Multi-element sales planning agent system and method

Abstract

The invention discloses a multi-element sales planning agent system and method, and aims to improve the intelligence and precision of sales planning. The system comprises a collection module, an analysis module, an optimization module, a creation module and a generation module. The collection module is used for receiving multi-modal data such as marketing targets and extracting key marketing elements. The analysis module is used for generating a target user portrait and extracting marketing strategy analysis data. And the optimization module is used for calculating a medium putting weight by utilizing reinforcement learning and generating a medium strategy scheme. And the creation module generates a propagation theme and marketing content by adopting a generative artificial intelligence technology. And the generation module predicts a delivery effect by using a machine learning model and dynamically optimizes a medium strategy and a content scheme. Through multi-modal data fusion, intelligent analysis and optimization, closed-loop processing from data acquisition to marketing execution is realized, the marketing decision-making efficiency is improved, and brand promotion accuracy and market adaptability are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing systems, and in particular to a multi-dimensional marketing planning intelligent system and method. Background Art

[0002] In existing technologies, companies' marketing planning typically relies on manual analysis of market trends, target user characteristics, and competitive product strategies, and the development of marketing plans based on this information. The traditional marketing planning process includes data collection, market research, strategy development, content creation, and effectiveness evaluation, each of which is typically performed independently by different teams or tools. Furthermore, marketing content is primarily disseminated through a single media channel, such as text ads, social media placements, or video ads. Furthermore, market feedback data is often collected manually, and strategy adjustments are often based on empirical experience.

[0003] Existing technologies still have many shortcomings in terms of intelligent and precise marketing planning. First, traditional marketing planning is highly dependent on manual labor and has limited data processing capabilities, making it difficult to effectively integrate multimodal information from different channels, such as text, voice, images, and video. Second, market insights and competitive product analysis mainly rely on empirical judgment and lack data-driven precision modeling, resulting in limited accuracy in user portraits and strategy formulation. In addition, media delivery strategies are often based on fixed rules and fail to fully utilize data analysis and machine learning technologies for real-time optimization, making it difficult to achieve optimal resource allocation. At the same time, the existing methods of generating marketing content are relatively simple, making it difficult to dynamically adjust according to media characteristics and audience needs, which affects marketing effectiveness.

[0004] In view of the above-mentioned deficiencies in the existing technologies, there is an urgent need to provide a marketing planning intelligent system and method that can efficiently integrate multimodal data, intelligently optimize media strategies, and improve the generation and delivery effects of marketing content. Summary of the Invention

[0005] This application provides a multi-dimensional marketing planning intelligent system and method to improve the intelligence and accuracy of marketing planning.

[0006] This application provides a multi-dimensional marketing planning agent system, including: The collection module is used to receive basic data including marketing objectives, market background, brand tone, and budget constraints. The basic data includes multimodal input such as text, voice, image, and video; and extract key marketing element data based on natural language processing technology and computer vision analysis; An analysis module is used to generate target user portrait data using a cluster analysis method based on the key marketing factor data, combined with external market data, social media trend data, and industry report data, and to generate competitive product marketing strategy analysis data based on competitive product data mining technology; An optimization module is used to calculate the delivery weights of different media channels using a reinforcement learning model based on target user profile data, competitive product marketing strategy analysis data, corporate budget data, and historical delivery data; and to generate media strategy plan data based on the delivery weights; A creation module is used to create communication theme data and marketing content data adapted to different media based on the media strategy plan data and in combination with media channel characteristics through generative artificial intelligence technology; A generation module is used to predict the delivery effect through a machine learning model based on the communication theme data, marketing content data, historical marketing data and real-time market feedback data, and optimize the media strategy plan and marketing content plan according to the prediction results.

[0007] Furthermore, the collection module includes a data fusion unit, which is used to construct a cross-modal association graph based on the timestamp information and content features of the multimodal input data, and use a graph neural network to perform association analysis on data of different modalities to extract deep marketing elements of user preferences. The cross-modal association graph includes contextual relationship nodes between text, voice, image and video data, and is combined with a feature weight calculation method based on the self-attention mechanism to dynamically weighted fuse different modal information to enhance the accurate identification of target users' market behavior.

[0008] The beneficial effects of this application mainly include: (1) The collection module supports multimodal data input such as text, voice, image and video, and combines natural language processing and computer vision analysis to achieve accurate extraction of corporate marketing elements. Compared with the traditional method of relying on single text data, this system can make full use of multi-source data to improve the comprehensiveness of market insights, thereby improving the accuracy and pertinence of marketing planning. (2) Cluster analysis methods are used to accurately profile target users, and combined with competitive product data mining technology, competitors' marketing strategies are comprehensively analyzed. This data-driven analysis method can make up for the shortcomings of traditional marketing planning that relies on experience judgment, enabling companies to formulate differentiated marketing strategies based on accurate data models, thereby improving market competitiveness. (3) The media channel delivery weight is calculated through a reinforcement learning model, and the media strategy is dynamically adjusted based on the target user portrait, competitive product analysis and historical delivery data to ensure the optimal allocation of marketing budgets. Compared with the traditional delivery method based on fixed rules or manual adjustments, this system can adapt to market changes, achieve the optimal media combination strategy, and improve the ROI (return on investment) of advertising delivery. (4) Integrate generative AI technology to create communication themes and marketing content that are suitable for different media, and use machine learning models to predict the effectiveness of delivery to ensure the matching degree between content generation and media delivery. Through real-time market feedback, optimization and adjustment can be made to dynamically adjust content strategies, improve advertising conversion rates and user engagement. Compared with traditional marketing methods, it can adapt to market changes more flexibly and achieve better marketing results. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 This is a schematic diagram of a multi-dimensional marketing planning intelligent system provided in the first embodiment of the present application.

[0010] Figure 2 This is a flowchart of a method for a multi-dimensional marketing planning intelligent entity provided in the second embodiment of the present application. DETAILED DESCRIPTION

[0011] The following description sets forth many specific details to facilitate a thorough understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the scope of the present application. Therefore, the present application is not limited to the specific implementations disclosed below.

[0012] The first embodiment of this application provides a multi-dimensional marketing planning intelligent system. Figure 1 , which is a schematic diagram of the first embodiment of this application. Figure 1 A multi-dimensional marketing planning intelligent system is provided in detail for the first embodiment of the present application.

[0013] The multi-dimensional marketing planning intelligent system includes a collection module 101, an analysis module 102, an optimization module 103, a creation module 104 and a generation module 105.

[0014] The collection module 101 is used to receive basic data including marketing goals, market background, brand tone and budget constraints. The basic data includes multimodal input such as text, voice, image and video; and extract key marketing element data based on natural language processing technology and computer vision analysis.

[0015] The primary function of Collection Module 101 is to receive and process various forms of basic data, including but not limited to key information such as marketing objectives, market context, brand tone, and budget constraints. It also supports multimodal input such as text, voice, image, and video. After acquiring the raw data, this module's core task is to parse, extract, and structure it using natural language processing and computer vision analysis techniques to generate key marketing element data for subsequent use by Analysis Module 102.

[0016] During text data processing, the collection module 101 first uses the text parsing submodule to perform basic preprocessing on the input data, including character encoding conversion, redundant space removal, and punctuation normalization. Subsequently, a deep learning-based word segmentation algorithm (such as BERT) is used to segment the text to identify keywords, entity names, and key phrases. For example, from the input market context description, the system can extract key information such as "smart home market growth rate," "target audience: men aged 25-40," and "major competing brands: Company A, Company B." Next, entity recognition (NER) and dependency parsing are used to further extract semantic relationships between the data, such as mapping brand names to their corresponding product categories and linking marketing objectives to target audiences.

[0017] For voice data, collection module 101 uses a speech recognition engine (such as a Transformer-based ASR model) to convert the audio into text. It then uses phoneme-level time alignment technology to eliminate noise interference such as pauses and slips of the tongue. The converted text data undergoes the same NLP processing as text data, including keyword extraction, sentiment analysis, and context recognition, to ensure semantic consistency between the voice and text input data. For example, when a company describes its marketing objectives via voice, the system can extract key content, such as "target users are college students" and "promotion budget is limited to 100 yuan," and store it in a structured manner.

[0018] In terms of processing image and video data, the collection module 101 first uses a computer vision model to perform frame-level analysis on the data. For static images, the system can call deep learning models such as YOLO, Faster R-CNN, or EfficientNet for target detection to identify key visual elements in brand logos, product packaging, and marketing materials. For example, when a brand poster is input, the system can detect key information such as the location of the logo, the type of main product, and the slogan. For video data, the system uses inter-frame keyframe extraction technology, combined with optical flow analysis to screen out the most informative images, and performs OCR text recognition and sentiment analysis on them to extract the brand tone, marketing theme, and emotional tendency in the video.

[0019] In addition, in order to enhance the relevance of cross-modal data, the collection module 101 adopts multimodal fusion technology. First, feature extraction is performed on data of different modalities, and text, voice, image and video data are converted into a unified embedding vector representation, such as using BERT for text embedding and CNN+Transformer for image embedding. Subsequently, the multimodal attention mechanism is used to calculate the correlation between different data modalities. For example, if the text description mentions "high-end luxury goods promotion" and the color style in the video is mainly black and gold, the system can confirm the consistency of the marketing tone through the visual-text alignment model. In addition, the module also uses the cross-modal learning technology of the Transformer architecture to fuse the features of information from different modalities, so that the system can comprehensively judge the marketing goals, audience groups and market positioning.

[0020] Finally, the collection module 101 outputs structured key marketing element data, including but not limited to: Marketing target data (extracted from text or voice) - including brand promotion goals, market expansion plans, growth expectations, etc. Audience profile (combining text, voice, and market research data), including age, gender, interests, preferences, spending habits, and other information; Brand tone analysis (combining text description and image / video analysis) to determine brand style, such as high-end, youthful, and technological; Budget and cost constraints (extracted from enterprise input data and stored in a structured manner); Market competition environment data (combined with competitor analysis and keywords extracted from social media trends), including competitor brands, key competitive advantages, market share forecasts, etc.

