A large model-based user group knowledge transfer recommendation method and system

Abstract

This invention discloses a user group knowledge transfer recommendation method and system based on a large-scale model. The method includes: classifying users with rich interaction information in the training set as seed users, collecting their positive and negative interaction data, and extracting their structured semantic knowledge using a large language model; vectorizing and encoding the semantic knowledge and using a clustering algorithm to divide user groups, then constructing group prototype knowledge based on the group semantic knowledge; determining the group affiliation of remaining non-seed users based on the existing groups; transferring group knowledge to users through an adaptive feature fusion mechanism to generate enhanced user representations; and predicting user preferences and generating a recommendation list based on the enhanced user representations. This invention extracts structured semantic knowledge and constructs group prototypes through a large language model, and achieves effective transfer of group knowledge to individual users through an adaptive feature fusion mechanism. It solves the problems of insufficient group knowledge extraction and rigid knowledge transfer strategies in existing recommendation technologies, effectively improving the recommendation accuracy and diversity for new users and users with sparse interactions.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence recommendation system technology, and in particular to a group feature knowledge transfer method that utilizes the deep semantic understanding capabilities of large language models to extract implicit knowledge from the behavior of seed users and transfer it to new users. Background Technology

[0002] Recommender systems aim to solve the problem of information overload and provide personalized services to users. Traditional recommendation techniques, such as collaborative filtering, matrix factorization, and deep learning models, mainly rely on statistical modeling based on users' historical interaction data. However, these methods have inherent limitations:

[0003] Insufficient deep semantic understanding and user interest mining: Traditional models struggle to deeply understand the semantics of item content (such as news headlines and product descriptions), and their characterization of user interests remains at the level of surface interaction patterns, failing to capture underlying implicit preferences, reading styles, or avoidance tendencies. This results in poor recommendation performance when faced with new or semantically similar items, meaning that data sparsity and cold start problems remain severe.

[0004] The methods for extracting and transferring common knowledge within user groups are limited: Existing methods rely heavily on simple statistical aggregations (such as average click-through rates and popular items) when mining user group characteristics, lacking a systematic approach to uncovering deeper, structured common knowledge within the group (such as core interests and unique reading styles). Therefore, when providing recommendations to new users or users with sparse interactions, the models lack effective "group experience" as support, and knowledge transfer strategies often employ fixed weights, failing to dynamically adjust based on the richness of individual user behaviors.

[0005] In recent years, Large Language Models (LLMs) have brought new opportunities to recommender systems due to their powerful semantic understanding and knowledge generation capabilities. However, current mainstream methods for integrating LLMs into recommendations, whether using them as static text encoders to enhance input features or attempting to use them as recommenders to directly generate results, have significant shortcomings: the former, while stable, only utilizes the understanding capabilities of LLMs and fails to fully leverage their knowledge generation and reasoning potential to actively construct reusable, structured prototypes of collective knowledge; the latter, due to the inherent contradiction between the large number of LLM parameters and the sparse nature of recommendation data, faces challenges such as high risk of overfitting and prohibitive computational costs, making it difficult to apply stably in real-world scenarios. Summary of the Invention

[0006] The core technical problem this invention aims to solve is to provide a technical solution that deeply integrates the semantic understanding and knowledge generation capabilities of a large language model, performs structured modeling of users' deep semantic needs, and achieves adaptive knowledge transfer from group commonalities to individual characteristics, addressing the aforementioned problems of existing technologies. This solves the problems of insufficient user interest mining due to inadequate semantic understanding in existing recommendation technologies, as well as the limitations of methods for extracting group common knowledge and the rigidity of transfer strategies.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0008] A recommendation method based on knowledge transfer from a large user group using a large model includes the following steps:

[0009] S101: Identify seed users whose historical interaction counts reach a preset threshold from the training set, extract their semantic knowledge and construct user semantic vectors;

[0010] S102, preprocess the semantic vectors of seed users, and perform semantic group division through dimensionality reduction and clustering algorithms to obtain the user group set and group center vector;

[0011] S103, Construct a structured group prototype knowledge representation based on the semantic knowledge of users within each group, for subsequent group knowledge transfer;

[0012] S104. For non-seed users, generate user vectors based on their historical interaction data, calculate the similarity with the group center vector, and assign users whose similarity meets the threshold condition to the corresponding group.

[0013] S105 prepares group prototype knowledge for users assigned to the group, and realizes knowledge transfer from group knowledge to individual recommendations through adaptive feature fusion during the recommendation model training process, generates enhanced user representations and inputs them into the recommendation model for recommendation prediction.

[0014] Optionally, identifying seed users in step S101 includes: setting a threshold for the number of historical interactions. From the training set (U represents the training set of users, I represents the set of items, and R represents user-item interaction information) Users whose historical interaction count reaches a threshold are selected as the seed user set. The expression is:

[0015] ,

[0016] in For the training set of users, For users Historical interaction sequence, Indicates user Number of historical interactions.

