Auxiliary marketing method and system based on large language model
By introducing auxiliary marketing methods and systems based on large language models into the marketing system, combining undirected weighted graph technology and RAG system, the shortcomings in personalization and prediction accuracy of traditional marketing methods are solved, and more efficient and accurate marketing effects are achieved.
Patent Information
- Application Number
- CN202510219856.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-26
- Publication Date
- 2025-06-13
AI Technical Summary
The existing technology has poor results in marketing data prediction and personalized marketing, especially in small educational institutions. It is difficult for traditional marketing methods to achieve true personalized customization, resulting in marketing information being ignored or deleted by users.
The auxiliary marketing methods and systems based on large language models are adopted, combined with undirected weighted graph technology, text Embedding features rich in recommendation-related information are extracted, and high-potential users are recalled using a multi-way recall algorithm, and user purchase data is predicted through the fusion model, and intelligent optimization and intention recognition are carried out in combination with the RAG system.
It significantly improves the accuracy and personalization of marketing data prediction, avoids noise interference, improves marketing effectiveness, and provides more accurate and efficient marketing support for the education industry.
Smart Images
Figure CN120146889A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of marketing systems, and in particular, to an assisted marketing method and system based on a large language model. Background Art
[0002] The accurate prediction of marketing data plays a crucial role in the operation and decision-making in the education field and many other fields. In other industries, this prediction ability has become the key for enterprises to optimize inventory management, improve the accuracy and effectiveness of marketing activities, and thus significantly enhance the market competitiveness of brands; while in the education field, accurate marketing data prediction provides educational institutions with a key to open the door to market demands and student preferences, enabling them to improve teaching effects more pertinently, thereby strengthening brand influence and attracting more students to join. In addition, through in-depth analysis of marketing data, educational institutions can formulate more scientific and efficient operation strategies, so as to comprehensively improve operation efficiency and achieve the maximization of profits. In the education field, the application modes of marketing data prediction technologies vary according to the scale and resources of educational institutions. Small educational institutions, limited by technology and resources, often adopt traditional marketing means with lower costs, such as mass text messaging and market surveys, while large educational companies with a large user base and rich data resources rely on cutting-edge technologies such as big data, artificial intelligence, and machine learning to collect and analyze massive amounts of user behavior and market trend data, and combine advanced algorithms to provide accurate and timely prediction results for enterprises. With the rapid progress of artificial intelligence technology, robot customer service has become a key strategy for many enterprises to improve customer service quality and reduce operation costs. By virtue of the excellent memory and profound understanding ability of large models, the limitations of traditional dialogue analysis have been broken through. Through extensive learning and training, large models have not only built a huge knowledge system, but also mastered rich language patterns and accurate semantic rules, and can accurately grasp the deep meaning of each conversation, opening up a broad space for the further optimization and upgrading of robot customer service. Looking ahead, the assisted marketing system based on large language models shows extremely broad development potential in the education field. Facing the increasingly fierce market competition and the diversified trend of consumer demands, the demand of educational institutions for such advanced systems will continue to rise. With the continuous innovation of technology, future marketing data prediction will achieve a higher degree of personalization, and robot customer service will also move towards a higher level of intelligence, thus providing more comprehensive and accurate data support for educational institutions and helping the education industry enter a new era of more efficient and accurate development.