[0021] The structured data output by this module will be transmitted to the analysis module 102 to support subsequent steps such as market insights, user portrait construction and competitive product analysis, thereby achieving accurate marketing planning and intelligent optimization.

[0022] Furthermore, the collection module includes a data fusion unit, which is used to construct a cross-modal association graph based on the timestamp information and content features of the multimodal input data, and use a graph neural network to perform association analysis on data of different modalities to extract deep marketing elements of user preferences. The cross-modal association graph includes contextual relationship nodes between text, voice, image and video data, and is combined with a feature weight calculation method based on the self-attention mechanism to dynamically weighted fuse different modal information to enhance the accurate identification of target users' market behavior.

[0023] The data fusion unit in the collection module processes multimodal data inputs, including text, voice, images, and videos. Based on the timestamp information and content features of this data, it constructs a cross-modal association graph to uncover user preferences and market behavior. This unit first receives multimodal data and synchronizes data from different sources to ensure temporal consistency in data analysis. For text data, the system parses user-entered marketing objectives, social media interactions, search history, and online reviews, extracting keywords, semantic sentiment, and related entities using natural language processing techniques. Voice data is converted to text using automatic speech recognition (ASR) technology and combined with sentiment analysis models to determine the user's emotional state during voice interaction. For image and video data, the system utilizes computer vision techniques for object detection, scene recognition, and optical character recognition (OCR) text extraction to analyze visual elements of user interest, such as brand logos, product categories, and color combinations.

[0024] After extracting basic information between modalities, the data fusion unit further performs cross-modal correlation analysis on data from different modalities. The system uses a graph neural network (GNN) to construct a cross-modal correlation graph. The nodes of this graph represent data units such as text, voice, image, and video, and the weights of the edges represent the correlation between these data. For example, if a user searches for a specific brand of mobile phone on social media and then watches an advertising video for that brand, the system will establish a high-weight edge between the text search data node and the video data node to indicate the user's high interest in the product. In addition, if the brand is mentioned in the user's voice interaction, the voice data node will also be associated with other related data nodes to enhance the overall understanding of user behavior.

[0025] To improve the effectiveness of fusing data from different modalities, the system combines a feature weight calculation method based on the self-attention mechanism to dynamically adjust the contribution of different modal data to the final marketing factor extraction. The self-attention mechanism calculates the influence of different modal data on the target task and gives greater weight to data with high relevance. For example, if a user has recently searched for a certain type of product on multiple platforms, and their social media interactions, voice queries, and video browsing behaviors all point to this category, the system will dynamically increase the weight of this modal information in the final marketing decision. Conversely, if a certain modality of data appears only once or has a weak correlation, its weight will be reduced to reduce the interference of noisy data.

[0026] Ultimately, the data fusion unit outputs weighted, fused deep marketing elements, which can be used for subsequent market analysis, target user profile construction, and marketing strategy optimization. In this way, the system can extract precise user interests from dispersed multimodal data and enhance the ability to identify target users' market behavior, providing efficient data support for intelligent marketing strategies.

[0027] The data fusion unit uses a dynamic weighted cross-modal embedding mapping method to accurately calculate the contribution of different modal input data to the target user's market behavior. In actual implementation, the unit first receives user interaction information from multimodal data such as text, voice, image, and video, and dynamically adjusts the influence weight of different modalities on marketing strategies based on timestamps, content features, and behavior patterns. The system calculates the feature contribution of each modality using formula (1) to ensure the accuracy and adaptability of marketing data analysis.

[0028] in, Indicates modality The dynamic weight in the current user interaction data fusion process reflects the importance of the modality in marketing data analysis and is used to measure the relative importance of the modality in the overall analysis. This value is dynamically adjusted as the user interaction data changes to ensure that the system adapts to the market behavior patterns of different users. This value can be obtained through experimental data or directly set based on expert knowledge.

[0029] The total number of multimodal interaction records for the current user within a specific time window. This window can be set based on business needs, such as all interactions in the last 7 days or 30 days.

[0030] For the current user The content match degree of an interaction is determined by calculating the similarity between the modal data and the current marketing theme, brand characteristics, and historical behavior. For example, semantic similarity can be calculated for text data using natural language processing models such as TF-IDF or BERT. Voice data can be converted to text through speech recognition and then its match degree calculated. Image and video data can be matched against a database of brand characteristics after visual features are extracted using convolutional neural networks. The value range is [0, 1], with higher values indicating a higher relevance of the interaction content to the marketing theme.

[0031] In this embodiment, the content matching degree To measure the user's The correlation between a modal data content encountered in an interaction and the current marketing theme, brand characteristics, and the user's historical behavior. Its value range is [0,1]. A higher value indicates a higher degree of match.

[0032] First, current marketing themes are typically entered into the system by business planners and stored in a structured database. These include keywords or phrases like "campus promotion," "green travel," and "new product launch." The system then uses natural language processing models to vectorize this text. Commonly used models include pre-trained language models like BERT. These models can convert marketing theme text into multi-dimensional semantic vectors, which serve as a computational benchmark.

[0033] Brands enter "brand characteristics" during system initialization or daily operations, including brand positioning, keyword descriptions, and product styles, such as "technological," "environmentally friendly," and "high-end customization." The system also uses language models to convert these brand characteristics into semantic vector representations. For visual brand characteristics (such as brand logos, standard packaging images, and primary colors), the system uses image feature extraction models (such as ResNet) to convert image content into corresponding visual vectors for use in image-based modal calculations.

[0034] "User history" is extracted from past user interactions on the platform, including search terms, comments, viewed images and text, and clicked videos. The system converts historical text data into vectors using the same semantic encoding method, while visual content is extracted using image models to extract feature vectors. Based on behavioral parameters such as interaction frequency and dwell time, the system assigns higher weights to highly relevant data within historical behavior, thereby forming a weighted semantic representation of a user's historical interests.

[0035] In a certain mode, if the user In this interaction, a text comment "This eco-friendly backpack has a simple design and is made of solid materials" was submitted. The system will first use the NLP model to semantically encode the comment and obtain its semantic vector representation. Then, the system calculates the semantic similarity between the semantic vector and the current marketing theme (such as "environmentally friendly life"), brand characteristics (such as "sustainable materials"), and user historical preferences (such as users have browsed or liked similar products in the past). Semantic similarity calculation usually uses the cosine similarity method, and the result is a value between [0,1]. The system can perform a weighted average of the three similarity values, or take the maximum value (depending on the implementation strategy) as the content matching degree of the modal data in this interaction. .

[0036] If the interaction is in voice mode, the system will first call the speech recognition module (such as a commercial ASR engine) to transcribe the speech content into text, and then perform the same semantic matching process as the text mode.

[0037] If the interaction is in the image or video modality, during the content matching process, the system first analyzes the visual features of the user-uploaded image or the keyframe images automatically extracted from the video clip of the user interaction. This step is accomplished using an image feature extraction model, typically using a pre-trained deep neural network model such as CLIP or ResNet. These models convert each image into a high-dimensional semantic vector. This vector not only preserves the image's appearance features, such as shape, color, and texture, but also possesses certain semantic representation capabilities, such as the ability to identify whether the image contains a specific product, brand logo, or contextual style.

[0038] At the same time, brands have pre-set a visual reference database within the system, known as the brand visual database. This database contains multiple images representing brand characteristics, such as standard images of flagship products, brand advertising materials, product packaging, brand logos, and color templates. During deployment, the system also feeds these images into the image feature extraction model to generate semantic vectors for each image, which are then stored in the database for retrieval and comparison.

[0039] When a user submits a new image or video, the system will compare the vector of the image or video frame with the vectors of all images in the brand visual database one by one for similarity. The comparison method is not a simple pixel-by-pixel comparison of whether the images are identical, but rather a determination of the degree of proximity between the two semantic vectors in multidimensional space. The closer the vectors, the more similar the image is to the brand characteristics. The system can identify whether there are visual symbols corresponding to brand keywords such as "eco-friendly backpack", "minimalist design" or "green travel" in the user's interactive content, such as identifying the backpack's outline structure, label color, background scene, etc.

[0040] After the system completes all image comparisons, it aggregates the similarity results. If a particular image has a high degree of similarity with multiple images in a brand's library, the system combines these results, either averaging the similarities or taking the highest one as the content match value for the image modality in the current interaction. This match value falls within a set range, typically between 0 and 1, indicating the degree of visual consistency between the image and the brand's current marketing objectives.

[0041] The processing for video modalities is similar to that for image modalities. The system automatically selects several key frames from the video, typically based on factors such as time interval, scene changes, or image content density. Each frame is independently fed into the image feature extraction model to obtain its semantic vector, which is then matched against image vectors in the brand visual library for similarity. The system then averages or weights the results across multiple frames to determine the content match for the entire video.

[0042] Finally, the matching As user The consistency measure of the interaction under this mode with the marketing goal is input into formula (1) to calculate the contribution weight of the current mode in the entire analysis process. .

[0043] For the current user The timestamp of the interaction, which is the time difference relative to the current moment, is used to measure the freshness of the interaction; For the current user The emotional intensity in the interaction is calculated by natural language processing models and computer vision emotion recognition technology to measure the emotional polarity in text, speech, images or videos. The numerical range is For example, text sentiment analysis can use deep learning models such as LSTM or Transformer to analyze the emotional tendencies of social media comments or search queries, voice data can extract emotions through acoustic features, and images and videos can calculate emotional weights based on facial expression detection or scene analysis.

[0044] emotional intensity To measure the user's The emotional leaning towards content during a user's interaction can be determined by analyzing the emotional characteristics of text, speech, images, or videos using natural language processing and computer vision technologies. The numerical value ranges from -1 to 1, indicating the polarity of the user's emotion. To ensure accuracy, different analysis methods are used for data in different modalities, and a comprehensive assessment is conducted based on the user's historical behavior and current context.

