New media intelligent marketing method based on AI

Abstract

The invention relates to the technical field of artificial intelligence-driven digital marketing, in particular to an AI-based new media intelligent marketing method, which comprises the following steps of: 1, carrying out cross-platform user behavior fusion modeling; 2, constructing a multi-modal content gene pool: analyzing a visual focus area and a text emotional tendency of historical content, and generating a content gene vector; 3, dynamic content generation and real-time optimization, wherein initial marketing content is generated according to the space-time behavior matrix retrieval gene bank; user interaction behaviors are monitored in real time, elements in a visual focus area are dynamically replaced, and the emotion intensity of the copywriting is adjusted; and step 4, closed-loop strategy optimization: a dynamic effect index is calculated by fusing multi-platform indexes, a gene variation mechanism is triggered when the index continuously decreases, and cross-platform compensation content is generated based on failed user behaviors. By analyzing the cross-platform behavior and the content preference of the user, the user demand can be accurately predicted, and the content more conforming to the personalized preference of the user is provided.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence-driven digital marketing technologies, and in particular to an AI-based new media intelligent marketing method. Background Art

[0002] In the field of AI-driven digital marketing, especially in cross-platform new media intelligent marketing scenarios, existing technologies have the following structural flaws: 1. The fragmentation of user behavior across platforms leads to distorted portraits; Mainstream marketing systems rely on single-platform behavioral data (such as Douyin viewing history and WeChat Moments interactions) to build user interest models. These models only weight behaviors within the platform (e.g., video completion rate weight > like weight), failing to capture users' interconnected behaviors across multiple platforms like Weibo and Xiaohongshu.

[0003] Due to the closed data interfaces and privacy policy restrictions of each platform, the user's cross-platform behavior chain is forcibly severed. For example, if user A watches a new energy vehicle video on Douyin and searches for similar products on Xiaohongshu the next day, the existing system cannot establish this association, resulting in: One-sided portrait: Douyin categorizes users as having “mild interest” and ignores the strong purchase intent on Xiaohongshu. Resource mismatch: Pushing low-level traffic-generating content to users instead of conversion-oriented content.

[0004] Platform data barriers are inherent to the industry ecosystem, and a single service provider cannot forcibly obtain raw cross-platform data. Existing technologies lack a quantifiable mechanism for temporal correlations between cross-platform behaviors, a problem that continues to worsen in the new media multi-platform environment.

[0005] 2. Content generation is disconnected from user feedback; AIGC-based content generation tools (such as Jasper and Copy.ai) generate initial marketing content based on user profiles, which is then manually reviewed and statically delivered. Optimization relies on periodic performance reports (such as 24-hour conversion rate).

[0006] Technical flaws: Content generation and real-time user interaction are completely disconnected, manifested as: Visual focus failure: When the user's eyes leave the key area (such as product display), the system cannot dynamically replace elements; Lack of copy adaptability: The density of emotional words cannot be adjusted in real time according to the user's reading speed (for example, strong attractive copy is not triggered when sliding quickly).

[0007] Existing AI content engines rely on a one-way generation architecture, lacking a closed loop of real-time behavior awareness and content self-optimization. This shortcoming is particularly pronounced in highly interactive scenarios like short videos, resulting in a drop in average user duration of over 30% (industry benchmark data).

[0008] Therefore, there is an urgent need for new media intelligent marketing methods based on AI to solve the above problems. Summary of the Invention

[0009] Based on the above objectives, the present invention provides an AI-based new media intelligent marketing method, including: Step 1: Cross-platform user behavior fusion modeling: Collect users' content interaction behaviors, implicit attention behaviors, and cross-platform association behaviors through the interfaces of each platform; Determine the platform weight factor based on user cross-platform activity and construct a spatiotemporal behavior matrix using a time decay algorithm; Step 2: Construction of multimodal content gene library: Analyze the visual focus areas and text sentiment tendencies of historical content to generate content gene vectors; Step 3: Dynamic content generation and real-time optimization: Search the gene library based on the spatiotemporal behavior matrix to generate initial marketing content; Monitor user interaction behavior in real time, dynamically replace elements in the visual focus area, and adjust the emotional intensity of the copy; Step 4: Closed-loop strategy optimization: The dynamic effect index is calculated by integrating multi-platform indicators. When the index drops continuously, the gene mutation mechanism is triggered, and cross-platform compensatory content is generated based on failed user behavior.

[0010] Preferably, the specific steps of the time decay algorithm include: a: Identify the timestamps of users’ actions on the same target on different platforms and calculate the duration of the behavior interval; b: Set the attenuation coefficient baseline value based on the user's historical cross-platform behavior frequency. The higher the behavior frequency, the smaller the attenuation coefficient baseline value; c: Calculate the association strength using the exponential decay model: for every unit increase in the behavior interval, the association strength is multiplied by the decay coefficient; d: When the interval between actions exceeds the maximum interval in the user's history, the action chain break detection is started: Extract the mean and standard deviation of the user's cross-platform behavior intervals over the past 30 days; If the current interval exceeds the mean plus three standard deviations, the behavior chain is determined to be broken, and the association strength is reset to a preset ratio of the initial value; e: Weighted fusion of the association strength of each platform and the platform weight factor, and output to the spatiotemporal behavior matrix.

[0011] Preferably, the identification of the visual focus area includes: a: Generate a gaze heat map of the content keyframe through a convolutional neural network. The heat map represents the gaze probability in pixel values. b: Perform regional clustering on the heat map and merge adjacent high-probability pixels to form candidate regions; c: Calculate the probability density integral of each candidate area and select the area with the largest integral value as the visual focus area; d: Record the center coordinates and coverage of the visual focus area. When the user's real-time vision deviates from this area: Get the shortest distance between the sight track and the boundary of the focus area; If the distance continues to increase and exceeds the dynamic threshold, it is determined as a gaze-away event.

