Personalized recommendation-oriented big language model social data label attribute generation method

Abstract

The invention relates to the technical field of social data processing, in particular to a personalized recommendation-oriented big language model social data tag attribute generation method. Comprising the following steps: constructing a multi-modal social feature incidence matrix; generating an initial label set adapted to the social scene; and establishing a tag hierarchical relationship through an improved semantic dependency tree, calculating an aging attenuation coefficient based on a behavior occurrence time interval, and counting a proportion of users with the same tag in a core social circle as a group association value. According to the method, the complexity of the multi-modal data in the social scene can be adapted by constructing the multi-modal social feature incidence matrix, so that the generated tag system is more fit with the multivariate features of the social data. According to the method, the tag hierarchical relationship is automatically constructed through the improved semantic dependency tree, the aging attenuation coefficient is calculated, and the deep analysis capability of social semantics is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of social data processing, in particular to a large language model social data label attribute generation method for personalized recommendation. BACKGROUND

[0002] In the scenario of personalized recommendation driven by social networks, the accurate capture of user interests depends on the deep analysis of multi-modal social data including text, behavior, relationship, and multimedia. Traditional label generation methods are mostly limited to single business scenarios such as transaction data management or simple rule aggregation, and are difficult to cope with the dynamic nature, complexity, and scenario-based needs of social data. How to build a dynamic label system that integrates multi-modal features, has multi-dimensional attribute levels, time effectiveness, and group association has become a core challenge to improve recommendation accuracy.

[0003] For example, Chinese patent CN202311400143.7, a business transaction data management method, system and medium, proposes a business data classification management based on label attributes to improve data security through statistical analysis and rule review. However, its application scenario is limited to closed business transaction data and does not involve the time sequence features of user interaction behavior in social scenarios such as interaction frequency decay and the hierarchical structure of relationship networks such as core social circles and peripheral relationships. Moreover, the label attributes only include business dimensions such as transaction type and amount, which cannot adapt to the dynamic attributes required by social recommendations such as interest timeliness and group association. For another example, Chinese patent CN202210633230.6, a multi-source heterogeneous data label generation method and generation system, supports label generation for multi-source data including manual and automatic, and the core is limited to business knowledge accumulation, such as generating static labels through SQL scripts or logical configurations. In the social scenario, its defects are obvious: multi-source data is limited to structured business data and does not integrate social-specific unstructured data such as user comment semantics, picture and video content; label generation relies on pre-set rules and lacks deep understanding of social semantics by large language models such as the hierarchical association between hiking and outdoor sports which needs to be defined manually; it does not consider the time decay of labels and is difficult to capture changes in interest.

[0004] Although the above technical solutions have corresponding design advantages, the above technical solutions still have the following technical defects: first, it cannot adapt to social multi-modal data: the technical solution of Chinese patent CN202311400143.7 is only for closed business transaction data, and does not integrate multi-modal information such as text semantics, user relationship network and multimedia content specific to social scenes, so the label system does not match the complexity of social data; second, it lacks dynamic analysis capability of social semantics: the label generation of Chinese patent CN202210633230.6 relies on manual rules or static logic configuration, and does not use large language models to deeply analyze social semantics, so it cannot automatically build hierarchical associations of labels, nor set a time decay mechanism for labels, making it difficult to capture dynamic changes in user interests; third, the design of social group association attributes is missing: the label attributes of Chinese patents CN202311400143.7 and CN202210633230.6 are limited to a single classification dimension, and do not consider the group association characteristics such as the proportion of users with the same label in the core social circle, so it cannot support differentiated needs of content recommendation and friend recommendation. In view of this, we propose a large language model social data label attribute generation method for personalized recommendation. SUMMARY

[0005] The purpose of the present application is to provide a large language model social data label attribute generation method for personalized recommendation to solve the problems raised in the background art.

[0006] To solve the above technical problems, the present application provides a large language model social data label attribute generation method for personalized recommendation, comprising the following steps: S100, constructing a multi-modal social feature correlation matrix: extracting features from user text data (such as user published content, comments), behavior data (such as likes, forwards, interaction frequency), relationship data (such as friend links, group chat participation), and multimedia data (such as image semantic labels, video key frame text descriptions), mapping them to the same feature space through feature normalization processing, calculating the co-occurrence correlation between different modal data, and generating a weighted fusion feature set; S200, generating an initial label set adapted to the social scene: dividing the core social circle (such as the set of users whose interaction frequency and reply rate reach the set threshold in the last 90 days) based on the interaction intensity of user social relationships, extracting common features (such as interaction frequency, common topic proportion) of the core social circle to construct a dynamic prompt template, inputting the fusion feature set into the large language model fine-tuned on social corpus according to the template format, and generating an initial label set containing appearance probability description; S300, multi-dimensional quantification of label attributes: build label hierarchical relationship through improved semantic dependency tree (such as "outdoor hiking" belonging to "sports"), calculate time decay coefficient based on time interval of behavior occurrence (recent behavior coefficient is higher than long-term behavior), and calculate the proportion of users with the same label in the core social circle as the group association value, forming a label attribute set containing hierarchical attributes, time decay attributes, and group association attributes; S400, dynamically adjusting label and attribute weights: set a real-time monitoring window, trigger initial weight assignment of new labels when a user generates new social behavior (such as publishing new content, high-frequency interaction, joining a new group chat and speaking), regularly adjust the weight according to subsequent interaction data (such as other user interaction, secondary forwarding), and decay the weight of inactive labels (such as no associated behavior for more than a set period) according to the set period; S500, output personalized recommendation adaptation result: according to the preset label attribute weight distribution according to the recommendation scenario (such as content recommendation / friend recommendation) (time decay attribute weight proportion is higher in content recommendation, and group association attribute weight proportion is higher in friend recommendation), calculate the matching score of the recommended object and the label, and output the result and the core contribution label according to the score.