[0003] Currently, the marketing strategies in the education industry are showing a diversified development trend. For small educational institutions with relatively limited resources and technology, they rely more on traditional and cost-effective marketing methods, such as SMS pushing and market research. Although these methods are straightforward, in a highly competitive market environment, their effects are often unsatisfactory. On the contrary, large educational institutions, relying on their large user base and rich data resources, attach more importance to the in-depth mining and analysis of data. By widely collecting multi-dimensional data such as user behavior and market trends, they use advanced statistical techniques for data cleaning, integration, and extract statistical features with predictive value. On this basis, they use mature machine learning algorithms such as logistic regression and decision trees to deeply mine and accurately predict the data, so as to formulate more refined and efficient marketing strategies. This marketing strategy based on big data and machine learning not only significantly improves the marketing effect but also better meets the personalized needs of consumers. In the application of robot customer service, small educational institutions usually adopt preset rules and keyword matching technology to build a detailed question-answering database and achieve efficient and accurate responses through rapid retrieval and matching. Large educational institutions, on the other hand, tend to use advanced large language models to deeply understand the semantics and emotions of user conversations and provide more intelligent and user-friendly answers, thus further enhancing the customer service experience and increasing user satisfaction and loyalty. However, traditional marketing methods, such as mass SMS sending and market surveys, often lack an in-depth understanding of users' prior knowledge and are difficult to achieve true personalized customization. They rely more on wide coverage and rough segmentation, ignoring the individual differences and unique needs among users, resulting in these marketing messages often being regarded as irrelevant spam by users and thus being ignored or deleted. This not only wastes the enterprise's marketing resources in vain but may also have a negative impact on the brand image and cause users' disgust and rejection. Compared with traditional methods, machine learning models have shown extraordinary strength in the field of data analysis. By deeply mining massive amounts of consumer problem-solving data, consumption records, and course browsing information, they can accurately grasp consumers' purchase preferences and consumption habits and then provide relatively personalized recommendations and services. With the sudden rise of large language models, their powerful natural language processing capabilities and in-depth understanding capabilities have further refreshed the understanding of personalized marketing. Large language models can not only more accurately understand complex and changing contexts but also keenly capture the subtle fluctuations in users' needs, effectively making up for the deficiencies of traditional models in terms of accuracy and personalized services. How to apply them to marketing is a problem that needs to be solved urgently. Therefore, the present invention proposes an auxiliary marketing method and system based on a large language model to solve the problems existing in the prior art. Summary of the Invention
[0004] In view of the above problems, the present invention proposes an assisted marketing method and system based on a large language model. The assisted marketing method and system based on a large language model integrate undirected weighted graph technology into the large model sample construction process, extract text Embedding features rich in recommendation-related information, and make full use of the correlation advantages of the graph. It effectively solves the problem of limited representation ability due to the separation from the actual recommendation business scenario, successfully avoids noise interference irrelevant to the recommendation task, and ensures a significant improvement in the application effect.
[0005] To achieve the purpose of the present invention, the present invention is implemented by the following technical solution: an auxiliary marketing method based on a large language model, characterized in that it includes the following steps:
[0006] S1: Collect logs from the APP, mobile, and h5 links.
[0007] S2: Use logs to count basic user features, use logs combined with undirected weighted graph technology to calculate basic and statistical features of products, and use multi-way recall algorithms to recall high-potential users;
[0008] S3: Based on the above features and the recalled users, the fusion model is used to predict the user's purchase data, and the prediction results are provided to the sales instructor;
[0009] S4: Based on the chat data collected from the logs, the user’s questions are intelligently optimized, and then the domain big model is used for intent recognition. Finally, different big model generation strategies are adopted according to different dialogue roles.
[0010] A further improvement is that: S2 comprises the following steps:
[0011] Use logs to count basic user features: Use time sliding window technology to mine users' dynamic test data and course browsing data, use Wilson interval smoothing technology to optimize features such as answer accuracy, use ComiRec algorithm to build user multi-interest Embedding features, and use domain big model technology to mine users' test preparation provinces, test preparation identities, and target tests.
[0012] Use logs to calculate basic and statistical features of products: Use user behavior data to construct an undirected weighted graph, in which the nodes of the graph represent products, and adjacent products are connected by edges. The weights of the edges are determined by the number of times these products co-occur. On this basis, adjacent product pairs with weights in a specific range are selected as positive samples, and negative sample pairs are formed by negative sampling from products that are not directly related. In order to extract the text embedding features of each product, a dedicated Token is designed, and with the help of fine-tuning of the large language model, product text embeddings containing rich recommendation information are finally generated.
[0013] A further improvement lies in that in S3, a fusion model of the MaskNet model and the WuKong model is used to predict user purchase data.
[0014] A further improvement lies in that S4 includes the following steps:
[0015] Based on the chat data collected from the logs, first, the questions of users are intelligently optimized based on the RAG technology;
[0016] Subsequently, a domain large model is used for intent recognition, and finally different large model generation strategies are adopted according to different dialogue roles;
[0017] When the dialogue party is a robot customer service, the domain large model is directly used to generate a standard answer and return it to the user;
[0018] When the dialogue party is a sales counselor, a personalized sales script is generated through the domain large model and provided to the sales counselor.
