Marketing strategy optimization method based on big language model analysis and evaluation driving

Abstract

The invention provides a marketing strategy optimization method based on big language model analysis and evaluation driving, and the method comprises the steps: obtaining historical data including user behavior data, label data and statistical data, carrying out the analysis, and constructing a basic marketing strategy; user content data are collected in real time through Kafka, deep analysis is conducted on user content through a large language model, and a personalized marketing scheme is matched from a knowledge base based on the deep analysis result of the large language model and basic rule matching; through a feedback circulation mechanism, feedback of the user to marketing activities is collected, and a feedback result is stored in a knowledge base and used for optimizing a marketing strategy and a large language model; meanwhile, the decision-making process is logged, textualized and vectorized and is stored in a vector library, and the decision-making process and the marketing scheme are optimized in combination with a continuously optimized large language model, so that dynamic adjustment and matching are realized. According to the invention, the accuracy and effect of the marketing strategy are improved by using the large language model.

Description

Technical Field

[0001] The present invention relates to the field of digital marketing technology, and in particular to a marketing strategy optimization method based on large language model analysis and evaluation drive. Background Art

[0002] In the field of digital marketing, understanding customer intentions in real time and responding to feedback quickly are the keys to developing effective marketing strategies. Existing methods Existing marketing strategies all rely on simple rule strategies, often using offline data to build models, relying on simple keyword matching or statistical techniques to complete the establishment of marketing strategies, and are unable to dig deep into the rich information in user-generated content in real time. Existing technologies cannot accurately hit user needs. Marketing strategies are traditional rule strategies. Marketing plans under different branches of marketing strategies are preset in advance based on user behavior and user tags through rule configuration. They cannot fully understand the user's existing intentions or needs, so there may be situations where marketing plans cannot be used. It is impossible to give a marketing plan as quickly as possible; currently, if there are changes in marketing strategies or data environments, the marketing strategy needs to be manually adjusted to match the marketing plan to the user, and an effective marketing plan cannot be automatically formulated to complete the marketing action. It is impossible to automatically upgrade the marketing strategy. The marketing strategy needs to be manually adjusted. It is impossible to adjust and optimize the marketing strategy in real time through an automatic feedback loop mechanism based on content behavior during execution.

[0003] Prior art 1, application number: CN202310646127.X discloses a private domain real-time marketing platform and marketing method, computer storage medium, belonging to the field of platform data management technology, the method includes the following steps: performing point-of-use processing of behavioral events in the private domain program, the point-of-use of behavioral events is triggered after the relevant behavioral events occur, and the private domain program is controlled to send preset data; pre-processing the preset data, and sending the pre-processed preset data to the preset rule engine; according to the marketing rules in the preset rule engine, judging whether there is a marketing action triggered in the preset rule engine; after any marketing action is triggered in the preset rule engine, the triggered marketing action is mounted on the page hook event of the private domain program; when the page hook event in the private domain program is triggered, the private domain program executes the marketing action corresponding to the page hook event. Although it can carry out real-time personalized precision marketing according to the marketing strategy, significantly improving the conversion efficiency of the private domain; but it can only rely on the rule engine to complete some simple strategy matching, and cannot deeply explore the real needs in the user-generated content; and if it is found that the user needs and environment have changed, the marketing strategy needs to be adjusted manually.

[0004] Prior art 2, application number: CN 202311010351.6 discloses a digital marketing method and marketing platform, which relates to the field of digital marketing technology. The data collection unit collects and integrates various data sources, the data storage unit stores and manages the collected data sources, the data analysis unit uses data analysis and mining technology to analyze and mine the collected data, the target audience positioning unit locates and segments the target audience according to the collected data and analysis results, and the management unit is used to create and manage content related to the target audience; the promotion unit is responsible for promoting marketing activities and content on multiple digital channels, and performs budget management and optimization to conduct digital marketing. Although the monitoring unit collects marketing data, evaluates the marketing effect online in real time, and generates corresponding management strategies based on the marketing effect, thereby evaluating the current marketing effect online in real time, and changing the marketing strategy in time based on the marketing effect, effectively improving the performance of marketing activities; however, it is impossible to fully understand the user's existing intentions or needs, so there may be a situation where the marketing plan cannot be used.

[0005] Prior art three, application number: CN 202310983051.X discloses a blockchain-based enterprise marketing planning system, the method includes the following steps: the enterprise selects multiple marketing platforms for marketing, and selects the main platform and the secondary platform at different stages according to the marketing data of the marketing market; by analyzing the multiple data sources of the multiple marketing platforms, the customer consumption portraits of the targets at different stages are planned; the enterprise analyzes the consumption characteristics and changing trends of consumers to achieve accurate judgment and prediction of the marketing market under social media; the enterprise determines the potential target customers based on the judgment and prediction, carries out precision marketing for the target customers, and increases the customer's in-depth understanding of the enterprise's products through product-oriented promotion; blockchain technology is applied to the current needs of the enterprise marketing market and the marketing feedback of real customer data. Although it has the characteristics of strengthening marketing strategies and improving market competitiveness; however, the marketing strategy needs to be manually adjusted to meet the matching of the marketing plan to the user, and it is impossible to automatically formulate an effective marketing plan to complete the marketing action.

[0006] At present, the existing technologies 1, 2 and 3 have the problem of not being able to deeply explore the real needs in the user-generated content; and if it is found that the user needs and environment have changed, the marketing strategy needs to be adjusted manually. Therefore, the present invention provides a marketing strategy optimization method based on large language model analysis and evaluation drive. Summary of the invention

[0007] In order to solve the above technical problems, the present invention provides a marketing strategy optimization method based on large language model analysis and evaluation drive, comprising the following steps:

[0008] Obtain historical data including user behavior data, tag data and statistical data, analyze and build basic marketing strategies; basic marketing strategies include triggering methods and marketing plans. The triggering methods are set based on user behavior, tags and large language models. When the user's real-time content meets the triggering rules, the corresponding marketing plan is automatically triggered; the marketing plan is stored in the database;

[0009] Collect user content data in real time through Kafka, and use the big language model to conduct in-depth analysis of user content, including the disassembly and reorganization of multi-dimensional information such as behavior tags, time, location, and text data; match personalized marketing plans from the knowledge base based on the in-depth analysis results and basic rule matching of the big language model;

[0010] Through the feedback loop mechanism, users’ feedback on marketing activities is collected, and the feedback results are stored in the knowledge base and used to optimize marketing strategies and large language models. At the same time, the decision-making process is logged, textualized and vectorized, and stored in the vector library. Combined with the continuously optimized large language model, the decision-making process and marketing plan are optimized to achieve dynamic adjustment and matching.

[0011] Optionally, the process of building a basic marketing strategy includes the following steps:

[0012] Integrate user behavior data, label data and statistical data and perform preprocessing to extract key features of user behavior including click frequency, browsing time and purchase path, and construct feature vectors;

[0013] Use clustering algorithms to group user behavior data and identify the behavior patterns and preferences of different user groups; build user portraits based on clustering results, combined with label data and statistical data, including demographic characteristics, behavioral characteristics, and interest preferences;

[0014] Based on user behavior data, label data and large language models, a trigger rule engine is designed to set trigger conditions and logic, including behavior thresholds, time windows and label matching; based on user portraits and trigger rules, a basic marketing strategy is built, including trigger methods, marketing plans and response mechanisms.

[0015] Optionally, the process of setting the trigger method, marketing plan and response mechanism in the basic marketing strategy includes the following steps:

[0016] Based on user behavior data, set the trigger thresholds for key behaviors such as clicks, browsing, and purchases; determine the trigger time window through time series analysis; design multi-level tag matching logic based on the tag data in the user portrait; and construct a trigger matrix that includes multiple dimensions of behavior, time, and tags;

[0017] Combined with a large language model, it generates personalized marketing content including text, images and videos. Based on the channel preferences in the user portrait, marketing information reaches users through the best channel through multi-channel collaborative strategies.

[0018] Build a response engine based on real-time data streams, optimize trigger rules and marketing plans through real-time feedback data; predict users' next behavior based on historical data and real-time behavior, and design response strategies in advance.

[0019] Optionally, the process of matching personalized marketing plans from the knowledge base includes the following steps:

[0020] Using the Kafka distributed stream processing platform, users’ content data from various channels is collected in real time, including behavior data, time data, location data, and text data.

[0021] Use the large language model to conduct multi-dimensional in-depth analysis of the collected user content data, including behavior tags, time patterns, location features, and text intent; through the sequence modeling and attention mechanism of the large language model, the user data can be accurately disassembled and reorganized to build a comprehensive user portrait;

[0022] Based on the in-depth analysis results of the large language model and combined with the basic matching rules predefined in the basic marketing strategy, potential personalized marketing plans are screened from the knowledge base, including behavior matching, time matching, location matching, and intention matching. A knowledge graph of personalized marketing plans is constructed, and the relationship between personalized marketing plans is modeled using graph neural networks to explore potential related plans.

[0023] The optional process of building a comprehensive user profile includes the following steps:

[0024] The large language model calculates the attention weight of each data point in real time based on the contextual information of the user's behavior. Through the hybrid architecture of the temporal convolutional network and Transformer, the large language model dynamically reorganizes the user's behavior sequence and identifies the key turning points in the user's behavior.

[0025] Deeply integrate users’ text comments, behavior trajectories, time series, and geographic location multimodal data to identify implicit associations between different modal data; identify users’ core behavior patterns through a global attention mechanism; and deeply analyze the detailed characteristics of user behavior at the local level through a local attention mechanism;

[0026] Capture the user's scene information in real time and analyze it in correlation with user behavior.

[0027] Optionally, the process of identifying key turning points in user behavior includes the following steps:

[0028] Initially encode the user's behavior sequence. Each behavior data point is converted into a high-dimensional vector representation, which contains the characteristics of the behavior itself and also embeds time information and context information. The behavior sequence is converted into a continuous vector sequence by encoding.

[0029] The temporal convolutional network is used to extract hierarchical features of behavior sequences. Through a multi-layer convolutional structure, the temporal dependencies in the behavior sequences are extracted layer by layer. Each layer of convolution extracts local features of the input sequence and captures longer time spans through dilated convolutions.

