Self-feedback recommendation method, device and system based on large language model and medium
Through the self-feedback recommendation method based on the large language model, using user feature database and behavioral data to optimize the recommendation content, the problems of high labor costs and insufficient recommendation diversity in the existing technology are solved, and the accuracy and diversity of personalized recommendations are achieved. promote.
Patent Information
- Application Number
- CN202411210628.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-30
- Publication Date
- 2024-12-03
AI Technical Summary
When the existing recommendation system is large in number of users and data, it requires high labor costs to write recommendation copy, and it is difficult to ensure the accuracy and diversity of recommendations, which affects the recommendation effect.
Using a self-feedback recommendation method based on a large language model, randomly select target features from the pre-constructed user feature library, match products and prompt templates, generate personalized recommendation copy, and update the feature library according to user behavior data to improve the diversity of recommendations and precision.
While saving labor costs, it improves the accuracy and diversity of personalized recommendations, generates recommendation copy through large language models and collects user feedback to optimize subsequent recommendation content.
Smart Images

Figure CN119067756A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of financial technology, and in particular to a self-feedback recommendation method, device, system and medium based on a large language model. Background Art
[0002] With the rapid development of mobile terminals and the Internet, more and more companies have launched their own apps for product sales or business processing, and provide personalized recommendation content on the app to recommend products. At present, the implementation methods of recommendation systems mainly include: user-based collaborative filtering, project-based collaborative filtering, recommendation system based on tree index structure, sorting stage based on click rate estimation, and combination recommendation.
[0003] Although existing methods can achieve personalized content or product recommendations, with the increasing number of users and data scale, the manpower costs required by existing recommendation methods are getting higher and higher. For example, more manpower costs are required to write recommendation copy for different groups of people, and it is difficult to ensure the accuracy and diversity of the recommendations, which affects the recommendation effect. Summary of the invention
[0004] In view of the above-mentioned deficiencies in the prior art, the purpose of the present invention is to provide a self-feedback recommendation method, device, system and medium based on a large language model that can be applied to financial technology or other related fields. Its main purpose is to improve the accuracy and diversity of personalized recommendations while saving labor costs.
[0005] The technical solution of the present invention is as follows:
[0006] A first aspect of the present invention provides a self-feedback recommendation method based on a large language model, comprising:
[0007] Randomly select the target features for the current recommendation round from the pre-built user feature library;
[0008] Matching the target product with the target prompt template in the preset product library and the preset prompt template library respectively according to the target feature;
[0009] Assembling the target product and the target prompt template and inputting them into a large language model, and outputting a recommendation copy corresponding to the target product and the target prompt template;
[0010] Pushing a recommendation page with the recommendation copy to the user, and collecting the user's behavior data on the recommendation page;
[0011] Feedback features for the recommended page are generated according to the behavior data and stored in the user feature library for use in matching selection in subsequent recommendation rounds.
[0012] In one embodiment, before randomly selecting the target feature of the current recommendation round from the pre-built user feature library, the method further includes:
[0013] Feature extraction is performed on user portraits to obtain multiple user features and build a corresponding user feature library.
[0014] In one embodiment, matching the target product with the target prompt template in a preset product library and a preset prompt template library respectively according to the target feature includes:
[0015] Confirm whether there are feedback features of historical recommendation rounds in the current user feature library;
[0016] If not, randomly selecting a target product and a target prompt template from a preset product library and a preset prompt template library respectively according to the target characteristics;
[0017] If so, based on the target features and feedback features of historical recommendation rounds, product attribute matching is performed with all products and style attribute matching is performed with all templates to determine the target product and target prompt template with the highest matching degree.
[0018] In one embodiment, before assembling the target product and the target prompt template and inputting them into a large language model and outputting a recommendation text corresponding to the target product and the target prompt template, the method further includes:
[0019] In response to the hyperparameter adjustment instruction, the hyperparameters of the large language model are dynamically adjusted.
[0020] In one embodiment, the dynamically adjusted hyperparameters include a temperature parameter, top k highest probabilities, a probability kernel, and / or a random number.
[0021] In one embodiment, collecting the user's behavior data on the recommendation page includes:
[0022] Detecting an operation event of a user on the recommendation page within a preset time;
[0023] The operation events within the preset time are counted and calculated to obtain corresponding behavior data.
[0024] In one embodiment, generating feedback features for the recommendation page according to the behavior data includes:
[0025] Extracting the user's behavior characteristics, emotional characteristics, and target conversion characteristics based on the behavior data;
[0026] The behavior feature, the emotion feature and the target conversion feature are integrated to serve as feedback features for the recommendation page.