[0012] Preferably, the adjustment execution process of the copywriting emotional intensity is: a: Capture the user's screen scrolling speed in real time and calculate its deviation rate from the user's historical average speed; b: When the deviation rate is lower than the first critical value, the emotional reinforcement mode is activated: Identify the sentiment polarity of the currently displayed paragraph; When inserting sentiment words of the same polarity and higher intensity into subsequent paragraphs, the insertion density is negatively correlated with the offset rate; c: When the offset rate is higher than the second critical value, the interactive wake-up mode is started: Insert a dynamic multiple-choice question component at the top of the next screen, with the options strongly related to the marketing theme; Generate branch copy in real time based on user selection results; d: The first critical value and the second critical value are dynamically set through a machine learning model: Input user historical stay time distribution and conversion rate data; Output the velocity threshold bounds that maximize content engagement.

[0013] Preferably, the gene variation mechanism includes: a: Calculate the correlation coefficient between each dimension of the content gene vector and the dynamic effect index; b: Screening weakly correlated dimensions whose correlation coefficients are lower than the adaptive threshold; c: Apply directional perturbations to weakly correlated dimensions: If it is a continuous dimension, perform Gaussian random perturbation within the value range; If it is a discrete dimension, select an alternative template from the variation library whose similarity is greater than the preset value; d: After generating variant content, conduct A / B testing on a small user group, and select the variant with the greatest improvement in performance index and launch it in full.

[0014] Preferably, the specific steps of generating cross-platform compensation content are: a: Extract the spatiotemporal behavior matrix subset of users who failed marketing on platform P and mark the failure reason label; b: Input the subset of behavior matrix into the content generation model of platform Q and perform differential deconstruction: Compare the differences in users' historical behavior patterns on the P and Q platforms; Identify potential unmet interest dimensions of users on platform Q; c: Generate compensatory content gene vector: Strengthen the visual elements that are unique to Platform Q; Weaken the copywriting elements related to the reasons for failure; d: Compensatory content is synthesized through generative adversarial networks and pushed first when the user logs into the platform Q next time.

[0015] Preferably, the process of constructing the dynamic effect index includes: a: Define the core indicator set: conversion rate, sharing rate, cross-platform communication depth, and user stay time; b: Train the weight distribution model through historical data: Input: product type, user profile classification, and current hot event characteristics; Output: initial weight coefficient of each indicator; c: Adjust weights in real time during content delivery: Every time a preset number of user interaction samples are added, the covariance matrix between indicators is recalculated; Scale the weight coefficients proportionally according to the magnitude of the covariance change; d: Perform a logarithmic transformation on the normalized weighted summation result to eliminate the influence of long-tail distribution.

[0016] Preferably, the determination of the dynamic threshold comprises: a: Establish a user's sightline stability model: Extract the fluctuation characteristics of the user's gaze dwell time in historical browsing; Calculate the standard deviation of the gaze trajectory as the baseline fluctuation; b: Real-time monitoring of gaze acceleration. If the acceleration exceeds a multiple of the baseline fluctuation, threshold adaptation is triggered: The larger the multiple, the smaller the threshold is to quickly respond to anomalies; The smaller the multiple, the larger the threshold is to avoid false triggering.

[0017] Preferably, the generation logic of the branch copy is: a: Build a decision tree content template library, where each branch node corresponds to a user selection path, i.e., a branch path; b: After the user selects an option: Match preset templates based on option semantics; If the match fails, the GPT model is called to generate coherent copy in real time based on the option content; c: Embed tracking tags in branch copy, and subsequent conversion behaviors are attributed to this branch path.

[0018] Preferably, the platform preference identification in the differentiated deconstruction includes: a: Extract the behavioral feature dimensions unique to platform Q; b: Calculate the user's deviation from the platform's average level in this dimension; c: If the deviation exceeds the significance threshold, the dimension is added to the list of compensatory content enhancements.

[0019] Beneficial effects of the present invention: 1. The present invention collects user behavior data from each platform, determines the platform weight factor based on the platform activity, and uses the time decay algorithm to establish a spatiotemporal behavior matrix, thereby realizing the temporal correlation of cross-platform behaviors.

[0020] 2. This invention generates a content gene vector by analyzing the visual focus areas and textual sentiment of historical content. It then generates initial marketing content based on the spatiotemporal behavior matrix. By monitoring user behavior in real time, the system dynamically adjusts content, replacing elements in the visual focus area and automatically adjusting the emotional intensity of the copy based on the user's scrolling speed.

[0021] 3. In this invention, when the dynamic effect index of content marketing continues to decline, the system automatically triggers a gene mutation mechanism based on the change in the effect index to perform perturbation optimization on the content. By calculating the correlation between each dimension of the content gene vector and the effect index, weakly correlated dimensions are screened and targeted perturbations are applied, thereby achieving fine-grained content optimization. After A / B testing and verification, the variant with the greatest improvement in the effect index is selected and fully deployed. This mechanism ensures that when the content effect declines, the system can automatically adjust the content to maintain continuous optimization of marketing results.

[0022] 4. In this invention, when a user's marketing campaign fails on a particular platform, the system extracts a subset of the behavior matrix based on the user's cross-platform behavior data and performs a differentiated deconstruction. This identifies the user's potential interests across different platforms and generates compensatory content gene vectors. By using a generative adversarial network, the system prioritizes these compensatory content the next time the user logs in, thereby improving the accuracy of cross-platform marketing, avoiding wasted marketing resources, and increasing user engagement and conversion rates. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0024] Figure 1 is a flow chart of the steps of the method of the present invention; Figure 2 Flow chart of the steps of constructing the dynamic effect index of the method of the present invention; Figure 3 Flow chart of the steps for identifying platform preference in the differential deconstruction method of the present invention. DETAILED DESCRIPTION

[0025] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It is also noted that, to provide a more detailed description, the following embodiments are best and preferred embodiments, and those skilled in the art may employ alternative methods for implementing certain known technologies. Furthermore, the accompanying drawings are intended only to provide a more detailed description of the embodiments and are not intended to limit the present invention.

