Multi-scene session type recommendation method based on large language model agent framework

Abstract

The invention provides a multi-scene session type recommendation method based on a large language model agent framework, and relates to the technical field of recommendation systems. The method specifically comprises the following steps: constructing a big language model agent framework-based recommendation model consisting of a dialogue manager, an evaluation agent, an article agent, a learning agent and an execution agent; for any user, article recommendation is carried out on the user by adopting a recommendation model based on a large language model agent framework according to the dialogue input by the user, and if the user is not a new user, article recommendation is carried out; if yes, generating a recommendation result for the user by adopting a fine-tuned ID-based collaborative filtering recommendation algorithm or a fine-tuned interaction history-based sequence recommendation algorithm; if the user is a new user, generating a recommendation result of the user by adopting a recommendation algorithm based on content similarity retrieval; and after receiving the recommendation result of the user, the user submits user feedback to correct the recommendation result of the user until the user is satisfied.

Description

Technical Field

[0001] The present invention relates to the technical field of recommendation systems, and in particular to a multi-scenario conversational recommendation method based on a large language model proxy framework. Background Art

[0002] Conversational recommendation systems aim to gradually explore users' interests and preferences through multiple rounds of natural language-based dialogues, thereby recommending items that may be of interest to users. Existing conversational recommendation systems generally conform to an overall framework. The user interaction module is responsible for understanding the user's natural language input and is also responsible for making system responses. The dialogue strategy management module is responsible for deciding how to respond based on the current state, such as whether to continue asking questions or make recommendations directly. The recommendation engine recommends suitable items based on the user's input or the extracted user profile.

[0003] At present, conversational recommendation can be divided into two research directions. One direction is attribute-based conversational recommendation. This recommendation method focuses on modeling the dialogue strategy management module. The research center focuses on how to achieve the most accurate recommendation in the least number of dialogue rounds. The general practice is to use reinforcement learning methods to train the dialogue strategy management module. A simple dialogue module is used to put the recommendation results into a fixed template with result slots to form a system reply. The other direction is generative recommendation. This direction pays more attention to providing users with a smooth dialogue experience, while flexibly integrating the relevant information of the recommended items into the reply text to improve the interpretability of the recommendation results. This type of model usually uses sequence models to build dialogue modules. The recently emerging large language model has also further enhanced the language capabilities of generative dialogue recommendation systems.

[0004] In previous work, attribute-based conversational recommendations relied on templates to conduct conversations, which resulted in poor user conversational experience and lack of explainability of recommendation results. Although generative recommendations pay attention to the user's conversational experience, in the current era of large models, the parameters of the conversational model used are small, which affects the quality of text generation. It does not consider how to combine large language models with recommendation algorithms to build a conversational recommendation system based on large models. Summary of the invention

[0005] In view of the above-mentioned deficiencies in the prior art, the present invention constructs a general interactive system framework based on a large language model, and proposes a multi-scenario conversational recommendation method based on a large language model agent framework, aiming to achieve transparency and scalability of multi-scenario recommendations.

[0006] The present invention proposes a multi-scenario conversational recommendation method based on a large language model proxy framework, which includes the following process:

[0007] Build a recommendation model based on a large language model proxy framework;

[0008] Obtain several groups of user rating data and item metadata, and save all the obtained user rating data and item metadata to the recommendation model based on the large language model proxy framework;

[0009] For any user, based on the dialogue input by the user, the recommendation model based on the large language model proxy framework is used to recommend items to the user. If the user is not a new user, the fine-tuned ID-based collaborative filtering recommendation algorithm or the fine-tuned interaction history-based sequence recommendation algorithm is used to generate recommendation results for the user; if the user is a new user, the recommendation algorithm based on content similarity retrieval is used to generate recommendation results for the user;

[0010] The fine-tuning methods of the ID-based collaborative filtering recommendation algorithm and the interaction history-based sequential recommendation algorithm are the same, which are: for the algorithm model predefined in the ID-based collaborative filtering recommendation algorithm or the interaction history-based sequential recommendation algorithm, fine-tuning is achieved by specifying the user interaction history data set path and modifying the predefined algorithm model name, and saving the fine-tuned algorithm model weight;

[0011] After receiving the recommendation result, the user submits the user feedback to the recommendation model based on the large language model proxy framework. The recommendation model based on the large language model proxy framework modifies the recommendation result according to the user feedback and re-recommends items to the user according to the modified result until the user is satisfied.

[0012] Preferably, the recommendation model based on the large language model agent framework includes: a dialogue manager, a judgment agent, an item agent, a learning agent and an execution agent; wherein the dialogue manager, the judgment agent, the item agent, the learning agent and the execution agent are all agents;

[0013] The dialogue manager is used to interact with the user in the chat scene, decide whether to switch the recommendation scene according to the content input by the user, and if necessary, use the content input by the user as the recommendation requirement and switch to the recommendation scene; if not, use the content input by the user as user feedback and transmit it to the evaluation agent; generate the recommendation scene description by judging whether the user ID exists, the number of dialogue rounds of the user and the recommendation requirement, and transmit it to the execution agent; receive the user portrait from the learning agent and the item description from the item agent respectively, and generate the user's recommendation result according to the user portrait and the item description, and send it to the user and the evaluation agent respectively;

[0014] The judging agent is used to evaluate the difference between the user feedback and the user's recommendation result, and to modify the user's recommendation result according to the difference;

[0015] The item agent is used to obtain item metadata and extract text information and image information respectively, then encode the extracted text information and image information respectively, fuse the encoded text feature representation and image feature representation, generate an item portrait and transmit it to the learning agent; use natural language generation technology to convert the item ID in the recommendation result candidate set into an item description and transmit it to the dialogue manager;

[0016] The learning agent is used to obtain the user's dialogue interaction history and user rating data, integrate all items that have been interacted with the user from the user's dialogue interaction history and user rating data, generate a user portrait based on the item portrait of the item, and then transmit the user portrait to the dialogue manager; calculate the similarity between items using the item portrait and transmit it to the execution agent;