[0027] A second aspect of the present invention provides a self-feedback recommendation device based on a large language model, comprising:
[0028] The feature selection module is used to randomly select the target features of the current recommendation round from the pre-built user feature library;
[0029] A matching module, used for matching a target product with a target prompt template in a preset product library and a preset prompt template library respectively according to the target feature;
[0030] A copywriting generation module, used for assembling the target product and the target prompt template and inputting them into a large language model, and outputting a recommendation copy corresponding to the target product and the target prompt template;
[0031] A push collection module, used to push a recommendation page with the recommendation copy to the user, and collect the user's behavior data on the recommendation page;
[0032] The self-feedback module is used to generate feedback features for the recommended page according to the behavior data and store them in the user feature library for use in matching selection in subsequent recommendation rounds.
[0033] A third aspect of the present invention provides a self-feedback recommendation system based on a large language model, the system comprising at least one processor; and
[0034] a memory communicatively connected to the at least one processor; wherein,
[0035] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the above-mentioned self-feedback recommendation method based on a large language model.
[0036] A fourth aspect of the present invention provides a non-volatile computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are executed by one or more processors, the one or more processors can execute the above-mentioned self-feedback recommendation method based on a large language model.
[0037] Beneficial effects: The present invention discloses a self-feedback recommendation method, device, system and medium based on a large language model. Compared with the prior art, the embodiment of the present invention randomly selects the target features of the current recommendation round from a pre-built user feature library; matches the target product and the target prompt template in the preset product library and the preset prompt template library respectively according to the target features; assembles the target product and the target prompt template and inputs them into the large language model, outputs the corresponding recommendation copy; pushes the recommendation page with the recommendation copy to the user, and collects the user's behavior data on the recommendation page; generates feedback features for the recommendation page based on the behavior data and stores them in the user feature library for matching selection in subsequent recommendation rounds. By matching the corresponding products and prompt templates with the randomly selected target features, the diversity of recommendations can be improved, and the corresponding recommendation copy is generated by using the large language model and user feedback is collected, so that the accuracy of personalized recommendations is improved while saving labor costs. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:
[0039] Figure 1 A flow chart of a self-feedback recommendation method based on a large language model provided by an embodiment of the present invention;
[0040] Figure 2 A flow chart of step S102 in the self-feedback recommendation method based on a large language model provided in an embodiment of the present invention;
[0041] Figure 3 A flow chart of step S104 in the self-feedback recommendation method based on a large language model provided in an embodiment of the present invention;
[0042] Figure 4 A flow chart of step S105 in the self-feedback recommendation method based on a large language model provided in an embodiment of the present invention;
[0043] Figure 5 A schematic diagram of functional modules of a self-feedback recommendation device based on a large language model provided by an embodiment of the present invention;
[0044] Figure 6 A schematic diagram of the hardware structure of a large language model-based self-feedback recommendation system provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0045] In order to make the purpose, technical solution and effect of the present invention clearer and more specific, the present invention is further described in detail below. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. The embodiments of the present invention are described below in conjunction with the accompanying drawings.
[0046] With the rapid development of mobile terminals and the Internet, more and more companies have launched their own apps for product sales or business processing, and provide personalized recommendation content on the app to recommend products. At present, the implementation methods of recommendation systems mainly include: user-based collaborative filtering, project-based collaborative filtering, recommendation system based on tree index structure, sorting stage based on click rate estimation, and combination recommendation.
[0047] Although existing methods can achieve personalized content or product recommendations, with the current increasing number of users and data scale, the manpower costs required for existing recommendation methods are increasing. For example, more manpower costs are required to write recommendation copy for different groups of people, and it is difficult to ensure the accuracy and diversity of the recommendations, which affects the recommendation effect.
[0048] In order to solve the above problems, the present invention proposes a self-feedback recommendation method based on a large language model. Figure 1 , Figure 1 A flowchart of an embodiment of a self-feedback recommendation method based on a large language model provided by the present invention. The system architecture of the self-feedback recommendation method based on a large language model provided in this embodiment may include a first terminal device, a second terminal device, a third terminal device, a network and a server. The network is used to provide a medium for a communication link between the first terminal device, the second terminal device, the third terminal device and the server. The network may include various connection types, such as wired and / or wireless communication links, etc.
[0049] The user may use the first terminal device, the second terminal device, and the third terminal device to interact with the server through the network to receive or send messages, etc. Various communication client applications may be installed on the first terminal device, the second terminal device, and the third terminal device, such as knowledge reading applications, web browser applications, search applications, instant messaging tools, email clients, and / or social platform software, etc. (only as examples).
[0050] The first terminal device, the second terminal device, and the third terminal device may be various electronic devices having display screens and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, desktop computers, and the like.
[0051] The server may be a server that provides various services, such as a background management server that provides support for the content browsed by the user using the first terminal device, the second terminal device, and the third terminal device (for example only). The background management server may analyze and process the received data such as the user request, and feed back the processing results (such as web pages, information, or data obtained or generated according to the user request) to the terminal device. The server may be a cloud server, also known as a cloud computing server or a cloud host, which is a host product in the cloud computing service system to solve the defects of difficult management and weak business scalability in traditional physical hosts and VPS services ("Virtual Private Server", or "VPS" for short). The server may also be a server for a distributed system, or a server combined with a blockchain.
