Intelligent agent recommendation method and device and computer readable storage medium
By obtaining user query information and context information, combined with search enhancement generation model and large model technology, the problem of weak data sparseness and generalization in the agent recommendation technology is solved, improving the accuracy and diversity of recommendations, and generating more in line with users. Personalized recommendation results for demand.
Patent Information
- Application Number
- CN202410932865.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-12
- Publication Date
- 2024-11-15
AI Technical Summary
The existing agent recommendation technology has problems of sparse data and weak generalization, making it difficult to provide better solutions.
By obtaining user query information and context information, fuse the input data, and using search enhancement generation model and large model technology, search text related to user interests from the database, dynamically adjust the search strategy, fine-tune the recommended intelligent model, and improve model training Efficiency and recommendation accuracy.
It improves the accuracy, diversity and generalization of agent recommendations, can better understand user needs, and generate personalized and customized recommendation results.
Smart Images

Figure CN118964590A_ABST
Abstract
Description
Technical Field
[0001] The present application belongs to the field of artificial intelligence technology, and specifically relates to an intelligent agent recommendation method, device and computer-readable storage medium. Background Art
[0002] With the rapid development of the Internet, the problem of information overload has become increasingly serious. Users often feel overwhelmed when faced with a huge amount of items, content, services, etc. to choose from.
[0003] Intelligent agent recommendation technology is an application technology based on artificial intelligence big models, which aims to provide users with personalized and customized recommended content, products or services. This technology relies on big data analysis and machine learning algorithms, and generates recommendation results that meet users' interests and needs by mining and analyzing user behavior data. This technology can combine classic content-based recommendation, collaborative filtering recommendation, deep learning recommendation and other types of methods with big model technology to generate customized recommendation lists for users, and can optimize recommendation algorithms and strategies through common performance evaluation indicators to improve the effectiveness of intelligent agent recommendation technology and user satisfaction.
[0004] However, the existing intelligent agent recommendation technology is still in its early stages of development and has the disadvantages of sparse data and weak generalization. How to provide a better solution is a technical problem that technicians in this field urgently need to solve. Summary of the invention
[0005] In view of the above analysis, an embodiment of the present invention aims to provide an intelligent agent recommendation method to improve the accuracy, diversity and generalization of intelligent agent recommendations.
[0006] In a first aspect of the present application, a method for recommending an intelligent agent is provided, comprising:
[0007] Obtain user query information and context information;
[0008] Fusing the user query information and the context information to generate fused input data;
[0009] Inputting the input data into the recommendation agent model for prediction to obtain a prediction result;
[0010] Generating final recommendation data according to the prediction results;
[0011] The recommendation agent model is constructed in the following way:
[0012] Acquire initial training set data, wherein the initial training set data includes the user's historical behavior data, context information, and recommended domain data;
[0013] Inputting the initial training set data into the large model for feature extraction to obtain a pre-trained recommendation agent model;
[0014] A retrieval enhancement generation model is used to retrieve texts related to user interests from a database, and the candidate set obtained by retrieval enhancement is used as a target training sample; the retrieval enhancement generation model is a model with a self-feedback mechanism, which dynamically adjusts the retrieval strategy according to the generated self-feedback results and updates the generated retrieval results; the pre-trained recommendation agent model is fine-tuned using the target training sample to construct a recommendation agent model.