[0030] Transformer calculates the correlation between each behavior data point and other data points through the self-attention mechanism, and identifies the global pattern in the behavior sequence; the global attention mechanism captures the key turning points in user behavior; under the synergy of the temporal convolutional network and Transformer, the behavior sequence is dynamically reorganized; the representation of the behavior sequence is dynamically adjusted to identify the key turning points in the behavior sequence;

[0031] The global attention mechanism identifies the main turning points in the behavior sequence, and the local attention mechanism deeply analyzes the detailed features of the turning points, captures the key changes in user behavior, and constructs a user portrait.

[0032] Optionally, the process of building a knowledge graph for a personalized marketing plan includes the following steps:

[0033] Identify key entities including user behavior types, time patterns, location preferences, and text intent from user portraits and marketing strategies. Use key entities as nodes in the knowledge graph. Use deep analysis of large language models to identify and extract relationships between entities. Use relationships as edges in the knowledge graph. Assign attributes to each entity and relationship.

[0034] The constructed knowledge graph is converted into graph structure data, where nodes represent entities, edges represent relationships, and the attributes of nodes and edges represent the characteristics of entities and relationships; the graph structure data is mapped to a low-dimensional vector space to capture the local and global relationships between nodes and generate an embedding vector for each node; through the message passing mechanism of the graph neural network, the information of neighbor nodes is aggregated, the embedding vector of each node is updated, and the complex association relationship between nodes is captured;

[0035] Based on the output of the graph neural network, potential association rules are mined from the graph structure, and personalized marketing plans that match the user portrait are screened out from the knowledge base, so that the plan matches behavior, time, location and intention.

[0036] Optionally, the process of optimizing decision-making processes and marketing plans includes the following steps:

[0037] Extract key features from the logs of the decision-making process, including user behavior patterns, time series features, and geographic location information; use word embedding to convert the extracted key features into high-dimensional vector representations to capture the semantic information and contextual relationships in the text; store the vectorized data in a vector library to form a knowledge base;

[0038] Using the continuously optimized big language model, the data in the vector library is analyzed and inferred in real time; by comparing historical decision data and current user feedback, the big language automatically adjusts and optimizes decision parameters; based on the optimized big language model, user behavior and marketing plans are matched in real time;

[0039] Through the feedback loop mechanism, users’ feedback data on marketing activities is collected in real time, and the feedback data is combined with historical data in the vector library to continuously optimize the large language model and decision-making process.

[0040] Optionally, feedback data includes click-through rate, conversion rate and user satisfaction indicators.

[0041] Optionally, the process of converting the extracted key features into high-dimensional vector representations using word embedding includes the following steps:

[0042] The multimodal data of user behavior patterns, time series features and geographic location information are integrated to form a unified feature space; more time-related features are extracted through the sliding window technology of the time series;

[0043] The preprocessed features are mapped to a high-dimensional semantic space using deep semantic embedding. The deep semantic relationship between key features is captured by using self-supervised learning. The contextual relationship of key features is modeled using the Transformer architecture. The dynamic changes and mutual influence of features in different contexts are captured through the multi-head self-attention mechanism.

[0044] Based on semantic capture and context modeling, nonlinear mapping technology is used to map key features from semantic space to high-dimensional vector space.

[0045] The present invention starts with solving user intentions and ends with generating marketing plans. By identifying user intentions and matching corresponding marketing plans through a knowledge base, all operations of the present invention can be achieved. All steps of the present invention are automated and can be updated in real time, avoiding errors and problems caused by human operations. The present invention can automatically generate marketing plans, greatly reducing the investment in human resources. After reaching specific users, a feedback loop mechanism is executed, the strategy execution history is stored in the knowledge base, and each step of the strategy is optimized and adjusted so that subsequent strategies can more effectively meet the needs of the market and customers.

[0046] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0047] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0049] Figure 1 This is a flow chart of a marketing strategy optimization method based on large language model analysis and evaluation-driven in Example 1 of the present invention;

[0050] Figure 2 This is a schematic diagram of a marketing strategy optimization method based on large language model analysis and evaluation drive in Example 1 of the present invention;

[0051] Figure 3 A process diagram for constructing a basic marketing strategy in Example 2 of the present invention;

[0052] Figure 4 This is a process diagram for setting the trigger mode, marketing plan and response mechanism in the basic marketing strategy in Example 3 of the present invention;

[0053] Figure 5 This is a process diagram of matching personalized marketing plans from a knowledge base in Embodiment 4 of the present invention;

[0054] Figure 6 A process diagram for constructing a comprehensive user portrait in Embodiment 5 of the present invention;

[0055] Figure 7 A process diagram for identifying key turning points in user behavior in Embodiment 6 of the present invention;

[0056] Figure 8 A process diagram for identifying implicit associations between different modal data in Example 7 of the present invention;

[0057] Fig. 9 A process diagram of constructing a knowledge graph of a personalized marketing solution in Example 8 of the present invention;

[0058] Fig.10 A process diagram for optimizing the decision-making process and marketing plan in Example 9 of the present invention;

[0059] Fig.11This is a process diagram of converting the extracted key features into high-dimensional vector representations using word embedding in Example 10 of the present invention. DETAILED DESCRIPTION

[0060] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0061] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the embodiments of the present application. The singular forms of "a", "said" and "the" used in the embodiments of the present application are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used in this article refers to and includes any or all possible combinations of one or more associated listed items.

[0062] When the following description relates to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. On the contrary, they are only examples of devices and methods consistent with some aspects of the present application. In the description of the present application, it should be understood that the terms "first", "second", "third", etc. are only used to distinguish similar objects, and do not have to be used to describe a specific order or sequence, nor can they be understood as indicating or implying relative importance. For those of ordinary skill in the art, the specific meanings of the above terms in the present application can be understood according to the specific circumstances.

[0063] Example 1: Figure 1 As shown, the embodiment of the present invention provides a marketing strategy optimization method based on large language model analysis and evaluation drive, comprising the following steps:

[0064] S100: Acquire historical data including user behavior data, tag data and statistical data, analyze and build a basic marketing strategy; the basic marketing strategy includes a trigger method and a marketing plan, the trigger method is set based on user behavior, tags and a large language model, and when the user's real-time content meets the trigger rules, the corresponding marketing plan is automatically triggered; the marketing plan is stored in the database;

[0065] S200: Collect user content data in real time through Kafka, and use the big language model to conduct in-depth analysis of user content, including the disassembly and reorganization of multi-dimensional information such as behavior tags, time, location and text data; match personalized marketing plans from the knowledge base based on the in-depth analysis results and basic rule matching of the big language model;

[0066] S300: Through the feedback loop mechanism, users’ feedback on marketing activities is collected, and the feedback results are stored in the knowledge base, which is used to optimize marketing strategies and large language models. At the same time, the decision-making process is logged, textualized, and vectorized, and stored in the vector library. Combined with the continuously optimized large language model, the decision-making process and marketing plan are optimized to achieve dynamic adjustment and matching.

[0067] The working principle and beneficial effects of the above technical solution are as follows: This embodiment realizes the acquisition of historical data including user behavior data, label data and statistical data, analyzes and constructs a basic marketing strategy; the basic marketing strategy includes a trigger method and a marketing plan, and the trigger method is set based on user behavior, label and large language model. When the user's real-time content meets the trigger rule, the corresponding marketing plan is automatically triggered; the marketing plan is stored in the database; secondly, user content data is collected in real time through Kafka, and the user content is deeply analyzed using the large language model, including the disassembly and reorganization of multi-dimensional information such as behavior labels, time, location and text data; based on the deep analysis results of the large language model and the basic rules of the basic marketing strategy, personalized marketing plans are matched from the knowledge base; finally, through the feedback loop mechanism, user feedback on marketing activities is collected, the feedback results are stored in the knowledge base, and used to optimize the marketing strategy and the large language model; at the same time, the decision-making process is logged, textualized and vectorized, and stored in the vector library, combined with the continuously optimized large language model, the decision-making process and marketing plan are optimized to achieve dynamic adjustment and matching (for specific principles, refer to the attached Figure 2). Step S100 of the above scheme constructs a basic marketing strategy, which provides a solid foundation for personalized marketing and ensures the pertinence and effectiveness of the marketing strategy; by setting trigger rules, it reduces manual intervention and improves marketing efficiency and response speed. Step S200 uses Kafka to collect user content data in real time to ensure the timeliness and integrity of the data; uses a large language model to perform multi-dimensional analysis of user content, including behavioral tags, time, location, and text data, to accurately grasp user needs; based on the analysis results and basic rules, matches personalized marketing plans from the knowledge base to improve marketing accuracy and user satisfaction. Significance: Through real-time data collection and analysis, quickly respond to user needs and enhance user experience; through in-depth analysis and personalized matching, achieve precision marketing and improve conversion rate and user stickiness. Step S300: Feedback loop and optimization. Through the feedback loop mechanism, collect user feedback on marketing activities, and timely understand the user's acceptance and satisfaction with the marketing strategy; store the feedback results in the knowledge base for optimizing the marketing strategy and the large language model, and realize dynamic adjustment and optimization of the strategy; log, text and vectorize the decision-making process, store it in the vector library, and optimize the decision-making process and marketing plan in combination with the continuously optimized large language model. Significance: Through feedback loops and knowledge base updates, the marketing strategy can be continuously optimized to improve the effectiveness and adaptability of the strategy; through decision logging and vectorization, data-driven decision optimization can be realized to improve the scientificity and accuracy of the decision; through dynamic adjustment and matching, ensure that the marketing strategy can respond to market changes and user needs in a timely manner, and enhance the market competitiveness of the enterprise.