[0007] As a further improvement of the technical solution, in the S100, the feature extraction is performed on the text data, behavior data, relationship data and multimedia data of the user, and the feature is normalized and mapped to the same feature space, including the following steps: S110.1, text data feature extraction: perform word segmentation processing on user published content and comments, extract semantic vectors using a pre-trained language model, and generate text feature vectors through pooling operation; Further, the word segmentation processing uses a general Chinese word segmentation tool (such as jieba), and performs word segmentation on the posts and comments published by the user, and filters stop words (such as filtering meaningless words such as "de"); Meanwhile, the pre-trained language model in this embodiment selects a basic model suitable for social text (such as a social field fine-tuning version of BERT), inputs the segmented text sequence, and outputs the semantic vector of each token; In addition, the pooling operation in this embodiment uses mean pooling to take the average of token-level semantic vectors to generate a unified dimension text feature vector (such as 768 dimensions) as the basis for subsequent fusion.

[0008] S110.2, behavior data feature extraction: quantize likes, forwards, and comments into numerical values, count the frequency of each behavior according to the time window, and construct a behavior feature vector; S110.3, relationship data feature extraction: build a social relationship graph with users as nodes and interaction frequency as edge weight, and generate node embedding vectors using a graph embedding algorithm as relationship feature vectors; S110.4 Multimedia Data Feature Extraction: Extract key features from images and videos to generate multimedia feature vectors; S110.5 Feature Normalization Processing: A normalization algorithm is used to process the text feature vector, behavior feature vector, relation feature vector, and multimedia feature vector respectively, mapping each modality feature to the same feature space.

[0009] As a further improvement to this technical solution, in step S100, the co-occurrence correlation degree between different modal data is calculated to generate a weighted fusion feature set, including the following steps: S120.1, Relevance Calculation: Calculate text feature vectors With behavioral feature vector Cosine similarity as a text-behavior correlation : ; in, Represents the vector dot product. The L2 norm of a vector; Calculate the relation feature vector With multimedia feature vectors Mutual information entropy, as a relation-multimedia correlation degree : ; in, Represents the joint probability distribution; and These are the marginal probability distributions; It is a relational feature vector The set of possible values; It is a multimedia feature vector The set of possible values; S120.2, Weight Matrix Generation: Based on the correlation degree of each modality data, construct a 4×4 modality weight matrix. : ; in, The degree of correlation between text and relational features is expressed, and the calculation method is the same as... ; The correlation between text and multimedia features is represented, and the calculation method is the same as... ; This indicates the degree of correlation between behavior and relational features, and the calculation method is the same as... ; The correlation between behavior and multimedia features is represented, and the calculation method is the same as... ; These are the symmetric values ​​corresponding to the degree of correlation; Furthermore, the correlation between text and relational features. Correlation between text and multimedia features Correlation between behavioral and relational characteristics Correlation between behavior and multimedia features Cosine similarity was used for all calculations, along with text-behavior correlation. The calculation logic is consistent. For example, the correlation between text and multimedia features. For example, the calculation formula is as follows: ; in, Represents the vector dot product. This represents the L2 norm of a vector.

[0010] Furthermore, because the associations between modal features are symmetrical (e.g., the association between text and relation is equivalent to the association between relation and text), therefore: (Relationship-text correlation degree equals text-relationship correlation degree); (Relationship-behavior correlation degree equals behavior-relationship correlation degree); (Multimedia - Relationship Degree equals Relationship - Multimedia Degree); (Multimedia-behavior correlation degree equals behavior-multimedia correlation degree); The aforementioned symmetric correlation degree does not require additional calculation; the corresponding correlation degree values ​​can be directly reused.

[0011] S120.3, Feature Fusion: Normalized text feature vectors behavioral feature vector relational feature vectors Multimedia feature vectors By weight matrix Perform weighted summation to generate a fused feature vector. : ; in, These are the weighting coefficients, and ; Furthermore, modal weighting coefficients Calculated in the following way: First, calculate the sum of the correlation degrees of each mode: ; ; ; ; Then, normalization is performed: ; ; ; ; S120.4, Feature Verification: Calculate the fusion feature vector Reconstruction error of each original modal eigenvector : ; in, This represents the squared Euclidean distance between the fused feature vector and the original modality feature vector; This represents the error weighting coefficient, with an initial value of 0.25. like If the threshold is exceeded, the weight matrix will be adjusted. And recalculate the fused feature vector until... The requirements are met.

[0012] As a further improvement to this technical solution, the process of dividing the core social circle based on the interaction intensity of users' social relationships in step S200 includes the following steps: S210.1 Construct a user social network graph, where nodes represent users, edges represent the interaction relationships between users, and the edge weight is the interaction frequency within a preset time window; S210.2 Calculate the interaction intensity index for each user based on the social network graph, and the interaction intensity index is obtained by integrating interaction frequency, interaction timeliness and interaction diversity; S210.3 Rank all users’ interaction intensity in descending order and select the top N% of users to form the initial core social circle; S210.4 Perform connectivity optimization on the initial core social circle, remove isolated users with sparse interactions with other members, and form the final core social circle.