[0017] The execution agent is used to select the recommendation algorithm to be called according to the recommendation scenario description, and based on the similarity between the items, use the recommendation algorithm to generate a candidate set of recommendation results and transmit it to the item agent;

[0018] Preferably, the user rating data includes: user ID, user name, item ID rated by the user, effective rating rate, rating text, rating timestamp and rating time;

[0019] The item metadata includes: item ID, item description information, item price, image path, item category, item name, and item purchaser's browsing history or purchase history;

[0020] Preferably, the process of using the recommendation model based on the large language model proxy framework to recommend items to the user is:

[0021] Obtain the conversation content input by the user and perform text analysis. When the conversation content is a recommendation requirement, switch the recommendation model based on the large language model proxy framework to the recommendation scenario;

[0022] Obtain the user ID of the user, and determine whether the user ID exists in the stored user rating data. If so, the user is a historical user, and the user rating data of the user is obtained; if not, the user is a new user, and the user's registration information is obtained and used as the user rating data of the user;

[0023] Generate a recommendation scenario description based on the user's user rating data, number of conversation turns, and recommendation requirements;

[0024] Extract the text information and image information of each item from the item metadata stored in the recommendation model based on the large language model proxy framework, encode the extracted text information and image information respectively, fuse the encoded text feature representation and image feature representation, and generate the item portrait of each item;

[0025] Integrate all items that the user has interacted with based on the user's conversation interaction history and user rating data, and generate a user profile based on the item portraits of all items that the user has interacted with.

[0026] For historical users, based on the user's user rating data, recommendation scenario description and item portrait, a fine-tuned ID-based collaborative filtering recommendation algorithm or a fine-tuned interaction history-based sequential recommendation algorithm is used to generate a candidate set of recommendation results for the user; for new users, based on the user's user rating data, recommendation scenario description and item portrait, a content similarity retrieval-based recommendation algorithm is used to generate a candidate set of recommendation results for the user;

[0027] Use natural language generation technology to convert the item IDs in the recommendation result candidate set into item descriptions, and generate the user's recommendation results based on the user portrait and item description and send them to the user;

[0028] The user gives feedback on the generated recommendation results, and the recommendation results are modified according to the user feedback, and items are re-recommended to the user according to the modified results until the user is satisfied;

[0029] Preferably, the method of using the fine-tuned ID-based collaborative filtering recommendation algorithm to generate a candidate set of recommendation results for the user is:

[0030] Construct a user interaction history dataset of the user based on the user rating data and item metadata of the user;

[0031] The method for constructing the user interaction history data set is as follows: according to the user rating data of the user, the interaction history between each item in the item metadata and the user is obtained; for any item, a user-item evaluation key-value pair is constructed with the user ID as the primary key and the interaction history between the user and the item as the value; and then the user interaction history data set of the user is constructed using the user-item evaluation key-value pairs of all items and the user; wherein the interaction history is the effective evaluation rate or evaluation text of the user on the item;

[0032] Based on the user's user interaction history dataset, an initial random embedding representation is assigned to the user and all items respectively;

[0033] Update the user embedding representation of the user and the item embedding representation of all items by iterating and weighted summing the initial random embedding representations of the user and all items:

[0034] According to the updated user embedding representation and the item embedding representation of all items, the rating matrix of the user is constructed, which is expressed as:

[0035]

[0036] in represents the user's rating matrix for all items; e u Indicates user embedding representation; e i Item embedding representation representing all items;

[0037] Remove items that the user has rated from the obtained rating matrix, and sort the items in the rating matrix from high to low according to the predicted scores of the remaining items to generate a candidate set of recommendation results;

[0038] Preferably, the updating method of the user embedding representation is:

[0039]

[0040] in represents the updated user embedding representation; u represents the user; k represents the iteration round; i represents the item; N u Represents the normalization coefficient of the user; N i Represents the normalization coefficient of the item; Indicates the embedding representation of items in the current round;

[0041] The updating method of the embedding representation of the item is:

[0042]

[0043] in Represents the updated item embedding representation; Indicates the user embedding representation in the current round;

[0044] Preferably, the method of using the fine-tuned interaction history-based sequential recommendation algorithm to generate a candidate set of recommendation results for the user is:

[0045] Using the user rating data of the user, all items rated by the user are sorted in time, and an interaction sequence between the user and each item is generated respectively; wherein the elements in the interaction sequence are the user's interaction behaviors with the items;

[0046] Define the user's behavior sequence, expressed as:

[0047]

[0048] Where S u Represents the user's behavior sequence; represents the interaction sequence between the user and the first item; U represents the user; represents the interaction sequence between the user and the second item; Represents the interaction sequence between the user and the last item; |S u | represents the length of the user's behavior sequence;

[0049] The length of the user's behavior sequence is limited to n;

[0050] Encode each item in the user's behavior sequence for sequence encoding and position encoding, generate the item's sequence Embedding representation and position Embedding representation, and add the item's sequence Embedding representation and position Embedding representation to obtain the item's sequence interaction information;

[0051] The self-attention mechanism is used to learn the sequence interaction information and calculate the attention score of each item;

[0052] For each item in the user's behavior sequence, generate an item Embedding representation and calculate the relevance score of each item;

[0053] Converting a user's behavior sequence into a user behavior sequence of the user;

[0054] Calculate the similarity between the user behavior sequence output and the item embedding representation of all items to predict the user's rating for each item, delete all items rated by the user, and generate a candidate set of recommendation results based on the relevance scores and ratings of the remaining items;

[0055] Preferably, the relevance score of each item is expressed as:

[0056]

[0057] where r h,t represents the relevance score of the t-th user to the h-th item; represents the function of the t-th user in the b-th self-attention block, and is a function that depends on M; M is the Embedding interaction representation of a single item; M h The item Embedding representation of the h-th item;

[0058] Preferably, the user behavior sequence output is expressed as:

[0059]

[0060] Among them,t Represents the user behavior sequence output of the t-th user; <pad>Indicates filling; pad indicates the filling value; s t User representation of the t-th user; s t+1 The user representation of the next user predicted;

[0061] Preferably, the method of using the recommendation algorithm based on content similarity retrieval to generate a candidate set of recommendation results for the user is:

[0062] Use item metadata to build a candidate set of item IDs;

[0063] The method for constructing the candidate set of item IDs is as follows: for any item, obtain the commodity description information of the item from the item metadata, and use a text encoder to encode the commodity description information of the item to obtain the item content representation I of the item. content ; With item ID as primary key, the item content of the item represents I content Build the item ID candidate set using the item ID key-value pairs of all items.