[0068] The training process of the large language model in this embodiment is as follows: multimodal data including user behavior data, label data and statistical data are obtained from historical data, including user behavior trajectory, time series, geographic location, text comments, etc.; data cleaning and preprocessing, user behavior data is annotated, for example, user behavior is divided into categories such as browsing, purchasing, and commenting, and its time, location and other information are annotated; a large language model architecture suitable for multimodal data processing (a hybrid architecture of a temporal convolutional network and a Transformer) is selected, and the model is pre-trained using large-scale unsupervised data to learn the general semantic representation of the language; in the pre-training stage, the contextual relationship in the text data is learned through the self-attention mechanism, and the generalization ability of the model is improved through multi-task learning (such as masked language model, next sentence prediction, etc.); in marketing activities, user feedback data on marketing plans is collected in real time, for example Such as click-through rate, conversion rate, user satisfaction score, etc., store feedback data in the knowledge base and use it to optimize marketing strategies and large language models; use feedback data as new training data to continuously optimize the large language model, and continuously update the model parameters through incremental learning technology to enable it to adapt to changes in user behavior and feedback data. Log, text and vectorize the decision-making process and store it in the vector library. Combined with the continuously optimized large language model, optimize the decision-making process and marketing plan; use the test data set to evaluate the optimized model and calculate its accuracy, recall rate, F1 value and other indicators in tasks such as user behavior prediction and marketing plan recommendation; deploy the optimized model to the production environment, analyze user content data in real time, and match personalized marketing plans; collect user content data in real time through real-time data collection tools such as Kafka, and use the large language model for in-depth analysis. The training process of the large language model includes data preparation, model initialization and pre-training, fine-tuning, feedback loop and continuous optimization, model evaluation and deployment, etc. Through this process, the model can learn the deep characteristics of user behavior from multimodal data and continuously optimize through the feedback loop mechanism, thereby achieving dynamic adjustment and precise matching of marketing strategies; it not only improves the model's predictive and generalization capabilities, but also provides strong support for personalized marketing.

[0069] This embodiment solves the technical problems of marketing strategy optimization by combining a large language model, which is reflected in the following aspects: Initialization of marketing strategy formulation: Through the analysis of historical data, a basic marketing strategy is formed, which can be set according to different data intervals to ensure that it can reflect the current situation of the market and the basic needs of customers; Deep content data analysis: Use a large language model to conduct in-depth analysis of text data such as user-generated content to identify consumers' emotions, preferences, needs and behavior patterns; Personalized marketing suggestions: Based on the analysis results of the large language model, combined with the initialization marketing strategy rules, real-time personalized marketing plan suggestions for customers are automatically proposed; Feedback loop mechanism: It can collect user feedback after the execution of marketing activities, and the marketing strategy can be adjusted and optimized according to real-time feedback; Continuous evolution strategy: Through continuous analysis and adjustment, the marketing strategy can continue to evolve and better adapt to market changes and consumer trends; The dynamic optimization process helps to improve the effectiveness and efficiency of marketing activities. Optimization of the decision-making process: The process of the marketing link is logged, textualized, and vectorized, and stored in the vector library, and the historical decision-making process and results are synchronized to the knowledge base. Combined with the language understanding and thinking chain reasoning ability of the large language model, the decision-making process can be optimized.

[0070] This embodiment uses the powerful text processing and analysis capabilities of the large language model to improve the accuracy and effectiveness of marketing strategies. In combination with the marketing strategy optimization method of the large language model, the large language model is first used to conduct an in-depth analysis of user content data to identify consumer emotions, preferences, needs and behavior patterns. Then, based on these analysis results, targeted marketing plan suggestions can be automatically proposed. It also includes a feedback loop mechanism to continuously collect user feedback after the execution of marketing activities in order to adjust and optimize marketing strategies in real time. In this way, marketing strategies can continue to evolve and better adapt to market changes and consumer trends; they can adjust and optimize marketing strategies in real time according to market changes and consumer trends, and improve the effectiveness and efficiency of marketing activities.

[0071] In summary, this embodiment achieves precision marketing and continuous optimization through real-time data collection, in-depth analysis and feedback optimization, improves marketing effects and user satisfaction; enhances the market responsiveness and competitiveness of enterprises through data-driven decision optimization and dynamic adjustment; optimizes user experience through personalized matching and real-time response, and improves user stickiness and loyalty. This embodiment realizes the intelligence and personalization of marketing strategies through technical means, providing enterprises with strong market competitiveness and user satisfaction.

[0072] Example 2: Figure 3 As shown, based on Example 1, the process of constructing a basic marketing strategy provided by the embodiment of the present invention includes the following steps:

[0073] S101: Integrate user behavior data, label data and statistical data and perform preprocessing to extract key features of user behavior including click frequency, browsing time and purchase path, and construct feature vectors;

[0074] S102: Use clustering algorithms to group user behavior data and identify the behavior patterns and preferences of different user groups; based on the clustering results, combined with label data and statistical data, build user portraits, including demographic characteristics, behavioral characteristics, and interest preferences;

[0075] The calculation process of the clustering algorithm specifically includes:

[0076] Objective function (minimize weighted squared error):

[0077]

[0078] In the formula, J represents the objective function, which is the sum of the weighted square errors from all data points to their cluster centers; k represents the number of clusters, that is, the number of user groups; C i represents the i-th cluster, which contains all the data points belonging to the cluster; x represents the data point, which represents the user behavior feature vector; μ i represents the center of the i-th cluster, that is, the average value of all data points in the cluster; ‖x-μ i ‖ represents the data point x and the cluster center μ i The Euclidean distance between i represents the weight of the i-th cluster, reflecting the importance of the cluster in the overall objective function;

[0079] Cluster center update formula:

[0080]

[0081] In the formula, μ i represents the center of the i-th cluster, w i represents the weight of the i-th cluster, x represents the data point, and represents the user behavior feature vector;

[0082] The process of clustering algorithm:

[0083] Randomly select k data points as the initial cluster centers μ1,μ2,…,μ k , and initialize the weights w1,w2,…,w for each cluster k ;

[0084] For each data point x, calculate its weighted distance to all cluster centers and assign it to the cluster C with the closest weighted distance. i ;

[0085] According to the allocation results, update each cluster center μi is the weighted average of all data points in the cluster, and updates the weight w of each cluster i ;

[0086] Repeat the allocation and update steps until the cluster center no longer changes or the preset number of iterations is reached; through the complex equation of the extended K-means algorithm, user behavior data can be more finely grouped to identify the behavior patterns and preferences of different user groups. The extended equation introduces weight factors so that the clustering results can better reflect the importance of different user groups, thereby providing a more accurate basis for the formulation of personalized marketing strategies;

[0087] S103: Based on user behavior data, label data and large language models, design a trigger rule engine, set trigger conditions and logic, including behavior thresholds, time windows and label matching, etc.; build basic marketing strategies based on user portraits and trigger rules, including trigger methods, marketing plans and response mechanisms.

[0088] The working principle and beneficial effects of the above technical solution are as follows: This embodiment first integrates user behavior data, label data and statistical data and performs preprocessing to extract key features of user behavior including click frequency, browsing time and purchase path, and constructs a feature vector; secondly, a clustering algorithm is used to group user behavior data to identify the behavior patterns and preferences of different user groups; based on the clustering results, combined with label data and statistical data, a user portrait is constructed, including demographic characteristics, behavioral characteristics and interest preferences; finally, based on user behavior data, label data and a large language model, a trigger rule engine is designed to set trigger conditions and logic, including behavior thresholds, time windows and label matching, etc.; according to user portraits and trigger rules, a marketing strategy is constructed, including trigger methods, marketing plans and response mechanisms. Step S101 of the above scheme is data integration and preprocessing, which integrates user behavior data, label data and statistical data to ensure the comprehensiveness and consistency of the data and provide a solid foundation for analysis; preprocessing the integrated data, including data cleaning, deduplication, filling missing values, etc., to ensure the quality and availability of the data; extracting key features in user behavior, such as click frequency, browsing time and purchase path, etc., to construct feature vectors and provide input for clustering and analysis. Significance achieved: Through data integration and preprocessing, the quality and consistency of the data are ensured, and a reliable data foundation is provided for analysis and modeling; through feature extraction, feature vectors are constructed to enhance the expression ability of user behavior and provide rich input for clustering and analysis; through data integration and preprocessing, the data management process is optimized and the efficiency and effect of data processing are improved. Step S102 is user grouping and portrait construction, which uses clustering algorithms to group user behavior data, identify the behavior patterns and preferences of different user groups, and provide a basis for personalized marketing; based on the clustering results, combined with label data and statistical data, user portraits are constructed, including demographic characteristics, behavioral characteristics and interest preferences, etc., to provide detailed user descriptions for personalized marketing. Significance achieved: Through clustering algorithms, the behavior patterns and preferences of different user groups are identified to improve the accuracy and effectiveness of user segmentation; through the construction of user portraits, detailed user descriptions are provided to enhance the ability of personalized marketing and improve the accuracy and effectiveness of marketing; through user grouping and portrait construction, the allocation of marketing resources is optimized to improve resource utilization efficiency and marketing effectiveness. Step S103 triggers the rule engine and basic marketing strategy construction. Based on user behavior data, label data and large language models, the trigger rule engine is designed, and the trigger conditions and logic are set, including behavior thresholds, time windows and label matching, etc., to ensure the accurate triggering of marketing activities; according to user portraits and trigger rules, a basic marketing strategy is constructed, including trigger methods, marketing plans and response mechanisms, to ensure the comprehensiveness and systematicness of marketing activities.Significance achieved: Through the design of the trigger rule engine, the accurate triggering of marketing activities is ensured, and the accuracy and timeliness of marketing are improved; through the construction of basic marketing strategies, the comprehensiveness and systematicness of marketing activities are ensured, and the overall marketing effect and user satisfaction are improved; through the construction of trigger rules and basic marketing strategies, the strategy execution process is optimized and the efficiency and effectiveness of strategy execution are improved.

[0089] In summary, this embodiment uses data integration and preprocessing, user grouping and portrait construction, triggering rule engine and basic marketing strategy construction, and the basic marketing strategy can achieve data-driven precision marketing and personalized services.

[0090] Example 3: Figure 4 As shown, based on Example 2, the setting process of the trigger mode, marketing plan and response mechanism in the basic marketing strategy provided by the embodiment of the present invention includes the following steps:

[0091] S1031: Based on user behavior data, set trigger thresholds for key behaviors such as clicks, browsing and purchases; determine the trigger time window through time series analysis; design multi-level tag matching logic based on tag data in user portraits; construct a trigger matrix containing multiple dimensions such as behavior, time, and tags;

[0092] S1032: Combined with the large language model, generate personalized marketing content including text, images and videos; based on the channel preferences in the user portrait, through multi-channel collaborative strategies, marketing information reaches users in the best channel;

[0093] S1033: Build a response engine based on real-time data streams to optimize trigger rules and marketing plans through real-time feedback data; predict the user's next behavior based on historical data and real-time behavior, and design response strategies in advance.