[0013] As a further improvement to this technical solution, in step S200, common features of the core social circle are extracted to construct a dynamic prompt template. The fused feature set is then input into a large language model fine-tuned from the social corpus according to the template format, including the following steps: S220.1 Extraction of common features of the core social circle: Topic distribution feature extraction: Frequency analysis is performed on the content posted and interactive texts of core social circle members. Keywords with a frequency exceeding a preset threshold are selected. Co-occurrence analysis is used to generate a topic network containing topic words and association strengths. This topic network is then mapped to a topic distribution feature vector. Among them, the keywords correspond to The correlation strength is used as the weight value for the corresponding dimension. Interaction pattern feature extraction: Based on the behavioral data of core social circle members, the frequency of comments, reposts, and likes is statistically analyzed, and the relationship between behavioral sequences is calculated. The conversion probability between different behaviors is then calculated to generate an interaction pattern feature vector. ; Time feature extraction: The active time of core social circle members is segmented and statistically analyzed to identify high-frequency active periods and intensive behavior cycles, forming a time feature distribution table; S220.2, Dynamic prompt template construction: Design a basic prompting framework, including a task instruction segment (such as explicitly requiring the generation of user interest tags and their probability of occurrence), a feature input segment (reserving embedding positions for common features), and an output format segment (specifying the number of tags and the probability representation method). The extracted topic distribution features, interaction pattern features, and time features are converted into natural language descriptions, and then embedded into the feature input segments of the basic prompt framework according to the importance of the features to form an initial prompt template. The template length is dynamically adjusted based on the size of the core social circle. When the number of members exceeds the preset value, a feature sampling mechanism is used to retain key information. S220.3, Textualization of Feature Sets: Semantic parsing is performed on the weighted fusion feature set generated by S100 to extract the core identifiers of each modality feature (such as keywords in text features and high-frequency behavior types in behavioral features). The core feature identifiers are converted into structured descriptive text by a pre-trained text generation model. The descriptive text includes feature categories, weight percentages, and descriptions of relationships. The description text is length-adapted to ensure that it matches the capacity of the feature input segment of the dynamic prompt template; S220.4, Prompt Text Integration and Model Input: The textualized fusion feature set is inserted into the designated position of the dynamic prompt template according to the preset format, and the coherence of the feature description is enhanced by semantic connectors; The integrated prompt text undergoes a compliance check, and redundant information exceeding the input length limit of the large language model is removed, while core feature descriptions are retained. Input the optimized prompt text into a large language model that has been fine-tuned with social corpus, and set the generation parameters; S220.5 Initial Tag Set Generation and Processing: Receive candidate labels and their corresponding probabilities from the output of the large language model, and generate a preliminary label set by filtering through a threshold (retaining labels with probabilities higher than a set value); The initial tag set is deduplicated, tags with the same or similar semantics are merged (based on a pre-built thesaurus), and the occurrence probability of the merged tags is recalculated (by taking the mean or weighted value). The output contains the final initial tag set, which includes the tag names and their corresponding probabilities of occurrence.

[0014] As a further improvement to this technical solution, the step S300, which involves constructing a tag hierarchy relationship using an improved semantic dependency tree, includes the following steps: S310.1, Tag Semantic Quantization: After deduplicating the initial label set, a pre-trained language model is used to process each label. Convert to 3D semantic vector Calculate any two tags and semantic similarity : ; in, For vector dot product, These are the L2 norms of the vectors. ; S310.2, Basic Hierarchical Clustering: Hierarchical agglomerative clustering is used to construct the initial hierarchy, and each label is initialized as an independent node; in each iteration, the two nodes with the highest similarity are merged, and the merging condition is as follows: in To preset a similarity threshold, and Generate parent node The semantic vector of its parent node is a weighted average of the vectors of its child nodes. ; in, , The weight of the child node is the probability of the corresponding label appearing in the initial label set. parent node semantic vector; S310.3, Hierarchical Logic Verification: Generate parent nodes using a large language model The descriptive text is used to calculate the semantic inclusion degree between parent and child nodes. : ; in, For multilayer perceptrons, This is a vector concatenation operation. ; like If so, remove the hierarchical association and re-cluster; where, This is the inclusion threshold; S310.4, Dynamic Hierarchical Update: Set update cycle When adding a new tag The maximum similarity between the semantic vector and the existing hierarchical nodes When this occurs, hierarchical restructuring is triggered, and steps S310.2-S310.3 are repeated to update the semantic dependency tree; among which, To update the threshold, and ; S310.5, Definition of Tag Hierarchy Relationship Symbols: The semantic dependency tree finally generated by S310.1-S310.4, and its parent-child node hierarchy and semantic containment degree are included. Unified definition of tag hierarchy relationship .

[0015] As a further improvement to this technical solution, in step S300, a time decay coefficient is calculated based on the time interval between behavior occurrences, and the percentage of users with the same tag in the core social circle is used as the group association value. This includes the following steps: S320.1 Calculation of aging attenuation coefficient: Based on the user behavior data extracted in S100, the behavior time series corresponding to the target tag is determined. ,in For the first The time of the next action, among which ; set the current time Interval with each action time Divide the time intervals to form a set of mutually exclusive time intervals. ,in: Recent Range (Including day 0 and day 7, corresponding to behavior within the last 7 days); medium-term range (Excluding the 7th day but including the 30th day, corresponding to actions taken between the 7th and 30th days prior); Long-term range (Excluding the 30th day, corresponding to actions taken 30 days prior); Statistical analysis of the frequency of actions performed by the target label within each interval. And calculate the aging attenuation coefficient. : ; in, The preset base attenuation coefficient is determined based on the statistical law of attenuation of user behavior on social platforms; S320.2, Calculation of Group Association Value: Based on the core social circles identified in S200, the number of users carrying the target tags was counted. With total number of users Calculate the initial group association value : ; when At that time, similar social circle data was introduced for correction: among which, The preset threshold is used to determine whether the number of users in the core social circle is sufficient to support the calculation of the group association value; The similar social circles are determined by calculating the topic distribution feature vectors of the target social circle and other social circles. and interaction pattern feature vector The cosine similarity on the x-axis is determined by the formula: ; in, For the target social circle, For candidate social circles, The L2 norm of a vector; Social circle similarity; Select the first with the highest similarity Each social circle is considered as a similar social circle; Calculate the average tag association value of similar social circles And the initial correlation values ​​are corrected using a weighted formula: ; in, These are the weighting coefficients, and In the formula The average number of users in similar social circles; This represents the final group association value. S320.3, Attribute Set Integration: The tag hierarchy constructed in S310 The aging decay coefficient calculated by S320.1 Population association value calculated by S320.2 Integrate into a set of triplet tag attributes This is used for subsequent training of personalized recommendation models.