[0015] Optionally, the process of generating a candidate set by the retrieval enhancement generation model includes:
[0016] According to the user query information and context information, retrieve the text related to the user's interests from the database as the initial candidate set D;
[0017] D={d 1 , d 2 , ..., d n = Retrieval(Q, C)
[0018] Among them, d i Represents the i-th candidate text retrieved;
[0019] Screening and sorting the initial candidate set D, selecting the text most relevant to the user's interests as input and feeding it to the generation layer of the retrieval enhancement generation model to generate a screened candidate set D′;
[0020] D′={d′ 1 , d′ 2 , ..., d′ m}∈D
[0021] Among them, m represents the size of the candidate set after screening, and m< <n。
[0022] Optionally, the self-feedback process of the retrieval enhancement generation model includes:
[0023] In the training process of the retrieval enhancement generation model, a self-feedback result F is generated based on the aggregation of user interaction behavior indicators:
[0024] F={f 1 , f 2 , ..., f n}
[0025] Among them, f i represents the i-th feedback signal;
[0026] Based on the obtained self-feedback result F, the retrieval model is updated:
[0027] RM updated =Update(RetrievalModel, F)
[0028] Using the updated retrieval model and the self-feedback result F, the parameters of the adjusted model are obtained:
[0029] GM adjusted =Adjusted(GM,RM updated , F)
[0030] Among them, Update() is a model retrieval function, which is used to adjust the model parameters according to the classic retrieval model and the self-feedback result F; RetrievalModel is a classic retrieval model; RM updated is the updated retrieval model; GM is the model parameter; GM adjusted are the parameters of the adjusted model;
[0031] Dynamically adjust the retrieval strategy by iteratively optimizing the retrieval model and updating the generated retrieval results:
[0032] (RM next , GM next )=Iterate(RM updated , GM adjusted , D′, F)
[0033] Among them, Iterate() is an iterative optimization function, which is used to update the model according to the new candidate set and self-feedback results in each training iteration.
[0034] Optionally, the step of inputting the initial training set data into a large model for feature extraction to obtain a pre-trained recommendation agent model comprises:
[0035] Unlabeled data and user behavior data are input into the big model for pre-training to obtain general domain knowledge in the agent recommendation task; the big model is a Transformer model, and the pre-training process is expressed as:
[0036]
[0037] Among them, L pre is the pre-training loss, and P(y|x;θ) is the probability of generating the correct output y given the input x when the model parameters are θ.
[0038] Optionally, the adopting the target training sample to fine-tune the pre-trained recommendation agent model to construct the recommendation agent model includes:
[0039] The pre-trained recommendation agent model is fine-tuned using the pre-given query information q and context information c. The fine-tuning process is expressed as:
[0040]
[0041] Among them, L fin is the fine-tuning loss, and P(x|q, c; θ′) is the probability of generating the agent recommendation item x given the query information q and context information c when the model parameters are θ′.
[0042] Optionally, obtaining the context information includes:
[0043] Based on the user profile information and environmental context information, context information C is obtained:
[0044] C=(P u , E c )
[0045] Among them, P u The user portrait information is obtained by modeling and calculating by aggregating the user's historical behavior data. c It is the environmental context information.
[0046] Optionally, fusing the user query information and the context information to generate fused input data includes:
[0047] Fusing the user query information Q and the context information C to generate fused input data X;
[0048] X=f(Q,C)
[0049] Where, f() represents the data fusion function;
[0050] After generating the fused input data X, it also includes:
[0051] Preprocess the fused input data X to generate an input vector V suitable for large model processing;
[0052] V = g(X)
[0053] Among them, g() represents the preprocessing function.
[0054] Optionally, generating final recommendation data according to the prediction result includes:
[0055] Sort the prediction results by their recommendation scores to generate the final recommendation list R:
[0056] R = f(r 1 , r 2 , ..., r k)
[0057] Among them, r i It is the result of sorting according to the recommendation score of the predicted results.
[0058] Optionally, after generating the final recommendation data, the following steps are also included:
[0059] Obtain user interaction behavior indicators for each recommended data in the recommendation list in real time;
[0060] Reflective word segmentation is used to analyze the user interaction behavior indicators, and the generated recommendation list is evaluated to obtain an evaluation result;
[0061] The recommended strategy of the model is adjusted according to the evaluation results.
[0062] Optionally, the user interaction behavior indicator includes any one or any combination of the following: click behavior, browsing behavior, interaction time, scrolling behavior, search behavior, interactive elements, navigation behavior, social interaction, user feedback, error rate and exit rate.