[0019] An assisted marketing system based on a large language model, including a user portrait module, a product portrait module, a recall strategy and algorithm module, a purchase data prediction algorithm module, and a dialogue system based on a large language model;
[0020] The user portrait module is used to mine the basic features of users; the product portrait module is used to calculate the basic features and statistical features of products; the recall strategy and algorithm module is used to recall potential users according to the mined features using a recall algorithm; the purchase data prediction algorithm module is used to predict the purchase intention of users; the dialogue system based on a large language model is used to obtain the chat information of users in real time from the logs and perform intelligent optimization based on the RAG system, use a domain large model for intent recognition, and adopt different large model generation strategies according to different dialogue roles.
[0021] A further improvement lies in that the user portrait module includes a time sliding window unit, a Wilson interval smoothing optimization unit, a time decay function optimization unit, a sequence feature mining unit, and a large language model unit;
[0022] The time sliding window unit uses the time sliding window technology to deeply mine the learning duration, the number of questions answered, and the course browsing data of users, and constructs time sliding window features; the Wilson interval smoothing optimization unit uses the Wilson interval smoothing method to optimize the confidence of features such as the correct answer rate in the education field. The Wilson interval function is as follows:
[0023]
[0024] Where p represents the active behavior rate, n represents the total number of behaviors, and z represents the parameter in the normal distribution.
[0025] A further improvement lies in that: the time decay function optimization unit is used to optimize the features with strong time correlation in the education field by combining the time decay function 1 / [log(t)+1], where t represents the time distance from the occurrence time of the event to the current time; the sequence feature mining unit uses the ComiRec algorithm to deeply mine the behavioral sequence features of users; the large language model unit adopts the Baichuan-13B-Chat large language model to deeply mine the chat content between sales and users, predict relevant label information, and use the relevant label information as the model input features.
[0026] A further improvement lies in that: the product portrait module is used to collect the price, type and time basic information of products, calculate the statistical information of the click-through rate and purchase rate of different products, and based on the large language model technology, construct an undirected weighted graph through the user behavior sequence, and generate training samples accordingly, and then fine-tune the large language model to extract the product text Embedding features rich in recommendation relevance.
[0027] A further improvement lies in that: the recall strategy and algorithm module includes: user recall based on product activity, user recall based on search keywords, and user recall based on the DeepWalk algorithm; the user recall based on product activity means: calculating the activity of users for different products in the past year, identifying users with high activity but not yet purchased, and preferentially including them in the recall scope; the user recall based on search keywords means: collecting and analyzing the search information of users in the APP, identifying users who have searched for specific products or related keywords, and regarding them as potential recall targets; the user recall based on the DeepWalk algorithm means: constructing a user-product interaction graph, where users are nodes, and when two users have a common product click behavior, an edge is established between these two user nodes, and through the DeepWalk algorithm, the TopK similar users of users with purchase behavior within one month are recalled.
[0028] A further improvement lies in that: the purchase data prediction algorithm module, based on the user portrait module and the product portrait module, after successfully recalling the user to be predicted, uses the model for scoring to predict the purchase intention of the user. Among them, the core features come from the user portrait module and the product portrait module, and a fusion model is adopted, combining the MaskNet model and the WuKong model for prediction in various application scenarios, providing a data basis for subsequent marketing decisions.
[0029] The beneficial effects of the present invention are:
[0030] 1. The present invention integrates the undirected weighted graph technology into the large model sample construction process, extracts text Embedding features rich in recommendation-related information, fully utilizes the correlation advantage of the graph, effectively solves the problem of limited representation ability caused by being divorced from the actual recommendation business scenario, successfully avoids the noise interference unrelated to the recommendation task, and ensures a significant improvement in the application effect.
[0031] 2. The present invention uses the WuKong large model for the prediction of marketing data, and performs weighted fusion on the prediction results with the prediction results of the MaskNet model. Combining with the scoring mechanism, the prediction accuracy is significantly improved.
[0032] 3. The present invention combines the RAG system, uses the large language model to deeply analyze the user's intention, and intelligently generates accurate responses or customized sales scripts, bringing a more considerate and efficient communication experience to users, and opening up a new path for the marketing and customer service upgrade in the education industry. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 is a flowchart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0034] To deepen the understanding of the present invention, the following will further elaborate on the present invention in combination with embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the protection scope of the present invention.