[0026] See Figure 1-Figure 3 An embodiment of the present invention provides an AI-based new media intelligent marketing method. In step 1, in traditional digital marketing, data barriers between different platforms fragment user behavior, leading to a one-sided interest model. This method uses the interfaces of various platforms (such as Douyin, WeChat, and Xiaohongshu) to collect users' content interaction behaviors (such as likes, comments, and shares), implicit attention behaviors (such as browsing history and search history), and cross-platform association behaviors (for example, after following a product on one platform, searching for related information on another platform). Next, the system determines platform weight factors based on the user's activity on each platform and constructs a spatiotemporal behavior matrix using a time decay algorithm. The application of the time decay algorithm adjusts the importance of user behaviors based on their temporal differences, giving higher weights to behaviors closer to the current time, thereby more accurately reflecting the user's latest interests and needs. This method effectively integrates cross-platform data to form a more accurate user profile, resolving the problems of data fragmentation and resource mismatch.

[0027] In step 2, content generation requires not only understanding the emotional leaning of the text but also focusing on the impact of visual elements on user behavior. In this step, the system analyzes the visual focus areas of historical content (such as product display images, brand logos, and CTA button locations) and their corresponding textual sentiment (such as positive, negative, and inquiring), then generates a content gene vector. This gene vector integrates the visual and emotional elements of the content. By analyzing historical data, the system identifies which visual focus areas and textual sentiments are most effective in increasing user engagement and stores them in the content gene library. This approach allows the system to optimize visual and emotional elements in subsequent marketing content, thereby improving user engagement and conversion rates.

[0028] In step 3, when generating the initial marketing content, the system retrieves content elements suitable for the current user from the content gene library based on the spatiotemporal behavior matrix in step 1. Next, the system monitors the user's interactive behavior in real time, such as user dwell time, scrolling speed, click location, etc. When the user's gaze leaves the key area, the system automatically replaces the visual elements in that area (such as changing the display image or adjusting the product angle) to maintain the user's attention. At the same time, the system also adjusts the emotional intensity of the copy based on the user's reading behavior. For example, when the user scrolls quickly, the system adjusts the emotional tendency of the copy to make it more concise and powerful to enhance its appeal. This mechanism ensures that content can be optimized at any time based on the user's actual behavior, improving user engagement and the attractiveness of the content.

[0029] Step 4 integrates marketing effectiveness metrics from multiple platforms, such as user dwell time, engagement rate, and conversion rate, to monitor the content's marketing effectiveness by calculating a dynamic performance index. When this index continues to decline, the system triggers a genetic mutation mechanism. This mechanism analyzes user behavior on other platforms and generates cross-platform compensatory content. This compensatory content, tailored to user interests and behavioral characteristics across different platforms, provides content that differs from previous marketing campaigns, compensating for failed marketing strategies. Through this closed-loop strategy optimization, the system automatically adjusts content, ensuring that marketing campaigns remain effective despite changes in user behavior and improving conversion rates.

[0030] Through a number of innovative technologies such as cross-platform data fusion, dynamic content generation and optimization, and closed-loop strategy optimization, the accuracy and effectiveness of new media intelligent marketing have been effectively improved, with significant commercial value.

[0031] In one possible implementation, the system first uses a cross-platform interface to obtain the timestamps of user interactions with the same target (such as a product, advertisement, or brand) across different platforms. Each time a user performs an action related to the target (such as a click, browse, or like) on a different platform, an action timestamp is recorded. The system then calculates the time interval between these actions to determine the duration of the inter-action interval. This inter-action duration serves as an important basis for subsequent association strength calculations, reflecting the user's attention and sustained interest in the target.

[0032] The system sets a baseline value for the decay coefficient based on the user's historical behavior frequency. Frequent cross-platform user behavior typically indicates a higher interest in a target, resulting in a smaller baseline decay coefficient value, implying a longer-lasting interest and slower decay in association strength. Conversely, a lower frequency of behavior indicates a more intermittent user's attention to a target, resulting in a larger baseline decay coefficient value and a rapid decrease in association strength. This setting can be flexibly adjusted based on each user's personalized behavior history to improve model accuracy.

[0033] In step c, the association strength of the behavior is calculated using an exponential decay model. This model dynamically adjusts the weight of user behaviors based on the relationship between the time decay coefficient and the interval between behaviors. As the interval between behaviors increases, the association strength decreases exponentially. That is, the longer the interval between behaviors, the lower the relevance of user interests. This model can more accurately reflect the decline in user attention over time, thereby better predicting user response probability in real-time marketing.

[0034] In step d, when the interval between a user's behaviors exceeds its historical maximum, the system activates a behavior chain break detection mechanism. By extracting the mean and standard deviation of the user's cross-platform behavior intervals over the past 30 days, the system assesses whether the current behavior interval is abnormal. If the current interval exceeds the mean plus three standard deviations, the system determines a "behavior chain break" and resets the user's association strength to a preset percentage of the initial value (e.g., 20%). This mechanism can detect sudden shifts in user interest or interruptions in behavior, preventing erroneous marketing decisions based on outdated behavior.

[0035] In step e, the system combines the behavioral correlation strengths of each platform with the platform's weight factor for a weighted fusion. The platform weight factor reflects the importance of each platform in the user's overall interests. For example, a platform with high activity or a greater impact on users will be given a higher weight. By fusing these weighted correlation strengths, a spatiotemporal behavior matrix is ​​formed, which is used for subsequent marketing content optimization and personalized recommendations.