[0063] According to a second aspect of the present application, there is provided an intelligent agent recommendation device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, any of the above-mentioned intelligent agent recommendation methods is implemented.
[0064] According to a third aspect of the present application, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the computer program implements any of the above-mentioned intelligent agent recommendation methods.
[0065] The agent recommendation method provided in the present application obtains user query information and context information; fuses the user query information and context information to generate fused input data; inputs the input data into the recommendation agent model for prediction to obtain prediction results; and generates final recommendation data based on the prediction results. Wherein, the recommendation agent model is constructed in the following manner: obtain initial training set data, the initial training set data includes the user's historical behavior data, context information and recommendation field data; inputs the initial training set data into a large model for feature extraction to obtain a pre-trained recommendation agent model; uses a retrieval enhancement generation model to retrieve texts related to user interests from a database, and uses the candidate set obtained by retrieval enhancement as a target training sample; the retrieval enhancement generation model is a model with a self-feedback mechanism, dynamically adjusts the retrieval strategy according to the generated self-feedback results, and updates the generated retrieval results; uses the target training sample to fine-tune the pre-trained recommendation agent model to construct a recommendation agent model. The present application generates targeted target training samples from massive text data through retrieval enhancement generation technology, thereby improving the efficiency of model training. Moreover, by combining the generation capabilities of large model technology, more complex relationships between users and items can be learned, thereby further improving the accuracy, diversity and generalization of intelligent recommendations.
[0066] In addition, the present application also provides an intelligent agent recommendation device and a computer-readable storage medium having the above-mentioned technical effects. BRIEF DESCRIPTION OF THE DRAWINGS
[0067] In order to more clearly illustrate the embodiments of this specification or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of this specification. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.
[0068] Figure 1 A flowchart of the intelligent agent recommendation method provided in this application;
[0069] Figure 2 Schematic diagram of the process of optimizing feedback for the recommendation agent;
[0070] Figure 3 This is a structural block diagram of the intelligent agent recommendation device provided in this application. DETAILED DESCRIPTION
[0071] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. It should be noted that, in the absence of conflict, the embodiments in the present disclosure and the features in the embodiments can be combined, separated, interchanged and / or rearranged with each other. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.
[0072] The terms used here are for the purpose of describing specific embodiments, and are not intended to be restrictive. As used here, unless the context clearly indicates otherwise, the singular forms "one (kind, person)" and "said (the)" are also intended to include plural forms. In addition, when the terms "comprise" and / or "include" and their variations are used in this specification, it is explained that there are stated features, integral bodies, steps, operations, parts, assemblies and / or their groups, but it is not excluded that there are or add one or more other features, integral bodies, steps, operations, parts, assemblies and / or their groups. It should also be noted that, as used here, the terms "substantially", "approximately" and other similar terms are used as approximate terms and not as degree terms, so that they are used to explain the inherent deviations of the measured values, calculated values and / or the values provided that will be recognized by those of ordinary skill in the art.
[0073] This application provides an intelligent agent recommendation method, such as Figure 1 The flow chart of the agent recommendation method provided in this application is as shown in the following figure. The method specifically includes:
[0074] S101: Obtain user query information and context information.
[0075] It can be understood that user query information and context information provide input data for the retrieval and generation process of the intelligent agent recommendation technology.
[0076] The user query information Q may be a direct input statement of the user, such as a search keyword, a question or a paragraph statement. The model input layer uses the user query information as initial input data, where Q represents the user query information text.
[0077] Wherein, obtaining the context information includes: obtaining context information C based on user portrait information and environmental context information:
[0078] C=(P u , E c )
[0079] Among them, P uThe user portrait information is obtained by modeling and calculating by aggregating the user's historical behavior data. c is the environmental context information. c It can contain preference information such as current time, location, project type, etc.
[0080] The context information C includes the user's historical behavior, interest preferences, time information, location information, etc., which are not limited here.
[0081] S102: Fusing the user query information and the context information to generate fused input data.