[0064] The user profile of the user is encoded using a text encoder to generate a user profile representation U content ;

[0065] The user portrait of the user is used to represent U content and the content representation of each item in the item ID candidate set I content , calculate the similarity between the user and all items in the item ID candidate set;

[0066] The similarity is calculated as follows:

[0067]

[0068] where p u,i Indicates the similarity between users and items; Normalize means normalization;

[0069] All items in the item ID candidate set are sorted in descending order according to the similarity between the user and all items in the item ID candidate set, and the first P items in the sorted result are selected to generate a recommendation result candidate set.

[0070] The beneficial effects of adopting the above technical solution are:

[0071] Compared with the existing recommendation system, the recommendation model based on the large language model proxy framework constructed in the method of the present invention has improved interactivity and realized a high-availability conversational recommendation model. By setting two recommendation scenarios, namely, the user is a historical user and the user is a new user, when the user is a historical user, a fine-tuned ID-based collaborative filtering recommendation algorithm or a fine-tuned interaction history-based sequence recommendation algorithm is used to generate recommendation results for the user; when the user is a new user, a recommendation algorithm based on content similarity retrieval is used to obtain user-satisfied recommendation results, thereby making up for the performance bottleneck of a single recommendation algorithm.

[0072] The method of the present invention provides a new conversational recommendation system and method that combines a large language model with a recommendation algorithm. Prompt engineering technology is used to guide the large language model to learn user input cases in recommendation scenarios, and then a multi-scenario conversational recommendation system is constructed based on the large language model and recommendation algorithms in different recommendation scenarios.

[0073] The method of the present invention builds a recommendation scenario input sample data set based on the self-instruction generation technology, and then uses the Prompt project based on the generated recommendation scenario samples to enhance the recommendation scenario recognition capability of the open source large language model. The method of the present invention builds a universal proxy framework to adapt to different large language model bases and different recommendation tasks, and emphasizes the versatility and practicality of the recommendation system.

[0074] Although traditional recommendation algorithms have good recommendation performance, they do not have language interaction capabilities; and although large language models have good language understanding and interaction capabilities, they have insufficient recommendation accuracy in the recommendation field. Therefore, the method of the present invention realizes a multi-scenario interactive recommendation system through the idea of ​​understanding and using tools with large models and the use and development of multiple technical frameworks. Specifically, considering the versatility, the method of the present invention customizes the use paradigm of the large language model base, which lowers the hardware threshold for users to use the recommendation system. A variety of traditional recommendation algorithms and similarity-based retrieval are encapsulated into an API tool library to achieve transparency and scalability of multi-scenario recommendations. The method of the present invention also proposes a mechanism for correcting recommendation results based on user feedback, which improves the accuracy of recommendation results based on content similarity.

[0075] Compared with the prior art, the present invention builds a general recommendation model based on a large language model proxy framework based on a large language model. The client only needs to use very few hardware resources to use the multi-scenario interactive recommendation model. Based on the customized prompt project, more recommendation scenarios can be expanded, and the range of recommended items of the recommendation model is improved. Through performance testing, the recommendation model implemented by the present invention has good interactivity and performance, and meets the application requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 This is a flowchart of a multi-scenario conversational recommendation method based on a large language model proxy framework in this embodiment;

[0077] Figure 2 A flowchart of interactive recommendation in this embodiment;

[0078] Figure 3 This is a schematic diagram of a recommendation model based on a large language model proxy framework in this implementation. DETAILED DESCRIPTION

[0079] In order to facilitate the understanding of the present application, the specific embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings and embodiments. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thoroughly understood.

[0080] This implementation uses Agent as the core framework and regards the recommendation algorithm as a tool that can be understood and used by the large language model. It determines the current scenario based on the user's input and gives a satisfactory response to the user. It mainly includes two aspects: recommendation algorithm preparation and conversational recommendation system construction. In order to fully consider the versatility of the recommendation method, this implementation has two recommendation scenarios: the user already exists, that is, there is a history of item interaction, and the user does not exist, that is, a new user. When there is a history of user-item interaction, the recommendation system needs a satisfactory recommendation result. In the recommendation scenario for new users, a recommendation algorithm based on content similarity retrieval is used to obtain a recommendation result that satisfies the user.

[0081] This embodiment is a multi-scenario conversational recommendation method based on a large language model proxy framework, such as Figure 1 As shown, the method includes the following process:

[0082] Build a recommendation model based on a large language model proxy framework.

[0083] The recommendation model based on the large language model agent framework includes: a dialogue manager, a judgment agent, an item agent, a learning agent and an execution agent; and the recommendation process of the item recommendation model based on the large language model agent framework is divided into a chat scenario and a recommendation scenario.

[0084] In this embodiment, if Figure 2 As shown in the figure, the item recommendation model based on the large language model agent framework mainly adopts the operating principle of the recommendation agent framework. Based on the idea of ​​group chat nesting, it divides the chat scene and the recommendation scene, and clarifies the functions and inputs and outputs that each agent needs to complete. In the process of user interaction with the dialogue manager, if the intention of recommending a product is expressed, the dialogue manager switches the dialogue scene to the recommendation scene. When entering the recommendation scene, the learning agent corresponding to the user will obtain the user portrait based on the existing item interaction sequence, and will also obtain the user portrait based on the user's interaction history. The user portrait is input into the dialogue manager, and after the item candidate set is obtained through the recommendation algorithm, the items that the user may be most interested in are selected and output based on the similarity. If the user is not satisfied, the learning agent will readjust the user portrait based on the user's feedback, and then regenerate the recommended items until the recommendation scene is exited.