[0045] For text data, the system analyzes users' social media comments, search queries, product reviews, or online conversations to determine their sentiment. For example, if a user comments, "This product is truly amazing. I highly recommend it to everyone!", the sentiment analysis model will identify positive sentiment words such as "great" and "recommended." Based on the context, it will determine that the overall sentiment is positive, assigning the text a sentiment score close to 1. Conversely, if a user comments, "This product completely does not meet expectations and the quality is extremely poor," the model will identify negative terms such as "This product does not meet expectations at all and the quality is extremely poor," and assign a sentiment score close to -1. If the comment is more objective, such as "This phone has both advantages and disadvantages," the system will determine that the sentiment is neutral, with a sentiment intensity close to 0.

[0046] In voice data analysis, the system determines emotion based on recorded user voice interactions, such as voice assistant queries, customer service calls, or voice messages, by analyzing acoustic features such as pitch, rate, intonation, and energy distribution. For example, a user speaking quickly, with a rising pitch and significant high-frequency components, may indicate excitement or anger, and the system will assign a higher positive or negative emotion score. A low, slow voice may indicate frustration or disappointment, and the emotion score will tend to be negative. A steady voice rhythm and intonation indicate neutral emotion, with an emotion score close to 0.

[0047] For image and video data, the system primarily relies on facial expression detection, posture recognition, and scene analysis to determine emotion. If a user displays characteristics such as a smile, wide eyes, or raised eyebrows while watching an ad, the system identifies this as positive emotion and assigns a score close to 1. Frowning or a drooping mouth may indicate dissatisfaction or disgust, resulting in a lower score. The system also analyzes user viewing behavior, such as whether they watch the video in its entirety, fast-forward, or close ads, and combines this with visual sentiment analysis to adjust the final emotion intensity value.

[0048] For example, if a user, while watching a brand ad, not only comments "This product is great," but also uses a voice assistant to ask for product details and watches the entire ad video with a smile on their face, the system will combine the sentiment scores of the text, voice, and image to assign a higher positive sentiment value. If another user quickly skips the ad mid-play and leaves a comment in the comments section saying "The ad is too long, I'm not interested," with an impatient tone in their voice interaction, the system will combine these data and assign a lower sentiment intensity value.

[0049] Through this cross-modal sentiment analysis method, the system can more accurately identify users' emotional reactions to advertisements or content, and further optimize personalized recommendations and marketing strategies.

[0050] is the time decay coefficient, and the recommended value is between [0.01, 0.1]. A larger value makes the system pay more attention to recent interactions, while a smaller value makes historical data affect for a longer time.

[0051] It is the emotion nonlinear amplification factor, and the recommended value is between [1,3]. A higher value makes the impact of emotion changes more significant.

[0052] For this user in The frequency of modal interactions in interactions. For example, within a certain time window, a user may search for a keyword, watch a video of a specific brand, or like a certain type of advertisement multiple times. The higher the frequency, the greater the user's reliance on the modality.

[0053] is the maximum interaction frequency of all modes, used for normalization; Indicates that the user The number of interactions with a specific modality in a given interaction. This value is used to measure the degree of user dependence on different modal contents. The system counts the user's specific interaction behaviors in each modality (text, voice, image, video) and calculates their frequency (for example, the last 7 or 30 days). Specifically, the interaction frequency of the text modality is calculated by the number of keywords entered by the user into the search engine, the number of related comments posted on social media, the number of times the article is read, etc. The interaction frequency of the voice modality refers to the number of times the user uses voice search, the number of times the voice assistant is called, etc. The interaction frequency of the image modality includes the number of times the user views product images related to a certain brand, the number of times the brand posters or advertising images are clicked on social media platforms, etc. The interaction frequency of the video modality counts the number of times the user watches a brand's advertising video, whether the complete viewing time exceeds the set value (for example, 10 seconds), etc.

[0054] The system is based on the interaction behavior of these modes in the time window The interaction frequency of each modality is calculated and used for normalization calculation. represents the highest interaction frequency among all modes, i.e. ,in is the total number of modal interactions of the current user. Normalization is done using Min-Max normalization, mapping the interaction frequency to The normalization formula is: In this way, the modality with the highest interaction frequency is normalized to 1, and the values of other modalities are between 0 and 1, allowing the system to fairly compare the interaction frequencies of different modalities and ensure that when calculating the modal weights, the contribution of each modality can be measured according to the actual user interaction situation without deviation due to data scale issues.

[0055] is the modal activity adjustment coefficient, and the recommended value is between [0.5, 2]. A higher value gives more weight to the modal with frequent interactions.

[0056] is the interaction frequency impact index, and the recommended value is between [1,2].

[0057] For the current user The decision complexity in an interaction is calculated based on the user's behavior pattern in the interaction, including dwell time, number of clicks, and swipe speed. For example, a longer dwell time and multiple clicks indicate a higher user interest in the ad or content, while a quick swipe to skip may indicate low interest or negative feedback.

[0058] Indicates that the user The decision complexity in the interaction is used to measure the user's behavioral characteristics during the interaction process and reflect the user's attention to the advertisement or content and the difficulty of decision making. The system calculates based on the user's behavior pattern, mainly including the three key factors of stay time, number of clicks and sliding speed. When the system calculates the weighted sum of the above three factors, the dwell time has the highest weight, as a longer dwell time generally indicates that the user is willing to read the content in depth. The number of clicks is a secondary factor, influencing the user's intention to further interact.

[0059] To control the adjustment factor of the influence of decision complexity on weight, the recommended value is between [0.1, 1]. A higher value makes the influence of decision complexity on weight calculation more significant.

[0060] The analysis module 102 is used to generate target user portrait data based on the key marketing element data, combined with external market data, social media trend data and industry report data, using a cluster analysis method, and generate competitor marketing strategy analysis data based on competitor data mining technology.

[0061] The primary function of analysis module 102 is to segment target user groups and conduct in-depth analysis of competing marketing strategies based on the key marketing element data extracted by collection module 101, combined with external market data, social media trends, and industry reports. This supports optimization module 103 in developing more precise marketing plans. This module utilizes a variety of data processing and analysis techniques to structure input data, identify patterns, and mine trends, generating high-value marketing insights.

[0062] First, the analysis module 102 receives structured data from the collection module 101, which contains key information such as marketing objectives, brand tone, budget constraints, and market background. Simultaneously, the module collects market-related information from multiple external data sources, including industry reports, social media trends, user review data, search engine trends, and competitor marketing data. This data may come from public databases, commercial data vendors, internal enterprise data warehouses, or online crawled data sources. To ensure data consistency and availability, the system first performs data cleaning and standardization operations, including deduplication, filling missing values, normalizing numerical attributes, and performing noise filtering to improve data quality.

[0063] To generate user profiles, the analysis module 102 uses cluster analysis to group target users. First, text data is vectorized, for example using natural language processing models such as Word2Vec or BERT, to convert the text into a computable vector representation. The system then combines user behavior data (such as browsing history, purchase history, and social media interactions) to perform feature engineering, mapping user features into a high-dimensional space. Next, unsupervised learning algorithms such as K-means, DBSCAN, or GMM (Gaussian Mixture Model) are used to automatically classify user groups, identifying groups with similar interests, behavioral patterns, and spending habits. Furthermore, hierarchical clustering methods can be used to construct multi-level user segmentation at a more granular level. For example, for a promotion of electronic products, the system might identify three distinct target groups: "tech enthusiasts," "student users," and "corporate buyers," and further segment their purchasing motivations and spending power.

[0064] To analyze competing brands' marketing strategies, analysis module 102 uses data mining techniques to extract key marketing strategies from data from multiple competing brands, including advertising channels, promotional frequency, content style, and market feedback. This analysis involves various data processing techniques, such as using natural language processing to conduct sentiment analysis on competing brands' advertising copy to assess positive or negative consumer reactions. Social media data analysis can combine topic modeling (LDA, BERT Topic) with web crawling techniques to automatically extract competing brands' marketing keywords and related interactions on different platforms. Furthermore, this module can utilize computer vision technology to analyze competing brands' advertising materials. For example, image recognition techniques (ResNet, YOLO) can be used to detect brand logos, product designs, color schemes, and other information in competing brands' ads to identify their brand positioning strategies.

[0065] During the analysis process, the system can also utilize time series forecasting models (such as ARIMA, LSTM, or Prophet) to study market trends. For example, the system can analyze an industry's marketing investment over different quarters, as well as fluctuations in consumer purchasing intentions during specific time periods, to predict future market trends. Combining this information, the analysis module 102 generates a competitive product strategy analysis report and transmits the target user profile data and competitive product marketing strategy analysis data to the optimization module 103 to further optimize media strategies and marketing delivery plans.

[0066] Ultimately, the target user profile data output by analysis module 102 includes detailed information such as the user's age, gender, interests, behavioral patterns, and purchasing intentions, as well as their market segment, potential spending power, and brand loyalty. Competitive marketing strategy analysis data includes information such as the competitor's advertising platform, market share, promotional content style, and consumer reviews. By cross-referencing this data, the system can identify marketing opportunities in the current market and provide strong support for subsequent strategy optimization and content creation.

[0067] Furthermore, the analysis module includes a contextual awareness unit, which is used to construct a dynamic evolution model of user interest preferences based on the user's social media interaction content, search history and real-time environmental data. The dynamic evolution model of user interest preferences combines the user's recent behavior patterns, device usage habits and geographic location data to identify the user's marketing needs in a specific time period or scenario, and dynamically adjust the target user portrait data, so that the marketing strategy can adapt to the user's behavior changes and fluctuations in the market environment in real time.

[0068] The context-aware unit within the analytics module processes users' social media interactions, search history, and real-time contextual data to build a dynamic model of user interests and preferences. This allows accurate identification of user marketing needs and dynamic adjustments to target user profiles based on market changes. The core of this unit is integrating short- and long-term user behavior information, analyzing interaction patterns across time, devices, and scenarios to ensure marketing strategies adapt in real time to individual user needs.

[0069] During the operation of this unit, users' social media interactions are first analyzed, including text comments, likes, forwarded content, and brands or topics they follow. Natural language processing technology is used to identify sentiment, interests, and potential needs in the text. For example, if a user frequently follows fitness-related topics on social media, the system will adjust their recent interest tags to "healthy lifestyle" or "potential buyer of sports products." The frequency, duration, and update rate of social media interactions are also analyzed to determine the stability and weight of this interest.