[0036] By carefully analyzing the changing trends of users' cross-platform behavior and combining technologies such as exponential decay and behavior chain break detection, a more accurate and dynamic user behavior analysis method is provided, providing strong support for intelligent marketing.

[0037] In one possible implementation, a convolutional neural network (CNN) is first used to analyze video or image content to generate a gaze heatmap. This heatmap represents the probability of a user's gaze using different pixel values, with highlighted areas representing likely areas of gaze, and dimmed areas indicating a lower probability of gaze. Through automatic learning and feature extraction, convolutional neural networks can accurately identify areas of user visual attention within content. This is particularly effective in identifying potential points of user attention in complex content or long videos.

[0038] Next, the system clusters the generated heatmap, merging adjacent high-probability pixels to form multiple candidate regions. This process is accomplished using clustering algorithms (such as K-means or DBSCAN), aiming to effectively segment the area where the user is likely to be focused, avoiding interference from single pixels. These merged candidate regions help further identify areas of potential visual focus, providing clear regional definitions for subsequent refined calculations.

[0039] At this point, the system calculates the probability density integral for each candidate area. The magnitude of the integral represents the total probability of fixation on that area. Generally speaking, areas with higher probability density indicate greater user interest, so the system selects the area with the largest integral value as the visual focus area. This selection process ensures that the system captures the most important fixation areas, providing the most accurate user attention information for intelligent marketing.

[0040] The system records the center coordinates and coverage of the selected visual focus area and tracks the user's gaze in real time. By continuously monitoring the relationship between the gaze and the focus area, the system can identify whether the user has strayed from the focus area. When the closest distance between the gaze track and the focus area boundary continues to increase and exceeds a preset dynamic threshold, the system identifies a "gaze departure event." At this point, the marketing system can immediately adjust its strategy, such as reducing advertising displayed in that area or pushing content more closely aligned with the user's interests.

[0041] The identification of visual focus areas can not only effectively capture users' hot spots of interest through precise technical means, but also adjust marketing strategies based on users' dynamic behaviors, improve the effectiveness of new media marketing, and thus optimize user experience.

[0042] In one possible implementation, the system first captures the user's scrolling speed in real time while reading. This process relies on the user's touch behavior or mouse scrolling data. By sensing the user's sliding or scrolling speed, the system can calculate the deviation rate from the user's historical average speed based on historical records. A low deviation rate generally indicates that the user is reading slowly or lingering on a specific paragraph of content, while a high deviation rate may indicate that the user is quickly browsing or is interested in a certain part of the content. This information provides basic data for subsequent emotional regulation.

[0043] When the calculated drift rate falls below a preset first threshold, the system activates sentiment reinforcement mode. In this mode, the system first identifies the sentiment polarity (e.g., positive or negative) of the currently displayed paragraph. Next, the system inserts words with a higher intensity and consistent with the current sentiment polarity into subsequent content paragraphs, aiming to intensify the user's emotional experience. The insertion density of these sentiment words is negatively correlated with the user's drift rate. That is, if a user stays longer, the system will moderately increase the insertion density of these sentiment words to maintain interest and emotional engagement. This process, through precise sentiment analysis and intelligent adjustment, helps enhance the user's emotional response to content.

[0044] When the user's deviation rate exceeds the second threshold, the system activates interactive wake-up mode, prompting the user to engage. In this mode, the system inserts dynamic multiple-choice questions at the top of the next screen's content. These questions are highly relevant to the current marketing theme, designed to stimulate user reflection and encourage engagement. Based on the user's selection, the system generates different branching copy in real time, adjusting the subsequent content presentation based on the user's interests and reactions. This allows the system to form a more personalized and interactive connection with the user, enhancing their sense of engagement.

[0045] The first and second thresholds are not fixed but are dynamically set through a machine learning model. The system inputs historical user dwell time distribution and conversion rate data and analyzes this data using machine learning algorithms (such as regression analysis and decision trees) to determine the most effective offset rate thresholds. These thresholds are optimized to maximize content engagement and ensure that emotional reinforcement and interactive awakening modes are activated at the most appropriate time.

[0046] Through mechanisms such as intelligent perception of user behavior, emotional enhancement, and interactive awakening, it can not only improve user emotional participation, but also achieve more personalized and precise marketing strategies, thereby enhancing the attractiveness of content and marketing effectiveness.

[0047] In one possible implementation, the system first converts the content into a gene vector. Each dimension in the gene vector represents a distinct feature of the content, such as the copy's emotional tone, keywords, image style, and layout. Next, the system calculates the correlation coefficient between each dimension and the performance index based on dynamic performance indices (such as click-through rate, engagement rate, and conversion rate). The correlation coefficient measures the strength of the relationship between each content feature and marketing effectiveness, helping the system identify which content features have a significant impact on marketing effectiveness and which features may not contribute sufficiently.

[0048] Next, the system screens out weakly correlated dimensions whose correlation coefficients fall below a set adaptive threshold. This threshold is dynamically adjusted, and the system continuously optimizes it based on historical data and real-time feedback to ensure accurate selection of weakly correlated dimensions. This process allows the system to identify features with minimal impact on content effectiveness, avoiding unnecessary adjustments to these features in subsequent steps and conserving computing resources.

[0049] For weakly correlated dimensions, the system will apply directional perturbations to explore potential optimization space. The specific perturbation methods are as follows: For numerical features, such as article length and image brightness, the system applies Gaussian random perturbations within the feature's range. This perturbation simulates varying degrees of variation and explores the feature's impact on the performance index. The magnitude and range of the perturbation can be dynamically adjusted based on historical data to ensure that the perturbed features cover a wide range of possible scenarios.

[0050] For discrete features, such as color selection and title style, the system selects alternative templates from a pre-built variation library that are more similar than a preset threshold to the current feature. This variation library contains a variety of verified content templates with potential for optimization. The system selects the most suitable alternative based on feature similarity, ensuring the feasibility and effectiveness of the variant content.