[0017] Optionally, the step S101 of extracting semantic knowledge and constructing a user semantic vector includes:

[0018] (1) For each seed user Structured prompts are built based on their historical interaction data. It calls a large language model to extract structured user semantic knowledge. The expression is:

[0019] =LLM( ),

[0020] LLM This indicates the large model used. Specifically includes , For users Core preference knowledge, Knowledge based on user interests Knowledge based on user style preferences Knowledge of user avoidance patterns;

[0021] (2) Use a pre-trained text encoding model to encode user semantic knowledge into positive semantic vectors. and negative semantic vector The expression is:

[0022] , ,

[0023] in This indicates text encoding processing. This indicates a text concatenation operation.

[0024] (3) In order to obtain a single vector that can comprehensively represent the user's positive and negative preferences, the normalized positive semantic vector and the negative semantic vector are fused by difference to obtain the user semantic vector. The expression is:

[0025] ,

[0026] in This is a negative weighting coefficient used to control the intensity of suppression of negative preferences. The difference mechanism of "positive minus negative" simultaneously incorporates user preference information (positive) and avoidance information (negative), making the direction in the vector space more accurately point to the user's net preference region, which will be used for subsequent group segmentation.

[0027] Optionally, step S102 includes:

[0028] (1) Calculate the global center point of all seed users. and the residual vector corresponding to each user The expression is:

[0029] , ,

[0030] in The residual vector represents the number of seed users. It is used to eliminate the influence of global preferences and highlight the semantic differences between individual users. Dimensionality reduction techniques (such as Singular Value Decomposition (SVD) and Principal Component Analysis (PCA)) are used to reduce the dimensionality of the residual vector and extract the main components of semantic knowledge.

[0031] (2) Use clustering algorithm to perform semantic grouping on the dimensionality-reduced vectors to obtain user group sets. K represents the number of clusters, which can be preset based on the characteristics of the dataset or determined using clustering effectiveness metrics such as the elbow rule. Then, the center vector of each cluster is calculated. The expression is:

[0032] .

[0033] Where i is the group index, and its value ranges from 1 to K.

[0034] Optionally, the construction of a structured group prototype knowledge representation in step S103 includes: for each user group obtained by the clustering algorithm Integrate the semantic knowledge of all users within the group to construct group prototype knowledge. The expression is: The specific construction methods for each component are as follows:

[0035] Group preference prototype The method is constructed using a distance-center-based preference sampling approach. First, the preference sampling method for each user within the group is calculated. semantic vectors With the center The Euclidean distance is expressed as:

[0036] ,

[0037] in L2 norm represents the L2 norm of two vectors, which is the magnitude of the difference between them, i.e., the Euclidean distance. The vector with the closest distance is selected. User core preference knowledge As a prototype of group preferences .

[0038] Group Theme Prototype It is constructed using a discriminative feature selection method. First, based on the user's corresponding... The atomic topic tag set of all users in the statistical group For each topic Calculate its probability of occurrence within the population. and global occurrence probability The expression is:

[0039] , ,in The total number of users in the group, of which This represents the total number of seed users. Representing a group theme Number of occurrences Themes for all groups Number of times it appears.

[0040] theme The discriminant score is defined as:

[0041] ,

[0042] in A smoothing factor is used to prevent division by zero. Then, the scores with higher values ​​are selected. Each theme forms a group theme prototype. .

[0043] Group style archetype It is constructed using a discriminative feature selection method. First, based on the user's corresponding... A collection of texts analyzing the style preferences of all users within a statistical group. For each style preference Calculate its probability of occurrence within the population. and global occurrence probability The expression is:

[0044] , ,

[0045] in Representing a group Style Tendency Number of occurrences Style preferences for all groups Frequency of occurrence. The discriminative score for style is defined as:

[0046] ,

[0047] Then select the ones with higher scores. Each style forms a style description for the group.

[0048] It is a group avoidance prototype constructed by using a discriminative feature selection method. First, according to the corresponding Statistically analyze the set of avoidance pattern texts of all users within the group , for each avoidance pattern , calculate its occurrence probability within the group and the global occurrence probability , and the expression is:

[0049] , ,

[0050] where represents the group avoidance pattern the number of occurrences, is the number of occurrences of all group avoidance patterns . The discriminative score of the avoidance pattern is defined as:

[0051] ,

[0052] Then screen out the avoidance patterns with higher scores to form a negative preference description of the group.

[0053] Finally, the group prototype knowledge simultaneously includes discriminative features in text form (preferred samples, themes, styles, avoidances), which not only retains interpretable semantic knowledge but also provides a numerical representation for subsequent matching.

[0054] Optionally, step S104 includes:

[0055] (1) For non-seed users , obtain their list of historical interaction items . If the number of user historical items , then use the same text encoding model as in step S101 to generate a user vector , and the expression is:

[0056] ,

[0057] where is the minimum historical quantity threshold, represents the user the j-th item interacted with; at the same time, if the number of user historical items m < M, no group assignment is performed, and subsequent recommendations are only based on individual characteristics to improve the accuracy of knowledge transfer.

[0058] (2) Calculate the user vector With the central vector of each group The cosine similarity is expressed as:

[0059] ,

[0060] in Representing vectors respectively and The L2 norm (i.e., the Euclidean modulus). It is the product of the magnitudes of two vectors.