[0052] It should be noted that the self-feedback recommendation based on the large language model provided in the embodiment of the present invention can generally be performed by the first terminal device, the second terminal device or the third terminal device. Correspondingly, the self-feedback recommendation based on the large language model provided in the embodiment of the present invention can also be set in the first terminal device, the second terminal device or the third terminal device.
[0053] Alternatively, the self-feedback recommendation method based on a large language model provided in an embodiment of the present invention may also be generally performed by a server. Accordingly, the self-feedback recommendation device based on a large language model provided in an embodiment of the present invention may generally be arranged in a server. The self-feedback recommendation method based on a large language model provided in an embodiment of the present invention may also be performed by a server or server cluster that is different from the server and can communicate with the first terminal device, the second terminal device, the third terminal device and / or the server. Accordingly, the self-feedback recommendation device based on a large language model provided in an embodiment of the present invention may also be arranged in a server or server cluster that is different from the server and can communicate with the first terminal device, the second terminal device, the third terminal device and / or the server.
[0054] It should be understood that the numbers of the above terminal devices, networks and servers are only illustrative. Any number of terminal devices, networks and servers may be provided according to implementation requirements.
[0055] like Figure 1 As shown, the self-feedback recommendation method based on a large language model provided in this embodiment specifically includes the following steps:
[0056] S101. Randomly select target features for the current recommendation round from a pre-built user feature library.
[0057] In this embodiment, a pre-constructed user feature library stores multiple user features obtained based on analysis of basic user information, historical behavior data, etc. Since user features do not contain deep information, in order to improve the randomness and diversity of recommendations, each time a personalized recommendation is made, a random selection algorithm is used to randomly select the target features of the current recommendation round from the user feature library. The number of target features can be one or more, which can be specifically set in the random selection algorithm to meet actual recommendation needs.
[0058] Specifically, a suitable random selection algorithm can be selected to select target features from the user feature library. For example, it can be simple random sampling or more complex algorithms, such as stratified sampling or importance sampling, where simple random sampling is to randomly select features from the user feature library, and the probability of each user feature being selected is the same; stratified sampling is to divide the user features in the user feature library into different categories or levels, and random sampling can be performed within each level first, and then the results are merged; importance sampling is to sample according to the importance or weight of user features, and more important user features have a higher probability of being selected; specifically, the corresponding random selection algorithm can be flexibly called according to the needs of the recommendation scenario, and this embodiment does not limit this.
[0059] In one embodiment, before randomly selecting the target feature of the current recommendation round from the pre-built user feature library, the method further includes:
[0060] Feature extraction is performed on user portraits to obtain multiple user features and build a corresponding user feature library.
[0061] In this embodiment, feature extraction is performed on the user portrait of each user by feature extraction, thereby obtaining multiple user features of each user, and the user features of all users are stored in a database to construct a user feature library, so that when making targeted recommendations to different users, the target features of the corresponding users can be flexibly and efficiently randomly selected from the user feature library. In specific implementation, taking an insurance company as an example, after obtaining user authorization, various types of user data are collected, including but not limited to: basic information: name, age, gender, occupation, education level, etc.; contact information: telephone, email, address, etc.; insurance history: type of insurance purchased, number of policies, claim history, etc.; financial information: income level, asset status, investment preferences, etc.; health data: physical examination results, chronic disease history, living habits, etc.; behavioral data: website visit records, customer service interactions, purchase behavior, etc. Clean the collected data, including removing duplicate records, filling missing values, correcting erroneous data, etc.; use the cleaned user data to build user portraits through data mining technology, including user interests, preferences, needs, etc.; extract key features from user portraits, such as user demographic characteristics, economic status characteristics, health and medical characteristics, insurance behavior characteristics, psychological characteristics, technology usage characteristics, etc., and then build a feature-rich user feature library, which is conducive to improving the diversity of recommendations.
[0062] S102: Match the target product with the target prompt template in a preset product library and a preset prompt template library respectively according to the target feature.
[0063] In this embodiment, a preset product library is pre-constructed based on all product information, for example, a product library containing various insurance products is established, and each product has a detailed attribute description; and a preset prompt template library is pre-constructed according to different recommendation speech styles, and recommendation copy templates of various styles are stored in the preset prompt template library. Each template has a specific style, tone and content structure, such as cute type, academic type, serious type, lively type, etc.
[0064] According to the selected target features, appropriate target products and target prompt templates are matched in the preset product library and the preset prompt template library respectively. The target prompt template is the prompt template of the large language model. Prompt is an input form used to instruct the large language model what action should be taken or what output should be generated when performing a specific task. That is, prompt is a natural language input, similar to a command or instruction, which lets the large language model know what it needs to do. The role of the prompt template selected in this embodiment is to instruct the large language model to generate corresponding recommended content with a specific style, tone, content structure, etc. in the template.