[0035] Embodiment 1
[0036] According to Figure 1 as shown, this embodiment proposes an auxiliary marketing method based on a large language model, including the following steps:
[0037] Collect logs from the APP side, WeChat side, and h5 link side;
[0038] Use the logs to statistically analyze the basic user characteristics; use the time sliding window technology to mine the user's dynamic question-solving data and course browsing data; use the Wilson interval smoothing technology to optimize the features of the answer correct rate; use the ComiRec algorithm to construct the user's multi-interest Embedding features; use the domain large model technology to mine the user's features such as the province for preparation, identity for preparation, and target exam, etc.;
[0039] Calculate the basic features and statistical features of products using logs; construct an undirected weighted graph using user behavior data, where the nodes of the graph represent products, adjacent products are connected by edges, and the weights of the edges are determined according to the number of co-occurrences of these products. On this basis, select adjacent product pairs with weights in a specific interval as positive samples, and at the same time form negative sample pairs by negative sampling from products with no direct association. To extract the text Embedding features of each product, a dedicated Token ([EMB]) is designed, and with the help of fine-tuning the large language model, the product text Embedding containing rich recommendation information is finally generated.
[0040] Use the multi-channel recall algorithm to recall high-potential users.
[0041] According to the above features and the recalled users, use the fusion model of the MaskNet model and the WuKong model to predict user purchase data.
[0042] Provide the prediction results to the sales counselors.
[0043] According to the chat data collected from the logs, first, based on the RAG system, the user's questions are intelligently optimized, then the domain large model is used for intent recognition, and finally, different large model generation strategies are adopted according to the different dialogue roles: if the dialogue party is a robot customer service, directly use the domain large model to generate a standard answer and return it to the user; if the dialogue party is a sales counselor, use the domain large model to generate personalized sales scripts and provide them to the sales counselor.
[0044] Utilize the powerful understanding ability and context extraction ability of the large language model to accurately mine potential purchasing users, providing strong support for marketing decisions. Integrate the MaskNet model and the WuKong model of Meta Company to construct a new scoring model, which ensures the accuracy of the prediction results with its powerful generalization ability, feature crossing and selection ability. To further improve the model effect, within the existing system framework, integrate the undirected weighted graph technology into the sample construction process of the large model, solving the problem of insufficient Embedding representation ability of the large model due to insufficient consideration of the actual recommendation business scenario. Not only successfully extract highly relevant recommended text Embedding features, but also significantly make up for the lack of text feature expression in the original product portrait technology, achieving an important breakthrough in technology.
[0045] In view of the problems of high costs for sales personnel in marketing scenarios and the inability to respond to user inquiries in real time, first, the RAG technology is combined to intelligently optimize the user's questions, and then a large model specialized in the field is used for accurate intention recognition. Finally, different generation strategies are adopted according to the different dialogue roles: if the dialogue party is a robot customer service, the standard answer is directly generated using the large model in the field; if the dialogue party is a sales counselor, personalized sales scripts are generated through the large model in the field to assist in formulating more accurate sales strategies.
[0046] Example Two
[0047] According to Figure 1 As shown, this embodiment proposes an assisted marketing system based on a large language model, including the following core functions:
[0048] User profiling technology based on user behavior
[0049] Time sliding window technology: In order to deeply mine the dynamic question-solving data and course browsing data of users in the education field, the time sliding window technology is adopted. The time sliding window technology is used to deeply mine data such as the learning duration, the number of questions answered, and course browsing of users, and a large number of time sliding window features are constructed. The experimental results show that the introduction of this technology has greatly improved the performance of the model.
[0050] Wilson interval smoothing optimization: In order to improve the confidence of features related to the correct answer rate in the education field (the confidence of data with a small number of answers is low), the Wilson interval smoothing (confidence interval) method is used for optimization. The Wilson interval function is as follows:
[0051]
[0052] Where p represents the active behavior rate, n represents the total number of behaviors, and z represents the parameter in the normal distribution. For example, z = 1.96 represents a 95% confidence level.
[0053] Time decay function optimization: For features with strong time correlation such as the mock exam scores and question-solving correct rates of users in the education field, the time decay function 1 / [log(t)+1] is combined for optimization, where t represents the time distance from the time when the event occurred to the current time; this function ensures that the feature values decay over time, making them more in line with the actual behavior patterns of users.
[0054] Sequential feature mining: In order to deeply capture the multi-interest features of users, the ComiRec algorithm is used to deeply mine the sequential behavior features of users.
[0055] Large language model technology: Adopted the advanced Baichuan-13B-Chat large language model to deeply mine the chat content between sales and users, accurately predict relevant tag information such as the user's preparation province, preparation identity, and target exam, and use the relevant tag information as model input features.