[0094] The working principle and beneficial effects of the above technical solution are as follows: This embodiment first sets the trigger thresholds of key behaviors such as clicks, browsing and purchases based on user behavior data; determines the trigger time window through time series analysis; designs multi-level tag matching logic in combination with the tag data in the user portrait; constructs a trigger matrix containing multiple dimensions such as behavior, time, and tags; secondly, combines the large language model to generate personalized marketing content containing multiple modalities such as text, images, and videos; based on the channel preferences in the user portrait, through a multi-channel collaborative strategy, marketing information reaches users in the best channel; finally, constructs a response engine based on real-time data streams, optimizes trigger rules and marketing plans through real-time feedback data; predicts the user's next behavior based on historical data and real-time behavior, and designs response strategies in advance. The setting of the triggering method in step S1031 of the above scheme is to analyze the user behavior data, set the triggering thresholds of key behaviors such as clicks, browsing and purchases, ensure that the marketing activities are triggered when the user behavior reaches a certain level, and improve the accuracy and timeliness of marketing; determine the triggering time window through time series analysis, ensure that the marketing activities are triggered in the best time period after the user behavior occurs, and improve the user's response rate and conversion rate; design multi-level tag matching logic in combination with the tag data in the user portrait, ensure that the marketing activities can accurately match the user's interests and preferences, and improve the personalization and pertinence of marketing; construct a trigger matrix containing multiple dimensions such as behavior, time, and tags, realize the comprehensive evaluation and decision-making of multi-dimensional trigger conditions, and improve the flexibility and adaptability of trigger rules. The significance achieved: through the precise setting of triggering thresholds and time windows, ensure that the marketing activities reach users at the best time, improve user participation and conversion rate; through the construction of multi-level tag matching logic and triggering matrix, ensure that the marketing activities can accurately match the user's interests and preferences, provide personalized user experience, and enhance user stickiness; through the comprehensive evaluation of multi-dimensional trigger conditions, optimize the allocation of marketing resources and improve marketing efficiency and effectiveness. Step S1032: Setting up the marketing plan, combined with the large language model, generates multi-modal personalized marketing content including text, images, and videos, enriches the expression of marketing information, and improves user appeal and engagement; based on the channel preferences in the user portrait, through multi-channel collaborative strategies, ensures that marketing information reaches users in the best channel, and improves the coverage and reach of information. Significance achieved: Through multi-modal content generation and multi-channel collaborative strategies, provide rich and diverse marketing information, meet the diverse needs of users, and enhance user experience; through precise channel selection and multi-channel collaborative strategies, ensure that marketing information reaches users in the best channel, and improve the reach and conversion rate of information; through the application of large language models, realize the rapid generation of personalized content, optimize the content creation process, and improve the efficiency and quality of content creation.Step S1033: Setting up the response mechanism, building a response engine based on real-time data stream, optimizing trigger rules and marketing plans through real-time feedback data, ensuring the real-time and dynamic nature of marketing activities; predicting the user's next behavior based on historical data and real-time behavior, designing response strategies in advance, and improving the predictability and initiative of marketing activities. Significance achieved: Through the real-time data stream response engine, ensure that marketing activities can quickly respond to the user's real-time behavior, improve response speed and timeliness; through user behavior prediction, design response strategies in advance, enhance the predictability and initiative of marketing activities, and improve user participation and conversion rate; through real-time feedback data, dynamically optimize trigger rules and marketing plans, ensure continuous optimization and improvement of marketing strategies, and improve marketing effectiveness and user satisfaction.

[0095] In summary, through the setting of triggering methods, marketing plans and response mechanisms, the basic marketing strategy in this embodiment can achieve accurate triggering, personalized marketing and real-time response, thereby improving the accuracy of marketing, user experience and marketing effect. It not only relies on advanced technical means, but also needs to be combined with actual business scenarios to make reasonable parameter settings and result interpretations to ensure the effectiveness and practicality of marketing strategies.

[0096] Example 4: Figure 5 As shown, based on Example 1, the process of matching personalized marketing plans from a knowledge base provided by the embodiment of the present invention includes the following steps:

[0097] S201: Use the Kafka distributed stream processing platform to collect user content data from various channels in real time, including behavior data, time data, location data, and text data;

[0098] Among them, behavioral data: clicks, browsing and purchases, etc.; time data: the timestamp of the behavior; location data: GPS location and IP address, etc.; text data: comments, search terms and chat records, etc.;

[0099] S202: Use the large language model to conduct multi-dimensional in-depth analysis of the collected user content data, including behavior tags, time patterns, location features, and text intent; through the sequence modeling and attention mechanism of the large language model, the user data can be accurately disassembled and reorganized to build a comprehensive user portrait;

[0100] Among them, behavior tags: identify user behavior types and frequencies; time patterns: analyze the time distribution of user behaviors; location features: mine user location preferences; text intent: analyze the true intent of user text content;

[0101] S203: Based on the in-depth analysis results of the large language model and combined with the basic matching rules predefined in the basic marketing strategy, potential personalized marketing plans are screened from the knowledge base, including behavior matching, time matching, location matching, and intention matching; a knowledge graph of personalized marketing plans is constructed, and the association relationship between personalized marketing plans is modeled using a graph neural network to mine potential association plans;

[0102] Among them, behavior matching: matching corresponding plans according to user behavior characteristics; time matching: combining user time patterns to screen time-sensitive plans; location matching: recommending localized plans based on user geographic location; intent matching: providing precise personalized marketing plans based on user text intent.

[0103] The working principle and beneficial effects of the above technical solution are as follows: This embodiment first uses the Kafka distributed stream processing platform to collect content data of users in various channels in real time, including behavior data, time data, location data and text data; among them, behavior data: click, browse, purchase, etc.; time data: timestamp of behavior; location data: GPS positioning, IP address, etc.; text data: comments, search terms, chat records, etc.; secondly, the large language model is used to perform multi-dimensional in-depth analysis of the collected user content data, including behavior tags, time patterns, location features and text intent; through the sequence modeling and attention mechanism of the large language model, the user data can be accurately disassembled and reorganized to build a comprehensive user portrait; among them, behavior tags: identify user behavior types and frequencies; time patterns: analyze The time distribution law of user behavior; location characteristics: mining user's geographical location preferences; text intent: parsing the real intention of user text content; finally, based on the deep analysis results of the large language model, combined with the basic matching rules predefined by the basic marketing strategy, potential personalized marketing plans are screened from the knowledge base, including behavior matching, time matching, location matching and intention matching; a knowledge graph of personalized marketing plans is constructed, and the association relationship between personalized marketing plans is modeled using graph neural networks to mine potential association plans; among them, behavior matching: matching corresponding plans according to user behavior characteristics; time matching: screening time-effective plans in combination with user time patterns; location matching: recommending localized plans based on user geographic location; intention matching: providing accurate personalized marketing plans for user text intent. Step S201 of the above scheme uses the Kafka distributed stream processing platform to collect content data of users in various channels in real time; Kafka, as a distributed stream processing platform, can efficiently collect user behavior data, time data, location data and text data in multiple channels (such as websites, APPs, social media, etc.) to ensure the real-time and integrity of the data; through Kafka, data from different channels can be uniformly collected and integrated to form a comprehensive user data source. Significance achieved: Provides a high-quality data foundation for deep analysis and personalized matching, ensuring the accuracy and comprehensiveness of the analysis results; real-time collection of user behavior data can timely capture users' dynamic needs and behavior changes, and provide support for rapid response to user needs. Step S202 uses a large language model to perform multi-dimensional deep analysis on the collected user content data. Through the sequence modeling and attention mechanism of the large language model, it can perform deep analysis on the user's behavior, time, location and text data, and extract behavior labels, time patterns, location features and text intent; based on the analysis results, build a comprehensive user portrait, including the user's interest preferences, behavior habits, time patterns and geographic location preferences.Significance achieved: Through in-depth analysis, users' needs and behavior patterns can be understood more accurately, providing a scientific basis for personalized marketing; dynamic updates of user portraits can reflect changes in user behavior in real time, ensuring the timeliness and pertinence of marketing strategies. Step S203, based on the in-depth analysis results of the large language model and combined with the basic matching rules predefined in the basic marketing strategy, screens potential personalized marketing plans from the knowledge base; through behavior matching, time matching, location matching and intention matching, it is possible to screen out personalized marketing plans that are highly consistent with user characteristics from the knowledge base; using graph neural networks to model the correlation between personalized marketing plans, it is possible to explore potential related plans and improve the diversity and coverage of recommended plans. Significance achieved: Through precise matching, users can be provided with highly personalized marketing plans, improving user experience and conversion rates; through knowledge graph modeling, it is possible to discover potential user needs, recommend related marketing plans, and expand marketing coverage.

[0104] In summary, this embodiment realizes efficient real-time processing in the entire process from data collection to analysis and matching, ensuring the timeliness of the marketing plan; through multi-dimensional analysis and deep learning models, it can accurately capture user needs and behavioral characteristics and achieve highly personalized matching; through knowledge graphs and graph neural networks, it can explore potential related plans and improve the diversity and coverage of marketing plans. Significance achieved: Through precise matching and personalized recommendations, the conversion rate and user satisfaction of marketing plans can be significantly improved; based on dynamic analysis of user behavior and needs, it can provide users with marketing content that is more in line with their preferences and improve user experience; the entire process embodies the core concept of intelligent marketing, and realizes the automation and precision of marketing strategies through technical means, reduces labor costs, and improves marketing efficiency. Through the above steps, the matching process of personalized marketing plans not only achieves technological breakthroughs, but also brings practical value to enterprises and users, and promotes the intelligent and refined development of the marketing field.

[0105] Example 5: Figure 6 As shown, based on Example 4, the process of building a comprehensive user portrait provided by the embodiment of the present invention includes the following steps:

[0106] S2021: The large language model calculates the attention weight of each data point in real time based on the contextual information of user behavior. Through the hybrid architecture of the temporal convolutional network and Transformer, the large language model dynamically reorganizes the user behavior sequence and identifies the key turning points in user behavior. For example, when a user switches from browsing to purchasing, the model automatically identifies the change in behavior pattern and uses it as an important feature in the user portrait.

[0107] S2022: Deeply integrate multimodal data such as user text comments, behavior trajectories, time series, and geographic locations to identify implicit associations between different modal data; identify the user's core behavior patterns through a global attention mechanism; and deeply analyze the detailed characteristics of user behavior at the local level through a local attention mechanism; for example, identify the user's main consumption areas (such as electronic products) at the global level and analyze their preference for a certain brand (such as Apple) at the local level;

[0108] S2023: Capture the user's scene information (such as time, location, device type, etc.) in real time, and correlate it with user behavior for analysis; for example, when a user visits a shopping platform on weekends, the model will automatically recognize this scene and increase the weight of their shopping behavior in the portrait.