[0082] Fusing the user query information Q and the context information C to generate fused input data X;
[0083] X=f(Q,C)
[0084] Where f() represents the data fusion function. It can be a simple vector concatenation operation or other complex neural network structures, such as Bi-LSTM or Transformer. r etc., to capture the interaction information between query and context.
[0085] After generating the fused input data X, the method further includes: performing a preprocessing operation on the fused input data X to generate an input vector V suitable for large model processing;
[0086] V = g(X)
[0087] Among them, g() represents a preprocessing function. Usually, operations such as word embedding representation model or positional encoding can be used.
[0088] Specifically, the preprocessing operation may be performing word segmentation, encoding, and other processing on the fused input data X.
[0089] Through the above process, the input layer converts user query information and context information into an input vector V that can be processed by the large model, providing training data for the following retrieval and generation processes. This design ensures that the recommendation agent can fully understand the user’s current needs and preferences, thereby generating personalized and accurate recommendation results.
[0090] S103: Input the input data into the recommendation agent model for prediction to obtain a prediction result.
[0091] S104: Generate final recommendation data according to the prediction results.
[0092] Sort the prediction results by their recommendation scores to generate the final recommendation list R:
[0093] R = f(r 1 , r 2 , ..., r k )
[0094] Among them, r i It is the result of sorting according to the recommendation score of the predicted results.
[0095] The recommendation agent model is constructed in the following way:
[0096] Acquire initial training set data, which includes the user's historical behavior data, context information and recommended field data; input the initial training set data into a large model for feature extraction to obtain a pre-trained recommendation agent model; use a retrieval enhancement generation model to retrieve texts related to user interests from a database, and use the candidate set obtained by retrieval enhancement as a target training sample; the retrieval enhancement generation model is a model with a self-feedback mechanism, which dynamically adjusts the retrieval strategy according to the generated self-feedback results and updates the generated retrieval results; use the target training sample to fine-tune the pre-trained recommendation agent model to construct a recommendation agent model.
[0097] The process of retrieving the enhanced generation model to generate candidate sets includes:
[0098] According to the user query information and context information, retrieve the text related to the user's interests from the database as the initial candidate set D;
[0099] D={d 1 , d 2 , ..., d n = Retrieval(Q, C)
[0100] Among them, d i represents the i-th candidate text retrieved;
[0101] Screening and sorting the initial candidate set D, selecting the text most relevant to the user's interests as input and feeding it to the generation layer of the retrieval enhancement generation model to generate a screened candidate set D′;
[0102] D′={d′ 1 , d′ 2 , ..., d′ m}∈D
[0103] Among them, m represents the size of the candidate set after screening, and m< <n。
[0104] The self-feedback process of the retrieval-augmented generative model includes:
[0105] In the training process of the retrieval enhancement generation model, a self-feedback result F is generated based on the aggregation of user interaction behavior indicators:
[0106] F={f 1 , f 2 , ..., f n}
[0107] Among them, f i represents the i-th feedback signal;
[0108] Based on the obtained self-feedback result F, the retrieval model is updated:
[0109] RM updated =Update(RetrievalModel, F)
[0110] Using the updated retrieval model and the self-feedback result F, the parameters of the adjusted model are obtained:
[0111] GM adjusted =Adjusted(GM,RM updated , F)
[0112] Among them, Update() is a model retrieval function, which is used to adjust the model parameters according to the classic retrieval model and the self-feedback result F; RetrievalModel is a classic retrieval model; RM updated is the updated retrieval model; GM is the model parameter; GM adjusted are the parameters of the adjusted model;
[0113] Dynamically adjust the retrieval strategy by iteratively optimizing the retrieval model and updating the generated retrieval results:
[0114] (RM next , GM next )=Iterate(RM updated , GM adjusted , D′, F)
[0115] Among them, Iterate() is an iterative optimization function, which is used to update the model according to the new candidate set and self-feedback results in each training iteration.