[0056] According to the product portrait technology of different products
[0057] Collect basic information such as the price, type, and time of products.
[0058] Calculate statistical information such as the click-through rate and purchase rate of different products.
[0059] Based on large language model technology: Construct an undirected weighted graph through the user behavior sequence, generate training samples accordingly, and then fine-tune the large language model to extract product text Embedding features rich in recommendation relevance.
[0060] Recall strategy and algorithm
[0061] User recall based on product activity: Calculate the activity of users towards different products in the past year (such as browsing, clicking, collecting, etc.), identify those users with high activity but who have not yet made a purchase, and prioritize including these users in the recall scope.
[0062] User recall based on search keywords: Collect and analyze the search information of users within the APP, and identify those users who have searched for specific products or related keywords. Since these users have shown a clear purchase intention, they are regarded as potential recall targets.
[0063] User recall based on the DeepWalk algorithm of the graph: Construct a user-product interaction graph, where users are nodes, and if two users have a common product click behavior, an edge is established between these two user nodes. Through the DeepWalk algorithm, recall the TopK similar users of users who have made a purchase within one month.
[0064] Purchase data prediction algorithm
[0065] After successfully recalling the users to be predicted, the core step is to use the model for scoring to accurately predict the purchase intention of users. The core features of this process come from user portrait technology and product portrait technology. The adopted fusion model combines the advantages of the MaskNet model and the WuKong model, not only having the powerful generalization ability of the deep learning model but also possessing unique feature screening and feature cross - ability, thus achieving a significant improvement in prediction effects in various application scenarios. The application of this technology ensures the accuracy and reliability of the prediction results, providing a solid data foundation for subsequent marketing decisions.
[0066] Dialogue System Based on Large Language Model
[0067] Real-time obtain the chat information of users from the logs and perform intelligent optimization based on the RAG system. Then, use the domain large model for intent recognition. Finally, adopt different large model generation strategies according to the different dialogue roles. The specific process is as follows: If the current interaction object is a robot customer service, directly generate the dialogue and feedback it to the user; if the current is a sales counselor, generate a specially customized sales script and provide it to the sales counselor to help them carry out more effective marketing activities.
[0068] The present invention integrates the undirected weighted graph technology into the large model sample construction process, extracts the text Embedding features rich in recommendation-related information, makes full use of the correlation advantages of the graph, effectively solves the problem of limited representation ability caused by being separated from the actual recommendation business scenario, successfully avoids the noise interference irrelevant to the recommendation task, and ensures a significant improvement in the application effect. And the present invention uses the WuKong large model for the prediction of marketing data, and performs weighted fusion on the prediction results with the prediction results of the MaskNet model, combined with a scoring mechanism, so that the prediction accuracy is significantly improved. At the same time, the present invention combines the RAG system, uses the large language model to deeply analyze the user's intent, and intelligently generates accurate responses or customized sales scripts, bringing a more considerate and efficient communication experience to users, and opening up a new path for the marketing and customer service upgrade in the education industry.
[0069] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art of this industry should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.
Claims
1. An auxiliary marketing method based on a large language model, characterized in that: The following steps are involved: S1: Collect logs from the APP, mobile, and h5 links. S2: Use logs to count basic user features, use logs combined with undirected weighted graph technology to calculate basic and statistical features of products, and use multi-way recall algorithms to recall high-potential users; S3: Based on the above features and the recalled users, the fusion model is used to predict the user's purchase data, and the prediction results are provided to the sales instructor; S4: Based on the chat data collected from the logs, the user’s questions are intelligently optimized, and then the domain big model is used for intent recognition. Finally, different big model generation strategies are adopted according to different dialogue roles.
2. The assisted marketing method based on a large language model according to claim 1, characterized in that: The S2 comprises the following steps: Use logs to count basic user features: Use time sliding window technology to mine users' dynamic test data and course browsing data, use Wilson interval smoothing technology to optimize features such as answer accuracy, use ComiRec algorithm to build user multi-interest Embedding features, and use domain big model technology to mine users' test preparation provinces, test preparation identities, and target tests. Use logs to calculate basic and statistical features of products: Use user behavior data to construct an undirected weighted graph, in which the nodes of the graph represent products, and adjacent products are connected by edges. The weights of the edges are determined by the number of times these products co-occur. On this basis, adjacent product pairs with weights in a specific range are selected as positive samples, and negative sample pairs are formed by negative sampling from products that are not directly related. In order to extract the text embedding features of each product, a dedicated Token is designed, and with the help of fine-tuning of the large language model, product text embeddings containing rich recommendation information are finally generated.