[0109] The working principle and beneficial effects of the above technical solution are as follows: In this embodiment, first, the large language model calculates the attention weight of each data point in real time according to the contextual information of the user behavior; through the hybrid architecture of the time convolution network and the Transformer, the large language model dynamically reorganizes the user behavior sequence and identifies the key turning points in the user behavior; for example, when the user switches from browsing to purchasing, the change in this behavior pattern is automatically identified and used as an important feature in the user portrait; secondly, the multimodal data such as the user's text comments, behavior trajectory, time series and geographic location are deeply integrated to identify the implicit associations between different modal data; the core behavior pattern of the user is identified through the global attention mechanism; at the local level, the detailed characteristics of the user behavior are deeply analyzed through the local attention mechanism; for example, the main consumption areas of the user (such as electronic products) are identified at the global level, and their preference for a certain brand (such as Apple) is analyzed at the local level; finally, the scene information of the user (such as time, place, device type, etc.) is captured in real time, and it is associated with the user behavior for analysis; for example, when the user visits the shopping platform on the weekend, the model will automatically identify this scene and increase the weight of its shopping behavior in the portrait. Step S2021 of the above scheme, dynamic calculation and behavior sequence reorganization, can dynamically adjust the weight of each data point according to the contextual information of user behavior through the attention mechanism of the large language model; for example, when a user frequently browses a certain category of goods in a specific time period, the model will automatically increase the weight of the behavior data in that time period, so as to more accurately capture the user's immediate needs; through the hybrid architecture of the temporal convolutional network (TCN) and Transformer, the model can dynamically reorganize the user's behavior sequence and identify the key turning points in user behavior; for example, when the user switches from browsing to purchasing, the model will automatically identify the change in this behavior pattern and use it as an important feature in the user portrait. Significance achieved: Through dynamic calculation and behavior sequence reorganization, the model can capture subtle changes in user behavior in real time, avoiding the lag of traditional static modeling; user portraits are no longer static, but are dynamically updated as user behavior changes in real time, which can more accurately reflect the user's immediate needs and behavior patterns; by identifying the key turning points of user behavior, the model can predict the user's future behavior trends and provide stronger support for personalized recommendations and precision marketing.Step S2022: Multimodal data fusion and hierarchical attention analysis, deeply fuse the multimodal data such as the user's text comments, behavior trajectory, time series and geographic location, and identify the implicit associations between different modal data; for example, if the user's comments mention "like outdoor sports", and his behavior trajectory shows frequent visits to outdoor goods stores, the model will associate the two and generate a user label of "outdoor sports enthusiast"; at the global level, the core behavior pattern of the user is identified through the global attention mechanism; at the local level, the detailed characteristics of the user's behavior are deeply analyzed through the local attention mechanism; for example, the user's main consumption areas (such as electronic products) are identified at the global level, and their preference for a certain brand (such as Apple) is analyzed at the local level. Significance achieved: Through multimodal data fusion, the model can comprehensively characterize the user's behavior characteristics from multiple dimensions, avoiding the limitations of single-dimensional data; by identifying the implicit associations between multimodal data, the model can more accurately capture the deep motivation behind the user's behavior and provide richer features for the user portrait; the global attention mechanism can identify the user's core behavior pattern, and the local attention mechanism can deeply analyze the detailed characteristics of the user's behavior, thereby achieving a refined characterization of the user portrait. Step S2023: Scenario perception and contextual modeling. The model can capture the scene information of the user in real time (such as time, location, device type, etc.) and associate it with user behavior for analysis; for example, when a user visits a shopping platform on weekends, the model will automatically identify this scene and increase the weight of his shopping behavior in the portrait; the user portrait is no longer static, but dynamically updates its feature weight according to the user's real-time behavior and scene changes; for example, when a user frequently searches for a certain type of product within a certain period of time, the model will automatically adjust the interest tags in his portrait and predict his future behavior trends. Significance achieved: By capturing the scene information of the user in real time, the model can dynamically adjust the construction logic of the user portrait according to the specific scene, so as to more accurately reflect the user's immediate needs; the user portrait can be updated in real time according to the scene changes, avoiding the lag of traditional static modeling, and can more flexibly adapt to the diverse needs of users; through scene perception and dynamic update, the model can more accurately predict the user's future behavior trends, providing stronger support for personalized recommendations and precision marketing.

[0110] In summary, this embodiment improves the real-time and dynamic nature of user portraits through dynamic calculation and behavior sequence reorganization, and can capture subtle changes in user behavior in real time and identify key behavioral turning points; through multimodal data fusion and hierarchical attention analysis, it enhances the comprehensiveness and refinement of user portraits, and can comprehensively characterize user behavior characteristics from multiple dimensions; through scene perception and contextual modeling, it improves the scene perception ability and dynamic nature of user portraits, and can dynamically adjust the construction logic of user portraits according to specific scenarios. This embodiment, through the above steps, jointly constitutes a comprehensive, dynamic and accurate user portrait construction system, which not only breaks through the limitations of traditional technologies, but also provides more powerful support for personalized services, precision marketing and intelligent recommendations; it can understand user needs more deeply and predict user behavior more accurately, thereby providing users with a better service experience.

[0111] Example 6: Figure 7 As shown, based on Example 5, the process of identifying key turning points in user behavior provided by the embodiment of the present invention includes the following steps:

[0112] S20211: Initially encode the user's behavior sequence. Each behavior data point (such as browsing, clicking, purchasing, etc.) is converted into a high-dimensional vector representation, which contains the characteristics of the behavior itself and also embeds time information and context information. The behavior sequence is converted into a continuous vector sequence by encoding;

[0113] S20212: The temporal convolutional network is used to extract hierarchical features of behavior sequences. Through a multi-layer convolutional structure, the temporal dependencies in the behavior sequences are extracted layer by layer. Each layer of convolution extracts local features of the input sequence and captures longer time spans through dilated convolutions.

[0114] S20123: Transformer calculates the correlation between each behavior data point and other data points through the self-attention mechanism, and identifies the global pattern in the behavior sequence; the global attention mechanism captures the key turning points in user behavior, such as the change in behavior pattern from browsing to purchasing; under the synergy of the temporal convolutional network and Transformer, the behavior sequence is dynamically reorganized; the representation of the behavior sequence is dynamically adjusted to identify the key turning points in the behavior sequence;

[0115] When users switch from browsing to purchasing, the change in behavior pattern is automatically identified and used as an important feature in the user portrait.

[0116] S20124: The global attention mechanism identifies the main turning points in the behavior sequence, and the local attention mechanism deeply analyzes the detailed features of the turning points, captures the key changes in user behavior, and constructs a user profile.

[0117] The working principle and beneficial effects of the above technical solution are as follows: This embodiment first performs initial encoding on the user's behavior sequence, and each behavior data point (such as browsing, clicking, purchasing, etc.) is converted into a high-dimensional vector representation, which contains the characteristics of the behavior itself and also embeds time information and context information; the encoded behavior sequence is converted into a continuous vector sequence; secondly, the temporal convolutional network is used to perform hierarchical feature extraction on the behavior sequence, and the time dependency in the behavior sequence is extracted layer by layer through a multi-layer convolution structure; each layer of convolution extracts local features of the input sequence and captures a longer time span through dilated convolution; then the Transformer calculates the relationship between each behavior data point and other data points through the self-attention mechanism. The correlation between data points identifies the global pattern in the behavior sequence; the global attention mechanism captures the key turning points in user behavior, such as the change in behavior pattern from browsing to purchasing; under the synergy of the temporal convolutional network and Transformer, the behavior sequence is dynamically reorganized; the representation of the behavior sequence is dynamically adjusted to identify the key turning points in the behavior sequence; among them, when the user switches from browsing to purchasing, the change in this behavior pattern is automatically identified and used as an important feature in the user portrait; finally, the global attention mechanism identifies the main turning points in the behavior sequence, and the local attention mechanism deeply analyzes the detailed features of the turning points, captures the key changes in user behavior, and constructs a user portrait. Step S20211 of the above scheme is the initial encoding of the behavior sequence, which converts the user's behavior data (such as browsing, clicking, purchasing, etc.) into a high-dimensional vector representation, which not only contains the characteristics of the behavior itself, but also embeds time information and context information; through this encoding method, the behavior sequence is converted into a continuous vector sequence, which provides a basis for subsequent feature extraction and analysis. Significance achieved: The initial encoding provides a unified representation for subsequent hierarchical feature extraction and global pattern recognition, ensuring the continuity and consistency of the behavior sequence; it lays the foundation for analysis, enabling the system to better understand the dynamic changes in user behavior. Step S20212 The hierarchical feature extraction of the temporal convolutional network extracts the temporal dependencies in the behavior sequence layer by layer through a multi-layer convolutional structure; each layer of convolution extracts local features of the input sequence and captures longer time spans through dilated convolutions; this hierarchical feature extraction method enables the model to identify changing trends in user behavior at different time scales. Significance achieved: The hierarchical feature extraction of the temporal convolutional network can effectively capture the temporal dependencies in user behavior and identify local patterns and trends in behavior sequences; it provides rich local features for global pattern recognition and enhances the expressiveness and interpretability of the model.Step S20123 Transformer's global attention mechanism and dynamic reorganization. Transformer calculates the correlation between each behavior data point and other data points through the self-attention mechanism, and identifies the global pattern in the behavior sequence. The global attention mechanism captures the key turning points in user behavior, such as the change in behavior pattern from browsing to purchasing. Under the synergy of the temporal convolutional network and Transformer, the behavior sequence is dynamically reorganized, the representation of the behavior sequence is dynamically adjusted, and the key turning points in the behavior sequence are identified. Significance achieved: Transformer's global attention mechanism can capture the global pattern and key turning points in the behavior sequence, enhancing the model's global understanding of user behavior; through dynamic reorganization, the representation of the behavior sequence can be adjusted in real time, and the key changes in user behavior can be identified, thus providing an important basis for building accurate user portraits. Step S20124 Collaborative analysis of global and local attention mechanisms. The global attention mechanism identifies the main turning points in the behavior sequence, and the local attention mechanism deeply analyzes the detailed features of the turning points to capture the key changes in user behavior. Through this collaborative analysis, the key turning points in user behavior can be more accurately identified and used as important features in user portraits. Significance achieved: The collaborative analysis of global and local attention mechanisms can effectively capture the key turning points and detailed features in user behavior, enhancing the accuracy and interpretability of the model; it provides important support for building a comprehensive and accurate user portrait, enabling the system to better understand user behavior patterns and preferences.