[0070] The user's search history provides another important source of data. This data includes the user's search behavior on e-commerce platforms, search engines, and applications, and is combined with click behavior to analyze the user's attention to different brands, product categories, and marketing themes. For example, when a user searches for a specific brand of electronic products multiple times in a short period of time and browses multiple related review articles or purchase pages, the system will infer that the user may be in the decision-making stage and adjust their interest model to make them more inclined to this category of products in future marketing content recommendations. In addition, the temporal information of the search history is used as a time decay factor to ensure that newer search behaviors have a greater impact on the user profile, while the influence of older historical search records gradually weakens.

[0071] Real-time environmental data also plays a key role in the dynamic evolution of user interests and preferences. These data include the user's geographic location, device usage, weather conditions, and frequency of social activities. For example, when a user enters a shopping mall or airport, the system can adjust its marketing recommendation strategy based on the business environment of the place. If the user has recently appeared frequently in high-end shopping malls or brand stores, it can be inferred that the user has a high spending power and is suitable for pushing advertisements of high-end brands. In addition, device usage habits also affect the construction of user portraits. The system infers the user's preferences for different media content by identifying the types of devices commonly used by users, such as mobile phones, tablets, or smart speakers. For example, users who use smart speakers more frequently may be more inclined to accept voice advertisements, while users who use high-end smartphones may pay more attention to high-definition visual advertisements.

[0072] The dynamic evolution model of user interest preferences combines users' recent behavior patterns, device usage habits, and geolocation data to accurately model users at multiple levels and dimensions. Recent behavior patterns are primarily based on short-term searches, social interactions, and purchase records. Machine learning methods are used to calculate interest shift trends. For example, if a user shifts their focus from travel products to fitness equipment in a short period of time, the system will reduce the weight of the former and prioritize fitness-related ads. Device usage habits identify users' daily media consumption habits by tracking their application usage, the frequency of smart device network connections, and multi-screen interaction patterns over a long period of time, ensuring that advertising push matches their content consumption patterns. Geolocation data is not only used for scene recognition but can also be combined with users' travel habits to infer their marketing needs during different time periods. For example, if a user is in the office at a specific time each week, career-related marketing content can be pushed during that time, while entertainment or travel ads can be prioritized on weekends.

[0073] Throughout the system's operation, the context-aware unit continuously monitors changes in user behavior and dynamically adjusts target user profile data, enabling marketing strategies to adapt in real time to market fluctuations. For example, if a user's interest in a particular brand decreases, the system reduces the exposure weight of related marketing content to avoid ineffective push notifications while simultaneously enhancing recommendations for content more relevant to their current behavior. This model not only improves marketing precision but also reduces user perception of ad push distractions, thereby enhancing the overall conversion rate of marketing strategies and user experience.

[0074] Furthermore, the analysis module includes a competitor influence factor evaluation unit, which is used to measure the relative influence of the marketing strategies of multiple competitors based on target user portrait data. The relative influence calculation process combines the advertising frequency, user interaction rate, brand attention change trend and marketing content style matching of competitors to identify the target users' sensitivity to different competitors, and adjust the competitor marketing strategy analysis data accordingly, so that the system can accurately identify the advantages of competitors.

[0075] The Competitive Impact Factor Assessment Unit within the Analysis Module is responsible for measuring the relative influence of multiple competing products' marketing strategies and accurately identifying competitive advantages based on target user profile data, thereby optimizing marketing strategies. In practice, this unit integrates multi-dimensional data sources, including competitor ad frequency, user engagement rates, brand awareness trends, and marketing content style matching, to determine target users' sensitivity to different competing products and dynamically adjust competitive marketing strategy analysis data.

[0076] The frequency of competitor advertising is a key dimension of this unit's analysis. This data is sourced from ad monitoring platforms, public social media information, and search engine advertising records. The system first calculates the frequency of competitor exposure on major advertising platforms, including the number of times search engine bidding ads appear, the push frequency of social media feed ads, and the placement of pre-roll ads on video platforms. This data is normalized and compared with the target brand's advertising to assess the competitor's share of user visibility. If the competitor's advertising exposure is significantly higher than that of the target brand, it indicates that the brand may face significant market competition. The system will adjust the target brand's advertising strategy accordingly to increase brand exposure.

[0077] User interaction rate is another key metric used to measure the effectiveness of competitor marketing campaigns among target user groups. The system calculates the actual user engagement with competitor content by capturing user comments, likes, reposts, favorites, and other behaviors on social media platforms, brand official websites, and third-party e-commerce websites. Competitors with high interaction rates often mean that users have shown interest in their marketing content and may convert this into purchasing behavior in the future. The system conducts sentiment analysis on user comments based on natural language processing technology, distinguishing between positive and negative comments, in order to more accurately measure the market acceptance of competitors. For example, if a competitor's advertising volume is large but there are many negative comments, the system may reduce the influence weight of the competitor in the target user portrait, thereby avoiding misleading marketing strategies.

[0078] The changing trend in brand attention is a key parameter used by the system to assess the market influence of competing products. This metric is calculated by monitoring the changes in the popularity of competing products on social media, news reports, search engine trends, and industry forums over a period of time. The system first obtains the search index of competing brands on major platforms and, combined with the discussion volume on social media, uses time series analysis to extract the fluctuation trend of brand attention. If a competitor's recent search volume and social media discussion volume increase sharply, it indicates that its market influence is growing and it may have launched new marketing campaigns or product upgrades. The system will adjust the target brand's marketing strategy based on this trend to address the market changes brought about by the competitor. In addition, the system will analyze the drivers of this increase in attention, such as whether it is caused by promotional activities, product innovation, or spokesperson marketing, so as to adjust the marketing plan accordingly.

[0079] The marketing content style matching is used to evaluate the degree of fit between the advertising content of competitors and the preferences of target users. This analysis involves the advertising copy, video style, color scheme and brand tone of competitors. The system uses computer vision and natural language processing technology to extract text features, image elements and video themes from the advertising content of competitors, and compares them with the target user portrait data. For example, if a competitor adopts a youthful visual style on social media, and the target users have a high acceptance of this style, the system will identify the competitor's advantages in visual marketing and recommend that the target brand adopt a similar style for content optimization. In addition, the system will also collect statistics on keywords, color preferences, and music types in competitor advertisements, and calculate the attractiveness of the content to users based on historical user interaction data, and finally derive a marketing content style matching score.

[0080] After completing the calculation of the relative influence of the competitor's marketing strategy, the system will integrate various analytical data, grade the target users' sensitivity to different competitors, and dynamically adjust the competitor's marketing strategy analysis data. If a competitor has a high influence in multiple dimensions, the system may recommend that the target brand adjust its own advertising strategy, such as increasing advertising exposure, optimizing content style, or strengthening interactive mechanisms to improve market competitiveness. On the contrary, if a competitor's market performance is poor, the system will reduce the impact of the competitor on the target user's marketing strategy to ensure that marketing resources are concentrated in more competitive market areas. In this way, the competitor influence factor evaluation unit can not only accurately identify the market advantages of competitors, but also provide brands with data-driven optimization strategies to ensure that marketing plans maintain optimal results in a competitive environment.

[0081] The optimization module 103 is used to calculate the delivery weight data of different media channels through a reinforcement learning model based on target user portrait data, competitor marketing strategy analysis data, enterprise budget data and historical delivery data; and generate media strategy plan data based on the delivery weight data.

[0082] The main function of the optimization module 103 is to use a reinforcement learning model to calculate the distribution weights of different media channels based on target user profile data, competitive product marketing strategy analysis data, corporate budget data, and historical distribution data. Based on the calculation results, it generates media strategy plan data to optimize marketing distribution effects and improve return on investment. The implementation of this module requires multiple steps, including data integration, feature extraction, distribution strategy modeling, reinforcement learning optimization, and plan generation.

[0083] During the operation of optimization module 103, it first integrates data from analysis module 102, including target user profile data and competitive marketing strategy analysis data. Target user profile data describes the characteristics of different user groups, such as age, gender, interests, purchase history, and social media interaction frequency. Competitive marketing strategy analysis data includes the proportion of competing brands' advertising on various media channels, advertising content style, and user feedback. This data, along with the company's own budget constraints, past advertising history, and corresponding market performance data, is input into optimization module 103 to establish a data-driven advertising optimization model.

[0084] After the data integration is completed, the optimization module 103 extracts features from various types of data and constructs feature vectors for decision-making. For the target user portrait, the system will normalize the user's social media behavior and purchase history, extract behavioral features, such as click-through rate, conversion rate, user retention rate and other indicators, and combine the cluster analysis results to assign a unique feature code to each user group. For competitor marketing strategy data, the system will calculate key indicators such as competitor advertising exposure rate, interaction rate, user feedback sentiment analysis, etc. to infer its market competitiveness. Enterprise budget data is used to set optimization constraints to ensure that the system does not exceed the established budget when formulating delivery strategies. Historical delivery data is mainly used to train models. The system will extract information such as the delivery channels, investment funds, actual conversion effects, etc. of past marketing activities to evaluate the long-term effectiveness of different media channels.

[0085] During the delivery strategy modeling phase, optimization module 103 employs reinforcement learning technology to construct a delivery strategy optimization model based on a Markov decision process. The core of reinforcement learning is to find the optimal delivery plan through trial and error. This method optimizes through three components: state, action, and reward. The system's state space consists of user profiles, competitive product analysis, budget constraints, and historical delivery performance, while the action space represents the delivery ratio adjustment plan for different media channels. The reward function is designed based on marketing effectiveness, such as ROI (return on investment), brand exposure, and user engagement.

[0086] During the optimization process, the system uses deep reinforcement learning algorithms, such as Deep Q-Learning (DQN) and policy gradients, to train the model. Initially, the model simulates different delivery scenarios, tries different media combinations, and calculates rewards based on historical data. As training progresses, the system continuously adjusts the media delivery ratio to achieve the most optimal marketing delivery plan. For example, if a media channel has a high CTR and CVR but a low cost per acquisition (CPA), the system will automatically increase the delivery weight of that channel. If a channel generates high exposure but a low conversion rate, the system may reduce the delivery budget for that channel.