[0016] As a further improvement to this technical solution, in S400, a real-time monitoring window is set up. When a user generates a new social behavior, the initial weight assignment of the newly added tag is triggered, including the following steps: S410.1, Binding and Construction of Monitoring Window: Based on the time interval of aging decay in S300, a corresponding monitoring window is constructed: The short-term monitoring window covers 1-7 days and is linked to the recent range of the S300. The mid-term monitoring window covers 8-30 days and is linked to the mid-term interval of the S300. The window starts at the current system time and triggers the initial weight calculation when it ends; S410.2 Define the triggering conditions for new social behaviors: Based on the behavior data type of S100, the following three types of triggering behaviors are defined: The semantic similarity between the topic tags carried in user-posted content and existing tags in the S300 tag hierarchy is greater than or equal to the preset similarity threshold in S300. ; The cumulative number of comments, reposts, and likes per unit of time is greater than or equal to the average total interaction frequency of core social circle members in S200 within the same unit of time. After a user joins a new social circle, the number of times they post within 72 hours is greater than or equal to the average posting frequency of that social circle, and their common tags are associated with it; S410.3 Establish the rules for assigning initial weights: Assign values ​​to preceding features based on behavior type: Content publishing trigger: Weight associated with the group association value of S300 ; High-frequency interaction trigger: The conversion probability of interaction behavior associated with S200; Social circle joining trigger: Weighted correlation of the similarity between the new social circle and the core social circle.

[0017] As a further improvement to this technical solution, in step S400, the weights are periodically adjusted based on subsequent interaction data, and the weights of inactive tags decay according to a set period, including the following steps: S420.1, Binding Weight Adjustment Cycle: The weight adjustment period is linked to the time interval of time decay in S300: the short-term adjustment period is 7 days, matching the recent interval. The medium-term adjustment period is 30 days, matching the medium-term range. Adjust the time to synchronize with the end time of the S410 monitoring window; S420.2, Layered processing of subsequent interactive data: The impact of interactions on weights is differentiated based on the type of behavioral data in S100: the impact of comment behavior is correlated with the semantic content of tags in S300. The impact of forwarding behavior on the conversion probability of interaction patterns in S200; the impact of liking behavior on the final group correlation value in S300. ; S420.3 Calculate the current tag weight. The current tag weight is calculated by adding historical weight decay to the new interaction gain: historical weight is calculated using the time decay coefficient of S300. attenuation( The smaller the value, the more significant the attenuation); the new interactive gain is calculated based on the cumulative effect of the hierarchical influence in S420.2 (the higher the influence, the greater the gain). S420.4, Implement inactive tag decay mechanism: Inactive tag determination: Tags that have no valid interaction data as described in S420.2 within two consecutive adjustment periods; Attenuation method: Adopting the same interval weighted logic as the S300 time-related attenuation (short-term cycle attenuation is faster than medium-term cycle); Elimination rule: When the weight decays to below the probability threshold of the initial label set in S200, it is marked as a label to be eliminated and will no longer participate in the recommendation calculation of S500.

[0018] As a further improvement to this technical solution, in step S500, the output of personalized recommendation and adaptation results includes the following steps: S510.1, Contextualized Tag Attribute Weight Preset: Based on the recommended scenario type, the tag attribute set generated by S300 is pre-set. Preset weight allocation rules: Content recommendation scenario: Time-delay coefficient Weighting coefficients Final group association value Weighting coefficients Tag hierarchy Weighting coefficients satisfy ; Friend recommendation scenario: weighting coefficients satisfy ; in, The weight values ​​can be fine-tuned based on the dynamically adjusted tag weights of S400, and the fine-tuning logic is positively correlated with the tag activity level. S510.2 Matching Score Calculation: The matching score between the recommended object and the tag. The formula is: ; in, The time-related decay coefficient for the object to be recommended is reused from S320.1. The computational logic; For the group association value corresponding to the object to be recommended, the calculation logic of G in S320.2 is reused; The hierarchical matching degree between the object to be recommended and the tag is based on the semantic inclusion degree in S310. Sure( ); S510.3, Result Sorting and Core Label Extraction: By match score The objects to be recommended are sorted in descending order to form a recommendation sequence; The core contribution tag is the tag corresponding to the attribute with the highest weight coefficient in the score composition (such as in content recommendation scenarios). At its highest point, take and (Associated tags), and mark the attribute parameter values ​​corresponding to the tags.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a multimodal social feature association matrix, integrates user text data, behavioral data, relationship data, and multimedia data, and maps them to the same feature space. This can adapt to the complexity of multimodal data in social scenarios, making the generated tag system more consistent with the diverse features of social data. 2. This invention generates initial tags based on a large language model fine-tuned from social corpus, and automatically constructs tag hierarchy relationships and calculates time-related decay coefficients through an improved semantic dependency tree. This reduces reliance on manual rules, enhances the ability to deeply analyze social semantics, and helps capture dynamic changes in user interests.

[0020] 3. This invention introduces group association value as one of the tag attributes and pre-allocates tag attribute weights according to the recommendation scenario, which makes up for the limitations of a single classification dimension and can support the differentiated needs of different scenarios such as content recommendation and friend recommendation, thereby enhancing the adaptability of the recommendation.

[0021] 4. This invention sets up a real-time monitoring window to dynamically adjust the weights of tags and attributes, so that the tag weights can be updated in real time with the user's new social behaviors and interaction data, avoiding the lag of static weights and making the tags more reflective of the user's current interest status. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the method steps of the present invention. Detailed Implementation

[0023] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention. Embodiment 1

[0024] As Figure 1 shown, this embodiment provides a method for generating social data tag attributes of a large language model for personalized recommendation, including the following steps: S100. Construct a multi-modal social feature association matrix: Extract features from the user's text data (such as user-posted content, comments), behavioral data (such as likes, forwards, interaction frequencies), relationship data (such as friend links, group chat participation), and multimedia data (such as picture semantic tags, video key frame text descriptions), map them to the same feature space through feature normalization processing, calculate the co-occurrence association degree between different modal data, and generate a weighted fusion feature set; In this step, in S100, extracting features from the user's text data, behavioral data, relationship data, and multimedia data and mapping them to the same feature space through feature normalization processing includes the following steps: S110.1. Text data feature extraction: Perform word segmentation on the user-posted content and comments, use a pre-trained language model to extract semantic vectors, and generate text feature vectors through pooling operations; As a further explanation of this step, in this embodiment, the word segmentation processing uses a general Chinese word segmentation tool (such as jieba) to segment the user-posted posts and comments, and filter out stop words (such as filtering meaningless words like "de"); At the same time, the pre-trained language model selected in this embodiment is a basic model suitable for social texts (such as a fine-tuned version of BERT in the social field), input the segmented text sequence, and output the semantic vectors of each token; In addition, the pooling operation in this embodiment uses average pooling, takes the average value of the token-level semantic vectors, and generates a text feature vector with a unified dimension (such as 768 dimensions) as the basis for subsequent fusion.