[0061] (3) Choice and The group with the highest similarity is selected as the candidate affiliation group for the user; that is, the group that makes the similarity the highest is selected. Reaching the largest group index :

[0062] ,

[0063] in This represents the value of the independent variable that maximizes the objective function. This is the index of the candidate group to which they belong. Let the maximum similarity be:

[0064] ,

[0065] Set similarity threshold ,like Then assign the user to the index as group Otherwise, the user will not be assigned to a group.

[0066] Optionally, step S105 includes a data preparation stage, a knowledge transfer stage, and a recommendation prediction stage:

[0067] (1) The data preparation stage includes: preparing group prototype knowledge text representations for users who have been successfully assigned to a group, adding the topic prototype, style prototype, preference prototype, and avoidance prototype of the user's group as text features to the user data, and converting these group prototype texts into fixed-dimensional numerical feature vectors through a pre-trained text encoder (such as BERT, LLaMA, etc.): , , These are vector representations of group preferences, themes, styles, and avoidance prototypes, respectively.

[0068] (2) The knowledge transfer stage includes:

[0069] First, during the training of the recommendation model, adaptive fusion weights are calculated. The expression is:

[0070] ,

[0071] in For maximum weight, For the number of historical user interactions, The attenuation coefficient is... It is an exponential function with the natural constant e as the base; through weight design, it is realized that as the number of historical interactions increases, Exponential decay gradually transitions to dependence on individual characteristics.

[0072] Secondly, the group prototype difference is calculated to obtain the group prior. The expression is:

[0073] ,

[0074] ,

[0075] ,

[0076] in This is a positive prior representation. This represents the negative prior group. is the negative weighting coefficient, used to control the strength of negative preference suppression, and n represents the number of types represented by the selected positive vector. It is the sigmoid activation function.

[0077] Simultaneously, a neural network gating function is used to dynamically adjust the influence strength of group knowledge based on individual user characteristics. The expression is:

[0078] ,

[0079] in and For learnable parameters, The sigmoid function is the gated function, and its output value ranges from [0,1]. The larger the value, the stronger the influence of the group's knowledge. The user's underlying semantic vector (obtained by the user data encoder) is used for enhancement. Finally, individual features and group knowledge are fused via residual connections to generate an enhanced user representation. To achieve knowledge transfer, the expression is:

[0080] .

[0081] (3) The recommendation prediction stage includes: encoding into item vector representation Candidate project information and enhanced user representation Input recommendation model prediction layer Calculate the matching score of user u for item i. The expression is:

[0082] ,

[0083] The candidate items are sorted according to their scores, and the top-N items with the highest scores are selected to generate a recommendation list. Further, the recommendation results for each user u can be obtained.

[0084] Furthermore, the present invention also provides a recommendation system based on large-scale user group knowledge transfer, comprising an interconnected microprocessor and a memory, wherein the microprocessor is programmed or configured to execute the recommendation method based on large-scale user group knowledge transfer.

[0085] Furthermore, the present invention provides a computer-readable storage medium storing a computer program or instructions that are programmed or configured to execute the recommendation method for user group knowledge transfer based on a large model via a processor.

[0086] Furthermore, the present invention also provides a computer program product, including a computer program or instructions that are programmed or configured to execute the recommendation method for user group knowledge transfer based on a large model via a processor.

[0087] Compared with existing technologies, this invention has the following advantages: By introducing the deep semantic understanding capabilities of a large language model, this invention systematically extracts four-dimensional structured semantic knowledge—preferences, topics, styles, and avoidance patterns—from users' historical behavior to construct interpretable user profiles; based on semantic similarity, users are divided into interest groups with discriminative characteristics, and common knowledge prototypes at the group level are mined; on this basis, an adaptive fusion mechanism is designed to dynamically adjust the contribution weights of group knowledge and individual characteristics according to the richness of users' historical behavior, achieving a smooth transfer from group commonalities to individual characteristics. This method significantly improves the recommendation quality for users with sparse behavior, effectively alleviates the data sparsity problem, and maintains the personalized expression capabilities of users with rich behavior, outperforming traditional methods in terms of recommendation accuracy, interpretability, and scenario adaptability. Attached Figure Description

[0088] Figure 1 This is a flowchart illustrating the overall method of an embodiment of the present invention.

[0089] Figure 2 A graph is constructed for seed user identification and semantic vector mapping.

[0090] Figure 3 A diagram for group segmentation and prototype construction.

[0091] Figure 4 Assignment map to non-seed user groups.

[0092] Figure 5 This is a prediction graph for knowledge transfer and recommendation. Detailed Implementation

[0093] To enable those skilled in the art to better understand the technical solution of this invention, the technical solution of this invention will be further described in detail below, taking news recommendation as an example and in conjunction with the accompanying drawings of the embodiments of this invention. It should be noted that the method of this invention is not limited to news recommendation, but can also be used for recommending various items such as goods, videos, audio, and papers.

[0094] The detailed implementation process for each step is as follows:

[0095] like Figure 1 As shown, the user group knowledge transfer recommendation method based on a large model in this embodiment includes the following steps:

[0096] S101: Identify seed users whose historical interaction counts reach a preset threshold from the training set, extract their semantic knowledge and construct user semantic vectors;

[0097] S102, preprocess the semantic vectors of seed users, and perform semantic group division through dimensionality reduction and clustering algorithms to obtain the user group set and group center vector;

[0098] S103, Construct a structured group prototype knowledge representation based on the semantic knowledge of users within each group, for subsequent group knowledge transfer;

[0099] S104. For non-seed users, generate user vectors based on their historical interaction data, calculate the similarity with the group center vector, and assign users whose similarity meets the threshold condition to the corresponding group.