[0065] Specifically, when matching products, the similarity between user features and product attributes can be compared; or products can be recommended according to preset rules; or user data can be used to train models to predict user preferences for products, etc. When matching prompt templates, user features can be matched with the attributes of the prompt template according to preset matching rules. For example, young users may prefer a lively and humorous tone; high-income users may pay more attention to high-end and professional product descriptions, etc. The matching method can be flexibly selected according to actual needs, and this embodiment does not limit this.
[0066] S103: Assemble the target product and the target prompt template and input them into a large language model, and output a recommendation text corresponding to the target product and the target prompt template.
[0067] In this embodiment, after matching the target product and the target prompt template, they are assembled and input into the large language model as input parameters. Large language models (LLMs) are artificial intelligence models trained using a large amount of data, which can understand and generate natural language text. Based on the learning ability and context understanding ability of the large language model, when making personalized recommendations, the text describing the recommended items can be automatically generated based on the given target product and target prompt template to provide a more attractive product description. Since different products need to write corresponding recommendation words for different groups of people, the same product needs to be recommended to different sales objects using different words, which means an increase in the scale of the words library. As the number of users and the scale of data are increasing, the cost of human writing recommendation texts is getting higher and higher, and the quality is uncontrollable. This embodiment uses a large language model and a matching prompt template to quickly generate recommendation words with corresponding styles, effectively saving labor costs.
[0068] In one embodiment, before assembling the target product and the target prompt template and inputting them into a large language model and outputting a recommendation text corresponding to the target product and the target prompt template, the method further includes:
[0069] In response to the hyperparameter adjustment instruction, the hyperparameters of the large language model are dynamically adjusted.
[0070] In this embodiment, before using the large language model to generate personalized recommendation text, the hyperparameters of the large language model can also be dynamically adjusted by inputting hyperparameter adjustment instructions. The specific adjusted hyperparameters include temperature parameter Temperature, the top k highest probabilities Top_k, probability kernel Top_p and / or random numbers. Wherein, Temperature is a hyperparameter used to control the randomness of the language model output. When the temperature value is high (close to or greater than 1), the model tends to select words with lower probabilities, thereby increasing the diversity of the generated text. When the temperature value is low (close to 0), the model tends to select words with the highest probability, and the generated text is more inclined to common and possible outputs; Top-k means that when generating the next word, only the k words with the highest probability are considered, which can reduce the randomness of the generated text, so that the generated results are more concentrated on the most likely words, thereby reducing the probability of irrelevant or unreasonable words appearing; Top_p is used to control the cumulative probability threshold considered when generating text; random numbers are usually used to determine which word to select, especially when using the above parameters, random numbers help decide whether to select a word or skip it. Therefore, by dynamically adjusting the above hyperparameters, the large language model can generate more natural and diverse recommendation copy, dynamically control the accuracy, fun, and creativity of the generated content, increase the diversity of the content generated by the large language model, and ensure diversity while mass-producing recommendation scripts, so that users no longer feel visually fatigued by the same text.
[0071] S104: Push a recommendation page with the recommendation copy to the user, and collect user behavior data on the recommendation page.
[0072] In this embodiment, the recommendation copy generated based on the large language model is further rendered to obtain a recommendation page for the target product, and the recommendation page is pushed to the user through a variety of preset channels, such as sending a customized email containing the recommendation copy and the product link to the user, or sending a push notification containing the recommendation copy and a direct link to the recommendation page through a mobile application installed by the user; or placing an advertising page containing the recommendation copy on the user's social media timeline; or displaying the recommendation copy and the product to the user through a recommendation module or banner within the mobile application; or sending a text message containing the recommendation copy and a short link to the user's mobile phone number; or using a chatbot to provide personalized recommendations when interacting with the user, etc.
[0073] After the push is completed, we further collect user behavior data on the recommendation page, such as click-through rate, page views, user dwell time, bounce rate, conversion rate, scroll depth, number of interactions, etc. The user's behavior data is used to reflect the effect of the current round of recommendations, to form user self-feedback, and to facilitate the optimization of subsequent recommendation content.
[0074] S105: Generate feedback features for the recommendation page according to the behavior data and store them in the user feature library for use in matching selection in subsequent recommendation rounds.
[0075] In this embodiment, the user's feedback features on the recommendation page are generated through the collected behavioral data. The user's satisfaction and attractiveness to the recommended content can be analyzed through the feedback features. The feedback features are stored in the user feature library for use in matching and selection in subsequent recommendation rounds, so that a new prompt template can be matched next time to change the style of the recommendation copy, or the recommendation copy of this style can be re-delivered, etc. The large language model is used to generate corresponding recommendation copy and collect user feedback, thereby saving labor costs while also improving the accuracy of personalized recommendations.