[0051] Once the variant content is generated, the system selects a small user group for A / B testing. In this A / B test, the system compares the variant content with the original content, evaluating their impact on user behavior and dynamic performance index. By collecting and analyzing user responses in real time, the system determines which version of the content performs better in terms of performance index. Performance indexes typically include user behavior data such as click-through rate, dwell time, and conversion rate. The system uses this data as a basis to select the best-performing variant.

[0052] Ultimately, based on the results of the A / B test, the system will select the variation with the highest performance index and fully deploy it. At this point, the optimized content will be displayed on a wider scale, ensuring maximum marketing effectiveness.

[0053] By dynamically adjusting content features, implementing targeted perturbations, and making data-driven optimization choices, the marketing effectiveness of content is effectively improved, maintaining high flexibility and accuracy in an ever-changing market environment, and ultimately increasing user engagement and conversion rates.

[0054] In one possible implementation, the system first extracts the spatiotemporal behavior matrix of users whose marketing campaigns failed from platform P. This spatiotemporal behavior matrix refers to user behavior data on platform P, including data in multiple dimensions such as timestamps, geographic location, browsing duration, and click history. Based on the pre-set goals of the marketing campaign (such as purchases and registrations), the system then marks user behaviors that fail to achieve the intended goals. These failed behaviors are assigned failure reason labels, such as "incomplete purchase," "low engagement rate," and "not following." Using these labels, the system can distinguish different failure reasons, providing data support for subsequent analysis and optimization.

[0055] In this step, the system inputs a subset of the spatiotemporal behavior matrix of failed users on platform P into the content generation model of platform Q. In this way, the system compares the differences in historical behavior patterns of users on platform P and platform Q. Platform P and platform Q may have different user interfaces, interaction methods, or content presentations, so the user behavior patterns will also be different. The system will identify the differences between platform P and platform Q through differential deconstruction analysis, especially potential interest dimensions, which may not be fully met on platform Q. For example, users do not pay attention to certain types of content or do not perform certain behaviors on platform Q. Through this analysis, the system can identify and fill in the user's unmet interest points on platform Q, helping to improve personalized content recommendations.

[0056] Next, the system will generate a compensatory content gene vector based on the results of the first two steps. The content gene vector represents various characteristic dimensions of the content, such as visual elements, copywriting style, color matching, etc. When generating compensatory content, the system will: Based on the user preferences of Platform Q, strengthen the visual elements on Platform Q that users often interact with or tend to pay more attention to (such as specific image styles, video types, or interface designs).

[0057] Based on the failure reason labels, the system will avoid using copywriting elements that may cause user churn (such as overly direct promotional information or overly commercial language), thereby reducing potential negative impacts.

[0058] These adjustments will make the generated compensatory content more in line with the preferences of platform Q users and to some extent make up for the reasons for marketing failure on platform P.

[0059] Finally, the system utilizes a generative adversarial network (GAN) to synthesize compensatory content. Through adversarial training, GANs generate more natural and highly personalized content. During this process, the generator generates compensatory content, while the discriminator evaluates its authenticity and adaptability. After repeated optimization, the generated compensatory content is deemed most appropriate for the user's needs. The system prioritizes this compensatory content the next time the user logs into Platform Q, increasing user engagement and conversion rates.

[0060] This AI-based new media intelligent marketing method uses cross-platform behavioral analysis and generative adversarial networks to optimize content, effectively improving personalized recommendations and precise content matching, thereby enhancing the effectiveness of marketing activities and increasing long-term user engagement.

[0061] In one possible implementation, we first need to define key metrics for measuring marketing effectiveness: conversion rate, share rate, cross-platform reach, and user retention time. The specific meaning of each metric is as follows: Conversion rate: refers to the proportion of users who complete the target behavior (such as purchase, registration), reflecting the impact of marketing activities on user behavior.

[0062] Sharing rate: refers to the frequency with which users share marketing content with others, reflecting the spreadability and virality of the content.

[0063] Cross-platform communication depth: measures the dissemination effect of content across multiple platforms and illustrates the scope of content dissemination.

[0064] User dwell time: refers to the time users stay on the platform, reflecting the attractiveness of the content and user engagement.

[0065] These indicators are defined to comprehensively capture the different dimensions of marketing activities and provide data support for subsequent model training.

[0066] Furthermore, the system will train a weight distribution model through historical data. The specific input includes: Product type: Different types of products may have different appeal among different user groups, so product type needs to be included as an input to influence the model's predictions.

[0067] User portrait classification: including basic user information (such as age, gender, region, etc.), behavioral habits, etc. These features help the model identify user needs and then determine the weight of each indicator.

[0068] Characteristics of current hot events: External factors such as holidays and social hot spots often have an impact on user behavior and therefore need to be considered in the model.

[0069] The output of the model is the initial weight coefficient of each indicator, which represents the contribution of each indicator to the dynamic effect index under specific conditions.

[0070] During the actual content delivery process, the system monitors user interactions in real time. Every time a preset number of user interaction samples are added, the covariance matrix between indicators is recalculated. The covariance matrix reflects the interrelationships and influences between various indicators. By analyzing covariance changes, the system can determine the relative importance of each indicator.

[0071] Based on the changes in covariance, the system will proportionally scale the weight coefficients of each indicator. For example, if the covariance between a certain indicator and other indicators changes significantly, the system will adjust the weight of that indicator to ensure the accuracy and timeliness of the dynamic effect index.

[0072] In the final stage of data processing, the system normalizes the weighted sum to eliminate the influence of data of different magnitudes and achieve a more balanced contribution across indicators. The system then applies a logarithmic transformation to the normalized weighted results to eliminate the influence of long-tail distributions. Long-tail distributions often cause extreme values ​​to have a disproportionate impact on the overall results. The logarithmic transformation smooths the impact of these extreme values, resulting in a more stable and reliable dynamic performance index.