[0116] Based on any of the above embodiments, after generating the final recommendation data, the present application also includes: obtaining user interaction behavior indicators for each recommendation data in the recommendation list in real time; using reflective word segmentation to analyze the user interaction behavior indicators, evaluating the generated recommendation list, and obtaining an evaluation result; and adjusting the recommendation strategy of the model according to the evaluation result.
[0117] It can be understood that the user interaction behavior indicators include any one or any combination of the following: click behavior, browsing behavior, interaction time, scrolling behavior, search behavior, interactive elements, navigation behavior, social interaction, user feedback, error rate and exit rate.
[0118] By obtaining user feedback in real time, we design a Critic Token to analyze user interaction information and adjust the recommendation strategy based on this feedback information, so that the recommendation agent can more flexibly adapt to the changing needs of users. Figure 2 The process of feedback optimization for the recommendation agent is shown in the schematic diagram. The process includes the following steps:
[0119] S201: Initial search.
[0120] According to the user query Q and context information C provided by the input layer, the texts related to the user's interests are retrieved from the large-scale text database as the initial candidate set D. Then the candidate set is obtained based on the vector retrieval model of similarity. Its specific definition is as follows.
[0121] D={d 1 , d 2 , ..., d n = Retrieval(Q, C)
[0122] Among them, d i Represents the i-th candidate text retrieved.
[0123] S202: Candidate set screening.
[0124] Screening and sorting the initial candidate set D, selecting the text most relevant to the user's interests as input and feeding it to the generation layer of the retrieval enhancement generation model to generate a screened candidate set D′;
[0125] D′={d′ 1 , d′ 2 , ..., d′ m}∈D
[0126] Among them, m represents the size of the candidate set after screening, and m< <n。
[0127] Among them, the screening process of text fragments can use user-related sorting algorithms, such as based on click rate, based on user logs, etc.
[0128] S203: Self-feedback mechanism and optimized retrieval stage.
[0129] In the model training, the present invention introduces the critic token into the model gradient training process, and then generates the final feedback signal F based on the aggregation of user interaction behaviors (such as clicks, browsing time, etc.), and uses it to adjust the search results. The self-feedback signal F is defined as follows.
[0130] F={f 1 , f 2 , ..., f n}
[0131] Among them, f i represents the i-th feedback signal;
[0132] Based on the obtained self-feedback result F, the retrieval model is updated:
[0133] RM updated =Update(RetrievalModel, F)
[0134] Among them, Update() is a model retrieval function, which is used to adjust model parameters according to the classic retrieval model and the self-feedback result F; RetrievalModel is a classic retrieval model; such as a probability model, a vector space model, etc.
[0135] In addition, the updated retrieval model and the self-feedback result F are used to obtain the parameters of the adjusted model:
[0136] GM adjusted =Adjusted(GM,RM updated , F)
[0137] RM updated is the updated retrieval model; GM is the model parameter; GM adjusted are the parameters of the adjusted model.
[0138] S204: Iterative optimization process.
[0139] Dynamically adjust the retrieval strategy by iteratively optimizing the retrieval model and updating the generated retrieval results:
[0140] (RM next ,GM next )=Iterate(RM updated ,GM adjusted , D′, F)
[0141] Among them, Iterate() is an iterative optimization function, which is used to update the model according to the new candidate set and self-feedback results in each training iteration.
[0142] Through the above process, the retrieval enhancement generation model based on the self-feedback mechanism can dynamically adjust the retrieval strategy and improve the accuracy and efficiency of the agent recommendation technology. In addition, the model can flexibly adapt to user needs and provide diversified and customized recommendation services.