[0076] In one embodiment, Figure 2 As shown, matching the target product with the target prompt template in the preset product library and the preset prompt template library respectively according to the target feature includes:
[0077] S201, confirm whether there is feedback feature of historical recommendation rounds in the current user feature library;
[0078] S202: If not, randomly selecting a target product and a target prompt template from a preset product library and a preset prompt template library respectively according to the target feature;
[0079] S203: If so, based on the target features and feedback features of historical recommendation rounds, product attribute matching is performed with all products and style attribute matching is performed with all templates to determine the target product and target prompt template with the highest matching degree.
[0080] In this embodiment, when following up on the target feature matching of the target product and the target prompt template, it is possible to first confirm in the user feature library whether there are user feedback features obtained in historical recommendation rounds. For example, for old users, the recommendation system has previously recommended certain products with corresponding recommendation texts, and collected user feedback features. At this time, there are user self-feedback features in the user feature library. However, for new users, since it is a cold start, there are no historical feedback features in the user feature library.
[0081] Therefore, for two different situations, in this embodiment, in the cold start scenario, that is, when there is no feedback feature in the user feature library, the target product and the target prompt template are randomly selected from the preset product library and the preset prompt template library based on the target feature, so that the randomness of the recommendation words and products is increased by random selection when the information is insufficient. When there is a feedback feature in the user feature library, the product attributes are matched with all products based on the selected target features and the feedback features of the historical recommendation rounds to determine the target product with the highest matching degree, and the style attributes are matched with all templates to determine the target prompt template with the highest matching degree. The specific matching method can be achieved through matching algorithms such as collaborative filtering, content-based recommendation, and hybrid recommendation. Since the feedback feature can reflect the user's preference for the historical recommended content, the matching results can be further optimized on the basis of the feedback received from the user, and more reference information can be provided for product attribute matching and style attribute matching, thereby further optimizing and adjusting the matching results, improving the effect of personalized recommendation, and further improving the conversion rate.
[0082] In one embodiment, Figure 3 As shown, the collected user behavior data on the recommendation page includes:
[0083] S301, detecting an operation event of a user on the recommendation page within a preset time;
[0084] S302: Count and calculate the operation events within the preset time to obtain corresponding behavior data.
[0085] In this embodiment, after the recommendation page is pushed to the user, the operation events input by the user on the recommendation page within the preset time can be detected, such as detecting click events, sliding events, staying events, jumping events, interactive events, collection and saving events, search events, etc., and the corresponding behavior data can be obtained by counting and calculating the operation events detected within the preset time. For example, the click rate is the ratio of the number of times the user clicks on the recommended content to the number of times the recommended content is displayed; the page views are the total number of times the user browses on the recommendation page; the user stay time is the total time the user stays on the recommendation page; the bounce rate is the proportion of users who only browse one page on the recommendation page and then leave; the conversion rate is the ratio of the number of times the user completes a specific target action (such as purchase, registration, etc.) on the recommendation page to the total number of visits; the scroll depth is the depth of the user scrolling on the page; the number of interactions is the number of times the user interacts with the recommended content, including likes, comments, shares, etc.; the collection / save behavior is the number of times the user adds the recommended content to favorites or bookmarks, etc., so as to provide multi-dimensional user behavior data for the extraction of feedback features and provide comprehensive and accurate feedback information.
[0086] In one embodiment, Figure 4As shown, the generating of feedback features for the recommended page according to the behavior data includes:
[0087] S401, extracting the user's behavior characteristics, emotional characteristics and target conversion characteristics according to the behavior data;
[0088] S402: The behavior feature, the emotion feature and the target conversion feature are integrated as feedback features for the recommendation page.
[0089] In this embodiment, by collecting and analyzing the behavioral data of the user's interaction with the recommendation page, effect feedback of different dimensions is obtained. Specifically, the user's behavioral characteristics are extracted through behavioral pattern recognition, such as using cluster analysis and other methods to identify the user's behavioral pattern, such as "quick browser", "deep reader", etc.; the user's emotional characteristics are extracted through sentiment analysis, such as sentiment analysis of the user's comments and feedback, extracting positive feedback, negative feedback or neutral feedback of emotional tendency, etc.; the user's target conversion characteristics are extracted through conversion analysis, such as analyzing the conversion of the user's target action (such as purchase, registration) under the guidance of the recommendation copy as the user's target conversion characteristics, etc. The extracted features of different dimensions are subjected to feature fusion, such as feature splicing, that is, directly merging different feature sets together to form a new feature vector; or feature weighting, that is, assigning different weights to different features to reflect their importance to the fusion result; or model-based feature fusion, that is, using machine learning models (such as random forests, gradient boosting trees, etc.) to learn the nonlinear relationship between features to achieve fusion, etc.
[0090] By integrating features from different dimensions to obtain user feedback on the recommendation page, we can more comprehensively understand user preferences and needs, dynamically adjust recommended content, respond to changes in user needs, and provide more accurate recommendations that are more in line with user expectations, thereby improving the overall user experience and business results.
[0091] It should be noted that there is not necessarily a certain order between the above-mentioned steps. A person skilled in the art can understand, based on the description of the embodiments of the present invention, that in different embodiments, the above-mentioned steps may have different execution orders, that is, they may be executed in parallel, may be executed interchangeably, and so on.