[0073] Through multi-dimensional data analysis and real-time feedback mechanisms, not only can the accuracy of marketing effectiveness evaluation be improved, but strategies can also be adjusted in real time according to different scenarios to ensure the efficiency and adaptability of marketing activities.

[0074] In a possible implementation, first, the system needs to establish a user line of sight stability model, the purpose of which is to analyze the stability of the user's line of sight when browsing content, especially the response in terms of content attractiveness and interactivity.

[0075] By analyzing historical user browsing data, the system extracts information about how long each user spends on a page and calculates the fluctuations in gaze dwell time. These fluctuations help the system understand user attention patterns, such as prolonged dwelling on certain content or pauses at specific locations, which reflect the user's level of interest in the content.

[0076] Based on historical user data, we calculate the standard deviation of their gaze trajectory as a baseline fluctuation. This standard deviation represents the typical fluctuation in the user's gaze movement and reflects their typical behavior. A larger standard deviation indicates a more erratic gaze movement on the page, while a smaller standard deviation indicates a more stable gaze.

[0077] The system monitors the user's gaze speed and acceleration in real time through sensors or front-end data capture technology on the user's device. If the user's gaze jumps significantly or moves frequently, the acceleration will increase accordingly, potentially indicating a strong interest in the content or unusual behavior (such as a sudden shift in attention).

[0078] When the acceleration of the line of sight exceeds a certain multiple of the baseline fluctuation, the system triggers a dynamic threshold adaptation mechanism. This mechanism dynamically adjusts the threshold according to the magnitude of the acceleration: When the acceleration is significantly greater than the baseline fluctuation, it indicates an abnormal change in gaze, which may indicate that the user's attention has shifted elsewhere or that there is a system problem. Therefore, the threshold is narrowed to allow the system to respond more quickly to such anomalies and make timely adjustments.

[0079] When the acceleration changes slightly, that is, the line of sight changes relatively smoothly, it indicates that the user is still browsing normally. The system will increase the threshold to avoid overly sensitive reactions and reduce the risk of false triggering.

[0080] By introducing the line of sight stability model and dynamic threshold adaptation mechanism, it can respond to changes in user behavior in real time, optimize marketing strategies, and significantly improve the accuracy of the system and user experience, which has high practical value.

[0081] In one possible implementation, the system first builds a decision tree-based content template library. A decision tree is a model that generates different content based on user selection paths. Each branch node represents a user's selection path, and different user choices lead to different paths. This content template library contains a variety of preset copy templates. The system determines the path and selects the appropriate copy template based on user interactions (such as clicks and selections). This approach ensures personalized and precise marketing content, generating customized marketing messages in real time based on user needs and behaviors.

[0082] When a user makes a selection in marketing content, the system matches the appropriate pre-set template based on the semantics of that selection. The system performs semantic matching based on the user's selected keywords or categories, finding the template from the template library that best matches that selection and displaying it. For example, if a user selects "Summer Clothing," the system will select the relevant summer clothing promotional copy template from the decision tree.

[0083] If the system can't find a fully matching template, it will invoke the GPT model to generate a coherent copy based on the user's selection in real time. Based on the context and intent of the user's selection, the GPT model uses natural language processing technology to generate copy that aligns with the user's interests, thus filling gaps in the template library and ensuring coherence and appeal.

[0084] The system embeds a tracking tag in each generated branch copy. This tracking tag is a unique identifier used to track and record the user's behavioral path. When users perform subsequent actions or conversions (such as purchases and registrations), the system can associate these actions with the branch path, clearly indicating that these conversions were guided by a specific copy branch. In this way, the system can accurately evaluate the effectiveness of different branch copies and optimize subsequent marketing strategies.

[0085] Through technical means such as intelligent decision trees, semantic matching, GPT copy generation and tracking tags, highly personalized and automated content generation can be achieved, which not only improves marketing efficiency but also provides accurate data support for subsequent effect analysis and strategy optimization.

[0086] In one possible implementation, user behavior typically varies across different social or media platforms (such as Platform Q). Platform Q has unique user behavior patterns, such as browsing frequency, interaction methods, and clickthrough habits. Extracting Platform Q-specific behavioral characteristic dimensions involves conducting in-depth analysis of Platform Q user behavior to identify specific behavioral dimensions on Platform Q. For example, a platform may have users who frequently exhibit a preference for a certain content type (e.g., videos, images, comments, etc.). The system extracts these behavioral characteristics and forms dimensions for subsequent analysis.

[0087] Next, the system needs to evaluate each user's performance in these behavioral dimensions and compare it with the average level of the platform. Specifically, the system calculates the difference between the user's performance in this particular dimension and the average level of the overall users of platform Q, that is, the deviation. Deviation can be measured by statistical methods such as standard deviation, mean difference, etc., to measure the degree to which user behavior differs from other users on the platform. For example, if the average daily video viewing time of users on platform Q is 30 minutes, and a user's viewing time is 60 minutes, then the user's deviation in the video viewing time dimension is 30 minutes, which means that the user deviates from the average behavior pattern of the platform.

[0088] When a user's deviation in a certain dimension exceeds the significance threshold preset by the system, it means that the user's behavior is significantly different from that of regular users on Platform Q. Based on this difference, the system will add this specific dimension to the compensatory content reinforcement list. The compensatory content reinforcement list is an adjustment mechanism for differences in user behavior. In subsequent content pushes, the system will strengthen or adjust the content related to this dimension to compensate for or optimize the user's experience in this dimension. For example, if a user has a large deviation in video viewing time, the system may push more short video content to the user, or recommend videos that they are more interested in, so that they are closer to the average user behavior pattern of Platform Q, thereby improving user engagement and platform activity.