[0085] The dialogue manager is used to interact with the user in a chat scenario, decide whether to switch the recommendation scenario based on the content input by the user, and if necessary, use the content input by the user as the recommendation requirement and switch to the recommendation scenario; if not, use the content input by the user as user feedback and transmit it to the evaluation agent; generate a recommendation scenario description by judging whether the user ID exists, the number of user dialogue rounds and the recommendation requirement, and transmit it to the execution agent; receive the user portrait from the learning agent and the item description from the item agent respectively, and generate the user's recommendation result based on the user portrait and the item description and send it to the user and the evaluation agent respectively.

[0086] In this embodiment, the conversation manager (Chat Manager) is the manager of the entire session, responsible for the interaction with the user and the processing of different output and input information, using ChatGPT as the base.

[0087] The evaluation agent is used to evaluate the difference between the user feedback and the user's recommendation result, and to modify the user's recommendation result according to the difference.

[0088] In this embodiment, the critic agent is used to evaluate the difference between the recommended product and the user feedback, and put the difference into the memory module to correct the user's recommendation result, so as to make more accurate personalized recommendations in the next recommendation, using ChatGPT as the base.

[0089] The item agent is used to obtain item metadata and extract text information and image information respectively, then encode the extracted text information and image information respectively, fuse the encoded text feature representation and image feature representation, generate an item portrait and transmit it to the learning agent; use natural language generation technology to convert the item ID in the recommendation result candidate set into an item description and transmit it to the dialogue manager.

[0090] In this implementation, the item agent (ItemAgent) is responsible for the extraction and fusion of bimodal item information and the generation of item portraits, using InternLM-xcomposer as the base to support bimodal information processing. Natural language generation (NLG) technology is used to convert the structured information of item metadata into fluent natural language text.

[0091] The learning agent is used to obtain the user's dialogue interaction history and user rating data, integrate all items that have interacted with the user from the user's dialogue interaction history and user rating data, generate a user portrait based on the item portrait of the item, and then transmit the user portrait to the dialogue manager; use the item portrait to calculate the similarity between items and transmit it to the execution agent.

[0092] In this implementation, the learning agent (LearnAgent) is responsible for generating user portraits, using ChatGPT as the base.

[0093] The execution agent is used to select the recommendation algorithm to be called according to the recommendation scenario description, and based on the similarity between items, use the recommendation algorithm to generate a candidate set of recommendation results and transmit it to the item agent.

[0094] In this implementation, the action agent (ActionAgent) is the role that calls the recommendation algorithm and is also the core role of this framework. Its main function is to obtain a candidate set of recommended items, using ChatGLM3 as the base.

[0095] In this embodiment, if Figure 3 As shown in the figure, the operation process of the item recommendation model based on the large language model agent framework in the recommendation scenario is as follows: the item agent processes the image and text information in the item metadata through the powerful capabilities of the multimodal large language model, and converts the multimodal information into a concise text description. The learning agent integrates the user's conversation history and existing interaction history information through prompt engineering and thought chain reasoning technology to obtain the user's preference portrait. The dialogue manager gives a recommendation scenario description by judging whether the user ID exists and the number of conversation rounds of the user in the recommendation system. The execution agent matches the recommendation algorithm API based on the recommendation scenario description and runs the appropriate recommendation algorithm to obtain a preliminary candidate set of recommended item IDs. The item agent converts the item ID into an item description. The dialogue manager makes accurate recommendations based on the content based on the powerful capabilities of the large language model. The judging agent gives modification suggestions based on user feedback and puts the modification suggestions into the memory module, so that more accurate personalized recommendations can be made in the next recommendation.

[0096] The dialogue manager, judging agent, item agent, learning agent and execution agent are all Agents, and the composition of the Agent includes: role definition prompt, large language model base, memory module Memory and tools.

[0097] In this embodiment, the recommendation model based on the large language model agent framework is implemented in the form of multiple large language models and multiple agents, where one agent is actually a higher-level application of the large language model, including a role definition prompt, a large language model base, a memory module memory and tools; the tools are non-essential components.

[0098] The role definition prompt is used to define the role of each agent and specify the functions and behaviors performed by each agent in the recommendation model based on the large language model agent framework; and define the predefined text or instructions used by the agent when interacting with the user or other agents.

[0099] The large language model base is used to provide the Agent with language understanding and language generation capabilities, and supports the execution of the role definition prompt and the update of the memory module Memory.

[0100] In this embodiment, the large language model base is the core of the agent, providing the agent with the ability to understand and generate natural language. Different agents will use different large language models as bases to adapt to their specific task requirements.

[0101] The memory module Memory is used to store and manage the Agent's dialogue history and user information, and provide data support for the role definition prompt.

[0102] In this implementation, the memory module allows the agent to store and retrieve past interaction information for use in subsequent conversations, which is crucial for maintaining contextual coherence and personalized recommendations.

[0103] The tool is used to execute tasks according to the instructions of Prompt with the support of the large language model base and update the memory module Memory.

[0104] In this implementation, tools are optional components, but they can extend the capabilities of the agent. For example, the agent may use external APIs to obtain real-time data, or use specific algorithms to handle recommendation tasks.

[0105] In practical applications, each agent may selectively implement the above components according to its role and task requirements. For example, an agent focused on generating user portraits may focus more on memory modules and large language model bases, while an agent responsible for calling recommendation algorithms may need to integrate specific tools to perform its tasks.

[0106] The interactive recommendation model implemented in this embodiment pays more attention to the universality and practicality of the recommendation model, and has good interactivity and accuracy after deployment, use and performance testing.

[0107] Obtain several groups of user rating data and item metadata, and save all the obtained user rating data and item metadata to a recommendation model based on a large language model proxy framework.

[0108] The user rating data includes: user ID, user name, item IDs rated by the user, effective rating rate, rating text, rating timestamp and rating time.

[0109] The item metadata includes: item ID, item description information, item price, image path, item category, item name, and the browsing history or purchase history of the item purchaser.

[0110] In this embodiment, the Amazon Beauty and Amazon Clothing datasets are used as training data for the subsequent recommendation algorithm when the user already exists. The Amazon dataset contains two types of data, one is Meta item metadata, including item ID, item description information, item price, image path, item category, item name, and the items that the user who purchased the item browsed or purchased at the same time. The other is the user rating dataset, which contains user ID, user name, item ID rated by the user, effective rating rate, rating text, rating timestamp, and rating time. In the user rating dataset, a user will rate multiple items. By using a self-written dataset processing script to sort the multiple items rated by the user according to time, the user's item interaction sequence can be obtained. Each element of the item interaction sequence is an item ID.