[0092] Another embodiment of the present invention provides a self-feedback recommendation device based on a large language model, such as Figure 5 As shown, the device 1 comprises:
[0093] A feature selection module 11 is used to randomly select target features for the current recommendation round from a pre-built user feature library;
[0094] A matching module 12, used to match a target product with a target prompt template in a preset product library and a preset prompt template library respectively according to the target feature;
[0095] A copywriting generation module 13, used for assembling the target product and the target prompt template and inputting them into a large language model, and outputting a recommendation copywriting corresponding to the target product and the target prompt template;
[0096] The push collection module 14 is used to push the recommendation page with the recommendation copy to the user and collect the user's behavior data on the recommendation page;
[0097] The self-feedback module 15 is used to generate feedback features for the recommended page based on the behavior data and store them in the user feature library for matching selection in subsequent recommendation rounds. The module referred to in the present invention refers to a series of computer program instruction segments that can complete specific functions, which is more suitable for describing the self-feedback recommendation execution process based on a large language model than a program. For the specific implementation of each module, please refer to the corresponding method embodiment above, which will not be repeated here.
[0098] In one embodiment, the device 1 further comprises:
[0099] The feature library construction module is used to extract features from user portraits, obtain multiple user features and construct a corresponding user feature library.
[0100] In one embodiment, the matching module 12 includes:
[0101] A feedback confirmation unit, used to confirm whether there are feedback features of historical recommendation rounds in the current user feature library;
[0102] A random matching unit, used for randomly selecting a target product and a target prompt template from a preset product library and a preset prompt template library respectively according to the target feature if the target product and the target prompt template do not exist;
[0103] The attribute matching unit is used to match product attributes with all products and style attributes with all templates based on the target features and feedback features of historical recommendation rounds, if any, to determine the target product and target prompt template with the highest matching degree.
[0104] In one embodiment, the device 1 further comprises:
[0105] A hyperparameter adjustment module is used to dynamically adjust the hyperparameters of the large language model in response to the hyperparameter adjustment instruction.
[0106] In one embodiment, the dynamically adjusted hyperparameters include a temperature parameter, top k highest probabilities, a probability kernel, and / or a random number.
[0107] In one embodiment, the push collection module 14 includes:
[0108] A detection unit, used to detect an operation event of a user on the recommendation page within a preset time;
[0109] The statistical unit is used to count and calculate the operation events within the preset time to obtain corresponding behavior data.
[0110] In one embodiment, the self-feedback module 15 includes:
[0111] A feature extraction unit, used to extract the user's behavior features, emotional features and target conversion features according to the behavior data;
[0112] The feature fusion unit is used to fuse the behavior feature, the emotion feature and the target conversion feature as feedback features for the recommendation page.
[0113] Another embodiment of the present invention provides a self-feedback recommendation system based on a large language model, such as Figure 6 As shown, the system 10 includes:
[0114] One or more processors 110 and memory 120, Figure 6 A processor 110 is used as an example for description. The processor 110 and the memory 120 may be connected via a bus or other means. Figure 6 The example of connecting through bus is taken in the following.
[0115] The processor 110 is used to complete various control logics of the system 10, and it can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), a single-chip microcomputer, an ARM (Acorn RISC Machine) or other programmable logic device, a discrete gate or transistor logic, a discrete hardware component, or any combination of these components. In addition, the processor 110 can also be any traditional processor, microprocessor or state machine. The processor 110 can also be implemented as a combination of computing devices, for example, a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors combined with a DSP and / or any other such configuration.
[0116] The memory 120, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer executable programs and modules, such as program instructions corresponding to the self-feedback recommendation method based on a large language model in the embodiment of the present invention. The processor 110 executes various functional applications and data processing of the system 10 by running the non-volatile software programs, instructions and units stored in the memory 120, that is, implements the self-feedback recommendation method based on a large language model in the above method embodiment.
[0117] The memory 120 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application required by at least one function; the data storage area may store data created according to the use of the system 10, etc. In addition, the memory 120 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 120 may optionally include a memory remotely arranged relative to the processor 110, and these remote memories may be connected to the system 10 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.
[0118] One or more units are stored in the memory 120. When executed by one or more processors 110, the self-feedback recommendation method based on the large language model in any of the above method embodiments is executed, for example, the following steps are executed:
[0119] Randomly select the target features for the current recommendation round from the pre-built user feature library;
[0120] Matching the target product with the target prompt template in the preset product library and the preset prompt template library respectively according to the target feature;
[0121] Assembling the target product and the target prompt template and inputting them into a large language model, and outputting a recommendation copy corresponding to the target product and the target prompt template;
[0122] Pushing a recommendation page with the recommendation copy to the user, and collecting the user's behavior data on the recommendation page;
[0123] Feedback features for the recommended page are generated according to the behavior data and stored in the user feature library for use in matching selection in subsequent recommendation rounds.