[0087] After training is complete, optimization module 103 generates media strategy data, including the optimal delivery ratio for each media channel, budget allocation plan, and expected delivery results. This plan not only considers the current market environment but also incorporates competing product strategies and user behavior trends to maximize advertising effectiveness. Furthermore, this module allows for dynamic adjustment of delivery strategies. That is, during the delivery process, if market feedback deviates significantly from the predicted results, the system will reassess the delivery strategy and iteratively optimize it using feedback data from generation module 105.

[0088] Finally, the media strategy plan data output by the optimization module 103 will be transmitted to the creation module 104 to guide the creation of marketing content, so that the advertising content can adapt to the delivery needs of different media platforms, thereby improving the overall marketing effect.

[0089] The optimization module uses a dynamic delivery optimization method based on reinforcement learning, calculating the delivery weights of different media channels through a Markov decision process. The core of this method is to use historical delivery data and market feedback data to make adaptive adjustments, so that advertising can maximize revenue on different media platforms while effectively controlling advertising budgets and avoiding unnecessary resource waste. The system calculates the optimal media delivery weights using formula (2) and continuously updates them during each delivery round, allowing the optimization process to dynamically adapt to changes in the market environment.

[0090] The optimization module uses a dynamic delivery optimization method based on reinforcement learning to calculate the delivery weight data of different media channels. The optimization method is based on the Markov decision process and calculates the optimal media delivery strategy through the following formula 2: in, For media channels In the delivery round The core basis for weight adjustment is the revenue of the media channel in the previous round, user click-through rate, conversion rate and delivery cost and other factors.

[0091] For media channels In the previous round The media delivery weight at that time is set by the system in the initial training phase and is continuously updated iteratively during the subsequent optimization process.

[0092] During the initial training phase of the optimization module, Set delivery weight Initial values are set based on comprehensive performance indicators related to the media in historical advertising data, using predefined calculation rules. Specifically, the system first extracts historical performance data for each media channel within a specific time window (e.g., the past 30 days) from an existing marketing database. This data includes, but is not limited to, impressions, click-through rates (CTRs), conversion rates (CVRs), average cost per acquisition (CPA), and total budget invested. This data is typically obtained by companies from previous advertising platforms (such as Baidu Information Feed, Tencent Guangdiantong, and Douyin platforms) and is structured and ready for analysis.

[0093] During the initial phase of the optimization module, the system sets an initial delivery weight for each media channel, which serves as the starting input for the reinforcement learning model. To ensure this initial strategy is data-driven and engineering-feasible, the system uses the last 30 days of historical delivery data for each channel to perform a combined score based on click-through rate, conversion rate, and unit cost per conversion. Click-through rate reflects user engagement with the ad, conversion rate reflects user purchase intent, and unit cost per conversion measures the platform's financial investment in generating conversions. These three metrics are combined to evaluate the effectiveness of the channel's delivery.

[0094] To facilitate calculation and control, these three metrics were first normalized to values between 0 and 1. This normalization method involved linearly stretching the maximum and minimum values of the corresponding metrics across all channels over the past 30 days, thereby mapping the data for each channel to the same range. The unit conversion cost item was inversely proportional during normalization, giving channels with lower costs higher scores.

[0095] The system then weights these three metrics using default weighting parameters to create an initial score. The default weightings are: 30% for click-through rate, 50% for conversion rate, and 20% for cost-effectiveness. These weights can be adjusted by the platform administrator in the system configuration or automatically based on pre-set templates for different industries (for example, brand promotions prioritize exposure, while performance-based advertising prioritizes conversions). Based on this combined score, the system generates a vector of initial delivery weights for each channel and normalizes them to sum to 1 to ensure that the initial delivery strategy meets budget allocation constraints.

[0096] The system incorporates a penalty mechanism into the weighted synthesis process: When a conversion rate falls below a set threshold (e.g., below 80% of the average across all channels), or the unit conversion cost exceeds 120% of the average, the system automatically multiplies its score by a discount factor (recommended between 0.7 and 0.85) to reduce the channel's weight in the initial weight distribution. For example, if a channel has a click-through rate of 0.6 and a conversion rate of 0.3, but a high cost, the system will reduce its score by more than 20% after applying the aforementioned logic to prevent its high click-through rate from misleading resource allocation.

[0097] To prevent channels from being overlooked due to insufficient historical data, the system implements a minimum weighting mechanism. For channels with scores below 1% of the average, the system automatically sets their weight to a minimum value (recommended to be 1% of the total budget) to ensure that all channels have room for exploration and avoid cold start bias.

[0098] This initial delivery weight setting process has been verified in internal A / B experiments: compared with traditional average distribution strategies and strategies based solely on click-through rate, it increased the average conversion rate by approximately 12.3% in the first five rounds of optimization, shortened the model convergence time by approximately 15.6%, and effectively improved the efficiency of advertising budget utilization and the quality of platform strategies.

[0099] The final set initial delivery weight will be used as the delivery round The initial values at the time of the optimization are input into the optimization model, and are then iteratively updated during subsequent deployments using the feedback adjustment mechanism within the reinforcement learning algorithm. Therefore, the initial weights are not only the result of static settings but also reflect historical performance and provide an effective initial state for the optimization model, thereby ensuring the stability and convergence of model training.

[0100] is the learning rate; Determines the step size of the delivery weight adjustment, which affects the optimization speed and stability. If the value is too large, the system will adjust too quickly, which may lead to unstable convergence. If the value is too small, the optimization process will converge more slowly, affecting the actual delivery results. The recommended value range is [0.01, 0.1], and is usually adjusted based on experimental data to ensure that the delivery strategy can quickly adapt to market changes without causing drastic fluctuations.

[0101] For media channels In the delivery round The immediate revenue at that time is calculated based on market feedback data; this revenue is usually the weighted sum of factors such as ad click-through rate, conversion rate, brand exposure, and user interaction rate.

[0102] For all media channels in the delivery round The average revenue on the platform is used to measure the performance of the current media channel relative to the overall delivery environment.

[0103] The standard deviation of revenue is used to measure the revenue fluctuation of all media channels. A larger standard deviation indicates that the delivery effects of different media channels vary greatly. The optimization process requires more sensitive adjustment of weights to ensure a reasonable allocation of budgets. When calculating the standard deviation of revenue, it is first necessary to determine the revenue data of each media channel in the current delivery round. Determined by market feedback data, the main sources include the number of times users click on ads, the conversion rate brought by ads, changes in brand exposure, and user interaction behavior. For example, if a social media platform has a high number of ad impressions and a high click-through rate, its revenue will be relatively large; conversely, if a search engine has a low number of ad impressions or a poor conversion effect, its revenue may be lower. The system collects revenue data from all media channels and calculates the average revenue of all channels. .

[0104] After obtaining the revenue data for each media channel, the system calculates the deviation between each medium's revenue and the average revenue, measuring the gap between each channel's revenue and overall market performance. For example, if a medium's revenue is much higher than the average revenue, it means that the advertising effect of the medium is good and it is worth further increasing the advertising budget; if a medium's revenue is much lower than the average revenue, it means that the channel may not be an ideal advertising platform and the budget allocation should be reduced or the advertising content should be optimized. The system calculates the standard deviation based on the deviation of all media revenue data to determine the volatility of the overall market. If the revenue data of all media channels vary slightly, the standard deviation is low, indicating that the market performance is relatively balanced, and the system will adopt a more stable advertising strategy; if the revenue data vary greatly and the standard deviation is high, the system will more sensitively adjust the advertising weight, tilting the advertising budget towards high-yield channels to improve the overall advertising return on investment.

[0105] For media channels In the delivery round Click-through rate when For media channels in the delivery round Time The conversion rate represents the probability that a user will convert after clicking an ad. The click-through rate measures user interest in the ad, while the conversion rate measures the actual commercial value generated by the ad. The system derives these two values through user behavior analysis and market feedback data, and makes adjustments based on historical delivery data. For example, if a media channel has a high click-through rate but a low conversion rate, this may mean that the platform's ads are highly appealing, but users have low purchasing intent. The system may then incorporate conversion optimization strategies, such as changing ad content and optimizing the target user group, to improve the final conversion rate.

[0106] For media channels In the delivery round Unit delivery cost is used to control advertising budget; To prevent small positive numbers with zero denominator; It is the time decay adjustment factor; it controls the influence of time factors in weight adjustment. The recommended value range is [0.5, 2]. Larger values indicate higher sensitivity to time factors and greater influence of recent delivery data.

[0107] is the time decay coefficient; the recommended value range is [0.01, 0.1]. When the value is smaller, the system will consider historical data more, while when it is larger, the optimization will be mainly based on the latest delivery results.

[0108] For media channels In the delivery round The time of the most recent ad serving on ; The current delivery time.

[0109] This optimization method dynamically adjusts the weighting of different media channels based on market feedback, making advertising budget allocation more rational and ensuring optimal advertising results across different media platforms. Furthermore, the adaptive adjustment mechanism of reinforcement learning enables the system to continuously adapt to market changes, improving the flexibility and effectiveness of marketing strategies and ultimately boosting the overall return on investment for advertising.

[0110] The creation module 104 is used to create communication theme data and marketing content data adapted to different media based on the media strategy plan data and in combination with the characteristics of the media channel through generative artificial intelligence technology.

[0111] The creation module 104 primarily uses generative AI technology to create communication theme data and marketing content tailored to different media, based on the media strategy data generated by the optimization module 103 and the characteristics of various media channels. This module ensures that marketing content is both consistent with the brand's tone and adaptively optimized based on the characteristics of different delivery channels, thereby improving advertising effectiveness and user engagement.

[0112] During the execution of creation module 104, the system first receives the media strategy data provided by optimization module 103. This data includes the distribution weights of various media channels, the interests and preferences of the target user group, the results of the competitive marketing strategy analysis, and the company's budget allocation. To ensure that the generated marketing content matches the strategy, the module parses the media strategy data and extracts key information, such as the target audience's age group, social media usage habits, visual preferences, language style, and content format requirements of different media platforms. For example, short video platforms (such as TikTok and Kuaishou) are suitable for short video content with high interactivity and entertainment, while search engine advertising (such as Baidu promotion) requires text content optimized for keywords to increase user click-through rate. In addition, television advertising tends to express brand stories, while outdoor advertising requires a design style with strong visual impact and concise and clear information.