3. The assisted marketing method based on a large language model according to claim 1, characterized in that: In S3, a fusion model of the MaskNet model and the WuKong model is used to predict user purchase data.
4. The assisted marketing method based on a large language model according to claim 1, characterized in that: The S4 comprises the following steps: Based on the chat data collected from the logs, firstly, the user's questions are intelligently optimized based on the RAG technology; Then, the domain big model is used for intent recognition, and finally different big model generation strategies are adopted according to different dialogue roles; When the other party is a robot customer service, it directly uses the domain model to generate a standard answer and returns it to the user; When the interlocutor is a sales counselor, the domain big model is used to generate personalized sales scripts and provide them to the sales counselor.
5. An auxiliary marketing system based on a large language model, applied to an auxiliary marketing method based on a large language model as described in any one of claims 1 to 4, characterized in that: Including user portrait module, product portrait module, recall strategy and algorithm module, purchase data prediction algorithm module, and dialogue system based on large language model; The user portrait module is used to mine the basic characteristics of users; the product portrait module is used to calculate the basic characteristics and statistical characteristics of products; the recall strategy and algorithm module is used to recall high-potential users using the recall algorithm based on the mined characteristics; the purchase data prediction algorithm module is used to predict the user's purchase intention; the dialogue system based on the large language model is used to obtain the user's chat information from the log in real time and perform intelligent optimization based on the RAG system, use the domain large model for intent recognition, and adopt different large model generation strategies according to different dialogue roles.
6. The auxiliary marketing system based on a large language model according to claim 5, characterized in that: The user portrait module includes a time sliding window unit, a Wilson interval smoothing optimization unit, a time decay function optimization unit, a sequence feature mining unit, and a large language model unit; The time sliding window unit uses the time sliding window technology to deeply mine the user's learning time, number of questions and course browsing data to construct the time sliding window feature; the Wilson interval smoothing optimization unit uses the Wilson interval smoothing method to optimize the confidence of the correct answer rate feature in the education field. The Wilson interval function is as follows: Where p represents the active behavior rate, n represents the total number of behaviors, and z represents the parameter in the normal distribution.
7. The auxiliary marketing system based on a large language model according to claim 6, characterized in that: The time decay function optimization unit is used to optimize the features with strong time correlation in the education field in combination with the time decay function 1 / [log(t)+1], where t represents the distance between the time when the event occurs and the current time; the sequence feature mining unit uses the ComiRec algorithm to deeply mine the user's behavior sequence features; the large language model unit uses the Baichuan-13B-Chat large language model to deeply mine the chat content between sales and users, predict related label information, and use the related label information as the model input feature.
8. The auxiliary marketing system based on a large language model according to claim 5, characterized in that: The product portrait module is used to collect basic information such as price, type and time of products, calculate the statistical information of click-through rate and purchase rate of different products, and build an undirected weighted graph through user behavior sequence based on the large language model technology, and generate training samples based on this, and then fine-tune the large language model to extract product text Embedding features that are rich in recommendation relevance.
9. The auxiliary marketing system based on a large language model according to claim 5, characterized in that: The recall strategy and algorithm modules include: user recall based on product activity, user recall based on search keywords, and user recall based on the DeepWalk algorithm; the user recall based on product activity refers to: calculating the user's activity for different products in the past year, identifying users with high activity but no purchase, and giving priority to them in the recall range; the user recall based on search keywords refers to: collecting and analyzing the user's search information in the APP, identifying users who have searched for specific products or related keywords, and treating them as potential recall targets; the user recall based on the DeepWalk algorithm refers to: constructing a user-product interaction graph, in which users are nodes. When two users have a common product click behavior, an edge is established between the two user nodes. Through the DeepWalk algorithm, the TopK similar users of the user who has a purchase behavior within a month are recalled.
10. The auxiliary marketing system based on a large language model according to claim 5, characterized in that: The purchase data prediction algorithm module is based on the user portrait module and the product portrait module. After successfully recalling the users to be predicted, the model is used to score to predict the user's purchase intention. The core features come from the user portrait module and the product portrait module. The core model adopts a fusion model of the MaskNet model and the WuKong model, and makes predictions in various application scenarios to provide a data basis for subsequent marketing decisions.