[0124] In one embodiment, before randomly selecting the target feature of the current recommendation round from the pre-built user feature library, the method further includes:
[0125] Feature extraction is performed on user portraits to obtain multiple user features and build a corresponding user feature library.
[0126] In one embodiment, matching the target product with the target prompt template in a preset product library and a preset prompt template library respectively according to the target feature includes:
[0127] Confirm whether there are feedback features of historical recommendation rounds in the current user feature library;
[0128] If not, randomly selecting a target product and a target prompt template from a preset product library and a preset prompt template library respectively according to the target characteristics;
[0129] If so, based on the target features and feedback features of historical recommendation rounds, product attribute matching is performed with all products and style attribute matching is performed with all templates to determine the target product and target prompt template with the highest matching degree.
[0130] In one embodiment, before assembling the target product and the target prompt template and inputting them into a large language model and outputting a recommendation text corresponding to the target product and the target prompt template, the method further includes:
[0131] In response to the hyperparameter adjustment instruction, the hyperparameters of the large language model are dynamically adjusted.
[0132] In one embodiment, the dynamically adjusted hyperparameters include a temperature parameter, top k highest probabilities, a probability kernel, and / or a random number.
[0133] In one embodiment, collecting the user's behavior data on the recommendation page includes:
[0134] Detecting an operation event of a user on the recommendation page within a preset time;
[0135] The operation events within the preset time are counted and calculated to obtain corresponding behavior data.
[0136] In one embodiment, generating feedback features for the recommendation page according to the behavior data includes:
[0137] Extracting the user's behavior characteristics, emotional characteristics, and target conversion characteristics based on the behavior data;
[0138] The behavior feature, the emotion feature and the target conversion feature are integrated to serve as feedback features for the recommendation page.
[0139] An embodiment of the present invention provides a non-volatile computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are executed by one or more processors, the self-feedback recommendation method based on a large language model in any of the above method embodiments is executed, for example, the following steps are performed:
[0140] Randomly select the target features for the current recommendation round from the pre-built user feature library;
[0141] Matching the target product with the target prompt template in the preset product library and the preset prompt template library respectively according to the target feature;
[0142] Assembling the target product and the target prompt template and inputting them into a large language model, and outputting a recommendation copy corresponding to the target product and the target prompt template;
[0143] Pushing a recommendation page with the recommendation copy to the user, and collecting the user's behavior data on the recommendation page;
[0144] Feedback features for the recommended page are generated according to the behavior data and stored in the user feature library for use in matching selection in subsequent recommendation rounds.
[0145] In one embodiment, before randomly selecting the target feature of the current recommendation round from the pre-built user feature library, the method further includes:
[0146] Feature extraction is performed on user portraits to obtain multiple user features and build a corresponding user feature library.
[0147] In one embodiment, matching the target product with the target prompt template in a preset product library and a preset prompt template library respectively according to the target feature includes:
[0148] Confirm whether there are feedback features of historical recommendation rounds in the current user feature library;
[0149] If not, randomly selecting a target product and a target prompt template from a preset product library and a preset prompt template library respectively according to the target characteristics;
[0150] If so, based on the target features and feedback features of historical recommendation rounds, product attribute matching is performed with all products and style attribute matching is performed with all templates to determine the target product and target prompt template with the highest matching degree.
[0151] In one embodiment, before assembling the target product and the target prompt template and inputting them into a large language model and outputting a recommendation text corresponding to the target product and the target prompt template, the method further includes:
[0152] In response to the hyperparameter adjustment instruction, the hyperparameters of the large language model are dynamically adjusted.
[0153] In one embodiment, the dynamically adjusted hyperparameters include a temperature parameter, top k highest probabilities, a probability kernel, and / or a random number.
[0154] In one embodiment, collecting the user's behavior data on the recommendation page includes:
[0155] Detecting an operation event of a user on the recommendation page within a preset time;
[0156] The operation events within the preset time are counted and calculated to obtain corresponding behavior data.
[0157] In one embodiment, generating feedback features for the recommendation page according to the behavior data includes:
[0158] Extracting the user's behavior characteristics, emotional characteristics, and target conversion characteristics based on the behavior data;
[0159] The behavior feature, the emotion feature and the target conversion feature are integrated to serve as feedback features for the recommendation page.
[0160] As an example, non-volatile storage media can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) as external cache memory. By way of illustration and not limitation, RAM can be obtained in many forms such as synchronous RAM (SRAM), dynamic RAM, (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), Synchlink DRAM (SLDRAM), and direct Rambus RAM (DRRAM). The disclosed memory components or memories of the operating environment described herein are intended to include one or more of these and / or any other suitable types of memory.