[0111] For any user, based on the dialogue input by the user, a recommendation model based on a large language model proxy framework is used to recommend items to the user. If the user is not a new user, a fine-tuned ID-based collaborative filtering recommendation algorithm or a fine-tuned interaction history-based sequence recommendation algorithm is used to generate recommendation results for the user; if the user is a new user, a recommendation algorithm based on content similarity retrieval is used to generate recommendation results for the user.

[0112] The fine-tuning methods of the ID-based collaborative filtering recommendation algorithm and the interaction history-based sequential recommendation algorithm are the same, which are: for the algorithm model predefined in the ID-based collaborative filtering recommendation algorithm or the interaction history-based sequential recommendation algorithm, fine-tuning is achieved by specifying the user interaction history data set path and modifying the predefined algorithm model name, and the fine-tuned algorithm model weights are saved.

[0113] In this implementation, the ID-based collaborative filtering recommendation algorithm uses the LightGCN recommendation algorithm provided by RecBole, which is a recommendation algorithm code library for researchers that is easy to develop and reproduce. The algorithm is fine-tuned based on the LightGCN recommendation algorithm code and training framework provided by RecBole, that is, the model can be fine-tuned directly by adding the dataset path parameter and model name after the run.py file to obtain the weight after fine-tuning.

[0114] In this embodiment, the sequence recommendation algorithm based on interaction history adopts the SASRec recommendation algorithm provided by RecBole. The algorithm is fine-tuned based on the SASRec recommendation algorithm code and training framework provided by RecBole, which is consistent with the method of fine-tuning the ID-based collaborative filtering recommendation algorithm, using the same data set, except that the model parameter of run.py is changed to --model=SASRec.

[0115] After receiving the recommendation results, the user submits user feedback to the recommendation model based on the large language model proxy framework. The recommendation model based on the large language model proxy framework modifies the recommendation results according to the user feedback, and re-recommends items to the user according to the modified results until the user is satisfied.

[0116] The process of using the recommendation model based on the large language model proxy framework to recommend items to the user is as follows:

[0117] The conversation content input by the user is obtained and text analysis is performed. When the conversation content is a recommendation requirement, the recommendation model based on the large language model proxy framework is switched to the recommendation scenario.

[0118] Get the user ID of the user, and determine whether the user ID exists in the stored user rating data. If so, the user is a historical user, and the user rating data of the user is obtained; if not, the user is a new user, and the user's registration information is obtained as the user rating data of the user.

[0119] Generate a recommendation scenario description based on the user's user rating data, number of conversation rounds, conversation content, and recommendation requirements.

[0120] In this embodiment, the recommendation scenario description usually includes the user ID, the user's current query or behavior, such as the number of conversation turns and the content of the conversation, the user's interests, preferences, historical behaviors, etc., the user's historical interaction records with the system, such as clicks, ratings, purchases, etc., and other contextual information, such as time, location, device type, and other environmental factors that may affect user preferences.

[0121] The text and image information of each item is extracted from the item metadata stored in the recommendation model based on the large language model proxy framework, and then the extracted text information and image information are encoded respectively, and the encoded text feature representation and image feature representation are fused to generate the item portrait of each item.

[0122] According to the user's conversation interaction history and user rating data, all items that the user has interacted with are integrated, and a user profile is generated based on the item portraits of all items that the user has interacted with.

[0123] For historical users, based on the user's user rating data, recommendation scenario description and item portrait, a fine-tuned ID-based collaborative filtering recommendation algorithm or a fine-tuned interaction history-based sequential recommendation algorithm is used to generate a candidate set of recommendation results for the user; for new users, based on the user's user rating data, recommendation scenario description and item portrait, a content similarity retrieval-based recommendation algorithm is used to generate a candidate set of recommendation results for the user.

[0124] In this embodiment, the user ID based user representation U id Recommendation algorithm for collaborative filtering, user representation U based on item ID interaction sequence seq Used in the sequential recommendation algorithm based on interaction history, items in the item metadata are selected as the item ID candidate set for the recommendation algorithm based on content similarity retrieval.

[0125] The method of using the fine-tuned ID-based collaborative filtering recommendation algorithm to generate a candidate set of recommendation results for the user is:

[0126] A user interaction history dataset of the user is constructed based on the user rating data and item metadata of the user.

[0127] The method for constructing the user interaction history data set is as follows: according to the user rating data of the user, the interaction history between each item in the item metadata and the user is obtained; for any item, a user-item evaluation key-value pair is constructed with the user ID as the primary key and the interaction history between the user and the item as the value; and then the user interaction history data set of the user is constructed using the user-item evaluation key-value pairs of all items and the user; wherein the interaction history is the effective evaluation rate or evaluation text of the user to the item.

[0128] In this embodiment, each row in the user interaction history data set represents a user, and the user ID is unique and is used to distinguish different users. The value in each row represents the interaction history between the user and the item, including the user's rating, purchase times, click times, etc. of each item.

[0129] Based on the user's user interaction history dataset, an initial random embedding representation is assigned to the user and all items respectively.

[0130] The user embedding representation of the user and the item embedding representation of all items are updated by iterating and weighted summing the initial random embedding representations of the user and all items.

[0131] The updating method of the user embedding representation is:

[0132]

[0133] in represents the updated user embedding representation; u represents the user; k represents the iteration round; i represents the item; N u Represents the normalization coefficient of the user; N i Represents the normalization coefficient of the item; Represents the item embedding representation in the current round.

[0134] The updating method of the embedding representation of the item is:

[0135]

[0136] in Represents the updated item embedding representation; Indicates the user embedding representation in the current round.

[0137] Construct the rating matrix of the user based on the updated user embedding representation and the item embedding representation of all items.

[0138] The scoring matrix is ​​expressed as:

[0139]

[0140] in represents the user's rating matrix for all items; e u Indicates user embedding representation; e i Item embedding representation representing all items.

