0.5, it is judged as the decision-making stage.

[0013] Furthermore, the method also includes: collecting user interaction data with the advertisement in real time, including gaze coordinate sequences, click locations, and hover durations; generating a dynamic attention heatmap based on the interaction data, and calculating personalized attention weights for each product element in the advertisement based on the dynamic attention heatmap. , in, For the first The attention weights of each product element are determined and normalized. This refers to the duration of attention for that element, which is the total time a user actually spends paying attention to a particular product element. Total exposure duration refers to the total time from when the advertisement is shown to the user until the end of this exposure. The core score is determined based on the role and importance of the product element in the advertisement, with a value range of [0.8, 1.2]. When the product is the core selling point, the value is 1.2; when it is an auxiliary related product, the value is 1; and when it is a decorative background element, the value is 0.8.

[0014] One or more technical solutions provided in this application have at least the following technical effects or advantages: By leveraging multimodal analysis to achieve precise matching between products and scenarios, recommending related products based on functional and visual complementarity, utilizing generative AI to automatically create and optimize scenario images, and achieving dynamic personalized optimization of advertising content through user attention tracking and phased strategies, the relevance, attractiveness, conversion efficiency, and scalability of advertising are comprehensively improved. Attached Figure Description

[0015] Figure 1 This is a flowchart of an AI-based intelligent advertising promotion method according to an embodiment of the present invention. Detailed Implementation

[0016] To facilitate understanding of the present invention, a more complete description of this application will be given below with reference to the accompanying drawings, which illustrate preferred embodiments of the invention. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to enable a more thorough and complete understanding of the disclosure of the present invention.

[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0018] Example 1: As Figure 1 As shown, this is an AI-based intelligent advertising promotion method.

[0019] S1: Perform visual and textual analysis on the promoted products, extract the visual feature tags and key selling points of the products, and generate a set of product visual slices and a set of product keywords. The visual analysis includes: using image segmentation technology, cutting different functional and appearance parts of the product image into independent visual slices, with each slice corresponding to several visual feature labels, forming a product visual slice set; Specifically, visual analysis is performed on the main image of the promoted product. Through image segmentation technology, prominent parts of the product image with different functions or appearances, such as the screen, camera module, and body of a mobile phone, are cut into independent visual slices. Each slice corresponds to one or more visual feature tags, such as high-definition large screen, multi-camera module, and flowing light back panel, forming a set of product visual slices.

[0020] The text analysis includes: performing natural language processing on the descriptive text of the promoted products to extract key nouns, verbs and adjectives that describe the product's functions, attributes and usage scenarios, forming a set of product keywords.

[0021] Specifically, natural language processing is used to extract key nouns, verbs, and adjectives describing product functions, attributes, and usage scenarios, forming a set of product keywords. Simultaneously, semantic analysis identifies the core selling points emphasized in the copy and assigns these core words higher initial weights.

[0022] S2: Construct a scenario dataset based on scenario data, which includes a set of scenario keywords for each scenario; compare the product keyword set with the scenario keyword set, and calculate the scenario matching score between the promoted product and each scenario; The scenario matching score includes the core selling point scenario fit score, keyword co-occurrence frequency score, and historical marketing performance score. Specifically, a scenario dataset is constructed based on scenario data (including scenario description text, scenario images, and historical advertising data). Each scenario contains a corresponding set of scenario keywords, which are generated by extracting high-frequency demand words from the scenario description text through natural language processing. The product keyword set is compared with each scenario keyword set to obtain the core selling point scenario fit score, keyword co-occurrence frequency score, and historical marketing effectiveness score.

[0023] For both product promotion and candidate scenarios, extract the unique or most important core keywords: From all the product's selling points, identify the most representative and compelling word for consumers as the product's core keyword. For example, if the core selling point of a sports shoe is ultimate cushioning, then its core keyword would be cushioning. From all the needs of a scenario, identify the most essential and universal need as the scenario's core keyword. For example, the core need for urban commuting is comfort and lightweight design to prevent foot fatigue. Query the pre-defined core keyword-scenario mapping knowledge base to determine the logical relationship between the two core keywords. The knowledge base stores strong association rules such as waterproof for beach vacations, quiet operation for library studies, and large capacity for long-distance travel. When a product's core selling point exists to solve a key pain point or core need in a given scenario, it is considered a strong correlation, scoring [85, 100] points; when a product's core selling point can improve the user experience or solve an important problem in a given scenario, although not absolutely necessary, the combination is very reasonable and attractive, it is considered a medium correlation, scoring [60, 84] points; when a product and scenario have a basic logical connection and can be used, but its core advantages are not prominent in this scenario, it is considered a basic correlation, scoring [20, 59] points; when a product and scenario lack a logical connection or even conflict, it is considered a weak correlation or no correlation, scoring [0, 19] points. The above scores serve as the core selling point scenario fit score.