[0089] Through precise platform preference identification and a compensatory content reinforcement mechanism, we not only achieve personalized and efficient content recommendations, but also provide important data information for subsequent marketing decisions. These technical features greatly improve the accuracy and effectiveness of marketing activities.

[0090] The following is a detailed explanation using examples: This invention is applied to an e-commerce platform. The platform's user behavior analysis system can make personalized recommendations based on historical user behavior data, browsing history, and purchasing preferences, driving sales and improving user experience. The system uses multiple algorithms combined with user deviation analysis to optimize recommended content.

[0091] Specifically, the platform records user behavior data in real time, including user browsing, clicking, searching, purchasing, and other operations. Each behavior is recorded through data fields such as timestamp, product ID, and behavior type (browse, click, purchase, etc.).

[0092] Example of collected data: User A: Views product A (time: 10:05), clicks product B (time: 10:15), and purchases product C (time: 10:30) User B: Views product D (time: 10:10), clicks product E (time: 10:20), and purchases product F (time: 10:40) Generate user interest vectors using behavioral data. For each user, calculate their interest value in different product categories as the result of feature extraction.

[0093] In this embodiment, the e-commerce platform has five product categories (A, B, C, D, and E). Each user's interest vector is calculated based on the weighted weights of the user's behavior in each category.

[0094] For user A, in this embodiment, there are 2 views on category A, 1 click on category B, 1 purchase on category C, and no actions on categories D and E. The interest vector can be calculated as follows: I A =(0.4,0.2,0.2,0.1,0.1) Each element represents the interest level of the category (i.e., behavior frequency / total number of behaviors).

[0095] This paper uses deviation analysis to evaluate the difference between user behavior and the overall user behavior of the platform. Deviation measures the difference between user interests and the average user interests of the platform, and the formula is: ; in, is the interest of user A in category i, is the average interest of all platform users in category i.

[0096] For example, if the average interest of all platform users in category A is 0.25, then for user A, the deviation from category A is: ; The greater the deviation, the more the user deviates from the behavior of the majority of users on the platform, and the system needs to recommend more content that suits their unique interests.

[0097] The system adjusts its recommendation strategy based on the degree of deviation. Users with a large degree of deviation will receive more personalized product recommendations, while users with a small degree of deviation will be recommended products that are generally popular on the platform.

[0098] The product recommendation formula is: ; in, is the final recommendation list for user A, is the deviation of category i, is the weight of category i (set according to the importance of the product category), is a category Rating of the product.

[0099] Based on the final calculated recommendation list, the platform displays personalized recommended products to users. The recommendation results may include multiple product categories and are sorted based on the user's purchase history and browsing preferences.

[0100] The frequency of behavioral data collection is once per second to ensure real-time updates of user behavior.

[0101] When calculating the deviation, a threshold is set (such as 0.3). When the user's deviation is greater than the threshold, the system will prioritize pushing highly personalized product recommendations.

[0102] The interest vector of each user is normalized between 0 and 1 to avoid bias caused by class imbalance.

[0103] Weight adjustment: The weight of a category is dynamically adjusted based on the sales of the product. For example, a hot-selling product will have a higher weight.

[0104] In order to verify the superiority of the recommendation system of the present invention, a comparative experiment was conducted with a traditional recommendation system based on product click-through rate. The experimental group used the deviation algorithm of the present invention, while the control group used the traditional recommendation method based on user behavior frequency.

[0105] Experimental design: Experimental subjects: 1,000 e-commerce platform users Experimental period: 1 month Evaluation indicators: user engagement (click-through rate), purchase conversion rate, recommendation accuracy Experimental results: Experimental group (deviation recommendation): User engagement increased by 20% and purchase conversion rate increased by 15%.

[0106] Control group (traditional recommendations): User engagement increased by 5% and purchase conversion rate increased by 8%.

[0107] The experimental comparison results prove that the personalized recommendation system based on deviation analysis of the present invention can significantly improve user participation and purchase conversion rate compared with traditional methods, and is more in line with the personalized needs of users.

[0108] By analyzing the deviations in user behavior, this invention enables more accurate personalized recommendations, effectively improving user experience and conversion rates on e-commerce platforms. This system can provide product recommendations that best match each user's unique interests, addressing the inability of traditional recommendation systems to accurately meet user needs. The recommendation algorithm and system design of this invention are highly innovative and practical, and can significantly enhance the commercial benefits of e-commerce platforms.

[0109] The present invention encompasses any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention. To provide a thorough understanding of the present invention, specific details are described in detail below in connection with the preferred embodiments of the present invention, but those skilled in the art will be able to fully understand the present invention without these detailed descriptions. Furthermore, to avoid unnecessary confusion regarding the essence of the present invention, well-known methods, processes, procedures, components, and circuits have not been described in detail.

[0110] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. AI-based new media intelligent marketing method, characterized by: include: Step 1: Cross-platform user behavior fusion modeling: Collect users' content interaction behaviors, implicit attention behaviors, and cross-platform association behaviors through the interfaces of each platform; Determine the platform weight factor based on user cross-platform activity and construct a spatiotemporal behavior matrix using a time decay algorithm; Step 2: Construction of multimodal content gene library: Analyze the visual focus areas and text sentiment tendencies of historical content to generate content gene vectors; Step 3: Dynamic content generation and real-time optimization: Search the gene library based on the spatiotemporal behavior matrix to generate initial marketing content; Monitor user interaction behavior in real time, dynamically replace elements in the visual focus area, and adjust the emotional intensity of the copy; Step 4: Closed-loop strategy optimization: The dynamic effect index is calculated by integrating multi-platform indicators. When the index drops continuously, the gene mutation mechanism is triggered, and cross-platform compensatory content is generated based on failed user behavior.