[0118] In summary, this embodiment can effectively identify the key turning points in user behavior and use them as important features in user portraits. It not only enhances the expressiveness and interpretability of the model, but also provides strong technical support for building comprehensive and accurate user portraits. Ultimately, it can better understand the user's behavior patterns and preferences, and provide users with more personalized and accurate services.

[0119] Example 7: Figure 8 As shown, based on Example 5, the process of identifying implicit associations between different modal data provided by the embodiment of the present invention includes the following steps:

[0120] S20221: Extract semantic features from texts to capture users’ emotional tendencies, evaluation priorities, and potential intentions; model users’ behavior paths to identify the correlation and transfer patterns between behavior nodes; extract features from the time series of user behaviors to capture periodicity, trends, and mutation points; extract users’ geographic location features and combine them with spatial analysis models to identify users’ geographic preferences and activity ranges;

[0121] S20222: Build a cross-modal global attention network to map the features of different modalities into a unified high-dimensional space, identify the global correlation between modalities through the self-attention mechanism, and identify the overall behavior trend of users by combining the sentiment characteristics of text comments with the transfer pattern of behavior trajectories; use the multi-head attention mechanism to analyze the fine-grained correlation between modalities, combine the mutation points of time series with the activity range of geographical location, and identify the behavior preferences of users in specific scenarios;

[0122] S20223: Through multimodal autoencoders, implicit associations between modalities are mined, potential relationships between modalities are reconstructed, and the emotional tendencies of users in specific scenarios are identified by combining the sentiment characteristics of text comments with the activity range of geographic locations. The dynamic attention mechanism is used to reorganize multimodal features in real time to enhance the weights of key features.

[0123] Among them, the equation of the implicit association between modalities is mined through a multimodal autoencoder. The multimodal autoencoder is usually composed of multiple encoders and decoders, one encoder and one decoder for each modality; the encoder maps the input modal data to the latent space, and the decoder maps the latent representation back to the original modal space; by minimizing the reconstruction error, the model can learn the implicit association between modalities;

[0124] The goal of the multimodal autoencoder is to minimize the reconstruction error of each modality simultaneously and capture the correlation between modalities through a shared potential representation; assuming there are M modalities and the input of each modality is x m , the corresponding encoder is E m , the decoder is D m , the potential is represented by z, and the objective function It is expressed as:

[0125]

[0126] Where N is the number of samples; is the input of the mth mode of the ith sample; is the potential representation of the i-th sample; D m (z (j) ) is the reconstruction result of the decoder mapping the latent representation back to the m-th modality; is a regularization term used to constrain the complexity of the potential representation; λ is the regularization coefficient;

[0127] Joint Optimization of Latent Representations In order to capture the implicit correlation between modalities, the latent representation z needs to be jointly optimized through the encoders of all modalities; assuming that the latent representation z is jointly generated by the encoders of all modalities, it can be expressed as:

[0128] z=Concat(E1(x1),E2(x2),…,EM (x M ))

[0129] In the formula, Concat means concatenating the encoding results of all modes;

[0130] In order to further explore the implicit associations between modalities, a complex equation is introduced to describe the generation process of the potential representation; assuming that the potential representation z is generated by a nonlinear transformation, and the encoder output of each modality is combined by a weighted sum, expressed as:

[0131]

[0132] Where W m is the weight matrix of the mth mode; b is the bias vector; σ is a nonlinear activation function (such as ReLU or Tanh).

[0133] The encoding results of different modalities are combined by weighted sum, and the final potential representation z is generated by a nonlinear activation function;

[0134] Dynamic Attention Mechanism The dynamic attention mechanism is introduced. Assuming that the encoding result of each modality is weighted by an attention mechanism, it can be expressed as:

[0135]

[0136] In the formula, α m is the attention weight of the mth modality, calculated through an attention mechanism:

[0137]

[0138] In the formula, u m is the attention vector of the mth modality; v is the global attention vector; in this way, the model can dynamically adjust the weights of different modalities to enhance the representation ability of key features; through the multimodal autoencoder, the implicit associations between different modal data can be effectively mined; through the equation, the generation process of the potential representation can be described, and the weights of key features can be enhanced through the dynamic attention mechanism; it has broad application prospects in multimodal data analysis.

[0139] The working principle and beneficial effects of the above technical solution are as follows: the present embodiment first extracts semantic features from the text to capture the user's emotional tendency, evaluation focus and potential intention; models the user's behavior path to identify the correlation and transfer mode between behavior nodes; extracts features from the time series of the user's behavior to capture periodicity, trend and mutation points; extracts the user's geographical location features, and combines the spatial analysis model to identify the user's geographical preference and activity range; secondly, constructs a cross-modal global attention network to map the features of different modalities to a unified high-dimensional space, identifies the global correlation between modalities through the self-attention mechanism, and identifies the user's overall behavior trend by combining the emotional features of text comments with the transfer mode of behavioral trajectories; uses a multi-head attention mechanism to analyze the fine-grained correlation between modalities, combines the mutation points of the time series with the activity range of the geographical location, and identifies the user's behavioral preference in a specific scenario; finally, uses a multimodal autoencoder to mine the implicit correlation between modalities, remodels the potential relationship between modalities, and combines the emotional features of text comments with the activity range of the geographical location to identify the user's emotional tendency in a specific scenario; uses a dynamic attention mechanism to reorganize multimodal features in real time to enhance the weight of key features. Step S20221 of the above scheme extracts features from multimodal data. Through pre-trained language models such as BERT, the semantic information in the text is deeply analyzed to capture the user's emotional tendency, evaluation focus and potential intention; basic features at the emotional and intention levels are provided for cross-modal association analysis to help identify the user's core needs and behavioral motivations; the user's behavior path is dynamically modeled using graph neural networks (GNNs) to identify the correlation and transfer patterns between behavior nodes; the logical chain and pattern changes of user behavior are revealed to provide structured support for the analysis of global behavioral trends; time series features are extracted using time convolutional networks (TCNs) to capture periodicity, trends and mutation points; the dynamic change rules of user behavior are identified to provide an important basis for the time dimension for scene perception and behavior prediction; the user's geographical preferences and activity range are extracted through geographic information system (GIS) technology, and the user's geographical features are identified in combination with spatial analysis models; spatial dimension support is provided for cross-modal association analysis to help understand the user's behavior pattern in a specific geographical environment. Step S20222 Global and local association analysis of cross-modality maps the features of different modalities to a unified high-dimensional space, and the global correlation between modalities is identified through the self-attention mechanism. Significance: Grasp the user's behavioral trends as a whole, for example, combine the sentiment characteristics of text comments with the transfer pattern of behavioral trajectories to identify the user's overall consumption tendency or interest areas; through the multi-head attention mechanism, deeply analyze the fine-grained associations between modalities, for example, combine the mutation points of the time series with the activity range of the geographical location to identify the user's behavioral preferences in specific scenarios. Significance: Reveal the local characteristics and scenario-based preferences of user behavior, and provide precise support for the construction of personalized portraits.Step S20223: Implicit association mining and feature enhancement, using multimodal autoencoders to reconstruct the potential relationship between modalities, for example, combining the sentiment characteristics of text comments with the activity range of geographic locations to identify the user's sentiment tendency in a specific scenario. Significance: Mining implicit associations between modalities to reveal the deep motivations and emotional drivers behind user behavior; real-time reorganization of multimodal features to enhance the weight of key features, for example, when a user frequently visits a brand within a specific time period, the weight of the brand in the portrait is automatically increased. Significance: Dynamically adjust the feature weight of the user portrait to ensure the real-time and accuracy of the portrait, and meet the needs of personalized recommendations and behavior prediction.

[0140] This embodiment realizes the efficient fusion of multimodal data and the deep mining of implicit associations. The specific meanings are as follows: The fusion and association analysis of multimodal data can comprehensively characterize user characteristics from multiple dimensions such as emotion, behavior, time, and space, and build a more three-dimensional and dynamic user portrait. Enhance scene perception and behavior prediction capabilities. By capturing the scene information of the user (such as time, location, device type, etc.) and associating it with multimodal data, it can more accurately predict the user's behavior trends and needs. Support personalized services and decision optimization, implicit association mining and feature enhancement, provide strong technical support for personalized recommendations, precision marketing, and intelligent decision-making, and improve user experience and service efficiency. Promote the innovative application of multimodal technology in the field of user portraits. Through innovative multimodal fusion and association analysis methods, it provides a new technical paradigm for the industry and promotes the further development of user portrait technology.

[0141] Example 8: Fig. 9 As shown, based on Example 4, the process of constructing a knowledge graph of a personalized marketing plan provided by the embodiment of the present invention includes the following steps:

[0142] S2031: Identify key entities including user behavior type, time pattern, location preference and text intent from user portraits and marketing strategies. Key entities are used as nodes of the knowledge graph. Through deep analysis of large language models, identify and extract relationships between entities. Relationships are used as edges of the knowledge graph. Attributes are assigned to each entity and relationship.

[0143] Among them, relationships include the association between user behavior types and marketing plans, the matching of time patterns and time-effectiveness plans, etc., and attributes include behavior frequency, time distribution pattern, geographic location preference intensity, and confidence in text intent, etc.;

[0144] S2032: Convert the constructed knowledge graph into graph structure data, where nodes represent entities, edges represent relationships, and the attributes of nodes and edges represent the characteristics of entities and relationships; map the graph structure data to a low-dimensional vector space, capture the local and global relationships between nodes, and generate an embedding vector for each node; aggregate the information of neighbor nodes through the message passing mechanism of the graph neural network, update the embedding vector of each node, and capture the complex association relationship between nodes;

[0145] S2033: Based on the output of the graph neural network, potential association rules are mined from the graph structure, and personalized marketing plans that match the user portrait are screened out from the knowledge base, so that the plans match behavior, time, location, and intention.