[0024] The product keyword set and the scenario keyword set were cross-referenced and searched in the historical advertising copy database to count the frequency of these product keywords and scenario keywords appearing together in the same copy. The calculated co-occurrence frequency was converted into a score between 0 and 100. In product and scenario-related copy, comments, and discussions, the keywords almost always appear in pairs, which is judged as high co-occurrence, with a score of [90, 100]; product and scenario keywords are often mentioned together, showing a strong accompanying relationship, which is judged as moderate co-occurrence, with a score of [70, 89]; product and scenario keywords sometimes appear together, but not in a fixed combination, and their association depends on specific needs or context, which is judged as low co-occurrence, with a score of [40, 69]; product and scenario keywords are rarely or never effectively associated in the same context, which is judged as very little or no co-occurrence, with a score of [0, 30]. The above scores are used as the keyword co-occurrence frequency score.

[0025] Query the backend data of the advertising platform to find all advertising campaigns that have used this candidate scenario. Calculate the average click-through rate (CTR) of these campaigns and convert it into a score between 0 and 100. All ads that used this scenario with an average CTR significantly exceeding the overall market average (e.g., more than twice) are considered excellent, scoring [90, 100]. Scenarios with consistently better historical marketing performance than the overall market average (e.g., 50%-200% higher) are considered good, scoring [70, 89]. Scenarios with historical marketing performance roughly equal to or slightly fluctuating around the overall market average are considered average, scoring [40, 69]. Scenarios with consistently and significantly lower historical marketing performance than the overall market average are considered poor, scoring [0, 39]. These scores are used as the historical marketing performance score.

[0026] The core selling point scenario fit score, keyword co-occurrence frequency score, and historical marketing performance score are weighted and combined to obtain the final scenario matching score: , in, For a candidate scenario With a product promotion Scene matching score, The score is based on the relevance of the core selling points to the specific scenarios. The score is the co-occurrence frequency of keywords. Score the historical marketing effectiveness. , and For the corresponding weight coefficients, and ,like Focusing on core selling points, we ensure that the core advantages of the product are precisely aligned with the pain points of the scenarios.

[0027] S3: Select the target scenario based on the scenario matching score, and calculate the complementarity score of each related product based on the e-commerce platform data and the related product set of the promoted product, and sort the related products. The complementary scoring function includes complementary score, visual style consistency score, and marketing strategy bonus score; Specifically, based on transaction and browsing behavior data from e-commerce platforms, association rule algorithms are used to identify and promote sets of frequently purchased or viewed related products. For example, when promoting a coffee machine, coffee beans, a grinder, and a coffee cup are associated. The set of general related products is then filtered using the top N scenarios with the highest scenario matching scores. The filtering condition is that the related products must conform to the usage logic of that scenario. For example, in a morning at home scenario, a coffee cup is suitable, while takeaway paper cups are filtered out. The filtered related products are then ranked based on a complementarity score, which is calculated by weighting and integrating functional complementarity scores, visual style consistency scores, and marketing strategy bonus scores.

[0028] Analyze the core functions of the promoted product and the typical usage process of the target user to complete a task, and check whether the associated products are indispensable or highly relevant to this usage process. Associated products are direct consumables or objects of execution for the core functions of the promoted product, and are judged as strong functional complementarity, scoring [85, 100]. For example, promoting a coffee machine and associating it with coffee beans is a typical example of strong complementarity, because without coffee beans, the core functions of the coffee machine cannot be realized. Associated products are important auxiliary tools or supporting equipment for the core functions of the promoted product, which can enhance the experience or unlock more uses, and are judged as medium functional complementarity, scoring [55, 84]. For example, promoting a coffee machine and associating it with a coffee grinder or milk frother is medium complementarity, as they enable the coffee machine to make better coffee. Associated products belong to the same category or usage scenario as the promoted product, but have no direct functional collaboration, and are more about satisfying the user's diverse choices, and are judged as weak functional complementarity, scoring [1, 54]. For example, promoting coffee and associating it with another coffee machine or electric kettle of a different color, both in the same kitchen scenario, but with independent functions. If the functions of the related product and the promoted product are contradictory or useless in the current scenario, it is judged as having no functional complementarity and receives a score of 0. The above score serves as the functional complementarity score.