[0113] After parsing the media strategy data, creation module 104 invokes a generative AI model to automatically generate communication theme data suitable for different media. This process involves the generation of multimodal content, including text, images, and video. For text content, this module utilizes natural language generation (NLG) technologies, such as Baidu's ERNIE-GEN, which is based on a pre-trained Chinese language model, to combine brand tone, market trends, and user profiles to generate advertising copy, product descriptions, social media posts, or press releases. For example, if the target audience prefers concise and intuitive newsfeed ads, the system automatically generates short, concise ad headlines and introductory copy. For promotional articles requiring search engine optimization (SEO), long-form content with high information density is generated to improve search rankings and user reading experience.

[0114] For image content, creation module 104 utilizes deep learning-based image generation techniques, such as generative adversarial networks (GANs), to automatically generate visual assets such as posters, ad banners, and social media covers based on brand style and market trends. The system adjusts image resolution, color scheme, and layout based on the requirements of different media channels. For example, on the Xiaohongshu platform, the system may favor high-value images that appeal to the aesthetic tastes of young female users, while ads on Bilibili may adopt a more anime-style visual design. Furthermore, the system can automatically detect and optimize text within images to ensure compliance with text ratio requirements for different platforms, such as the text coverage limits in Douyin ads.

[0115] For video content, creation module 104 utilizes text-to-video (T2V) technology and computer vision models to automatically generate video ads that meet delivery requirements. The system first generates ad scripts based on natural language generation (NLG) technology, including video narration, subtitles, and shot descriptions. It then utilizes a video generation model, such as Baidu Wenxin Yige, to automatically generate short video ads that align with the brand's tone. If the ad requires a real person, the system can also incorporate deep synthesis technology to generate a virtual spokesperson, making the brand ad more interactive and personalized. Furthermore, this module supports optimization processes such as automatic video editing, subtitle generation, and background music matching to ensure that the final video meets the requirements of different media platforms. For example, short video platforms (such as Douyin and Kuaishou) generally require content to be between 15 and 60 seconds long. The system automatically breaks long-form ads into multiple shorter segments and customizes different versions for different audiences to increase viewing completion rates and user engagement.

[0116] After marketing content is generated, the creation module 104 performs a content quality assessment to ensure that the ad aligns with the brand's tone, marketing objectives, and media platform requirements. This assessment process combines artificial intelligence with a rules engine for automated review, including grammar checking, brand consistency testing, and visual aesthetic scoring. For example, text content is checked for grammatical correctness using natural language understanding models such as BERT, and its emotional alignment with the brand image is analyzed. Image content is checked for brand consistency using deep learning models such as CNN to avoid the use of low-quality or inappropriate image elements. Video content is analyzed using computer vision algorithms for key metrics such as color, composition, and caption accuracy, ensuring optimal quality in the final ad. Furthermore, this module can adjust the content style based on historical user behavior data. For example, ads on e-commerce platforms (such as Taobao and JD.com) may focus more on promotional information, while ads on social media platforms (such as Weibo and WeChat Moments) may emphasize brand storytelling and user resonance.

[0117] Finally, the communication theme data and marketing content data generated by creation module 104 are transmitted to generation module 105 for further delivery effect prediction and optimization. This process ensures that marketing content not only meets brand needs but also adapts to different media channels, maximizing advertising effectiveness while increasing user engagement and marketing conversion rates.

[0118] Furthermore, the creation module includes an adaptive content style adjustment unit, which is used to perform style adaptive optimization on the generated communication theme data and marketing content data based on the media strategy plan data and the personalized characteristics of the target user portrait data. The style adaptive optimization dynamically adjusts the content expression, tone, color scheme and visual layout based on the user's historical interaction behavior, cultural preferences, language habits and the specific format requirements of the media channel.

[0119] The Adaptive Content Style Adjustment Unit within the creation module is used to optimize advertising content, ensuring it not only aligns with the brand's tone but also adapts to the specific needs of different media channels and target user groups. During operation, this unit first receives media strategy plan data, including information such as the delivery platform's content format requirements, ad display rules, and user interaction preferences. Furthermore, the system acquires target user profile data, which includes the user's historical interaction behavior, cultural background, language habits, visual preferences, and personalized aesthetic tendencies. This ensures that the generated content best matches user needs, improving ad conversion rates and engagement effectiveness.

[0120] During the adaptive optimization of content style, the system will dynamically adjust to the cultural preferences and language habits of different user groups. For language style optimization, the system uses natural language processing technology to analyze users' historical interaction texts, including social media comments, search records, online shopping reviews, etc., to extract users' commonly used expressions, sentence structures and emotional tendencies. For example, for users who prefer formal expressions, the advertising copy generated by the system may use a precise and objective description method, while for users who like a humorous style, the system may add colloquial expressions and infectious copy. For multilingual users, the system can also combine the user's native language and reading habits to select the most suitable language for content delivery, while adjusting the copy structure to adapt to the semantic habits and expressions in different language and cultural backgrounds.

[0121] In terms of visual style optimization, the system uses computer vision to analyze users' social media pictures, likes, and viewing history to extract the user's preferred color style, composition method, and design elements. For example, if a user often browses minimalist design works on social media, the system will adjust the advertising material to make it conform to the minimalist visual style, including reducing complex backgrounds, optimizing color contrast, etc. For users who prefer highly saturated colors, the system will increase the color vividness of the advertising image to enhance the visual impact. For different media channels, the system can also adaptively adjust the visual layout of the advertisement. For example, when placing advertisements on short video platforms, the system may add dynamic text effects to attract user attention, while in social media information streams, it may optimize the layout of the image and text combination to make the information transmission more intuitive.

[0122] In response to the specific format requirements of media channels, the system will automatically adjust the structure of advertising content to make it conform to the optimal display method of the platform. For example, for short video platforms, the system will optimize the video length, subtitle display method, picture rhythm, etc. to ensure that users can quickly understand the advertising information within a limited time. For advertising on news websites, the system will enhance the readability of text information and optimize the attractiveness of the title to increase click-through rate. For search engine advertising, the system will prioritize adjusting the density and distribution of keywords to improve search rankings and ad exposure. In addition, the system can also adaptively adjust advertising content based on user interaction patterns on different platforms, such as adding voting options and interactive Q&A on social media to increase user engagement.

[0123] During the overall content optimization process, the system can also adjust the content display method based on the user's device usage habits. For example, for users who are accustomed to using mobile devices, the system will optimize the vertical screen display effect of the advertisement, adjust the layout to adapt it to the small-screen reading experience, and reduce loading time to improve fluency. For PC users, the system may add richer visual elements, such as multi-column layouts, interactive animations, etc., to enhance the browsing experience. In addition, for smart TVs or outdoor large-screen advertisements, the system can also adjust the content proportions, font size, and dynamic special effects to ensure that it maintains good visual readability when viewed from a distance.

[0124] Furthermore, the creation module includes a content context adaptation unit, which is used to dynamically adjust the content structure and presentation form of the communication theme data and marketing content data based on the user's current usage scenario, device characteristics and environmental variables. The adjustment process combines the user's interaction mode, the screen size of the terminal device, the network bandwidth status and the ambient light intensity to optimize the readability, loading speed and visual effect of the advertising content.

[0125] The Content Context Adaptation Unit within the Creation Module dynamically adjusts the presentation of marketing content based on user usage scenarios, device characteristics, and environmental variables, ensuring the optimal advertising experience in every situation. By comprehensively analyzing user interaction patterns, device characteristics, network bandwidth, and ambient lighting conditions, this unit optimizes the structure, layout, color, text format, and media elements of advertising content in real time, improving readability, loading speed, and visual quality of marketing content, thereby increasing user engagement and engagement.

[0126] During content adaptation, the system first detects the user's current interaction mode to determine whether they are actively browsing, scrolling quickly, lingering for an extended period, or multitasking. For active browsing scenarios, the system enhances the information density of the ad content, providing detailed product descriptions and multi-layered interactive content, such as embedded video demos, dynamic infographics, or 360-degree product displays, to enhance user understanding and interest. If the user is detected scrolling quickly, the system prioritizes key marketing information, such as the brand logo, promotional discounts, and short slogans, and simplifies visual elements to improve the ad's instantaneous recognition. For users who linger for extended periods, the system can appropriately add additional content, such as recommending related products and providing user reviews, to extend user engagement and increase brand influence. For multitasking scenarios, the system can deliver ads tailored to short attention spans, such as split-screen small windows or audio ads, to ensure that users receive the ad message even when distracted.

[0127] In addition to analyzing interaction patterns, the system also adjusts content based on the characteristics of the user's terminal device. For example, on large-screen devices, such as tablets or smart TVs, advertising content can adopt a multi-column layout to enhance visual hierarchy and utilize high-resolution images and high-definition video materials to increase visual impact. On small-screen devices, such as smartphones, the system optimizes the layout of advertising content, uses a single-column flow layout, and adjusts font size and button spacing to accommodate touch operations, ensuring that users can interact conveniently. In addition, on wearable devices, such as smart watches, the system provides ultra-simplified advertising content in the form of simple icons, short sentence push and voice broadcast to adapt to small-screen display and quick browsing needs.

[0128] In terms of network status analysis, the system will detect the user's network bandwidth and stability in real time to determine how to load advertising content. If it is detected that the user is in a high-speed Wi-Fi environment, the system can prioritize loading high-definition videos, high-resolution images, and complex animations to provide the best visual experience. In mobile networks or low-speed network environments, the system will automatically adjust the content loading strategy, such as using low-resolution images, reducing animation effects, or adopting a segmented loading mode to ensure that the advertising content can be displayed smoothly and avoid user loss due to slow loading. At the same time, the system can also detect the user's traffic usage and provide lightweight advertising in low-traffic mode, such as pushing only text and static images, to reduce user traffic consumption and improve advertising acceptance.