[0025] S110.2. Behavioral data feature extraction: Quantify the like, forward, and comment behaviors into numerical values, count the frequencies of each behavior according to a time window, and construct a behavioral feature vector; As a further explanation of this step, the behavior quantization rules in this embodiment are as follows: A like is recorded as 1, a forward is recorded as 2, and a comment is recorded as 3 (if the comment length exceeds 20 characters, an additional 1 is added); The time window is set to 30 days, and the time is divided into 5 sub-windows by week. The cumulative value of behavior in each sub-window is counted (e.g., 3 likes and 1 repost in the first week, the corresponding sub-window value is 3×1+1×2=5). The dimension of the behavioral feature vector is the same as the number of time sub-windows (e.g., 5 dimensions), and the vector value is the cumulative behavioral quantification result of each sub-window.

[0026] S110.3, Relationship Data Feature Extraction: Construct a social relationship graph with users as nodes and interaction frequency as edge weights, and use a graph embedding algorithm to generate node embedding vectors as relationship feature vectors; As a further explanation of this step, the social relationship graph in this embodiment is constructed as follows: the target user is the central node, direct friends are the first-level nodes, and users who interact ≥ 5 times in group chats are the second-level nodes; the edge weight is the total frequency of interaction in the past 90 days (e.g., if there are 20 interactions between friends, the weight is 20). Meanwhile, the graph embedding algorithm uses Node2Vec, sets the walk length to 10 and the window size to 5, and generates a 128-dimensional node embedding vector as a relation feature vector. The vector value reflects the user's position and association strength in the social network.

[0027] S110.4 Multimedia Data Feature Extraction: Extract key features from images and videos to generate multimedia feature vectors; As a further explanation of this step, the image feature extraction and video feature extraction in this embodiment specifically include: Image feature extraction: A pre-trained image feature extraction model (such as ResNet50) is used to encode the features of user-posted images, outputting 2048-dimensional original features, which are then reduced to 256-dimensional through principal component analysis; Video feature extraction: A key frame is extracted every 10 seconds, processed using the same image feature extraction method, and the average of the features from multiple frames is taken to generate a 256-dimensional video feature vector, which is consistent with the dimension of the image feature vector.

[0028] S110.5 Feature Normalization Processing: A normalization algorithm is used to process the text feature vector, behavior feature vector, relation feature vector, and multimedia feature vector respectively, mapping each modality feature to the same feature space.

[0029] As a further explanation of this step, this embodiment uses the Min-Max normalization algorithm to process the text, behavior, relation, and multimedia feature vectors separately, mapping each vector value to the [0,1] interval; the mapping formula is: ; in, For the original vector values, , These are the minimum and maximum values ​​of the modal feature, respectively, to ensure that different modal features are comparable within the same numerical range.

[0030] In this step, S100, the co-occurrence correlation between different modalities is calculated to generate a weighted fusion feature set, including the following steps: S120.1, Relevance Calculation: Calculate text feature vectors With behavioral feature vector Cosine similarity as a text-behavior correlation : ; in, Represents the vector dot product. The L2 norm of a vector; Calculate the relation feature vector With multimedia feature vectors Mutual information entropy, as a relation-multimedia correlation degree : ; in, Represents the joint probability distribution; and These are the marginal probability distributions; It is a relational feature vector The set of possible values; It is a multimedia feature vector The set of possible values; For example, taking a user as an example, their text feature vector (768-dimensional) Describes content related to "outdoor hiking", behavioral feature vector (5-dimensional) Display frequently liked posts about "hiking gear" in the past 30 days, and calculate the cosine similarity between the two. This indicates a high degree of correlation between text and behavior; Relationship feature vector of the user's "hiking group" (128-dimensional) reflects the close interaction within the group, and the feature vector of the images it publishes. (256-dimensional) Mostly hiking scenarios, calculating mutual information entropy. This indicates a moderate correlation between relationship networks and multimedia content.

[0031] S120.2, Weight Matrix Generation: Based on the correlation degree of each modality data, construct a 4×4 modality weight matrix. : ; in, The degree of correlation between text and relational features is expressed, and the calculation method is the same as... ; The correlation between text and multimedia features is represented, and the calculation method is the same as... ; This indicates the degree of correlation between behavior and relational features, and the calculation method is the same as... ; The correlation between behavior and multimedia features is represented, and the calculation method is the same as... ; These are the symmetric values ​​corresponding to the degree of correlation; S120.3, Feature Fusion: Normalized text feature vectors behavioral feature vector relational feature vectors Multimedia feature vectors By weight matrix Perform weighted summation to generate a fused feature vector. : ; in, These are the weighting coefficients, and ; Furthermore, modal weighting coefficients Calculated in the following way: First, calculate the sum of the correlation degrees of each mode: ; ; ; ; Then, normalization is performed: ; ; ; ; S120.4, Feature Verification: Calculate the fusion feature vector Reconstruction error of each original modal eigenvector : ; in, This represents the squared Euclidean distance between the fused feature vector and the original modality feature vector; This represents the error weighting coefficient, with an initial value of 0.25. like If the threshold is exceeded, the weight matrix will be adjusted. And recalculate the fused feature vector until... The requirements are met.

[0032] For example, taking a user named "Zhang San" on a social platform as an example, the complete implementation process for constructing a multimodal social feature association matrix is ​​as follows: First, we collected Zhang San's text data (posts and comments on "weekend hikes" published in the past 30 days), behavioral data (liked "hiking route" posts 12 times and forwarded them 5 times), relationship data (joined an "outdoor enthusiasts" group and interacted with 10 people in the group more than 30 times), and multimedia data (posted 3 hiking photos). Subsequently, the modal features were extracted and normalized according to steps S110.1-S110.5 to obtain... , , , ; Then, calculate the modal correlation degree according to S120.1 to generate the weight matrix. ; Next, the weight coefficients are calculated and features are fused according to S120.3 to obtain... ; Finally, verify the reconstruction error according to S120.4 and confirm. Once the threshold is met, a weighted fusion feature set is output for subsequent label attribute generation.