[0029] For the background elements of the promoted product, related products, and the scene, extract their visual style tags. These tags may include: modern minimalist, industrial, retro, Nordic, bright colors, black, white, and gray, wood, and metallic textures. Calculate the consistency score between the style tags of the related products and the dominant style of the promoted product and scene. When the style tags of all elements highly overlap, it is judged as highly consistent, with a score of [90, 100]. For example, promoting a white minimalist coffee machine, selecting a bright and clean kitchen, and associating it with ceramic coffee cups and wooden trays of the same color scheme, the style is highly unified. Most of the style tags overlap, with only a few non-core tags differing, but without conflict, it is judged as basically harmonious, with a score of [50, 89]. For example, promoting a white minimalist coffee machine, selecting a bright and clean kitchen, and associating it with a black metallic sugar bowl, although the colors are different, does not conflict with the minimalist style and is acceptable. There is a fundamental contradiction in the style tags, judged as a style conflict, with a score of [0, 49]. For example, promoting a minimalist white coffee machine while associating it with a brightly colored, exaggeratedly designed cartoon mug will create a strong visual conflict. The above score is a visual style consistency score.

[0030] Incorporating business strategy considerations allows for flexible alignment with current marketing objectives. Even if some products are not optimal in terms of functionality and aesthetics, they should still be prioritized to achieve specific KPIs. The current marketing campaign configuration table defines the scoring rules for various strategic objectives, including: New Product Promotion: If the related product is a key new product promoted this period, it will directly receive a fixed bonus. Bundled Promotion: If the promoted product and related products are participating in a buy-one-get-one-free or bundle promotion, bonus points will be awarded. High-Profit Products: If the profit margin of the related product exceeds a certain threshold, bonus points will be awarded. Clearance Products: If the related product needs to clear inventory, bonus points will be awarded. Cross-Selling Targets: If the company's strategy requires increasing the sales volume of a certain type of product, that type of product will be awarded bonus points. The bonus points for each strategy will be accumulated, but the maximum will not exceed 100 points. If the product is not involved in any special strategy, it will receive a base score of 50 points. The above scores will serve as marketing strategy bonus points.

[0031] Calculate the complementarity score: , in, To promote products Related products With scene Complementarity score, For functional complementarity scores, Score for visual style consistency. Add points to marketing strategies , and For the corresponding weight coefficients, and ,like Functionality is the core of user selection.

[0032] S4: Based on the promoted product, target scenario, top 5 related products, and visual feature tags, generate several candidate scenario images; parse the promotional product copy to obtain a semantic importance weight sequence, identify each candidate scenario image to obtain a visual importance weight sequence, calculate the difference evaluation value between the two weight sequences, and select the image with the smallest scenario difference evaluation value as the optimal scenario image.

[0033] The semantic importance weight sequence is obtained by parsing promotional product copy through natural language processing. The parsed elements include core selling point words, scene atmosphere words, and implicit related demand words. Specifically, the promoted product, selected scene elements, the top 5 related products, and visual feature tags are used as input descriptions to drive the generative AI model to generate several candidate scene images. The visual feature tags are associated with the product through structured data (JSON format), guiding the generative AI to highlight core visual features, avoid visual conflicts, and match materials and design styles when constructing scene images. For example, if a product slice includes a high-definition large screen, the algorithm will prioritize enhancing the clarity and color contrast of that area in the generated image. An improved difference assessment method was used to select candidate images, and the scene difference assessment value of each image was calculated. The promotional product copy was analyzed using a Natural Language Processing (NLP) model. The analyzed elements included core selling point keywords, scene atmosphere keywords, and implicit related demand keywords. Core selling point keywords were words modified by absolute adjectives or frequently repeated words, with initial weights set to [0.8, 1]. Scene atmosphere keywords described the emotional tone of the scene, with initial weights set to [0.5, 0.7]. Implicit related demand keywords were potential user needs inferred from the context but not directly mentioned, with initial weights set to [0.3, 0.5]. Finally, a standardized semantic importance weight sequence was generated. , in, It is a sequence of semantic importance weights. In the promotional copy, the first The semantic importance weights of each key element. An index is used to identify the first element in the set. There are 1 elements, with values ​​ranging from 1 to 1. , This represents the total number of key elements extracted from the product promotion copy. .

[0034] The visual importance weight sequence is obtained through visual analysis of each candidate scene image; Specifically, visual analysis is performed on the generated candidate images. Using object detection models and image segmentation techniques, the main elements in the images are identified, including promoted products, related products, and background scenes. A visual importance weight sequence is then generated based on their visual salience in the image. Visual salience is calculated based on area proportion, location centrality, and color contrast. , in, As visual importance weight, The area percentage is the ratio of the area of ​​the element detection box to the total area of ​​the canvas, and is normalized to [0,1]. The positional centrality is the reciprocal of the Euclidean distance between the element's center point coordinates and the canvas's geometric center, normalized to [0,1]. For color contrast, it is the difference between the average color brightness / saturation of the element area and the background, and is normalized to [0,1]. For the corresponding weights, and .