[0141] In this embodiment, the prediction part is to multiply the representations of the user and the item to obtain a rating matrix.

[0142] Remove items that the user has rated from the obtained rating matrix, and sort the items in the rating matrix from high to low according to the predicted scores of the remaining items to generate a candidate set of recommendation results.

[0143] The method of using the fine-tuned sequential recommendation algorithm based on interaction history to generate a candidate set of recommendation results for the user is:

[0144] The user rating data of the user is used to sort all items rated by the user in time, and an interaction sequence between the user and each item is generated respectively; wherein the elements in the interaction sequence are the user's interaction behaviors with the items.

[0145] In this embodiment, the user's interaction sequence may include various forms of user interactions with items, such as clicks, purchases, ratings, or browsing, etc. These interactions are used to capture the user's interests and preferences so that the model can learn and predict other items that the user may be interested in.

[0146] Define the user's behavior sequence, expressed as:

[0147]

[0148] Where S u Represents the user's behavior sequence; represents the interaction sequence between the user and the first item; U represents the user; represents the interaction sequence between the user and the second item; Represents the interaction sequence between the user and the last item; |S u |Indicates the length of the user's behavior sequence.

[0149] The length of the user's behavior sequence is limited to n.

[0150] In this embodiment, the behavior sequence is controlled to an appropriate length n, which is 16 by default, indicating that a user has interacted with 16 items. If the number exceeds 16, the behavior sequence is truncated, and if the number is less than 16, the behavior sequence is padded.

[0151] Each item in the user's behavior sequence is encoded for sequence encoding and position encoding to generate the sequence Embedding representation and position Embedding representation of the item, and the sequence Embedding representation and position Embedding representation of the item are added to obtain the sequence interaction information of the item.

[0152] The self-attention mechanism is used to learn the sequence interaction information and calculate the attention score of each item.

[0153] The function of the self-attention mechanism is expressed as:

[0154]

[0155] Where Attention(Q,K,V) represents the attention score; Q is the query vector, which is used to represent the result of multiplying the sequence interaction information with the query vector weight matrix; K is the key vector, which is used to represent the result of multiplying the sequence interaction information with the key vector weight matrix; V is the value vector, which is used to represent the result of multiplying the sequence interaction information with the value vector weight matrix; softmax represents normalization; QK T represents the similarity between Q and K; d represents the dimension size.

[0156] In this embodiment, the query vector represents the query demand or interest of the current user or item; the key vector represents the characteristics of each item in the sequence; and the value vector represents the actual information or Embedding representation of each item in the sequence. The overall functional representation of the self-attention mechanism means that for a given query vector Q, its correlation with all key vectors K is calculated, and then these correlation scores are normalized by the softmax function, and finally the value vector V is weighted and summed with the normalized scores, and the final output is a weighted representation of each item in the sequence according to its correlation with the query vector.

[0157] For each item in the user's behavior sequence, generate an item Embedding representation and calculate the relevance score of each item.

[0158] The relevance score of each item is expressed as:

[0159]

[0160] where r h,t represents the relevance score of the t-th user to the h-th item; represents the function of the t-th user in the b-th self-attention block, and is a function that depends on M; M is the Embedding interaction representation of a single item; M h Item Embedding representation representing the h-th item.

[0161] The user's behavior sequence is converted into the user's user behavior sequence output, expressed as:

[0162]

[0163] Among them, t Represents the user behavior sequence output of the t-th user; <pad>Indicates padding, used to indicate that this part of the output is invalid; pad indicates the padding value; s t User representation of the t-th user; s t+1 Represents the user representation of the predicted next user.

[0164] In this embodiment, three situations of user behavior sequence output are defined: when s t When it is equal to the pad value, the output <pad>, indicating that this part of the output is invalid. When 1 ≤ t < n, that is, within the effective length range of the sequence, the output is s t+1 , indicating the next user representation. When t is equal to the sequence length, the output is S u , indicating the complete behavior sequence of the user.

[0165] Calculate the similarity between the user behavior sequence output of this user and the item Embedding representations of all items to predict the rating of each item by this user. Delete all items that this user has rated, and generate a candidate set of recommendation results based on the relevance scores and ratings of the remaining items.

[0166] The method for generating a candidate set of recommendation results for this user using a recommendation algorithm based on content similarity retrieval is as follows:

[0167] Construct a candidate set of item IDs using item metadata.

[0168] The construction method of the candidate set of item IDs is as follows: For any item, obtain the product description information of this item from the item metadata, and use a text encoder to encode the product description information of this item to obtain the item content representation I of this item content ; Using the item ID as the primary key and the item content representation I of this item content as the value to construct the key-value pair of the item representation of this item; Use the key-value pairs of the item representations of all items to construct a candidate set of item IDs.

[0169] In this embodiment, the recommendation algorithm based on content similarity retrieval recommends based on the similarity between the user portrait representation U obtained from the context dialogue content and the item representation I in the candidate set of item IDs content . Among them, U content is obtained by the method of Prompt engineering. The large language model summarizes the text description of the user portrait based on the dialogue history, and then uses the bge model to perform an embedding representation on the user portrait. The candidate set of item IDs is in the format of key-value pairs. The product description information is encoded using a text encoder such as the bge model to obtain the item representation. bge is an open-source embedding encoding model proposed by the Beijing Academy of Artificial Intelligence.

[0170] Use a text encoder to encode the user portrait of this user to generate the user portrait representation U of this user content .

[0171] Use the user portrait representation U of this user content and the content representation I of each item in the candidate set of item IDs content to calculate the similarity between this user and all items in the candidate set of item IDs.

[0172] The similarity is calculated as follows:

[0173]

[0174] where p u,i Indicates the similarity between users and items; Normalize means normalization.

[0175] All items in the item ID candidate set are sorted in descending order according to the similarity between the user and all items in the item ID candidate set, and the first P items in the sorted result are selected to generate a recommendation result candidate set.

[0176] Output=TopP(p u,i )

[0177] Among them, Output represents the output candidate set of recommendation results; TopP represents the top P items in the selection sorting results.