[0100] S105 prepares group prototype knowledge for users assigned to the group, and realizes knowledge transfer from group knowledge to individual recommendations through adaptive feature fusion during the recommendation model training process, generates enhanced user representations and inputs them into the recommendation model for recommendation prediction.

[0102] like Figure 2 As shown, identifying seed users in step S101 includes: setting a threshold for the number of historical interactions. In this embodiment, T=3 is chosen. This threshold is determined based on the statistical characteristics of the dataset, ensuring that seed users have sufficient historical interaction information while also ensuring that the number of seed users meets the requirements for group segmentation. From the training set... (U represents the training set of users, I represents the set of items, and R represents user-item interaction information) Users whose historical interaction count reaches a threshold are selected as the seed user set. The expression is:

[0103] ,

[0104] in For the training set of users, For users Historical interaction sequence, Indicates user Number of historical interactions.

[0105] Step S101, which involves extracting semantic knowledge and constructing a user semantic vector, includes:

[0106] (1) For each seed user Structured prompts are built based on their historical interaction data. It calls a large language model to extract structured user semantic knowledge. In this embodiment, the large language model selected is glm-4.7-flash from Zhipu AI, and the expression is:

[0107] =LLM( ),

[0108] LLM This indicates the large model used. Specifically includes , For users Core preference knowledge, Knowledge based on user interests Knowledge based on user style preferences Knowledge of user avoidance patterns;

[0109] (2) Use a pre-trained text encoding model to encode user semantic knowledge into positive semantic vectors. and negative semantic vector The expression is:

[0110] , ,

[0111] in This indicates text encoding processing. This indicates a text concatenation operation. In this embodiment, the text encoder is the all-mpnet-base-v2 model of Sentence-BERT, because it performs well in semantic similarity tasks and has good generalization ability.

[0112] (3) In order to obtain a single vector that can comprehensively represent the user's positive and negative preferences, the normalized positive semantic vector and the negative semantic vector are fused by difference to obtain the user semantic vector. The expression is:

[0113] ,

[0114] in This is the negative weighting coefficient, used to control the intensity of suppression of negative preferences. In this embodiment, it is taken as... =0.5. This The difference mechanism of "positive minus negative" simultaneously incorporates user preference information (positive) and avoidance information (negative), making the direction in the vector space more accurately point to the user's net preference region, which will be used for subsequent group segmentation.

[0115] like Figure 3 As shown, step S102 includes:

[0116] (1) Calculate the global center point of all seed users. and the residual vector corresponding to each user The expression is:

[0117] , ,

[0118] in The number of seed users is represented by the residual vector, which is used to eliminate the influence of global preferences and highlight the semantic differences between individual users. Dimensionality reduction techniques (such as Singular Value Decomposition (SVD) and Principal Component Analysis (PCA)) are used to reduce the dimensionality of the residual vector to extract the main components of semantic knowledge. In this embodiment, TruncatedSVD is used to reduce the dimensionality of the residual vector. TruncatedSVD is chosen instead of PCA because it is suitable for dimensionality reduction of sparse matrices, has higher computational efficiency, and can handle the extraction of the main components of high-dimensional semantic vectors.

[0119] (2) Use clustering algorithm to perform semantic grouping on the dimensionality-reduced vectors to obtain user group sets. K represents the number of clusters, which can be preset based on the characteristics of the dataset or determined using clustering effectiveness indicators such as the elbow rule. In this embodiment, K is 8, determined based on the cluster silhouette coefficient. Then, the center vector of the group is calculated. The expression is:

[0120] .

[0121] Where i is the group index, and its value ranges from 1 to K.

[0122] The construction of a structured group prototype knowledge representation in step S103 includes: for each user group obtained by the clustering algorithm Integrate the semantic knowledge of all users within the group to construct group prototype knowledge. The expression is: The specific construction methods for each component are as follows:

[0123] Group preference prototype The method is constructed using a distance-center-based preference sampling approach. First, the preference sampling method for each user within the group is calculated. semantic vectors With the center The Euclidean distance is expressed as:

[0124] ,

[0125] in L2 norm represents the L2 norm of two vectors, which is the magnitude of the difference between them, i.e., the Euclidean distance. The vector with the closest distance is selected. User core preference knowledge As a prototype of group preferences This embodiment Take 5.

[0126] Group Theme Prototype It is constructed using a discriminative feature selection method. First, based on the user's corresponding... The atomic topic tag set of all users in the statistical group For each topic Calculate its probability of occurrence within the population. and global occurrence probability The expression is:

[0127] , ,in The total number of users in the group, of which This represents the total number of seed users. Representing a group theme Number of occurrences Themes for all groups Number of times it appears.

[0128] theme The discriminant score is defined as:

[0129] ,

[0130] in A smoothing factor is used to prevent division by zero. Then, the scores with higher values ​​are selected. Each theme forms a group theme prototype. This embodiment Take 5.