[0146] The working principle and beneficial effects of the above technical solution are as follows: This embodiment first identifies key entities including user behavior type, time pattern, geographic location preference and text intent from user portraits and marketing strategies, and the key entities serve as nodes of the knowledge graph; through in-depth analysis of the large language model, the relationship between entities is identified and extracted, and the relationship serves as the edge of the knowledge graph; attributes are assigned to each entity and relationship; among which, the relationship includes the association between user behavior type and marketing plan, the matching of time pattern and timeliness plan, etc., and the attributes include behavior frequency, time distribution law, geographic location preference strength and text intent confidence, etc.; secondly, the constructed knowledge The recognition graph is converted into graph structure data, where nodes represent entities, edges represent relationships, and the attributes of nodes and edges represent the characteristics of entities and relationships; the graph structure data is mapped to a low-dimensional vector space to capture the local and global relationships between nodes and generate an embedding vector for each node; through the message passing mechanism of the graph neural network, the information of neighboring nodes is aggregated, the embedding vector of each node is updated, and the complex association relationships between nodes are captured; finally, based on the output of the graph neural network, potential association rules are mined from the graph structure, and personalized marketing plans that match the user portrait are screened out from the knowledge base, so that the plan matches behavior, time, location and intention. Step S2031 of the above scheme is to construct the entities and relationships of the knowledge graph. By extracting key entities (such as user behavior type, time pattern, geographic location preference, and text intent) from user portraits and marketing strategies, the nodes of the knowledge graph are constructed to ensure that the basic data of the knowledge graph is comprehensive and accurate; through in-depth analysis of large language models, the relationships between entities are identified (such as the association between user behavior and marketing plans, and the matching of time patterns and timeliness plans), the edges of the knowledge graph are constructed to form logical connections between entities; attributes are given to each entity and relationship (such as behavior frequency, time distribution pattern, geographic location preference intensity, and confidence in text intent) to enhance the expressiveness of the knowledge graph and provide rich feature support for subsequent in-depth analysis. The significance achieved: Through the construction of multi-dimensional entities and relationships, a comprehensive and dynamic user portrait is formed, which can accurately portray the personalized characteristics of users; a high-quality data foundation is provided for graph neural network modeling and association rule mining, ensuring the structuring and semanticization of the knowledge graph. Step S2032: Graph structure representation and embedding of graph neural network, converting the knowledge graph into graph structure data, where nodes represent entities, edges represent relationships, and attributes of nodes and edges represent features, forming a high-dimensional graph data structure; using graph embedding technology (such as GraphSAGE, GAT, etc.), mapping graph structure data to a low-dimensional vector space, capturing local and global relationships between nodes, and generating an embedding vector for each node, reducing computational complexity and improving expression capabilities; using the message passing mechanism of the graph neural network, aggregating information about neighboring nodes, updating the embedding vector of each node, capturing complex associations between nodes, and enhancing the expression capabilities of the graph structure.Significance achieved: Through the message passing mechanism of the graph neural network, it is possible to capture complex associations between entities, such as the implicit association between user behavior and marketing plans, and improve the semantic expression ability of the knowledge graph; through the graph embedding technology, the high-dimensional graph structure data is reduced and converted into a low-dimensional vector, which is convenient for calculation and analysis, while retaining the global and local characteristics of the graph structure. Step S2033 Association rule mining and personalized marketing plan recommendation, based on the output of the graph neural network, using association rule mining algorithms (such as Apriori, FP-Growth, etc.), to mine potential association rules from the graph structure, such as "user behavior type A is highly associated with marketing plan B", "time pattern C matches timeliness plan D", etc.; according to the mined association rules, screen out personalized marketing plans that match the user portrait from the knowledge base to ensure behavior matching, time matching, location matching and intention matching of the plan. Significance achieved: Through association rule mining, the potential relationship between user portraits and marketing plans can be accurately identified, highly personalized marketing plans can be provided, and user satisfaction and conversion rates can be improved; based on real-time data and dynamic adjustments of graph neural networks, personalized marketing plans can be continuously optimized to ensure the real-time and adaptability of the plans and enhance marketing effectiveness.

[0147] In summary, this embodiment constructs a comprehensive and dynamic knowledge graph, providing a high-quality data foundation for in-depth analysis; through the graph structure representation and embedding of the graph neural network, it captures the complex associations between entities, improving the expression ability and computing efficiency of the knowledge graph; through association rule mining and personalized solution recommendation, it realizes accurate personalized marketing, improves the accuracy of marketing solutions and user satisfaction. Finally, it realizes a closed loop from data to knowledge, from knowledge to decision-making, providing strong technical support for personalized marketing.

[0148] Example 9: Fig.10 As shown, based on Example 1, the process of optimizing the decision-making process and marketing plan provided by the embodiment of the present invention includes the following steps:

[0149] S301: Extract key features from the logs of the decision-making process, including user behavior patterns, time series features, and geographic location information; use word embedding to convert the extracted key features into high-dimensional vector representations to capture the semantic information and contextual relationships in the text; store the vectorized data in a vector library to form a knowledge base;

[0150] S302: Use the continuously optimized big language model to perform real-time analysis and reasoning on the data in the vector library; by comparing historical decision data and current user feedback, the big language automatically adjusts and optimizes decision parameters; based on the optimized big language model, user behavior and marketing plans are matched in real time;

[0151] S303: Through the feedback loop mechanism, collect user feedback data on marketing activities in real time, including indicators such as click-through rate, conversion rate and user satisfaction; combine the feedback data with the historical data in the vector library to continuously optimize the large language model and decision-making process.

[0152] The working principle and beneficial effects of the above technical solution are as follows: This embodiment first extracts key features from the logs in the decision-making process, including user behavior patterns, time series features, and geographic location information; word embedding is used to convert the extracted features into high-dimensional vector representations to capture semantic information and contextual relationships in the text; the vectorized data is stored in a vector library to form a knowledge base; secondly, a continuously optimized large language model is used to perform real-time analysis and reasoning on the data in the vector library; by comparing historical decision data with current user feedback, the large language automatically adjusts and optimizes decision parameters; based on the optimized large language model, user behavior and marketing plans are matched in real time; finally, through a feedback loop mechanism, user feedback data on marketing activities is collected in real time, including indicators such as click-through rate, conversion rate, and user satisfaction; the feedback data is combined with historical data in the vector library to continuously optimize the large language model and decision-making process. Step S301 of the above scheme extracts and vectorizes key features such as user behavior patterns, time series features, and geographic location information from decision logs, which can fully capture the user's dynamic behavior and preferences; these features are converted into high-dimensional vectors through word embedding technology, which not only retains the semantic information of the original data, but also captures the contextual relationship in the text, making the data richer and more three-dimensional; the vectorized data is stored in the vector library to form a knowledge base. Significance achieved: Through feature extraction and vectorization, the decision-making process is transformed from traditional experience-driven to data-driven, which improves the scientificity and accuracy of decision-making; the establishment of the vector library enables historical data to be effectively stored and reused, providing a valuable reference for future decisions. Step S302 Real-time analysis and reasoning of the large language model, using the continuously optimized large language model, real-time analysis and reasoning of the data in the vector library can quickly respond to market changes and user needs; by comparing historical decision data and current user feedback, the large language model automatically adjusts and optimizes decision parameters, making decisions more accurate and flexible; based on the optimized large language model, real-time matching of user behavior and marketing plans ensures the efficient execution of marketing activities. Significance achieved: The application of large language models makes the decision-making process more intelligent, and can automatically adapt to market changes and user needs, improving the efficiency and effectiveness of decision-making; real-time matching of user behavior and marketing plans can achieve personalized marketing and improve user experience and satisfaction. Step S303 feedback loop mechanism, through the feedback loop mechanism, real-time collection of user feedback data on marketing activities, including indicators such as click-through rate, conversion rate and user satisfaction, to provide real-time data support for decision optimization; the feedback data is combined with historical data in the vector library to continuously optimize the large language model and decision-making process, so that decisions and marketing plans are more in line with user needs. Significance achieved: The feedback loop mechanism enables the decision-making process and marketing plan to be continuously improved, constantly adapt to changes in the market and users, and maintain competitiveness; through real-time feedback and optimization, marketing activities are more user-centric, improving user participation and loyalty.

[0153] In summary, the process of optimizing decision-making processes and marketing plans in this embodiment realizes a complete closed loop from data-driven to intelligent decision-making and then to continuous improvement. It jointly promotes the efficiency and precision of decision-making and marketing, and ultimately achieves the improvement of corporate competitiveness and user satisfaction.

[0154] Example 10: Fig.11 As shown, based on Example 9, the process of converting the extracted key features into high-dimensional vector representations by using word embedding provided in the embodiment of the present invention includes the following steps:

[0155] S3011: Fuse multimodal data such as user behavior patterns, time series features, and geographic location information to form a unified feature space; extract more time-related features through the sliding window technology of the time series;

[0156] S3012: Use deep semantic embedding to map preprocessed features into a high-dimensional semantic space, and use self-supervised learning to capture the deep semantic relationship between key features; use the Transformer architecture to model the contextual relationship of key features; and use the multi-head self-attention mechanism to capture the dynamic changes and mutual influence of features in different contexts;

[0157] S3013: Based on semantic capture and context modeling, nonlinear mapping technology is used to map key features from semantic space to high-dimensional vector space.