2. The AI-based new media intelligent marketing method according to claim 1, characterized in that: The specific steps of the time decay algorithm include: a: Identify the timestamps of users’ actions on the same target on different platforms and calculate the duration of the behavior interval; b: Set the attenuation coefficient baseline value based on the user's historical cross-platform behavior frequency. The higher the behavior frequency, the smaller the attenuation coefficient baseline value; c: Calculate the association strength using the exponential decay model: for every unit increase in the behavior interval, the association strength is multiplied by the decay coefficient; d: When the interval between actions exceeds the maximum interval in the user's history, the action chain break detection is started: Extract the mean and standard deviation of the user's cross-platform behavior intervals over the past 30 days; If the current interval exceeds the mean plus three standard deviations, the behavior chain is determined to be broken, and the association strength is reset to a preset ratio of the initial value; e: Weighted fusion of the association strength of each platform and the platform weight factor, and output to the spatiotemporal behavior matrix.

3. The AI-based new media intelligent marketing method according to claim 1, characterized in that: The identification of the visual focus area includes: a: Generate a gaze heat map of the content keyframe through a convolutional neural network. The heat map represents the gaze probability in pixel values. b: Perform regional clustering on the heat map and merge adjacent high-probability pixels to form candidate regions; c: Calculate the probability density integral of each candidate area and select the area with the largest integral value as the visual focus area; d: Record the center coordinates and coverage of the visual focus area. When the user's real-time vision deviates from this area: Get the shortest distance between the sight track and the boundary of the focus area; If the distance continues to increase and exceeds the dynamic threshold, it is determined as a gaze-away event.

4. The AI-based new media intelligent marketing method according to claim 1, characterized in that: The adjustment process of the emotional intensity of the copywriting is as follows: a: Capture the user's screen scrolling speed in real time and calculate its deviation rate from the user's historical average speed; b: When the deviation rate is lower than the first critical value, the emotional reinforcement mode is activated: Identify the sentiment polarity of the currently displayed paragraph; When inserting sentiment words of the same polarity and higher intensity into subsequent paragraphs, the insertion density is negatively correlated with the offset rate; c: When the offset rate is higher than the second critical value, the interactive wake-up mode is started: Insert a dynamic multiple-choice question component at the top of the next screen, with the options strongly related to the marketing theme; Generate branch copy in real time based on user selection results; d: The first critical value and the second critical value are dynamically set through a machine learning model: Input user historical stay time distribution and conversion rate data; Output the velocity threshold bounds that maximize content engagement.

5. The AI-based new media intelligent marketing method according to claim 1, characterized in that: The gene variation mechanisms include: a: Calculate the correlation coefficient between each dimension of the content gene vector and the dynamic effect index; b: Screening weakly correlated dimensions whose correlation coefficients are lower than the adaptive threshold; c: Apply directional perturbations to weakly correlated dimensions: If it is a continuous dimension, perform Gaussian random perturbation within the value range; If it is a discrete dimension, select an alternative template from the variation library whose similarity is greater than the preset value; d: After generating variant content, conduct A / B testing on a small user group, and select the variant with the greatest improvement in performance index and launch it in full.

6. The AI-based new media intelligent marketing method according to claim 1, characterized in that: The specific steps of generating cross-platform compensation content are: a: Extract the spatiotemporal behavior matrix subset of users who failed marketing on platform P and mark the failure reason label; b: Input the subset of behavior matrix into the content generation model of platform Q and perform differential deconstruction: Compare the differences in users' historical behavior patterns on the P and Q platforms; Identify potential unmet interest dimensions of users on platform Q; c: Generate compensatory content gene vector: Strengthen the visual elements that are unique to Platform Q; Weaken the copywriting elements related to the reasons for failure; d: Compensatory content is synthesized through generative adversarial networks and pushed first when the user logs into the platform Q next time.

7. The AI-based new media intelligent marketing method according to claim 1, characterized in that: The process of constructing the dynamic effect index includes: a: Define the core indicator set: conversion rate, sharing rate, cross-platform communication depth, and user stay time; b: Train the weight distribution model through historical data: Input: product type, user profile classification, and current hot event characteristics; Output: initial weight coefficient of each indicator; c: Adjust weights in real time during content delivery: Every time a preset number of user interaction samples are added, the covariance matrix between indicators is recalculated; Scale the weight coefficients proportionally according to the magnitude of the covariance change; d: Perform a logarithmic transformation on the normalized weighted summation result to eliminate the influence of long-tail distribution.

8. The AI-based new media intelligent marketing method according to claim 3, characterized in that: Determining the dynamic threshold includes: a: Establish a user's sightline stability model: Extract the fluctuation characteristics of the user's gaze dwell time in historical browsing; Calculate the standard deviation of the gaze trajectory as the baseline fluctuation; b: Real-time monitoring of gaze acceleration. If the acceleration exceeds a multiple of the baseline fluctuation, threshold adaptation is triggered: The larger the multiple, the smaller the threshold is to quickly respond to anomalies; The smaller the multiple, the larger the threshold is to avoid false triggering.

9. The AI-based new media intelligent marketing method according to claim 4, characterized in that: The generation logic of the branch copy is: a: Build a decision tree content template library, where each branch node corresponds to a user selection path, i.e., a branch path; b: After the user selects an option: Match preset templates based on option semantics; If the match fails, the GPT model is called to generate coherent copy in real time based on the option content; c: Embed tracking tags in branch copy, and subsequent conversion behaviors are attributed to this branch path.

10. The AI-based new media intelligent marketing method according to claim 6, characterized in that: The platform preference identification in the differentiation deconstruction includes: a: Extract the behavioral feature dimensions unique to platform Q; b: Calculate the user's deviation from the platform's average level in this dimension; c: If the deviation exceeds the significance threshold, the dimension is added to the list of compensatory content enhancements.

Images