[0131] Group style archetype It is constructed using a discriminative feature selection method. First, based on the user's corresponding... A collection of texts analyzing the style preferences of all users within a statistical group. For each style preference Calculate its probability of occurrence within the population. and global occurrence probability The expression is:

[0132] , ,

[0133] in Representing a group Style Tendency Number of occurrences Style preferences for all groups Frequency of occurrence. The discriminative score for style is defined as:

[0134] ,

[0135] Then select the ones with higher scores. Each style forms a style description for the group, as illustrated in this embodiment. Take 3.

[0136] To construct a prototype for group avoidance, a discriminative feature selection method is used. First, based on the user's corresponding... A collection of texts illustrating avoidance patterns of all users within a statistical group. For each avoidance pattern Calculate its probability of occurrence within the population. and global occurrence probability The expression is:

[0137] , ,

[0138] in Representing a group Avoidance mode Number of occurrences Avoidance mode for all groups Number of occurrences. The discriminative score for avoidance pattern is defined as:

[0139] ,

[0140] Then select the ones with higher scores. An avoidance mode forms a negative preference description of the group. In this embodiment Take 3.

[0141] Finally, the group prototype knowledge Simultaneously containing discriminative features in text form (preference samples, themes, styles, avoidances) not only retains interpretable semantic knowledge but also provides a numerical representation for subsequent matching.

[0142] As Figure 4 shown, step S104 includes:

[0143] (1) For non-seed users , obtain their list of historical interaction items . If the number of the user's historical items , then use the same text encoding model as in step S101 to generate a user vector based on the historical item information , and the expression is:

[0144] ,

[0145] where is the minimum historical quantity threshold. In this embodiment, take M = 2. Taking M = 2 is considered for the extremely sparse scenario. Even if there are only 2 historical interactions, a meaningful user vector can be obtained through item encoding. represents the jth item that the user has interacted with; at the same time, if the number of the user's historical items m < M, no group assignment is performed, and subsequent recommendations are only based on individual characteristics to improve the accuracy of knowledge transfer.

[0146] (2) Calculate the cosine similarity between this user vector and each group center vector , and the expression is:

[0147] <...>,

[0148] where respectively represent the L2 norms (i.e., Euclidean norms) of the vectors and , is the product of the norms of the two vectors.

[0149] (3) Select the group with the highest similarity to as the candidate belonging group of this user, that is, find the group index that makes the similarity reach the maximum:

[0150] ,

[0151] in This represents the value of the independent variable that maximizes the objective function. This is the index of the candidate group to which they belong. Let the maximum similarity be:

[0152] ,

[0153] Set similarity threshold ,like In this embodiment, we take =0.3, then assign the user to the index as group Otherwise, the user will not be assigned to a group.

[0154] like Figure 5 Step S105 includes the data preparation stage, the knowledge transfer stage, and the recommendation prediction stage:

[0155] (1) The data preparation stage includes: preparing group prototype knowledge text representations for users who have been successfully assigned to a group, adding the topic prototype, style prototype, preference prototype, and avoidance prototype of the user's group as text features to the user data, and converting these group prototype texts into fixed-dimensional numerical feature vectors through a pre-trained text encoder (such as BERT, LLaMA, etc.): , , These are vector representations of group preferences, themes, styles, and avoidance prototypes, respectively.

[0156] (2) The knowledge transfer stage includes:

[0157] First, during the training of the recommendation model, adaptive fusion weights are calculated. The expression is:

[0158] ,

[0159] in To achieve the maximum weight, this embodiment uses 0.4. To ensure that the maximum contribution of group knowledge does not exceed 50% of individual characteristics, based on the number of users' historical interactions, over-reliance on group knowledge should be avoided. The attenuation coefficient is set to 8 in this embodiment. It is an exponential function with the natural constant e as the base; through weight design, it is realized that as the number of historical interactions increases, Exponential decay gradually transitions to dependence on individual characteristics.

[0160] Secondly, the group prototype difference is calculated to obtain the group prior. The expression is:

[0161] ,

[0162] ,

[0163] ,

[0164] in This is a positive prior representation. This represents the negative prior group. This is the negative weighting coefficient, set to 1 in this embodiment, used to control the suppression strength of negative preferences. The positive vector selected in this embodiment represents the group preference and the group topic, with n being 2. It is the sigmoid activation function.

[0165] Simultaneously, a neural network gating function is used to dynamically adjust the influence strength of group knowledge based on individual user characteristics. The expression is:

[0166] ,

[0167] in and For learnable parameters, The sigmoid function is the gated function, and its output value ranges from [0,1]. The larger the value, the stronger the influence of the group's knowledge. The user's underlying semantic vector (obtained by the user data encoder) is used for enhancement. Finally, individual features and group knowledge are fused via residual connections to generate an enhanced user representation. To achieve knowledge transfer, the expression is:

[0168] .

[0169] (3) The recommendation prediction stage includes: encoding into item vector representation Candidate project information and enhanced user representation Input recommendation model prediction layer Calculate the matching score of user u for item i. The expression is:

[0170] ,

[0171] The candidate items are sorted according to their scores, and the top-N items with the highest scores are selected to generate a recommendation list. Further, the recommendation results for each user u can be obtained.