[0143] Furthermore, after obtaining the candidate set for retrieval enhancement, the present application uses the large model technology to pre-train and fine-tune the candidate set. In the pre-training stage, the present application uses the Transformer model for model training. r The model is a deep learning model based on the self-attention mechanism. The model is initialized through a large-scale data set to obtain an initial model with certain recommendation capabilities. The key to this stage is to use a large-scale recommendation data set to perform initial training on the model to obtain the basic recommendation capabilities of the model. In the fine-tuning stage, retrieval technology is used to filter out samples related to user interests from the candidate set, and these samples are mixed with the original data set for training to further optimize the performance of the model. Specifically, through technologies such as inverted indexing, relevant data samples are retrieved from the candidate set based on the keywords of the user's query to construct a fine-tuning data set. Subsequently, the fine-tuning data set is mixed with the original data set, and the model is fine-tuned to improve the generalization ability and recommendation accuracy of the model.
[0144] Among them, in the pre-training stage, large-scale unlabeled data and user behavior data are used to pre-train the large model, with the goal of acquiring general domain knowledge in the intelligent agent recommendation task.
[0145] Inputting the initial training set data into the big model for feature extraction to obtain a pre-trained recommendation agent model includes: inputting unlabeled data and user behavior data into the big model for pre-training to obtain general domain knowledge in the agent recommendation task; the big model is a Transformer model, and the pre-training process is expressed as:
[0146]
[0147] Among them, L pre is the pre-training loss, and P(y|x;θ) is the probability of generating the correct output y given the input x when the model parameters are θ.
[0148] In the fine-tuning stage, the pre-trained large model will be fine-tuned using the user's given query q and context information c, with the goal of making the model better suited to specific recommendation tasks.
[0149] The pre-trained recommendation agent model is fine-tuned using the target training sample to construct a recommendation agent model including:
[0150] The pre-trained recommendation agent model is fine-tuned using the pre-given query information q and context information c. The fine-tuning process is expressed as:
[0151]
[0152] Among them, L fin is the fine-tuning loss, and P(x|q, c; θ′) is the probability of generating the agent recommendation item x given the query information q and context information c when the model parameters are θ′.
[0153] This application uses retrieval enhancement generation technology to generate targeted target training samples from massive text data to improve the efficiency of model training. In addition, by combining the generation capabilities of large model technology, it learns more complex relationships between users and items, thereby further improving the accuracy, diversity and generalization of intelligent agent recommendations.
[0154] In addition, the present application also provides an intelligent agent recommendation device, such as Figure 3 As shown in the structural block diagram of the intelligent agent recommendation device provided in the present application, the device includes a memory 31 and a processor 32, and the memory 31 stores a computer program. When the computer program is executed by the processor 32, any of the above-mentioned intelligent agent recommendation methods is implemented.
[0155] It is understandable that the recommendation agent can provide customized recommendation services to users based on artificial intelligence big model technology. It can build a preference retrieval model that can be customized by users by analyzing the user's historical behavior, contextual information and project characteristics, thereby generating a recommendation list that meets user needs. Therefore, the architecture of the agent recommendation device includes key components such as agent construction, prediction result generation, and evaluation and feedback. Among them, the agent construction process is responsible for converting the original data into feature representation, and then building a recommendation model based on big model technology or deep learning algorithms; the prediction result generation process generates a recommendation list based on the model prediction results; the evaluation and feedback process evaluates the recommendation list and adjusts the optimization model based on the evaluation results. Ultimately, the goal of the recommendation agent is to provide accurate and personalized customized recommendation services, improve user experience and satisfaction, and generate recommendations that better meet user needs.
[0156] Recommendation agent construction: Collect users’ historical behavior data, contextual information, and recommendation domain data, clean and format them for model training, and then use the pre-trained large model to extract features from the input data to build the recommendation agent model, which is defined as follows.
[0157] f(x;θ)=g(h(x,θ) h ), c; θ g )
[0158] Where x is the input data, h(x, θ h ) is the feature extraction function, c is the context information, g() is the recommendation generation function, and θ is the model parameter.
[0159] Prediction result generation: For new user query q and context information c, the constructed recommendation agent model is used to make predictions, and the prediction results can be expressed as shown below.
[0160]
[0161] in, is the prediction result recommended by the agent.
[0162] The final recommendation list R is generated based on the prediction results, which can be expressed as follows.