[0178] Use natural language generation technology to convert the item IDs in the recommendation result candidate set into item descriptions, and generate the user's recommendation results based on the user portrait and item description and send them to the user;

[0179] The user provides feedback on the generated recommendation results, and the recommendation results are modified according to the user feedback until the user is satisfied.

[0180] In order to achieve the universality of the recommendation framework and the recommendation method, it is necessary to expand the open source large language model base. The large language model base used in the development stage of this implementation is shown in Table 1.

[0181] Table 1 Large language model base table

[0182]

[0183] The development framework used in this embodiment is LangChain, which is a framework designed specifically for building large language model applications. Based on the LangChain framework, this embodiment uses a unified deployment solution for different large language models. The core idea of ​​this solution is the C / S architecture. The open source LLM is deployed on the server side and the access interface is opened to become a customized API interface. If it is a closed source model ChatGPT, the API interface is provided by the service provider. The middle layer is the component layer. The open API interface is customized into a request access base using LangChain and inherits the LLM class. Users can also use the recommendation system without high-performance computing resources. The client is responsible for processing and packaging input data, user qualification verification and prompt template customization. The development language used in this embodiment is python, and the recommendation algorithm fine-tuning uses RecBole, which is a recommendation algorithm code library for researchers that is easy to develop and reproduce. The vector retrieval acceleration technology based on content similarity retrieval uses Faiss, which is a library for efficient similarity search and clustering of dense vectors. The tool usage scheme uses a tool usage scheme based on the definition of Tsinghua ChatGLM3 large language model. The fine-tuned recommendation algorithm is encapsulated into the API tool library. The large language model matches the most appropriate recommendation algorithm according to the API description and recommendation scenario to obtain the recommendation result. The system front-end test uses Streamlit, a python toolkit for quickly generating interactive front-ends.

[0184] In order to facilitate the understanding of the present application, the specific embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings and embodiments. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thoroughly understood.< / pad> < / pad> < / pad>

Claims

1. A multi-scenario conversational recommendation method based on a large language model proxy framework, characterized in that: The method includes the following steps: Build a recommendation model based on a large language model proxy framework; Obtain several groups of user rating data and item metadata, and save all the obtained user rating data and item metadata to the recommendation model based on the large language model proxy framework; For any user, based on the conversation input by the user, the recommendation model based on the large language model proxy framework is used to recommend items to the user. If the user is not a new user, a fine-tuned ID-based collaborative filtering recommendation algorithm or a fine-tuned interaction history-based sequential recommendation algorithm is used to generate recommendation results for the user. If the user is a new user, a recommendation algorithm based on content similarity retrieval is used to generate recommendation results for the user; The fine-tuning methods of the ID-based collaborative filtering recommendation algorithm and the interaction history-based sequential recommendation algorithm are the same, which are: for the algorithm model predefined in the ID-based collaborative filtering recommendation algorithm or the interaction history-based sequential recommendation algorithm, fine-tuning is achieved by specifying the user interaction history data set path and modifying the predefined algorithm model name, and saving the fine-tuned algorithm model weight; After receiving the recommendation results, the user submits user feedback to the recommendation model based on the large language model proxy framework. The recommendation model based on the large language model proxy framework modifies the recommendation results according to the user feedback, and re-recommends items to the user according to the modified results until the user is satisfied.

2. According to claim 1, a multi-scenario conversational recommendation method based on a large language model proxy framework is characterized in that: The recommendation model based on the large language model agent framework includes: a dialogue manager, a judgment agent, an item agent, a learning agent and an execution agent; wherein the dialogue manager, the judgment agent, the item agent, the learning agent and the execution agent are all agents; The dialogue manager is used to interact with the user in the chat scene, decide whether to switch the recommendation scene according to the content input by the user, and if necessary, use the content input by the user as the recommendation requirement and switch to the recommendation scene; if not, use the content input by the user as user feedback and transmit it to the evaluation agent; generate the recommendation scene description by judging whether the user ID exists, the number of dialogue rounds of the user and the recommendation requirement, and transmit it to the execution agent; receive the user portrait from the learning agent and the item description from the item agent respectively, and generate the user's recommendation result according to the user portrait and the item description, and send it to the user and the evaluation agent respectively; The judging agent is used to evaluate the difference between the user feedback and the user's recommendation result, and to modify the user's recommendation result according to the difference; The item agent is used to obtain item metadata and extract text information and image information respectively, then encode the extracted text information and image information respectively, fuse the encoded text feature representation and image feature representation, generate an item portrait and transmit it to the learning agent; use natural language generation technology to convert the item ID in the recommendation result candidate set into an item description and transmit it to the dialogue manager; The learning agent is used to obtain the user's dialogue interaction history and user rating data, integrate all items that have been interacted with the user from the user's dialogue interaction history and user rating data, generate a user portrait based on the item portrait of the item, and then transmit the user portrait to the dialogue manager; calculate the similarity between items using the item portrait and transmit it to the execution agent; The execution agent is used to select the recommendation algorithm to be called according to the recommendation scenario description, and based on the similarity between items, use the recommendation algorithm to generate a candidate set of recommendation results and transmit it to the item agent.

3. According to claim 2, a multi-scenario conversational recommendation method based on a large language model proxy framework is characterized in that: The user rating data includes: user ID, user name, item IDs rated by the user, effective rating rate, rating text, rating timestamp and rating time; The item metadata includes: item ID, item description information, item price, image path, item category, item name, and the browsing history or purchase history of the item purchaser.