[0129] Detection of ambient light intensity is also an important factor in optimizing ad readability. Based on the light sensor data of the terminal device, the system can determine the user's current lighting environment and adjust the color mode and brightness contrast of the advertisement. In strong light environments, such as outdoors or in high-brightness screen environments, the system will enhance the contrast of the advertising content, improve font clarity, and avoid reading difficulties caused by excessive light. In weak light environments, such as at night or in low light conditions, the system can automatically switch to dark mode to reduce screen glare, improve user reading comfort, and reduce visual fatigue. In addition, if the smart device supports adaptive color temperature, the system can also adjust the color saturation of the advertisement in combination with the user's night mode or eye protection mode settings to adapt to the user's visual habits.

[0130] In order to improve the user experience, the system also combines AI intelligent recommendation algorithms to predict user usage scenarios based on historical interaction data. For example, if a user frequently uses a mobile device to browse news in the morning, the system can push advertising models that are adapted to news reading during this time period, such as short video ads embedded in the information flow or swiping card ads. If the user often watches long videos on a smart TV at night, the system can adjust the advertising display strategy, such as playing high-quality brand promotional videos before the video starts, or embedding interactive ads in the pause interface to improve the acceptability and conversion rate of the ads. For specific environmental changes, such as when a user enters a shopping mall, airport, or gym, the system can combine geographic location data to push personalized ads that are in line with the current scenario, such as pushing brand discount information in a shopping mall, recommending travel-related products in an airport environment, and pushing healthy food or sports equipment ads in the gym to improve the relevance and effectiveness of the ads.

[0131] By comprehensively analyzing user interaction patterns, device characteristics, network conditions, and environmental variables, the content context adaptation unit ensures that advertising content consistently delivers optimal readability, loading speed, and visual quality in different contexts, increasing user acceptance and brand engagement. This system not only optimizes ad display but also adapts to the personalized needs of different users, improving overall marketing precision and effectiveness, thereby effectively enhancing brand influence and improving market competitiveness.

[0132] The generation module 105 is used to predict the delivery effect through a machine learning model based on the communication theme data, marketing content data, historical marketing data and real-time market feedback data, and optimize the media strategy plan and marketing content plan according to the prediction results.

[0133] The primary function of Generation Module 105 is to predict advertising effectiveness using machine learning models based on thematic data, marketing content data, historical marketing data, and real-time market feedback. This module then optimizes media strategies and marketing content plans based on these predictions. This module dynamically evaluates actual advertising performance and, based on historical data and real-time market feedback, adjusts future advertising strategies and content optimization plans to improve marketing efficiency and return on investment.

[0134] During the operation of generation module 105, it first receives the communication theme data and marketing content data generated by creation module 104. This data includes advertising copy, visual assets, video content, brand tone information, and content adapted for different media. Simultaneously, this module also obtains historical marketing data, which is derived from statistical information on past advertising campaigns. This data includes key indicators such as click-through rate (CTR), conversion rate (CVR), user engagement, cost per advertisement (CPA), and audience behavioral responses. Furthermore, through a real-time market feedback mechanism, the system continuously collects performance data on current advertising campaigns, including user browsing behavior, dwell time, interaction, purchasing behavior, social media comment sentiment, and market response to competing products.

[0135] After data aggregation, the generation module 105 uses machine learning models to predict advertising effectiveness. First, it uses time series analysis methods (such as deep learning models like ARIMA and LSTM) to analyze historical marketing data to identify long-term trends and cyclical patterns, such as changes in advertising performance between quarters or fluctuations in user spending behavior during specific holidays. Subsequently, based on the degree of match between audience characteristics and advertising content, it uses supervised learning models (such as XGBoost) to classify and perform regression predictions on advertising effectiveness on different media platforms. This allows it to calculate the expected performance of ads across various channels, such as whether the click-through rate on a particular social media platform is higher than the industry average, or whether the conversion rate of a particular ad type is affected by the style of the content.

[0136] To further optimize the delivery strategy, generation module 105 uses reinforcement learning technology to construct an adaptive optimization model. This model continuously optimizes ad delivery parameters through trial and error, such as budget allocation, media channel weighting, and the frequency of use of different creatives. The reinforcement learning reward mechanism is based on key metrics of delivery effectiveness, such as ROI (return on investment), user retention, and brand exposure. The system dynamically adjusts strategies during the delivery process. For example, if a platform's ads have a high click-through rate but a low conversion rate, the system may reduce that platform's advertising budget and allocate more resources to platforms with higher conversion rates.

[0137] To optimize marketing content, the generation module 105 combines natural language processing and computer vision technologies to evaluate the market performance of advertising copy and visual materials. For example, the system can use sentiment analysis models to assess users' emotional reactions to advertising content, determining whether the ad copy has generated positive or negative feedback. Computer vision models (CNNs) can be used to analyze the color, composition, and element distribution of advertising images, comparing them with the characteristics of effective advertising materials to optimize the design style of future ads. For video ads, the system can also optimize details such as video length, subtitle placement, and content pacing based on user viewing behavior, such as video completion rate, skip rate, and pause rate, to improve viewability and conversion rates.

[0138] In the final stage of campaign optimization, the generation module 105 adjusts the media strategy and marketing content plan based on the forecast results and feeds the optimized plan back to the optimization module 103 and the creation module 104. The optimization module 103 uses the new data to adjust the media placement ratio to ensure optimal budget allocation, while the creation module 104 adjusts the marketing content based on market feedback, such as rewriting ad copy with low click-through rates or redesigning visual assets with low engagement rates. This entire process forms a closed-loop optimization system, enabling marketing plans to adapt to market changes in real time, improving overall campaign effectiveness and maximizing marketing returns.

[0139] Furthermore, the generation module includes a multi-stage delivery effect prediction unit, which uses a machine learning model to perform stage-by-stage prediction and optimization on data at different levels, wherein the machine learning model includes a delivery environment modeling submodule, a user behavior prediction submodule, and a dynamic content adjustment submodule; The delivery environment modeling submodule is used to receive historical marketing data, target user profile data, and media strategy plan data, analyze the advertising delivery characteristics of different media platforms, including exposure rate, user dwell time, and interaction mode, and generate an environment parameter vector to provide initial constraints for the delivery platform; The user behavior prediction submodule is used to receive the environmental parameter vector output by the delivery environment modeling submodule, and combine it with real-time market feedback data to predict the user's click-through rate, conversion rate and purchase intention. It uses time series analysis and deep learning models to extract the user's short-term and long-term behavior trends and output behavior prediction data; The dynamic content adjustment submodule is used to receive the behavior prediction data and adjust the text style, visual elements, and video editing strategy of the content in combination with the marketing content data generated by the delivery environment modeling submodule.

[0140] The Multi-Stage Delivery Effect Prediction Unit within the Generation Module is responsible for predicting and optimizing advertising effectiveness at various levels to improve the precision of marketing strategies and the return on investment of advertising. Based on a machine learning model and employing a multi-stage prediction approach, this unit dynamically optimizes advertising plans based on the characteristics of different media platforms, user behavior patterns, and market feedback data, achieving precise delivery and adaptive content adjustment.

[0141] During the operation of this unit, the delivery environment modeling submodule first receives historical marketing data, target user profile data, and media strategy data, and analyzes the characteristics of advertising delivery across different media platforms. Historical marketing data includes metrics such as exposure rate, click-through rate, conversion rate, user interaction, and cost-benefit ratio for different time periods and different advertising strategies. Target user profile data, derived from the collection and analysis modules, includes information such as user interests, social behavior, purchase history, and search history, enabling the system to optimize delivery strategies based on user needs. Media strategy data includes the ad placement attributes, content display rules, recommendation mechanisms, and user interaction patterns of each media channel, enabling the system to accurately identify the delivery characteristics of each media platform. For example, social media platforms typically display ads in the form of news feeds, while search engine ads rely on keyword bidding and ranking mechanisms. Based on this data, this submodule generates an environment parameter vector, a mathematical representation of the delivery platform's characteristics. This vector incorporates key parameters such as the delivery rules, user behavior patterns, and content interaction methods of each media channel, providing initial constraints for subsequent delivery strategies.

[0142] After modeling the delivery environment, the user behavior prediction submodule uses the environment parameter vector and real-time market feedback data to predict users' click-through rates, conversion rates, and purchase intent. Real-time market feedback data includes user interactions after ad delivery, such as whether users click on ads, view ad details, complete purchases, and share ad content. This submodule uses time series analysis and deep learning models to extract short-term and long-term user behavior trends and generate user behavior prediction data. Time series analysis is used to identify cyclical patterns in user behavior. For example, some users may be more likely to browse e-commerce ads on weekends and more interested in news content on weekdays. Deep learning models use algorithms such as convolutional neural networks (CNNs) or long short-term memory networks (LSTMs) to analyze long-term trends in user interests and identify potential purchase intent. For example, if a user has searched for a product multiple times over the past few days but has not yet completed a purchase, the system may predict a high conversion probability for this user and adjust the frequency of ad push and content presentation accordingly.

[0143] After generating user behavior prediction data, the dynamic content adjustment submodule uses this prediction data, combined with the marketing content data generated by the delivery environment modeling submodule, to adjust the ad's text style, visual elements, and video editing strategy to ensure that the ad content meets the specific requirements of different media platforms and attracts user interest. For example, when optimizing text style, the system adjusts the tone and expression of the ad copy based on the user's interaction history. If the user prefers formal and professional content, the system will use more precise terminology. If the user prefers a lighter, more humorous style, the system will adjust the ad content to make it more approachable. When optimizing visual elements, the system adjusts the ad's color scheme, image layout, and font style based on the user's social media browsing preferences. For example, if a user frequently likes images with highly saturated colors on social media, the system may select a more vibrant ad design style to increase visual appeal. When optimizing video editing, the system adjusts the video's tempo, subtitle style, and background music based on the user's viewing behavior. For example, if a user on a short video platform prefers to watch quickly edited videos rather than long narrative content, the system will adjust the editing rhythm of the advertising video to make it more in line with the user's viewing habits, thereby increasing the completion rate and interaction rate of the advertisement.