[0163] R = f(r 1 , r 2 , ..., r k )
[0164] Among them, r i It is the result of sorting according to the agent recommendation score.
[0165] Evaluation and feedback: Use click-through rate, conversion rate or other relevant indicators to evaluate the generated recommendation list R, and then further adjust and optimize the model based on the evaluation results.
[0166] In summary, this application, by combining retrieval enhancement generation and large model technology, can quickly screen out high-quality candidate sets related to user interests, generate content that better meets the user's personalized needs, and effectively improve the accuracy and efficiency of recommendation services. By constructing a recommendation agent, this application can better capture the user's dynamic preferences and complex behavioral characteristics, thereby generating customized recommendation services that better meet user needs and alleviate the data sparsity and inefficient sampling problems in the recommendation field. The intelligent agent recommendation technology proposed in this application can be adaptively expanded according to different application scenarios and data characteristics to meet the needs of user groups and business in different fields. It can be widely used in various fields such as e-commerce and social media to increase user recommendation satisfaction and the commercial value of the platform.
[0167] In addition, the present application also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements any of the above-mentioned intelligent agent recommendation methods.
[0168] Computer readable storage media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.
[0169] The professionals should also be further aware that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented with electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.
[0170] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.
[0171] The specific implementation methods described above further illustrate the purpose, technical solutions and beneficial effects of the present application in detail. It should be understood that the above description is only the specific implementation method of the present application and is not intended to limit the scope of protection of the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.
Claims
1. An agent recommendation method, It is characterized in that include: Obtain user query information and context information; Fusing the user query information and the context information to generate fused input data; Inputting the input data into the recommendation agent model for prediction to obtain a prediction result; Generating final recommendation data according to the prediction results; The recommendation agent model is constructed in the following way: Acquire initial training set data, wherein the initial training set data includes the user's historical behavior data, context information, and recommended domain data; Inputting the initial training set data into the large model for feature extraction to obtain a pre-trained recommendation agent model; A retrieval enhancement generation model is used to retrieve texts related to user interests from a database, and the candidate set obtained by retrieval enhancement is used as a target training sample; the retrieval enhancement generation model is a model with a self-feedback mechanism, which dynamically adjusts the retrieval strategy according to the generated self-feedback results and updates the generated retrieval results; the pre-trained recommendation agent model is fine-tuned using the target training sample to construct a recommendation agent model.
2. The agent recommendation method according to claim 1, It is characterized in that The process of generating a candidate set by the retrieval enhancement generation model includes: According to the user query information and context information, retrieve the text related to the user's interests from the database as the initial candidate set D; D={d 1 ,d 2 ,...,d n }=Retrieval(Q,C)} Among them, d i Represents the i-th candidate text retrieved; Screening and sorting the initial candidate set D, selecting the text most relevant to the user's interests as input and feeding it to the generation layer of the retrieval enhancement generation model to generate a screened candidate set D′; D′={d′ 1 ,d 2 ,...,d m }∈D Wherein, m represents the size of the candidate set after screening, and m<<n.
3. The agent recommendation method according to claim 2, It is characterized in that The self-feedback process of the retrieval enhancement generation model includes: In the training process of the retrieval enhancement generation model, a self-feedback result F is generated based on the aggregation of user interaction behavior indicators: F={f 1 ,f 2 ,...,f n } Among them, f i represents the i-th feedback signal; Based on the obtained self-feedback result F, the retrieval model is updated: RM updated =Update(RetrievalModel,F) Using the updated retrieval model and the self-feedback result F, the parameters of the adjusted model are obtained: GM adjusted =Adjusted(GM,RM updated ,F) Among them, Update() is a model retrieval function, which is used to adjust the model parameters according to the classic retrieval model and the self-feedback result F; RetrievalModel is a classic retrieval model; RM updated is the updated retrieval model; GM is the model parameter; GM adjusted are the parameters of the adjusted model; Dynamically adjust the retrieval strategy by iteratively optimizing the retrieval model and updating the generated retrieval results: (RM next ,GM next )=Iterate(RM updated ,GM adjusted ,D′,F) Among them, Iterate() is an iterative optimization function, which is used to update the model according to the new candidate set and self-feedback results in each training iteration.