[0035] Finally, a standardized sequence of visual importance weights is generated: , in, For visual importance weight sequence, For the first The visual importance weight of each major element in the scene graph. An index is used to identify the first element in the set. There are 1 elements, with values ​​ranging from 1 to 1. , This represents the total number of main elements extracted from the candidate scene graph. .

[0036] The text elements in the semantic importance weight sequence are matched with the visual elements in the visual importance weight sequence to form matching pairs, such as a high-definition large screen and a screen area in a scene image. The consistency between the semantic importance weight sequence and the visual importance weight sequence is compared by comparing the matching pairs, and a consistency score is calculated. The higher the consistency, the lower the scene difference evaluation value. , in, For consistency score, For the first Index of matching pairs The total number of matching pairs. For the first in the copy The semantic importance weight of each text element. For the scene diagram and the first The visual importance weight of visual elements that match each text element. This is used to sum all matching text-image pairs.

[0037] , in, The scene difference evaluation value, ranging from [0,1], is selected from the generated candidate images. The smallest image is selected as the optimal scene image.

[0038] In the underlying control of scene graph generation, rules must be followed to ensure a natural and harmonious layout. A built-in physical relationship knowledge base defines the spatial relationships between common objects. For example, support relationships: small or fragile objects must be placed on stable supports and cannot be suspended; usage relationships: user interactions with products must conform to ergonomics; scale relationships: the size proportions of figures and furniture should conform to common sense.

[0039] While adhering to physical rules, the design aims for visual balance, hierarchy, and a prominent focal point. Promoted products are placed near the intersection of the rule of thirds by default. Their positional relationship with core products is determined based on their functional complementarity score and visual style consistency score, ensuring that the placement satisfies physical relationships. The canvas proportion of each product element is determined by its semantic importance weight, with promoted products occupying the largest proportion (35-45% canvas area), followed by core related products (20-30%), and decorative background elements the smallest (≤20%), smoothed using the golden ratio. Promoted products typically occupy the largest proportion, followed by core related products, and then decorative elements. The final result is a scene image with optimized multi-product layout.

[0040] The technical solutions described in the embodiments of this application above have at least the following technical effects or advantages: This application utilizes multimodal AI fusion analysis technology to optimize the entire process of intelligent advertising promotion, from product understanding to material generation. It employs product visual slicing and text keyword extraction technologies, combined with a weighted evaluation of core selling point scenario relevance, keyword co-occurrence frequency, and historical marketing performance. This achieves precise quantitative matching between products and advertising scenarios, enhancing the relevance and conversion potential of advertising content. The related product recommendation integrates functional complementarity, visual style consistency, and collaborative consideration of marketing strategies, ensuring that the recommendation results both conform to user scenario logic and support business objectives. A generative AI model creates scene diagrams, and a weighted consistency evaluation mechanism automatically selects the optimal image, ensuring a high degree of consistency between visual prominence and information delivery in advertising materials. The introduction of a physical relationship knowledge base and aesthetic layout rules guarantees the rationality and visual appeal of the generated content, making the advertising images both realistic and attractive. This method transforms the traditional advertising production process, which relies on manual creativity, into an automated and scalable intelligent process, improving the efficiency, scalability, and personalization of advertising production.

[0041] Example 2: Example 1 achieved ad scene matching, related product recommendation, and static content generation based on multimodal AI, but it still has the shortcoming of not being able to dynamically adjust ad content according to the psychological changes of users after multiple exposures to maintain continuous attractiveness and conversion efficiency. This example further supplements and explains the content of Example 1.

[0042] The method further includes: constructing a set of core elements of advertising content, including a sequence of feature word weights, scene type identifiers, and a list of associated products; calculating the attention stability coefficient and interaction depth coefficient based on user attention trajectory data, and dividing the advertising stage into the cognition stage, the consideration stage, and the decision-making stage; constructing a variation dimension library based on user historical interaction data, including narrative angle variation and composition layout variation; and selecting the corresponding variation dimension combination through a variation strategy function according to the current stage and user preferences to generate personalized advertising content.