[0144] In practice, the system continuously adapts to market changes, optimizing advertising strategies and improving the relevance and appeal of advertising content, making marketing campaigns more precise and efficient. This multi-stage optimization approach, powered by machine learning, not only improves advertising effectiveness but also dynamically adjusts content to meet the needs of different user groups, improving overall advertising ROI and ensuring optimal dissemination across various media channels.

[0145] In the above embodiment, a multi-dimensional marketing planning agent system is provided. Correspondingly, this application also provides a multi-dimensional marketing planning agent method. Figure 2 , which is a flow chart of a method embodiment of a multi-dimensional marketing planning agent of the present application. Since this embodiment, namely the second embodiment, is basically similar to the first embodiment, the description is relatively simple. For relevant details, please refer to the partial description of the first embodiment. The method embodiment described below is merely illustrative.

[0146] A second embodiment of the present application provides a method for a multi-dimensional marketing planning agent, including: Step S201: Receive basic data including marketing objectives, market background, brand tone, and budget constraints, where the basic data includes multimodal input such as text, voice, image, and video; pre-process the basic data based on natural language processing technology and computer vision analysis to extract key marketing element data; Step S202: Based on the key marketing factor data, combined with external market data, social media trend data, and industry report data, a cluster analysis method is used to generate target user portrait data, and competitor marketing strategy analysis data is generated based on competitor data mining technology; Step S203: Based on the target user profile data, competitor marketing strategy analysis data, enterprise budget data, and historical delivery data, the reinforcement learning model is used to calculate delivery weight data for different media channels; and media strategy plan data is generated based on the delivery weight data. Step S204: Based on the media strategy data and in combination with the characteristics of the media channels, generative artificial intelligence technology is used to create communication theme data and marketing content data adapted to different media; Step S205: Based on the communication theme data, marketing content data, historical marketing data and real-time market feedback data, the delivery effect is predicted through a machine learning model, and the media strategy plan and marketing content plan are optimized according to the prediction results to improve the overall marketing effect.

[0147] Although the present application is disclosed as above with the preferred embodiments, it is not intended to limit the present application. Any person skilled in the art may make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims of the present application.

Claims

1. A multi-dimensional marketing planning agent system, characterized in that: include: The collection module is used to receive basic data including marketing objectives, market background, brand tone, and budget constraints. The basic data includes multimodal input such as text, voice, image, and video; and extract key marketing element data based on natural language processing technology and computer vision analysis; An analysis module is used to generate target user portrait data using a cluster analysis method based on the key marketing factor data, combined with external market data, social media trend data, and industry report data, and to generate competitive product marketing strategy analysis data based on competitive product data mining technology; An optimization module is used to calculate the delivery weights of different media channels using a reinforcement learning model based on target user profile data, competitive product marketing strategy analysis data, corporate budget data, and historical delivery data; and to generate media strategy plan data based on the delivery weights; A creation module is used to create communication theme data and marketing content data adapted to different media based on the media strategy plan data and in combination with media channel characteristics through generative artificial intelligence technology; A generation module is used to predict the delivery effect through a machine learning model based on the communication theme data, marketing content data, historical marketing data and real-time market feedback data, and optimize the media strategy plan and marketing content plan according to the prediction results.

2. The multi-dimensional marketing planning agent system according to claim 1, characterized in that: The collection module includes a data fusion unit, which is used to construct a cross-modal association graph based on the timestamp information and content features of the multimodal input data, and use a graph neural network to perform association analysis on data of different modalities to extract deep marketing elements of user preferences. The cross-modal association graph includes contextual relationship nodes between text, voice, image and video data, and is combined with a feature weight calculation method based on the self-attention mechanism to dynamically weight and fuse different modal information to enhance the accurate identification of target users' market behavior.

3. The multi-dimensional marketing planning agent system according to claim 2, characterized in that: The data fusion unit uses a dynamic weighted cross-modal embedding mapping method to perform feature extraction, time series analysis, and personalized optimization on input data of different modalities, and calculates the contribution of each modal feature to the target user's market behavior using the following formula 1: in, Indicates modality The dynamic weight in the current user interaction data fusion process reflects the importance of this modality in marketing data analysis; The total number of multimodal interaction records of the current user within a specific time window; For the current user The content matching degree in this interaction is obtained by calculating the similarity between the modal data and the current marketing theme, brand characteristics and historical behavior; For the current user The timestamp of the interaction, which is the time difference relative to the current moment, is used to measure the freshness of the interaction; For the current user The emotional intensity in the interaction is calculated by natural language processing models and computer vision emotion recognition technology to measure the emotional polarity in text, speech, images or videos. The numerical range is where positive values represent positive emotions, negative values represent negative emotions, and zero represents neutral emotions; is the time attenuation coefficient; is the emotion nonlinear amplification factor; For this user in frequency of modal interactions in subinteractions; is the maximum interaction frequency of all modes, used for normalization; is the modal activity adjustment coefficient; is the interaction frequency impact index; For the current user The decision complexity in this interaction is calculated based on the user's behavior pattern in this interaction, including dwell time, number of clicks, and sliding speed; It is a regulatory factor to control the impact of decision complexity on weight.

4. The multi-dimensional marketing planning agent system according to claim 1, characterized in that: The analysis module includes a contextual perception unit, which is used to construct a dynamic evolution model of user interest preferences based on the user's social media interaction content, search history and real-time environmental data. The dynamic evolution model of user interest preferences combines the user's recent behavior patterns, device usage habits and geographic location data to identify the user's marketing needs in a specific time period or scenario, and dynamically adjust the target user portrait data, so that the marketing strategy can adapt to the user's behavior changes and fluctuations in the market environment in real time.

5. The multi-dimensional marketing planning agent system according to claim 1, characterized in that: The analysis module includes a competitor influence factor evaluation unit, which is used to measure the relative influence of the marketing strategies of multiple competitors based on target user portrait data. The relative influence calculation process combines the advertising frequency, user interaction rate, brand attention change trend and marketing content style matching of competitors to identify the target users' sensitivity to different competitors, and adjust the competitor marketing strategy analysis data accordingly, so that the system can accurately identify the advantages of competitors.

6. The multi-dimensional marketing planning agent system according to claim 1, characterized in that: The optimization module uses a dynamic delivery optimization method based on reinforcement learning to calculate the delivery weight data of different media channels. The optimization method is based on the Markov decision process and calculates the optimal media delivery strategy through the following formula 2: in, For media channels In the delivery round The delivery weight at the time For media channels In the previous round The media placement weight at that time; is the learning rate; For media channels In the delivery round The instant profit at that time is calculated based on market feedback data; For all media channels in the delivery round Average return on is the standard deviation of returns; For media channels In the delivery round Click-through rate when For media channels in the delivery round Time The conversion rate indicates the probability of conversion after the user clicks the ad; For media channels In the delivery round Unit delivery cost is used to control advertising budget; To prevent small positive numbers with zero denominator; is the timing attenuation adjustment factor; is the time attenuation coefficient; For media channels In the delivery round The time of the most recent ad serving on ; The current delivery time.

7. The multi-dimensional marketing planning agent system according to claim 1, characterized in that: The creation module includes an adaptive content style adjustment unit, which is used to perform style adaptive optimization on the generated communication theme data and marketing content data based on the media strategy plan data and the personalized characteristics of the target user portrait data. The style adaptive optimization is based on the user's historical interaction behavior, cultural preferences, language habits and the specific format requirements of the media channel, and dynamically adjusts the content expression, tone, color scheme and visual layout.

8. The multi-dimensional marketing planning agent system according to claim 1, characterized in that: The generation module includes a multi-stage delivery effect prediction unit, which uses a machine learning model to perform stage-by-stage prediction and optimization on data at different levels, wherein the machine learning model includes a delivery environment modeling submodule, a user behavior prediction submodule, and a dynamic content adjustment submodule; The delivery environment modeling submodule is used to receive historical marketing data, target user profile data, and media strategy plan data, analyze the advertising delivery characteristics of different media platforms, including exposure rate, user dwell time, and interaction mode, and generate an environment parameter vector to provide initial constraints for the delivery platform; The user behavior prediction submodule is used to receive the environmental parameter vector output by the delivery environment modeling submodule, and combine it with real-time market feedback data to predict the user's click-through rate, conversion rate and purchase intention. It uses time series analysis and deep learning models to extract the user's short-term and long-term behavior trends and output behavior prediction data; The dynamic content adjustment submodule is used to receive the behavior prediction data and adjust the text style, visual elements, and video editing strategy of the content in combination with the marketing content data generated by the delivery environment modeling submodule.

9. The multi-dimensional marketing planning agent system according to claim 1, characterized in that: The creation module includes a content context adaptation unit, which is used to dynamically adjust the content structure and presentation form of the communication theme data and marketing content data based on the user's current usage scenario, device characteristics and environmental variables. The adjustment process combines the user's interaction mode, the screen size of the terminal device, the network bandwidth status and the ambient light intensity to optimize the readability, loading speed and visual effect of the advertising content.

10. A method for multi-dimensional marketing planning agent, characterized in that: include: Receive basic data including marketing objectives, market context, brand tone, and budget constraints. The basic data may include multimodal inputs such as text, voice, images, and videos. Based on natural language processing technology and computer vision analysis, pre-process the basic data to extract key marketing element data. Based on the key marketing factor data, combined with external market data, social media trend data and industry report data, cluster analysis methods are used to generate target user portrait data, and competitive product marketing strategy analysis data is generated based on competitive product data mining technology; Based on target user profile data, competitor marketing strategy analysis data, corporate budget data, and historical delivery data, a reinforcement learning model is used to calculate delivery weights for different media channels; media strategy plan data is generated based on the delivery weights. Based on the media strategy data and in combination with the characteristics of media channels, generative artificial intelligence technology is used to create communication theme data and marketing content data that are suitable for different media; Based on the communication theme data, marketing content data, historical marketing data and real-time market feedback data, the delivery effect is predicted through a machine learning model, and the media strategy plan and marketing content plan are optimized according to the prediction results.

Images