[0033] It should be added that, in order to calculate relational features With multimedia features To obtain the mutual information entropy, the continuous feature vectors need to be discretized first. This embodiment uses the equal-width binning method as the default solution: 128-dimensional relation features of Node2Vec output 64-dimensional multimedia features after ResNet50+PCA dimensionality reduction Calculate the global minimum (min) and maximum (max) for each dimension respectively; Division A series of equal-width intervals (interval width) ), mapping continuous feature values ​​to corresponding intervals (e.g. Corresponding interval 3); after statistical discretization and The co-occurrence frequency, as a joint probability The basis for the calculation.

[0034] If the feature distribution exhibits long-tail characteristics (such as interaction frequency features), the equal-frequency binning method can be switched to (ensuring that the sample size in each interval is equivalent). It can be dynamically adjusted according to the feature dimension (e.g., 5≤ ≤10), to adapt to the feature distribution patterns of different social scenarios.

[0035] S200. Generate an initial tag set adapted to social scenarios: Divide the core social circle based on the interaction intensity of users' social relationships (such as a set of users whose interaction frequency and response rate reach a set threshold in the past 90 days), extract the common features of the core social circle (such as interaction frequency and the proportion of common topics), construct a dynamic prompt template, input the fused feature set into the large language model that has been fine-tuned by social corpus according to the template format, and generate an initial tag set containing occurrence probability descriptions. In this step, the division of the core social circle based on the interaction intensity of users' social relationships in S200 includes the following steps: S210.1 Construct a user social network graph, where nodes represent users, edges represent the interaction relationships between users, and the edge weight is the interaction frequency within a preset time window; As a further explanation of this step, the construction of the user social network graph in this embodiment includes the following steps: First, define network nodes and edges: with the target user as the central node, other users as associated nodes, and edges representing the interaction relationships between users; Subsequently, the scope of interactive behavior was clarified: comments, reposts, and private messages were considered valid interactions (simple likes were not counted), and each valid interaction was counted as 1. Then, set the time window: the time window is the past 90 days (which can be adjusted according to the frequency of platform interaction), denoted as... ; Next, calculate the edge weights: edge weights (user and The formula for calculating the interaction weight is: ,in For the first Does the interaction exist? (1 if it exists, 0 otherwise) Finally, a network graph is generated: users are labeled with nodes, and edges are calculated. value.

[0036] S210.2 Calculate the interaction intensity index for each user based on the social network graph, and the interaction intensity index is obtained by integrating interaction frequency, interaction timeliness and interaction diversity; As a further explanation of this step, the calculation of the interaction intensity index in this embodiment includes the following steps: First, determine the weighting coefficients: set ( The weighting coefficients for interaction frequency, timeliness, and diversity are respectively, and the sum of the three is 1, based on the importance allocation of social interaction features. Subsequently, the normalized value of the interaction frequency is calculated: the maximum edge weight between the target user and all associated users is taken. ,calculate ; Then, calculate the timeliness factor: The number of days since the last interaction. (Aging attenuation coefficient), calculated using the following formula: ; Next, we will calculate the diversity of interactions: (Total number of valid interaction types, i.e., comments, reposts, and private messages) For users The number of interaction types, calculated ; Finally, integrate the metrics: calculate the intensity of interaction. : .

[0037] S210.3 Rank all users’ interaction intensity in descending order and select the top N% of users to form the initial core social circle; As a further explanation of this step, the selection of the initial core social circle in this embodiment includes the following steps: First, ranking interaction intensity: an interaction intensity metric for all associated users. Sort in descending order; Subsequently, the selection ratio N is set: the total number of associated users. When N=20, When N=15 (N is the selection ratio, which is dynamically adjusted according to the density of the social network); Then, determine the selection quantity: calculate the initial number of members based on the top N% of the total number of associated users (e.g., if there are 80 associated users and N=20, the selection quantity is...). people); Finally, the initial circle is selected: the top N% of users after sorting form the initial core social circle.

[0038] S210.4 Perform connectivity optimization on the initial core social circle, remove isolated users with sparse interactions with other members, and form the final core social circle.

[0039] As a further explanation of this step, the connectivity optimization of the initial core social circle in this embodiment includes the following steps: First, calculate the average interaction frequency: the average interaction frequency among members of the initial core social circle. ;in, This represents the average frequency of interaction among members of the initial circle. This is the sum of the edge weights between all members of the initial circle; is the total number of edges between the initial circle members (i.e., the number of pairs of members with actual interaction relationships); Subsequently, set the isolation threshold: The isolation threshold is (based on the empirical value of network connectivity, used to determine isolated users); Then, identify isolated users: Check the initial circle members one by one. If the interaction frequency of a user with all other members is < the isolation threshold, then the user is determined to be an isolated user; Finally, generate the final circle: Remove all isolated users, and the remaining members form the final core social circle.

[0040] In this step, in S200, extract the common features of the core social circle to construct a dynamic prompt template, and input the fusion feature set into the large language model fine-tuned with social corpus according to the template format, including the following steps: S220.1. Extraction of common features of the core social circle: Extraction of topic distribution features: Perform word frequency statistics on the published content and interaction texts of the core social circle members, screen out the keywords with occurrence frequencies higher than the preset threshold, generate a topic network including topic words and association strengths through co-occurrence analysis, and map the topic network into a topic distribution feature vector , where the topic words correspond to dimensions, and the association strength is used as the weight value for the corresponding dimensions; Extraction of interaction pattern features: Based on the behavior data of the core social circle members, count the occurrence frequencies and sequence relationships of comment, forward, and like behaviors, calculate the conversion probabilities between different behaviors, and generate an interaction pattern feature vector ; Extraction of time features: Segment and count the active time of the core social circle members, identify the high-frequency active periods and behavior-intensive cycles, and form a time feature distribution table; As a further explanation of this step, the extraction of topic distribution features in this embodiment includes the following steps: First, preprocess the text: Remove duplicates, filter out expressions and links from the published content and interaction texts of the core circle members, use a word segmentation tool (such as jieba) to segment words and remove stop words (such as "de", "le"); Subsequently, screen out topic words: Count the word frequencies, and retain the words with occurrence frequencies > 5% (preset ratio) of the total text as topic words; Then, calculate the co-occurrence intensity: Count the co-occurrence times of topic words ( is the number of times that topic word and topic word appear simultaneously in the same text), take ( is the maximum value of the co-occurrence intensities of topic word a with all other topic words); Next, generate the feature vector: topic distribution feature vector. The dimension corresponds to the number of keywords, and the dimension value is... ,in: Keywords The sum of co-occurrence intensities with all other keywords; Finally, output As a feature of topic distribution.