[0043] Specifically, after obtaining the optimized scene map with multiple product layouts, the feature word weight sequence, scene type identifier, and related product list are extracted to construct a set of core elements of the advertising content. The feature word ranking sequence that results in the smallest scene difference evaluation value is recorded as the feature word weight sequence; the scene type is identified by the features of the promoted product, and the scene keywords and weights are recorded as scene type identifiers; a list of auxiliary products strongly associated with the promoted product is extracted, and the importance weight of each associated product in the scene is recorded as the associated product list. The feature word weight sequence, scene type identifier, and associated product list are encoded into structured data. For example, the core product feature sequence is: [feature word A; weight 0.5, feature word B: weight 0.3,...]; the scene identifier is: {scene type: outdoor picnic, environmental elements: [lawn, sunshine], weight: 0.7}; the associated product set is: [snacks: weight 0.2, beverages: weight 0.1].

[0044] The attention stability coefficient is the ratio of the longest time a user gazes at the same area within a single exposure to the total exposure time, and its value ranges from [0,1]. The interaction depth coefficient is calculated based on user-initiated interaction behavior and has a value range of [0,1]. The user-initiated interaction behavior includes click behavior and hover behavior. Specifically, based on the constructed set of core elements of advertising content and combined with user attention trajectory data, ad exposure is divided into three stages: cognition, consideration, and decision-making by calculating the attention stability coefficient and interaction depth coefficient. User interaction data for each exposure is recorded, including gaze coordinate sequence, click location, hover duration, and scrolling behavior. The attention stability coefficient (S) is the ratio of the longest continuous gaze duration on the same area within a single exposure to the total exposure duration; the interaction depth coefficient (D) is the weighted sum of user's active interaction behaviors, with clicks scored as 3 points and hover as 1 point, normalized to the range [0,1]. The first exposure or S < 0.3 indicates the cognition stage, where user attention is scattered, interaction is minimal, and the goal is to quickly attract attention. When the number of exposures is ≥ 2 and 0.3 ≤ S < 0.6, D > 0.1, it indicates the consideration stage, where users develop interest and engage in active interaction, with the goal of deepening understanding. When the number of exposures is ≥ 3 and S ≥ 0.6, D > 0.5, it indicates the decision-making stage, where user attention is highly concentrated, interaction is frequent, and the goal is to stimulate conversion.

[0045] A variation dimension library is constructed based on historical user interaction data, and appropriate combinations of variation dimensions are selected for different stages through variation strategy functions. Personalized variation strategies are developed for ad content based on the characteristics of different exposure stages and user needs. The variation dimension library includes narrative perspective variation and composition layout variation.

[0046] Narrative perspective variation focuses on the sentence style and expression logic of advertising copy. By analyzing users' historical interaction behavior with ads of different sentence structures (such as click-through rate, text gaze duration, and sentence dwell time), it identifies users' preferred expression methods. For example, if a user clicks on ads containing numerical quantification multiple times and has a higher-than-average text gaze duration on such ads, it indicates that they prefer data-driven narratives; if a user hovers for a longer time on contextualized interrogative sentences, it indicates that they prefer context-evoking narratives.

[0047] Composition and layout variations target the spatial arrangement, proportion, and visual focus of product elements in advertisements. By analyzing user interaction behavior with different layouts, quantifiable layout parameters are extracted, including position coordinates, size ratios, line distribution, and visual focus. Position coordinates, with the canvas center as the origin, record the offset of core product elements and the distribution density of auxiliary products. Size ratios represent the percentage of the core product area relative to the total canvas area. Line distribution indicates whether the arrangement of product elements conforms to aesthetic rules and whether the lines are smooth. Visual focus uses historical data to determine the area most frequently viewed by users, prioritizing the placement of high-importance products there.

[0048] Based on the cognitive, consideration, and decision-making stages of ad exposure, and combined with the preference characteristics accumulated from users' historical interaction data, a dynamic combination of variation dimensions is selected: In the cognitive stage, when users are encountering the ad for the first time or their attention is highly scattered, the core objective is to quickly attract attention and clearly convey the core information. At this stage, the narrative perspective defaults to concise declarative sentences, avoiding complex rhetoric to ensure that the information is straightforward and easy to understand. The layout places the core product with the highest weight in the feature word weight sequence in the center of the canvas and enlarges it to occupy 35%-40% of the area, while auxiliary products are reduced to 15%-20% and scattered around, forming a layout with a prominent center and clear hierarchy. Too many interactive hotspots or dynamic effects are avoided to reduce the cognitive load on users. In the consideration stage, when users have developed initial interest, the core objective is to deepen understanding and provide detailed information. The narrative perspective is selected based on users' historical preferences, choosing between data-driven narratives or scene-evoking narratives. In the layout, the core product still maintains a high proportion (30%-35%), but the area proportion of auxiliary products is increased to 20%-25%, and the usage relationship between products is reflected through reasonable positional offsets, such as placing snacks near the tent entrance and drinks on the picnic mat. During the decision-making phase, users are deeply involved. The core objective is to strengthen memory and provide new stimuli to promote conversion. Narrative angles include: using urgency-based phrases, such as "Limited-time offer! Last 24 hours to buy!", or emotionally resonant phrases, such as "A picnic with family, how regrettable it would be to miss it!" The composition and layout involve shifting the core product's position, such as slightly adjusting it to the edge of the canvas to create visual novelty, or adjusting its proportion, such as enlarging it to 40%-45%, while simultaneously adjusting the order of auxiliary products. Finally, the mutation strategy function outputs a set of specific adjustment parameters. For example, in the current decision-making phase, the chosen narrative angle is: urgency-based phrases + composition and layout: core product shifted 30 pixels to the right + area occupancy 42% + auxiliary products arranged in order of use.