[0161] In summary, the self-feedback recommendation method, device, system and medium based on the large language model disclosed in the present invention include: randomly selecting the target feature of the current recommendation round from the pre-built user feature library; matching the target product and the target prompt template in the preset product library and the preset prompt template library respectively according to the target feature; assembling the target product and the target prompt template and inputting them into the large language model, and outputting the corresponding recommendation copy; pushing the recommendation page with the recommendation copy to the user, and collecting the user's behavior data on the recommendation page; generating feedback features for the recommendation page according to the behavior data and storing them in the user feature library for matching selection in subsequent recommendation rounds. By matching the corresponding products and prompt templates with the randomly selected target features, the diversity of recommendations can be improved, and the corresponding recommendation copy is generated by using the large language model and user feedback is collected, so that the accuracy of personalized recommendations is improved while saving labor costs.
[0162] Of course, those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing related hardware (such as a processor, a controller, etc.) through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium, and the computer program can include the processes of the above-mentioned method embodiments when executed. The storage medium can be a memory, a disk, a floppy disk, a flash memory, an optical storage device, etc.
[0163] It should be understood that the application of the present invention is not limited to the above examples. For ordinary technicians in this field, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.
Claims
1. A self-feedback recommendation method based on a large language model, characterized in that: include: Randomly select the target features for the current recommendation round from the pre-built user feature library; Matching the target product with the target prompt template in the preset product library and the preset prompt template library respectively according to the target feature; Assembling the target product and the target prompt template and inputting them into a large language model, and outputting a recommendation copy corresponding to the target product and the target prompt template; Pushing a recommendation page with the recommendation copy to the user, and collecting the user's behavior data on the recommendation page; Feedback features for the recommended page are generated according to the behavior data and stored in the user feature library for use in matching selection in subsequent recommendation rounds.
2. The self-feedback recommendation method based on a large language model according to claim 1, characterized in that: Before randomly selecting the target feature of the current recommendation round from the pre-built user feature library, the method further includes: Feature extraction is performed on user portraits to obtain multiple user features and build a corresponding user feature library.
3. The self-feedback recommendation method based on a large language model according to claim 1, characterized in that: The matching of the target product with the target prompt template in the preset product library and the preset prompt template library according to the target feature includes: Confirm whether there are feedback features of historical recommendation rounds in the current user feature library; If not, randomly selecting a target product and a target prompt template from a preset product library and a preset prompt template library respectively according to the target characteristics; If so, based on the target features and feedback features of historical recommendation rounds, product attribute matching is performed with all products and style attribute matching is performed with all templates to determine the target product and target prompt template with the highest matching degree.
4. The self-feedback recommendation method based on a large language model according to claim 1, characterized in that: Before assembling the target product and the target prompt template and inputting them into the large language model and outputting the recommendation copy corresponding to the target product and the target prompt template, the method further includes: In response to the hyperparameter adjustment instruction, the hyperparameters of the large language model are dynamically adjusted.
5. The self-feedback recommendation method based on a large language model according to claim 4, characterized in that: Dynamically adjusted hyperparameters include temperature parameter, top k highest probabilities, probability kernel and / or random numbers.
6. The self-feedback recommendation method based on a large language model according to claim 1, characterized in that: The collecting of user behavior data on the recommendation page includes: Detecting an operation event of a user on the recommendation page within a preset time; The operation events within the preset time are counted and calculated to obtain corresponding behavior data.
7. The self-feedback recommendation method based on a large language model according to claim 1, characterized in that: Generating feedback features for the recommendation page according to the behavior data includes: Extracting the user's behavior characteristics, emotional characteristics, and target conversion characteristics based on the behavior data; The behavior feature, the emotion feature and the target conversion feature are integrated to serve as feedback features for the recommendation page.
8. A self-feedback recommendation device based on a large language model, characterized in that: include: The feature selection module is used to randomly select the target features of the current recommendation round from the pre-built user feature library; A matching module, used for matching a target product with a target prompt template in a preset product library and a preset prompt template library respectively according to the target feature; A copywriting generation module, used for assembling the target product and the target prompt template and inputting them into a large language model, and outputting a recommendation copy corresponding to the target product and the target prompt template; A push collection module, used to push a recommendation page with the recommendation copy to the user, and collect the user's behavior data on the recommendation page; The self-feedback module is used to generate feedback features for the recommended page according to the behavior data and store them in the user feature library for use in matching selection in subsequent recommendation rounds.
9. A self-feedback recommendation system based on a large language model, characterized in that: The system includes at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the self-feedback recommendation method based on a large language model as described in any one of claims 1-7.
10. A non-volatile computer-readable storage medium, characterized in that: The non-volatile computer-readable storage medium stores computer-executable instructions, which, when executed by one or more processors, enable the one or more processors to execute the self-feedback recommendation method based on a large language model as described in any one of claims 1-7.
Citation Information
Cited By
UI self-feedback generation and test method and system based on large model
CN120872331A
A Method and System for UI Self-Feedback Generation and Testing Based on Large Models
CN120872331B
Commodity recommendation method and system based on agent and hotspot information collaboration
CN121146867A
Method and system for commodity recommendation based on agent and hotspot information cooperation
CN121146867B