[0041] Furthermore, the interaction pattern feature extraction in this embodiment includes the following steps: First, count the frequency of behavior: record the total number of comments and reposts by core members, and calculate the behavior percentage (e.g., comment percentage = number of comments / total number of behaviors). Next, the conversion count was calculated: the conversion count of consecutive actions within 7 days (such as the number of times a comment was made and then a share was made). Then, calculate the transition probability: ; in: For behavior Transform into behavior The probability of; Behavior occurring within 7 days Recurrence of behavior Total number of times; Indicates behavior The total number of occurrences during the statistical period; Next, we construct the feature vector: the interaction mode feature vector. The dimensions are "behavior proportion + conversion probability" (e.g., when there are two behaviors, comment and forward, the dimensions are 2 behavior proportions + 2 conversion probabilities = 4 dimensions). Finally, output As a feature of the interaction mode.

[0042] S220.2, Dynamic prompt template construction: Design a basic prompting framework, including a task instruction segment (such as explicitly requiring the generation of user interest tags and their probability of occurrence), a feature input segment (reserving embedding positions for common features), and an output format segment (specifying the number of tags and the probability representation method). The extracted topic distribution features, interaction pattern features, and time features are converted into natural language descriptions, and then embedded into the feature input segments of the basic prompt framework according to the importance of the features to form an initial prompt template. The template length is dynamically adjusted based on the size of the core social circle. When the number of members exceeds the preset value, a feature sampling mechanism is used to retain key information. As a further explanation of this step, the construction of the dynamic prompt template in this embodiment includes the following specific components: First, design the basic framework: including the task instruction section (clearly defining the generated interest tags and probability range, format requirements), the feature input section (reserving placeholders for "[topic features], [interaction mode], [time features]"), and the output format section (providing examples); Subsequently, the features are converted into natural language: for example, the topic features are described as "high-frequency discussion of XX in the core circle, co-occurrence intensity XX", the interaction pattern is described as "mainly XX behavior (proportion XX), XX conversion probability XX", and the time feature is described as "high-frequency active period is XX". Then, embedding by importance: natural language features are embedded into corresponding placeholders in the order of "topic features > interaction patterns > time features" (based on their direct impact on tag generation); Next, adjust the template length: If the core group has more than 50 members, select the top 50% of keywords based on word frequency, retain the behavior types that account for more than 10%, and ensure the total template length is within acceptable limits. 500 words; Finally, generate a dynamic prompt template.

[0043] S220.3, Textualization of Feature Sets: Semantic parsing is performed on the weighted fusion feature set generated by S100 to extract the core identifiers of each modality feature (such as keywords in text features and high-frequency behavior types in behavioral features). The pre-trained text generation model converts the core feature identifiers into structured descriptive text, which includes feature categories, weight percentages, and relational descriptions. Understandably, this embodiment uses a pre-trained text generation model (such as T5-small) to convert the text into descriptive text, as shown in the following example: "User text features are mainly 'marathon' (30%), and the most frequent behavior is liking fitness content (25%), with a high correlation between the two; and close interaction with 10 fitness enthusiasts (20%)." The text length is adapted to within 150 characters to ensure matching with the template.

[0044] The description text is length-adapted to ensure that it matches the capacity of the feature input segment of the dynamic prompt template; S220.4, Prompt Text Integration and Model Input: The textualized fusion feature set is inserted into the designated position of the dynamic prompt template according to the preset format, and the coherence of the feature description is enhanced by semantic connectors; The integrated prompt text undergoes a compliance check, and redundant information exceeding the input length limit of the large language model is removed, while core feature descriptions are retained. The optimized prompt text is input into a large language model that has been fine-tuned with social corpus, and the generation parameters are set. The social corpus is selected from publicly available interactive texts on social platforms (such as comments and private messages), and is deduplicated, desensitized by user identifiers, and filtered for low-quality content (such as filtering invalid text with too few characters).

[0045] S220.5 Initial Tag Set Generation and Processing: Receive candidate labels and their corresponding probabilities from the output of the large language model, and generate a preliminary label set by filtering through a threshold (retaining labels with probabilities higher than a set value); The initial tag set is deduplicated, tags with the same or similar semantics are merged (based on a pre-built thesaurus), and the occurrence probability of the merged tags is recalculated (by taking the mean or weighted value). The output contains the final initial tag set, which includes the tag names and their corresponding probabilities of occurrence.

[0046] As a further explanation of this step, the initial tag set generation and processing in this embodiment includes the following steps: First, filter the labels: retain the labels with a probability > 0.5 in the model output (the filtering threshold is set based on feature relevance) to form a preliminary label set; Subsequently, deduplication and merging are performed: based on a pre-built thesaurus (including public thesaurus and supplementary synonyms from the social domain), semantically similar tags are merged; Then, recalculate the probabilities: take the mean of the original probabilities of the merged labels (if the probabilities differ). 0.2); Finally, output the final initial label set, in the following format: "Label 1 (probability); Label 2 (probability); ...".