[0172] To validate the user group knowledge transfer recommendation method based on large language model semantic understanding in this embodiment, the MIND (Microsoft News Dataset) news dataset was used in the experiments, and the following widely accepted evaluation metrics in the recommendation field were selected: Group AUC (GAUC), Mean Reciprocal Rank (MRR), and Normalized Decay Cumulative Gain (NDCG@1, NDCG@5, NDCG@10) for the top 1 / 5 / 10 positions. This embodiment uses the large model GLM-4.7-Flash to extract user structured knowledge, and LLaMA-IISAN-NAML is used as the skeleton recommendation model to validate the method. LLaMA-IISAN-NAML is an augmented model that combines the pre-trained language model LLaMA1 with the IISAN module. IISAN (Intra- and Inter-modal Side Adapted Network) is a plug-and-play, highly efficient fine-tuning architecture that performs intra-modal and inter-modal adaptation simultaneously through decoupled side networks, achieving performance comparable to full-parameter fine-tuning while significantly reducing GPU memory usage and training time. Meanwhile, the following recommendation models based on the LLaMA1 semantic encoding backbone were selected as the comparison baseline:

[0173] LLaMA-NAML: The existing NAML model's GloVe word embedding and CNN feature extraction modules are replaced with LLaMA1 semantic encoding output, while the rest of the network structure is the same as NAML.

[0174] LLaMA-LSTUR: Replaces the original word embedding layer of the existing LSTUR model with the LLaMA1 semantic encoding results, while retaining the GRU short-term / long-term interest modeling module;

[0175] LLaMA-MINER: Replaces the news content Transformer feature extraction module of the existing MINER model with the semantic encoding output of the LLaMA1 pre-trained language model, while retaining the original MINER's PolyAttention user behavior encoding and TargetAwareAttention target awareness attention predictor.

[0176] LLaMA-NRMS: Replaces the word embedding layer of the existing NRMS model with LLaMA1 semantic encoding, while keeping the self-attention mechanism module unchanged;

[0177] LLaMA-Fastformer-NRMS: Based on the LLaMA-NRMS architecture, the standard self-attention mechanism on the user side is replaced with Fastformer additive attention (which reduces computational complexity through global query-key aggregation). LLaMA1 is still used as the semantic encoding backbone of news content. The encoding logic on the news side is consistent with LLaMA-NRMS, and the rest of the network structure remains unchanged.

[0178] LLaMA-IISAN-NAML: The skeleton recommendation model used in our method implementation.

[0179] The LLaMA1 series methods described above all use the pre-trained language model LLaMA1 as the semantic encoding backbone, replacing the original word embedding and feature extraction modules. This represents a news recommendation scheme based on a large language model, a concept currently available. The final comparison results are shown in Table 1.

[0180] Table 1. Experimental results comparing the performance of recommendation methods on the MIND dataset.

[0181] Model GAUC MRR NDCG@1 NDCG@5 NDCG@10 LLaMA-NAML 65.12 36.20 51.85 56.06 64.37 LLaMA-LSTUR 65.30 36.60 51.98 56.09 64.86 LLaMA-MINER 66.33 <![CDATA[ 37.07 ]]> <![CDATA[ 52.48 ]]> 56.78 65.39 LLaMA-NRMS 66.48 36.98 52.44 57.01 65.45 LLaMA-Fastformer-NRMS 65.74 36.33 52.05 55.94 64.85 LLaMA-IISAN-NAML 67.63 36.64 51.43 <![CDATA[ 57.39 ]]> <![CDATA[ 65.89 ]]> Ours <![CDATA[ 67.31 ]]> 37.52 54.30 57.74 66.51

[0182] As shown in Table 1, the method presented in this embodiment significantly outperforms the LLaMA1-based recommendation model in the experiment on the MIND news recommendation dataset. Overall, this method achieves the best performance on most evaluation metrics, especially in metrics measuring head ranking quality: NDCG@1 reaches 54.30, an improvement of 2.87 percentage points compared to the strongest baseline LLaMA-IISAN-NAML; MRR reaches 37.52, an improvement of 0.88; and NDCG@10 reaches 66.51, an improvement of 0.62. The nearly 3 percentage point improvement in NDCG@1 is particularly noteworthy, directly demonstrating that this method, through positive and negative semantic separation and adaptive group knowledge transfer, can effectively improve the model's accuracy in capturing users' core interests, achieving the technical effect of "letting users see what they want at first glance."

[0183] The technical advantage of this embodiment stems from its innovative dual-drive architecture of "large language model semantic understanding + group knowledge transfer". First, this embodiment fully utilizes the deep semantic understanding capabilities of GLM-4.7-Flash to perform fine-grained analysis of the historical interaction data of seed users, extracting structured semantic knowledge covering multiple dimensions such as explicit interest preferences, potential interest migration trajectories, reading behavior patterns, and content style tendencies, forming a standardized user profile representation. This differs from existing technologies (such as the LLaMA1 series of methods) that simply use the LLaMA1 model directly for semantic encoding. This division of labor makes the knowledge extraction process more refined and controllable.

[0184] Secondly, this embodiment achieves accurate user group segmentation through semantic similarity calculation and constructs a group prototype knowledge base. Then, through an adaptive feature fusion mechanism, it transfers group knowledge to non-seed users, generating enhanced user representations. This mechanism effectively solves the problem of sparsely interacting users lacking effective representations, achieving a "group-to-individual" knowledge transfer effect. In contrast, while methods like LLaMA-IISAN-NAML introduce the large language model LLaMA1, they lack a systematic group knowledge extraction and transfer strategy, failing to fully leverage the potential of large models in semantic understanding and knowledge generalization.