4. The agent recommendation method according to claim 1, It is characterized in that The step of inputting the initial training set data into the large model for feature extraction to obtain a pre-trained recommendation agent model comprises: Unlabeled data and user behavior data are input into the big model for pre-training to obtain general domain knowledge in the agent recommendation task; the big model is a Transformer model, and the pre-training process is expressed as: Among them, L pre is the pre-training loss, and P(y|x;θ) is the probability of generating the correct output y given the input x when the model parameters are θ.
5. The agent recommendation method according to claim 4, It is characterized in that The step of fine-tuning the pre-trained recommendation agent model by using the target training sample to construct the recommendation agent model comprises: The pre-trained recommendation agent model is fine-tuned using the pre-given query information q and context information c. The fine-tuning process is expressed as: Among them, L fin is the fine-tuning loss, and P(x|q,c;θ′) is the probability of generating the agent recommendation item x given the query information q and context information c when the model parameters are θ′.
6. The agent recommendation method according to any one of claims 1 to 5, It is characterized in that Acquiring the context information includes: Based on the user profile information and environmental context information, context information C is obtained: C=(P u ,E c ) Among them, P u The user portrait information is obtained by modeling and calculating by aggregating the user's historical behavior data. c It is the environmental context information.
7. The agent recommendation method according to any one of claims 1 to 5, It is characterized in that The fusing the user query information and the context information to generate fused input data includes: Fusing the user query information Q and the context information C to generate fused input data X; X=f(Q,C) Where, f() represents the data fusion function; After generating the fused input data X, it also includes: Preprocess the fused input data X to generate an input vector V suitable for large model processing; V = g(X) Among them, g() represents the preprocessing function.
8. The agent recommendation method according to any one of claims 1 to 5, It is characterized in that Generating final recommendation data according to the prediction result includes: Sort the prediction results by their recommendation scores to generate the final recommendation list R: R=f(r 1 ,r 2 ,...,r k ) Among them, r i It is the result of sorting according to the recommendation score of the predicted results.
9. The agent recommendation method according to claim 8, It is characterized in that After generating the final recommendation data, it also includes: Obtain user interaction behavior indicators for each recommended data in the recommendation list in real time; Reflective word segmentation is used to analyze the user interaction behavior indicators, and the generated recommendation list is evaluated to obtain an evaluation result; The recommended strategy of the model is adjusted according to the evaluation results.
10. The agent recommendation method according to claim 9, It is characterized in that The user interaction behavior indicators include any one or any combination of the following: click behavior, browsing behavior, interaction time, scrolling behavior, search behavior, interactive elements, navigation behavior, social interaction, user feedback, error rate and exit rate.
11. An intelligent agent recommendation device, It is characterized in that It comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the intelligent agent recommendation method according to any one of claims 1-10 is implemented.
12. A computer-readable storage medium, It is characterized in that A computer program is stored thereon, and when the computer program is executed by a processor, the intelligent agent recommendation method according to any one of claims 1-10 is implemented.
Citation Information
Patent Citations
Information recommendation method, device and equipment, storage medium and program product
CN116452263A
Application recommendation method and device, electronic equipment and storage medium
CN117196765A
Input recommendation method, device and equipment and readable storage medium
CN117784948A
Electricity utilization safety assessment method and system, storage medium and equipment
CN118071543A
Cited By
All-weather RAG intelligent agent automatic newspaper design method and device and electronic equipment
CN119476247A
Method, device and electronic equipment for all-weather RAG intelligent agent automatic newspaper design
CN119476247B
Development component recommendation method and device, equipment and medium
CN119557517A
Digital resource online editing method and system
CN119917745A
An online editing method and system for digital resources
CN119917745B