[0049] After obtaining the parameters output by the mutation strategy function, generative AI technology is used to generate personalized content that meets the user's current needs while retaining the core information of the advertisement, in conjunction with the core elements in the advertisement's DNA. Segments are rewritten based on the narrative perspective, preserving the characteristic words and weight sequences in the DNA; the product's position and size are adjusted according to the layout parameters to ensure scene consistency.

[0050] The technical solutions described in the embodiments of this application above have at least the following technical effects or advantages: This application achieves structured encapsulation and reusability of advertising content elements by constructing a core element set of advertising content, including feature word weight sequences, scene type identifiers, and a list of associated products. By analyzing user attention stability and interaction depth in real time, the advertising exposure process is scientifically divided into three stages: cognition, consideration, and decision-making. This allows advertising content optimization strategies to accurately match users' different psychological states and information needs throughout their decision-making journey. A variation dimension library is constructed based on historical user interaction data, and a stage-aware variation strategy function is designed to dynamically adapt the optimal narrative angle and layout parameters for different exposure stages. Through DNA-driven and staged strategy selection, the sustained attractiveness, information delivery efficiency, and final conversion effect of advertisements across multiple exposures are improved.

[0051] Example 3: Example 2 implemented a phased layout optimization strategy based on the core elements of advertising content and user attention trajectory data, but it still has shortcomings such as triggering static layout adjustments solely through heatmap changes and gaze duration, and lacking dynamic quantitative analysis of attention data. This example further supplements and explains the content of Example 2.

[0052] The method further includes: collecting user interaction data on the advertisement in real time, including gaze coordinate sequence, click position and hover duration; generating a dynamic attention heatmap based on the interaction data, and calculating the personalized attention weight of each product element in the advertisement based on the dynamic attention heatmap; Specifically, real-time data collection of user gaze coordinate sequences, click locations, hover durations, and scrolling behavior is used. This data is mapped onto the ad canvas, and an attention heatmap is generated using Gaussian kernel density estimation. The heatmap's color intensity reflects the distribution of user attention, with darker colors indicating higher levels of attention. Based on this dynamic attention heatmap, a personalized attention weight is calculated for each core product element in the ad. , in, For the first The attention weights of each product element are determined and normalized. This refers to the duration of attention for that element, which is the total time a user actually spends paying attention to a particular product element. Total exposure duration refers to the total time from when the advertisement is shown to the user until the end of this exposure. The core focus coefficient is determined based on the role and importance of product elements in the advertisement, with a value range of [0.8, 1.2]. When the product is the core selling point, the value is 1.2; for related products, the value is 1; and for decorative background elements, the value is 0.8. This value is derived from historical data and user behavior research. The optimal conversion rate is achieved when the core product's attention share is approximately 1.2 times that of the related products. Excessively high values ​​can lead to visual imbalance and cognitive burden. When the background element weight is below 0.8, the overall acceptance of the advertisement by users decreases.

[0053] Real-time monitoring of the calculated attention weights; if a certain weight is found to be set to a high weight ( Product elements with a density greater than 0.3 (not particularly high) will automatically trigger a layout fine-tuning mechanism, shifting them to the area with the highest attention. (or The weight is 50 × the difference between the product's weight and the average weight of the high-attention area, with a maximum offset of no more than 15% of the canvas side length. The canvas proportion of high-weight products will be temporarily increased by 5%-10%, while the proportion of other low-weight product elements will be compressed. However, a bottom line will be set to ensure that all elements, especially decorative background elements, can still maintain a minimum visible size to ensure the integrity of the advertising information.

[0054] For product elements identified as having high weight but currently insufficient attention, additional visual enhancements will be applied, such as moderately increasing their color saturation or adding a moderately frequent dynamic effect, to attract users' attention and guide their gaze. For product elements with lower weight and currently low attention, ensuring basic visual clarity is sufficient is sufficient; over-emphasis should be avoided.