[0185] This embodiment validates the method using LLaMA-IISAN-NAML as the backbone recommendation model. The results show that even without changing the underlying recommendation architecture, a significant improvement in recommendation performance can be achieved simply by introducing a user group knowledge transfer mechanism based on a large language model, validating the method's good scalability. It can be extended to diverse scenarios such as e-commerce product recommendations, video content recommendations, and music recommendations, providing a new solution for the technological upgrade of personalized recommendation systems.

[0186] Furthermore, this embodiment also provides a user group knowledge transfer recommendation system based on a large model, including a microprocessor and a memory interconnected, wherein the microprocessor is programmed or configured to execute the user group knowledge transfer recommendation method based on the large model.

[0187] Furthermore, this embodiment also provides a computer-readable storage medium storing a computer program or instructions that are programmed or configured to execute the recommendation method for user group knowledge transfer based on a large model via a processor.

[0188] Furthermore, this embodiment also provides a computer program product, including a computer program or instructions that are programmed or configured to execute the recommendation method for user group knowledge transfer based on a large model via a processor.

[0189] Those skilled in the art will understand that the technical solutions provided by this invention may take the form of a method, system, or computer program product. Therefore, this invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this invention may take the form of a computer program product embodied on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, produce an implementation of the flowchart... Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0190] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A recommendation method based on knowledge transfer from a large user group using a large model, characterized in that, Includes the following steps: S101: Identify seed users whose historical interaction counts reach a preset threshold from the training set, extract their semantic knowledge and construct user semantic vectors; S102, preprocess the semantic vectors of seed users, and perform semantic group division through dimensionality reduction and clustering algorithms to obtain the user group set and group center vector; S103, Construct a structured group prototype knowledge representation based on the semantic knowledge of users within each group, for subsequent group knowledge transfer; S104. For non-seed users, generate user vectors based on their historical interaction data, calculate the similarity with the group center vector, and assign users whose similarity meets the threshold condition to the corresponding group. S105 prepares group prototype knowledge for users assigned to the group, and realizes knowledge transfer from group knowledge to individual recommendations through adaptive feature fusion during the recommendation model training process, generates enhanced user representations and inputs them into the recommendation model for recommendation prediction.

2. The user group knowledge transfer recommendation method based on a large model according to claim 1, characterized in that, The step S101 of identifying seed users includes: setting a threshold for the number of historical interactions. From the training set (U represents the training set of users, I represents the set of items, and R represents user-item interaction information) Users whose historical interaction count reaches a threshold are selected as the seed user set. The expression is: ,in For the training set of users, For users Historical interaction sequence, Indicates user Number of historical interactions.

3. The user group knowledge transfer recommendation method based on a large model according to claim 1, characterized in that, The extraction of semantic knowledge and construction of user semantic vectors in step S101 includes the following details: (1) For seed users whose number of historical interactions in the training set reaches a preset threshold Structured prompts are built based on their historical interaction data. It calls a large language model to extract structured user semantic knowledge. The expression is: =LLM( ), LLM This indicates the large model used. Specifically includes , For users Core preference knowledge, Knowledge based on user interests Knowledge based on user style preferences Knowledge of user avoidance patterns; (2) Use a pre-trained text encoding model to encode user semantic knowledge into positive semantic vectors. and negative semantic vector The expression is: ， ， in This indicates text encoding processing. This indicates a text concatenation operation; (3) In order to obtain a single vector that can comprehensively represent the user's positive and negative preferences, the normalized positive semantic vector and the negative semantic vector are fused by difference to obtain the user semantic vector. The expression is: ， in This is a negative weighting coefficient used to control the strength of suppression of negative preferences. The "positive minus negative" difference mechanism incorporates both user preference information (positive) and avoidance information (negative), making the direction in the vector space more accurately point to the user's net preference region, which will be used for subsequent group segmentation.

4. The user group knowledge transfer recommendation method based on a large model according to claim 1, characterized in that, The step S102, which involves obtaining the user group set and the group center vector, includes: (1) Calculate the global center point of all seed users. and the residual vector corresponding to each user The expression is: ， ， in The residual vector represents the number of seed users. It is used to eliminate the influence of global preferences and highlight the semantic differences between individual users. Dimensionality reduction techniques (such as Singular Value Decomposition (SVD) and Principal Component Analysis (PCA)) are used to reduce the dimensionality of the residual vector and extract the main components of semantic knowledge. (2) Use clustering algorithm to perform semantic grouping on the dimensionality-reduced vectors to obtain user group sets. K represents the number of clusters, which can be preset based on the characteristics of the dataset or determined using clustering effectiveness metrics such as the elbow rule. Then, the center vector of each cluster is calculated. The expression is: ， Where i is the group index, and its value ranges from 1 to K.