4. According to claim 3, a multi-scenario conversational recommendation method based on a large language model proxy framework is characterized in that: The process of using the recommendation model based on the large language model proxy framework to recommend items to the user is as follows: Obtain the conversation content input by the user and perform text analysis. When the conversation content is a recommendation requirement, switch the recommendation model based on the large language model proxy framework to the recommendation scenario; Obtain the user ID of the user, and determine whether the user ID exists in the stored user rating data. If so, the user is a historical user, and the user rating data of the user is obtained; if not, the user is a new user, and the user's registration information is obtained and used as the user rating data of the user; Generate a recommendation scenario description based on the user's user rating data, number of conversation turns, and recommendation requirements; Extract the text information and image information of each item from the item metadata stored in the recommendation model based on the large language model proxy framework, encode the extracted text information and image information respectively, fuse the encoded text feature representation and image feature representation, and generate the item portrait of each item; Integrate all items that the user has interacted with based on the user's conversation interaction history and user rating data, and generate a user profile based on the item portraits of all items that the user has interacted with. For historical users, based on the user's user rating data, recommendation scenario description and item portrait, a fine-tuned ID-based collaborative filtering recommendation algorithm or a fine-tuned interaction history-based sequential recommendation algorithm is used to generate a candidate set of recommendation results for the user; for new users, based on the user's user rating data, recommendation scenario description and item portrait, a content similarity retrieval-based recommendation algorithm is used to generate a candidate set of recommendation results for the user; Use natural language generation technology to convert the item IDs in the recommendation result candidate set into item descriptions, and generate the user's recommendation results based on the user portrait and item description and send them to the user; The user provides feedback on the generated recommendation results, and the recommendation results are corrected based on the user feedback. Then, items are re-recommended to the user based on the corrected results until the user is satisfied.

5. According to claim 4, a multi-scenario conversational recommendation method based on a large language model proxy framework is characterized in that: The method of using the fine-tuned ID-based collaborative filtering recommendation algorithm to generate a candidate set of recommendation results for the user is: Construct a user interaction history dataset of the user based on the user rating data and item metadata of the user; The method for constructing the user interaction history data set is as follows: according to the user rating data of the user, the interaction history between each item in the item metadata and the user is obtained; for any item, a user-item evaluation key-value pair is constructed with the user ID as the primary key and the interaction history between the user and the item as the value; and then the user interaction history data set of the user is constructed using the user-item evaluation key-value pairs of all items and the user; wherein the interaction history is the effective evaluation rate or evaluation text of the user on the item; Based on the user's user interaction history dataset, an initial random embedding representation is assigned to the user and all items respectively; Update the user embedding representation of the user and the item embedding representation of all items by iterating and weighted summing the initial random embedding representations of the user and all items: According to the updated user embedding representation and the item embedding representation of all items, the rating matrix of the user is constructed, which is expressed as: in represents the user's rating matrix for all items; e u Indicates user embedding representation; e i Item embedding representation representing all items; Remove items that the user has rated from the obtained rating matrix, and sort the items in the rating matrix from high to low according to the predicted scores of the remaining items to generate a candidate set of recommendation results.

6. According to claim 5, a multi-scenario conversational recommendation method based on a large language model proxy framework is characterized in that: The updating method of the user embedding representation is: in represents the updated user embedding representation; u represents the user; k represents the iteration round; i represents the item; N u Represents the normalization coefficient of the user; N i Represents the normalization coefficient of the item; Indicates the embedding representation of items in the current round; The update method of the embedding representation of the item is: in Represents the updated item embedding representation; Indicates the user embedding representation in the current round.

7. According to claim 6, a multi-scenario conversational recommendation method based on a large language model proxy framework is characterized in that: The method of using the fine-tuned sequential recommendation algorithm based on interaction history to generate a candidate set of recommendation results for the user is: Using the user rating data of the user, all items rated by the user are sorted in time, and an interaction sequence between the user and each item is generated respectively; wherein the elements in the interaction sequence are the user's interaction behaviors with the items; Define the user's behavior sequence, expressed as: Where S u Represents the user's behavior sequence; represents the interaction sequence between the user and the first item; U represents the user; represents the interaction sequence between the user and the second item; Represents the interaction sequence between the user and the last item; |S u | represents the length of the user's behavior sequence; The length of the user's behavior sequence is limited to n; Encode each item in the user's behavior sequence for sequence encoding and position encoding, generate the item's sequence Embedding representation and position Embedding representation, and add the item's sequence Embedding representation and position Embedding representation to obtain the item's sequence interaction information; The self-attention mechanism is used to learn the sequence interaction information and calculate the attention score of each item; For each item in the user's behavior sequence, generate an item Embedding representation and calculate the relevance score of each item; Converting a user's behavior sequence into a user behavior sequence of the user; The similarity between the user behavior sequence output of the user and the item Embedding representation of all items is calculated to predict the user's rating of each item, delete all items rated by the user, and generate a candidate set of recommendation results based on the relevance scores and ratings of the remaining items.

8. According to claim 7, a multi-scenario conversational recommendation method based on a large language model proxy framework is characterized in that: The relevance score of each item is expressed as: where r h,t represents the relevance score of the t-th user to the h-th item; represents the function of the t-th user in the b-th self-attention block, and is a function that depends on M; M is the Embedding interaction representation of a single item; M h Item Embedding representation representing the h-th item.

9. According to claim 8, a multi-scenario conversational recommendation method based on a large language model proxy framework is characterized in that: The user behavior sequence output is expressed as: Among them, t Represents the user behavior sequence output of the t-th user; <pad>Indicates filling; pad indicates the filling value; s t User representation of the t-th user; s t+1 Represents the user representation of the predicted next user.< / pad> 10. According to claim 9, a multi-scenario conversational recommendation method based on a large language model proxy framework is characterized in that: The method of using the recommendation algorithm based on content similarity retrieval to generate a candidate set of recommendation results for the user is: Use item metadata to build a candidate set of item IDs; The method for constructing the candidate set of item IDs is as follows: for any item, obtain the commodity description information of the item from the item metadata, and use a text encoder to encode the commodity description information of the item to obtain the item content representation I of the item. content ; With item ID as primary key, the item content of the item represents I content Build the item ID candidate set using the item ID key-value pairs of all items. The user profile of the user is encoded using a text encoder to generate a user profile representation U content ; The user portrait of the user is used to represent U content and the content representation of each item in the item ID candidate set I content , calculate the similarity between the user and all items in the item ID candidate set; The similarity is calculated as follows: where p u,i Indicates the similarity between users and items; Normalize means normalization; All items in the item ID candidate set are sorted in descending order according to the similarity between the user and all items in the item ID candidate set, and the first P items in the sorted result are selected to generate a recommendation result candidate set.

Images