[0055] When the overall change in the heatmap exceeds 30%, or when it is observed that during a single exposure, the user's gaze duration on a specific product element significantly increases from less than 1 second to more than 3 seconds, it will be considered necessary to trigger dynamic adjustments to the layout.

[0056] After layout adjustments, a series of rationality checks will be performed to ensure that the adjusted ad layout still conforms to basic visual logic and user experience. These checks include: whether the core product's proportion on the canvas remains within the ideal range of 35%-45%; whether the proportion of related products remains within 20%-30%; whether the proportion of decorative background elements is controlled within 20%; whether there is any overlap between all product elements, requiring the overlap area to be less than 5% of the total canvas area; and whether the visual center of gravity of the entire ad image remains stable, with the center of gravity shift not exceeding 10% of the canvas width. If, after verification, the adjusted layout is found to be inconsistent with these core logics and constraints, it will automatically revert to the previous layout state, or only make necessary, minor adjustments to ensure that the final ad display effect optimizes attention guidance while maintaining overall aesthetics and clear information delivery.

[0057] The technical solutions described in the embodiments of this application above have at least the following technical effects or advantages: This application employs real-time user attention tracking and heatmap analysis technology to achieve dynamic and personalized optimization of ad layout. By intelligently migrating high-weight product elements to high-attention areas, adjusting visual proportions, and enhancing expressiveness, the ad content can respond to users' visual attention habits in real time, improving the efficiency of key information delivery and user engagement. This not only enhances the visual appeal and information focus of ads at different exposure stages but also provides a data-driven dynamic adjustment mechanism for improving click-through rates and conversion rates.

[0058] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An AI-based intelligent advertising promotion method, characterized in that, include: S1: Perform visual and textual analysis on the promoted products, extract the visual feature tags and key selling points of the products, and generate a set of product visual slices and a set of product keywords. S2: Construct a scenario dataset based on scenario data, which includes a set of scenario keywords for each scenario; compare the product keyword set with the scenario keyword set, and calculate the scenario matching score between the promoted product and each scenario; S3: Select the target scenario based on the scenario matching score, and calculate the complementarity score of each related product based on the e-commerce platform data and the related product set of the promoted product, and sort the related products. S4: Based on the promoted product, target scenario, top 5 related products, and visual feature tags, generate several candidate scenario images; parse the promotional product copy to obtain a semantic importance weight sequence, identify each candidate scenario image to obtain a visual importance weight sequence, calculate the difference evaluation value between the two weight sequences, and select the image with the smallest scenario difference evaluation value as the optimal scenario image.

2. The AI-based intelligent advertising promotion method as described in claim 1, characterized in that, The visual analysis includes: using image segmentation technology, cutting different functional and appearance parts of the product image into independent visual slices, with each slice corresponding to several visual feature labels, forming a product visual slice set; The text analysis includes: performing natural language processing on the descriptive text of the promoted products to extract key nouns, verbs and adjectives that describe the product's functions, attributes and usage scenarios, forming a set of product keywords.

3. The AI-based intelligent advertising promotion method as described in claim 1, characterized in that, The scenario matching score includes a core selling point scenario fit score, a keyword co-occurrence frequency score, and a historical marketing performance score. The core selling point scenario fit score is obtained by matching and scoring the product's core selling points with the core needs of the scenario. The keyword co-occurrence frequency score is obtained by statistically analyzing the co-occurrence frequency of product keywords and scenario keywords in historical advertising copy and converting it into a score. The historical marketing performance score is obtained by calculating the average click-through rate of candidate scenarios in historical advertising campaigns and converting it into a score. The scenario matching score is then calculated through a weighted fusion. ， in, For the scene With product promotion Scene matching score, The score is based on the relevance of the core selling points to the specific scenarios. The score is the co-occurrence frequency of keywords. Score the historical marketing effectiveness. , and For the corresponding weight coefficients, and .

4. The AI-based intelligent advertising promotion method as described in claim 1, characterized in that, The complementarity scoring system comprises a complementarity score, a visual style consistency score, and a marketing strategy bonus score. The functional complementarity score is calculated by analyzing the functional relationship between the promoted product and related products in their usage process. The visual style consistency score is calculated by assessing the degree of consistency between the related product and the promoted product, as well as the dominant style of the scenario. The marketing strategy bonus score considers the business strategy factors of the related products. The complementarity score is obtained through a weighted fusion calculation. , in, To promote products Related products With Scene Complementarity score, For functional complementarity scores, Score for visual style consistency. Add points to marketing strategies , and For the corresponding weight coefficients, and .