5. The user group knowledge transfer recommendation method based on a large model according to claim 1, characterized in that, The construction of a structured group prototype knowledge representation in step S103 includes: for each user group obtained by the clustering algorithm Integrate the semantic knowledge of all users within the group to construct group prototype knowledge. The expression is: The specific construction methods for each component are as follows: Group preference prototype The method is constructed using a distance-center-based preference sampling approach. First, the preference sampling method for each user within the group is calculated. semantic vectors With the center The Euclidean distance is expressed as: ， in The L2 norm of a vector is the magnitude of the difference between two vectors, i.e., the Euclidean distance. The vector with the closest distance is selected. User core preference knowledge As a prototype of group preferences ; Group Theme Prototype It is constructed using a discriminative feature selection method. First, based on the user's corresponding... The atomic topic tag set of all users in the statistical group For each topic Calculate its probability of occurrence within the population. and global occurrence probability The expression is: , ,in The total number of users in the group, of which This represents the total number of seed users. Representing a group theme Number of occurrences Themes for all groups Number of occurrences; theme The discriminant score is defined as: , in A smoothing factor is used to prevent division by zero. Then, the scores with higher values ​​are selected. Each theme forms a group theme prototype. ; Group style archetype It is constructed using a discriminative feature selection method, first based on the user's corresponding... A collection of texts analyzing the style preferences of all users within a statistical group. For each style preference Calculate its probability of occurrence within the population. and global occurrence probability The expression is: ， ， in Representing a group Style Tendency Number of occurrences Style preferences for all groups Frequency of occurrence. The discriminative score for style is defined as: ， Then select the ones with higher scores. Individual styles form a style description for the group; To construct a prototype for group avoidance, a discriminative feature selection method is used; firstly, based on the user's corresponding... A collection of texts illustrating avoidance patterns of all users within a statistical group. For each avoidance pattern Calculate its probability of occurrence within the population. and global occurrence probability The expression is: ， ， in Representing a group Avoidance mode Number of occurrences Avoidance mode for all groups Number of occurrences. The discriminative score for avoidance pattern is defined as: ， Then select the ones with higher scores. An avoidance pattern forms a description of the group's negative preferences; Ultimately, group prototype knowledge The inclusion of discriminative features in text form (preference samples, topics, styles, avoidance) preserves interpretable semantic knowledge and provides numerical representations that can be used for subsequent matching.

6. The user group knowledge transfer recommendation method based on a large model according to claim 1, characterized in that, Step S104 includes: (1) For non-seed users Get its historical interaction list If the number of historical projects of a user Then the same text encoding model as in step S101 is used. Generate user vectors based on historical project information The expression is: ， Among them is the minimum historical quantity threshold represents the j-th item with which the user has interacted; meanwhile, if the number of the user's historical items m < M, no group allocation is performed, and subsequent recommendations are only based on individual characteristics to improve the accuracy of knowledge transfer; (2) Calculate the user vector With the central vector of each group The cosine similarity is expressed as: ， in Representing vectors respectively and The L2 norm (i.e., the Euclidean modulus). It is the product of the magnitudes of two vectors; (3) Choice and The group with the highest similarity is selected as the candidate affiliation group for the user; that is, the group that makes the similarity the highest is selected. Reaching the largest group index : , in This represents the value of the independent variable that maximizes the objective function. This is the index of the candidate group to which they belong. Let the maximum similarity be: , Set similarity threshold ,like Then assign the user to the index as group Otherwise, the user will not be assigned to a group.

7. The user group knowledge transfer recommendation method based on a large model according to claim 1, characterized in that, The knowledge transfer stage described in step S105 includes: First, during the training of the recommendation model, adaptive fusion weights are calculated. The expression is: ， in For maximum weight, For the number of historical user interactions, The attenuation coefficient is... It is an exponential function with the natural constant e as the base; through weight design, it is realized that as the number of historical interactions increases, Exponential decay gradually transitions to dependence on individual characteristics; Secondly, the group prototype difference is calculated to obtain the group prior. The expression is: ， ， ， in This is a positive prior representation. This represents the negative prior group. This is a negative weighting coefficient used to control the strength of suppression of negative preferences. , , These are vector representations of group preferences, themes, styles, and avoidance prototypes, respectively, where n represents the number of types of positive vector representations chosen. It is the sigmoid activation function; Simultaneously, a neural network gating function is used to dynamically adjust the influence strength of group knowledge based on individual user characteristics. The expression is: ， in and For learnable parameters, The sigmoid function is the gated function, and its output value ranges from [0,1]. The larger the value, the stronger the influence of the group's knowledge. The user's basic semantic vector to be enhanced (obtained by the user data encoder) is used; finally, individual features and group knowledge are fused through residual connections to generate an enhanced user representation. To achieve knowledge transfer, the expression is: 。 8. A user group knowledge transfer recommendation system based on a large model, comprising interconnected microprocessors and memory, characterized in that, The microprocessor is programmed or configured to perform the user group knowledge transfer recommendation method based on any one of claims 1 to 7.

9. A computer-readable storage medium storing a computer program or instructions, characterized in that, The computer program or instructions are programmed or configured to execute, via a processor, the user group knowledge transfer recommendation method based on any one of claims 1 to 7.

10. A computer program product, comprising a computer program or instructions, characterized in that, The computer program or instructions are programmed or configured to execute, via a processor, the user group knowledge transfer recommendation method based on any one of claims 1 to 7.

Images