[0158] The working principle and beneficial effects of the above technical solution are as follows: the embodiment first fuses multimodal data such as user behavior patterns, time series features and geographic location information to form a unified feature space; extracts more time-related features through the sliding window technology of the time series; secondly, uses deep semantic embedding to map the preprocessed features to a high-dimensional semantic space, and captures the deep semantic relationship between key features by using self-supervised learning; uses the Transformer architecture to model the contextual relationship of key features; captures the dynamic changes and mutual influence of features in different contexts through the multi-head self-attention mechanism; finally, based on semantic capture and context modeling, nonlinear mapping technology is used to map key features from the semantic space to a high-dimensional vector space. Step S3011 of the above solution is multimodal data fusion and time series feature extraction, which forms a unified feature space by fusing multimodal data such as user behavior patterns, time series features and geographic location information; can integrate the advantages of different types of data to enhance the expressiveness and diversity of features; extracts more time-related features through the sliding window technology of the time series; can capture trends, periodicity and sudden changes in time series, and enhance the time sensitivity and prediction ability of features. Significance achieved: Multimodal data fusion and time series feature extraction make the feature space more comprehensive and rich, which can better reflect the real behavior of users and environmental changes; by enhancing the temporal correlation and diversity of features, the prediction accuracy and stability of the model are significantly improved, providing a more reliable foundation for subsequent decision optimization and marketing plan matching. Step S3012 Deep semantic embedding and context modeling, through self-supervised learning methods such as contrastive learning or masked language model, the preprocessed features are mapped to a high-dimensional semantic space; it can capture the deep semantic relationship between features and enhance the semantic expression ability of features; using the Transformer architecture, the contextual relationship of key features is modeled; through the multi-head self-attention mechanism, the dynamic changes and mutual influence of features in different contexts are captured, and the context sensitivity and adaptability of features are enhanced. Significance achieved: Deep semantic embedding and context modeling enable features to express not only surface numerical information, but also deep semantic and contextual relationships, enhancing the semantic understanding ability of the model; by enhancing the context sensitivity and adaptability of features, the generalization ability and robustness of the model are significantly improved, and it can better cope with complex and changing environments and user behaviors. Step S3013 nonlinear mapping and high-dimensional vector generation, based on semantic capture and context modeling, uses nonlinear mapping technology to map key features from semantic space to high-dimensional vector space; can fully express the complexity and diversity of features, enhance the expressive power and storage efficiency of vectors; through nonlinear mapping, generate high-dimensional vector representation, so that features can be more accurately and comprehensively expressed in high-dimensional space, enhance the discrimination and expressiveness of vectors.Significance achieved: Nonlinear mapping and high-dimensional vector generation enable features to be expressed more accurately and comprehensively in high-dimensional space, enhancing the expressiveness and discrimination of vectors. By enhancing the expressiveness and storage efficiency of vectors, the decision-making ability and response speed of the model are significantly improved, providing more powerful technical support for real-time decision-making and marketing plan matching.

[0159] In summary, this embodiment uses word embedding to convert the extracted key features into high-dimensional vector representations, which not only enhances the comprehensiveness, semantic understanding and expression capabilities of the features, but also significantly improves the prediction accuracy, generalization and decision-making capabilities of the model. It ensures the high expression ability and high storage efficiency of the feature vector, and provides solid technical support for decision optimization and marketing plan matching.

[0160] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention belong to the scope of equivalent technologies of the present invention, the present invention is also intended to include these modifications and variations.

Claims

1. A marketing strategy optimization method based on large language model analysis and evaluation drive, characterized in that: The following steps are involved: Obtain historical data including user behavior data, tag data and statistical data, analyze and build basic marketing strategies; basic marketing strategies include triggering methods and marketing plans. The triggering methods are set based on user behavior, tags and large language models. When the user's real-time content meets the triggering rules, the corresponding marketing plan is automatically triggered; the marketing plan is stored in the database; Collect user content data in real time through Kafka, and use large language models to conduct in-depth analysis of user content, including the disassembly and reorganization of multi-dimensional information such as behavior tags, time, location, and text data; Based on the deep analysis results of the large language model and basic rule matching, personalized marketing plans are matched from the knowledge base; Through the feedback loop mechanism, users’ feedback on marketing activities is collected, and the feedback results are stored in the knowledge base and used to optimize marketing strategies and large language models. At the same time, the decision-making process is logged, textualized and vectorized, and stored in the vector library. Combined with the continuously optimized large language model, the decision-making process and marketing plan are optimized to achieve dynamic adjustment and matching.

2. The marketing strategy optimization method based on large language model analysis and evaluation drive as claimed in claim 1, characterized in that: The process of building a basic marketing strategy includes the following steps: Integrate user behavior data, label data and statistical data and perform preprocessing to extract key features of user behavior including click frequency, browsing time and purchase path, and construct feature vectors; Use clustering algorithms to group user behavior data and identify the behavior patterns and preferences of different user groups; Based on the clustering results, combined with label data and statistical data, a user profile is constructed, including demographic characteristics, behavioral characteristics, and interest preferences; Based on user behavior data, label data and large language models, a trigger rule engine is designed to set trigger conditions and logic, including behavior thresholds, time windows and label matching; based on user portraits and trigger rules, a basic marketing strategy is built, including trigger methods, marketing plans and response mechanisms.

3. The marketing strategy optimization method based on large language model analysis and evaluation drive as claimed in claim 2, characterized in that: The process of setting up trigger methods, marketing plans, and response mechanisms in basic marketing strategies includes the following steps: Based on user behavior data, set the trigger thresholds for key behaviors such as clicks, browsing, and purchases; determine the trigger time window through time series analysis; design multi-level tag matching logic based on the tag data in the user portrait; and construct a trigger matrix that includes multiple dimensions of behavior, time, and tags; Combined with a large language model, it generates personalized marketing content including text, images, and videos; Based on the channel preferences in user portraits, marketing information reaches users through the best channels through multi-channel collaboration strategies; Build a response engine based on real-time data streams, optimize trigger rules and marketing plans through real-time feedback data; predict users' next behavior based on historical data and real-time behavior, and design response strategies in advance.

4. The marketing strategy optimization method based on large language model analysis and evaluation drive according to claim 1, characterized in that: The process of matching personalized marketing plans from the knowledge base includes the following steps: Using the Kafka distributed stream processing platform, users’ content data from various channels is collected in real time, including behavior data, time data, location data, and text data. Use a large language model to conduct multi-dimensional in-depth analysis of the collected user content data, including behavior tags, time patterns, location features, and text intent; Through the sequence modeling and attention mechanism of the large language model, the user data can be accurately disassembled and reorganized to build a comprehensive user portrait. Based on the in-depth analysis results of the large language model and combined with the basic matching rules predefined in the basic marketing strategy, potential personalized marketing plans are screened from the knowledge base, including behavior matching, time matching, location matching, and intention matching. A knowledge graph of personalized marketing plans is constructed, and the relationship between personalized marketing plans is modeled using graph neural networks to explore potential related plans.

5. The marketing strategy optimization method based on large language model analysis and evaluation drive as claimed in claim 4, characterized in that: The process of building a comprehensive user profile includes the following steps: The large language model calculates the attention weight of each data point in real time based on the contextual information of the user's behavior. Through the hybrid architecture of the temporal convolutional network and Transformer, the large language model dynamically reorganizes the user's behavior sequence and identifies the key turning points in the user's behavior. Deeply integrate users’ text comments, behavior trajectories, time series, and geographic location multimodal data to identify implicit associations between different modal data; identify users’ core behavior patterns through a global attention mechanism; and deeply analyze the detailed characteristics of user behavior at the local level through a local attention mechanism; Capture the user's scene information in real time and analyze it in correlation with user behavior.

6. The marketing strategy optimization method based on large language model analysis and evaluation drive as claimed in claim 5, characterized in that: The process of identifying key turning points in user behavior involves the following steps: Initially encode the user's behavior sequence. Each behavior data point is converted into a high-dimensional vector representation, which contains the characteristics of the behavior itself and also embeds time information and context information. The behavior sequence is converted into a continuous vector sequence by encoding. The temporal convolutional network is used to extract hierarchical features of behavior sequences. Through a multi-layer convolutional structure, the temporal dependencies in the behavior sequences are extracted layer by layer. Each layer of convolution extracts local features of the input sequence and captures longer time spans through dilated convolutions. Transformer calculates the correlation between each behavior data point and other data points through the self-attention mechanism, and identifies the global pattern in the behavior sequence; the global attention mechanism captures the key turning points in user behavior; under the synergy of the temporal convolutional network and Transformer, the behavior sequence is dynamically reorganized; the representation of the behavior sequence is dynamically adjusted to identify the key turning points in the behavior sequence; The global attention mechanism identifies the main turning points in the behavior sequence, and the local attention mechanism deeply analyzes the detailed features of the turning points, captures the key changes in user behavior, and constructs a user portrait.

7. The marketing strategy optimization method based on large language model analysis and evaluation drive as claimed in claim 4, characterized in that: The process of building a knowledge graph for personalized marketing solutions includes the following steps: Identify key entities including user behavior types, time patterns, location preferences, and text intent from user portraits and marketing strategies. Use key entities as nodes in the knowledge graph. Use deep analysis of large language models to identify and extract relationships between entities. Use relationships as edges in the knowledge graph. Assign attributes to each entity and relationship. The constructed knowledge graph is converted into graph structure data, where nodes represent entities, edges represent relationships, and the attributes of nodes and edges represent the characteristics of entities and relationships; the graph structure data is mapped to a low-dimensional vector space to capture the local and global relationships between nodes and generate an embedding vector for each node; through the message passing mechanism of the graph neural network, the information of neighbor nodes is aggregated, the embedding vector of each node is updated, and the complex association relationship between nodes is captured; Based on the output of the graph neural network, potential association rules are mined from the graph structure, and personalized marketing plans that match the user portrait are screened out from the knowledge base, so that the plan matches behavior, time, location and intention.

8. The marketing strategy optimization method based on large language model analysis and evaluation drive as claimed in claim 1, characterized in that: The process of optimizing decision-making processes and marketing plans includes the following steps: Extract key features from the logs of the decision-making process, including user behavior patterns, time series features, and geographic location information; use word embedding to convert the extracted key features into high-dimensional vector representations to capture the semantic information and contextual relationships in the text; store the vectorized data in a vector library to form a knowledge base; Use the continuously optimized large language model to perform real-time analysis and reasoning on the data in the vector library; By comparing historical decision data and current user feedback, Big Language automatically adjusts and optimizes decision parameters; based on the optimized Big Language model, it matches user behavior and marketing plans in real time; Through the feedback loop mechanism, users’ feedback data on marketing activities is collected in real time, and the feedback data is combined with historical data in the vector library to continuously optimize the large language model and decision-making process.

9. The marketing strategy optimization method based on large language model analysis and evaluation drive as claimed in claim 8, characterized in that: Feedback data includes click-through rate, conversion rate and user satisfaction indicators.

10. The marketing strategy optimization method based on large language model analysis and evaluation drive according to claim 8, characterized in that: The process of converting the extracted key features into high-dimensional vector representations using word embedding includes the following steps: The multimodal data of user behavior patterns, time series features and geographic location information are integrated to form a unified feature space; more time-related features are extracted through the sliding window technology of time series; Use deep semantic embedding to map the preprocessed features into a high-dimensional semantic space, and use self-supervised learning to capture the deep semantic relationship between key features. Use the Transformer architecture to model the contextual relationship of key features. Use the multi-head self-attention mechanism to capture the dynamic changes and mutual influence of features in different contexts. Based on semantic capture and context modeling, nonlinear mapping technology is used to map key features from semantic space to high-dimensional vector space.

Images