5. The AI-based intelligent advertising promotion method as described in claim 1, characterized in that, The semantic importance weight sequence is obtained by parsing promotional product copy using natural language processing. The parsed elements include core selling point keywords, scene atmosphere keywords, and implicit related demand keywords. Among them, core selling point keywords are words with a repetition rate of more than 10%, scene atmosphere keywords use adjectives describing the emotional aspects of the scene, and implicit related demand keywords infer the user's potential needs from the context, generating a standardized semantic importance weight sequence: ， in, It is a sequence of semantic importance weights. In the promotional copy, the first The semantic importance weights of each key element. This represents the total number of key elements extracted from the product promotion copy. .

6. The AI-based intelligent advertising promotion method as described in claim 1, characterized in that, The visual importance weight sequence is obtained through visual analysis of each candidate scene image. Using object detection models and image segmentation techniques, the main elements in the image are identified, including promoted products, related products, and scene background. Based on their visual salience in the image, a standardized visual importance weight sequence is generated. ， in, For visual importance weight sequence, For the first The visual importance weight of each major element in the scene graph. An index is used to identify the first element in the set. There are 1 elements, with values ​​ranging from 1 to 1. , This represents the total number of main elements extracted from the candidate scene graph. ; The visual saliency is calculated based on area proportion, positional centrality, and color contrast: ， in, As visual importance weight, For area percentage, For location centrality, For color contrast, For the corresponding weights, and .

7. The AI-based intelligent advertising promotion method as described in claim 1, characterized in that, The scene difference evaluation value includes: matching text elements in the semantic importance weight sequence with visual elements in the visual importance weight sequence to form matching pairs; comparing the consistency between the semantic importance weight sequence and the visual importance weight sequence through the matching pairs, and calculating the consistency score. ， in, For consistency score, For the first Index of matching pairs The total number of matching pairs. For the first in the copy The semantic importance weight of each text element. For the scene diagram and the first The visual importance weight of visual elements that match each text element. Summation is performed on all matching text-image pairs; The scenario difference assessment value is calculated based on the consistency score; the higher the consistency score, the lower the scenario difference assessment value. ， in, The scene difference evaluation value, ranging from [0,1], is selected from the generated candidate images. The smallest image is selected as the optimal scene image.

8. The AI-based intelligent advertising promotion method as described in claim 1, characterized in that, The method further includes: constructing a set of core elements of advertising content, including a sequence of feature word weights, scene type identifiers, and a list of associated products; calculating the attention stability coefficient and interaction depth coefficient based on user attention trajectory data, and dividing the advertising stage into the cognition stage, the consideration stage, and the decision-making stage; constructing a variation dimension library based on user historical interaction data, including narrative angle variation and composition layout variation; and selecting the corresponding variation dimension combination through a variation strategy function according to the current stage and user preferences to generate personalized advertising content.

9. The AI-based intelligent advertising promotion method as described in claim 8, characterized in that, The attention stability coefficient is the ratio of the longest time a user gazes at the same area within a single exposure to the total exposure time, and its value ranges from [0,1]. The interaction depth coefficient is calculated based on the user's active interaction behavior and has a value range of [0,1]. The user-initiated interactive behaviors include click behaviors and hover behaviors; The advertising exposure stages are divided according to the attention stability coefficient and the interaction depth coefficient: if it is the first exposure or the attention stability coefficient is <0.3, it is judged as the cognition stage; if the number of exposures is ≥2 and 0.3≤attention stability coefficient<0.6, and the interaction depth coefficient>0.1, it is judged as the consideration stage; if the number of exposures is ≥3 and the attention stability coefficient is ≥0.6, and the interaction depth coefficient>0.5, it is judged as the decision-making stage.

10. The AI-based intelligent advertising promotion method as described in claim 1, characterized in that, The method further includes: collecting user interaction data with the advertisement in real time, including gaze coordinate sequences, click locations, and hover durations; generating a dynamic attention heatmap based on the interaction data; and calculating personalized attention weights for each product element in the advertisement based on the dynamic attention heatmap. ， in, For the first The attention weights of each product element are determined and normalized. This refers to the duration of attention for that element, which is the total time a user actually spends paying attention to a particular product element. Total exposure duration refers to the total time from when the advertisement is shown to the user until the end of this exposure. The core score is determined based on the role and importance of the product element in the advertisement, with a value range of [0.8, 1.2]. When the product is the core selling point, the value is 1.2; when it is an auxiliary related product, the value is 1; and when it is a decorative background element, the value is 0.8.

Images