[0047] S300, Achieve multi-dimensional quantification of tag attributes: Construct a hierarchical relationship of tags through an improved semantic dependency tree (e.g., "outdoor hiking" belongs to "sports"), calculate the time decay coefficient based on the time interval of behavior occurrence (the coefficient for recent behavior is higher than that for long-term behavior), and statistically analyze the proportion of users with the same tags in the core social circle as the group association value to form a tag attribute set that includes hierarchical attributes, time attributes, and group association attributes. In this step, the construction of the tag hierarchy relationship in S300 through the improved semantic dependency tree includes the following steps: S310.1, Tag Semantic Quantization: After deduplicating the initial label set, a pre-trained language model is used to process each label. Convert to 3D semantic vector Calculate any two tags and semantic similarity : ; in, For vector dot product, These are the L2 norms of the vectors. ; As a further explanation of this step, the label semantic quantization in this embodiment includes the following steps: First, deduplication is performed on the initial tag set: based on the social domain thesaurus (such as mapping of near-synonyms such as "hiking" and "hiking trip", "fitness" and "exercising"), semantically duplicated tags are merged and unique tags are retained (such as merging "outdoor hiking" and "wilderness hiking" into "outdoor hiking"). Subsequently, a pre-trained language model was selected: the Sentence-BERT model, which is suitable for semantic encoding of short texts (because the model performs stably in phrase-level semantic similarity calculation and has a uniform output dimension), was adopted to convert each label into a semantic vector of fixed dimensions. Then, calculate the similarity of all tag pairs according to the semantic similarity formula, and output a structured similarity matrix (e.g., 5 tags correspond to a 5×5 matrix, and the matrix elements are the similarity values ​​of each pair of tags). Finally, the similarity matrix is ​​stored in CSV format for easy subsequent clustering processing, where the row and column indices are the label names and the cell values ​​are the corresponding similarities.

[0048] S310.2, Basic Hierarchical Clustering: Hierarchical agglomerative clustering is used to construct the initial hierarchy, and each label is initialized as an independent node; in each iteration, the two nodes with the highest similarity are merged, and the merging condition is as follows: ,in To preset a similarity threshold, and Generate parent node The semantic vector of its parent node is a weighted average of the vectors of its child nodes. ; in, , The weight of the child node is the probability of the corresponding label appearing in the initial label set. parent node semantic vector; As a further explanation of this step, the execution of the basic hierarchical clustering in this embodiment includes the following specific steps: First, initialize clustering units: treat each deduplicated label as an independent clustering unit (e.g., “outdoor hiking”, “marathon”, “yoga”, and “fitness” are 4 initial units respectively). Subsequently, a similarity threshold was set. Based on the general rules of semantic association of social tags, The initial value is set to 0.6 (this can be adjusted according to the scenario type, such as a workplace scenario). It can be increased to 0.7, and reduced to 0.5 in entertainment scenarios. Then, according to the clustering rules disclosed in the claims, the iterative merging units are selected each time from the similarity matrix. The tag pair with the largest value is merged into a parent node, and the semantic vector of the parent node is calculated according to the public formula (the weight of the child node is directly taken from the probability of occurrence in the initial tag set, such as "outdoor hiking" with a probability of 0.8 and "marathon" with a probability of 0.7, then the weights are 0.8 and 0.7 respectively). Finally, repeat the merging process until the similarity of the remaining label pairs is less than 1. This forms an initial hierarchical tree centered on the parent-child relationship (e.g., "outdoor hiking" and "marathon" are merged into "endurance sports", and "yoga" and "fitness" are merged into "indoor sports").

[0049] S310.3, Hierarchical Logic Verification: Generate parent nodes using a large language model The descriptive text is used to calculate the semantic inclusion degree between parent and child nodes. : ; in, For multilayer perceptrons, This is a vector concatenation operation. ; like If so, remove the hierarchical association and re-cluster; where, This is the inclusion threshold; Furthermore, in this embodiment, semantic inclusion degree is calculated. The 3MLP model design is as follows: Input layer: The dimension is equal to the dimension of the multimodal fusion feature (e.g., 128 dimensions, directly connected to the fused feature vector, reflecting the compatibility of the method), and directly connected to the fused feature vector; Hidden layer: 64 neurons, using the ReLU activation function (to alleviate the gradient vanishing problem, a well-known activation function in the field); Output layer: Set up 2 neurons, and activate them through Softmax to output the probability distribution of "existence / non-existence of semantic inclusion relationship".

[0050] The model training uses the cross-entropy loss function.

[0051] S310.4, Dynamic Hierarchical Update: Set update cycle When adding a new tag The maximum similarity between the semantic vector and the existing hierarchical nodes When this occurs, hierarchical restructuring is triggered, and steps S310.2-S310.3 are repeated to update the semantic dependency tree; among which, To update the threshold, and ; As a further explanation of this step, the dynamic hierarchical update in this embodiment includes the following specific steps: First, set the update cycle. Based on the frequency of new tags added on social media platforms, Set to 7 days (i.e., check for new tags every 7 days); Then, new tags are received: new tags are generated from the tag generation results triggered by new user behavior (e.g., if the user has recently mentioned "trail running" frequently, then "trail running" is a new tag). Then, calculate the maximum similarity between the new label and all nodes in the existing hierarchy, and set the update threshold. (like =0.6 =0.5, lowering the threshold for merging new tags); Next, if the maximum similarity is ≥ If the new label is added, the clustering and verification process of S310.2-S310.3 will be repeated to update the hierarchy tree; Finally, record the update log, including new tag names, merged nodes, and a comparison of the hierarchical structure before and after the update.

[0052] S310.5, Definition of Tag Hierarchy Relationship Symbols: The semantic dependency tree finally generated by S310.1-S310.4, and its parent-child node hierarchy and semantic containment degree are included. Unified definition of tag hierarchy relationship .

[0053] In this step, within S300, a time decay coefficient is calculated based on the time interval between behavior occurrences, and the percentage of users with the same tag in the core social circle is used as the group association value. This includes the following steps: S320.1 Calculation of aging attenuation coefficient: Based on the user behavior data extracted from S100, determine the behavior time series corresponding to the target label. ,in For the first The time of the next action, among which ; set the current time Interval with each action time Divide the time intervals to form a set of mutually exclusive time intervals. ,in: Recent Range (Including day 0 and day 7, corresponding to behavior within the last 7 days); medium-term range (Excluding the 7th day but including the 30th day, corresponding to actions taken between the 7th and 30th days prior); Long-term range (Excluding the 30th day, corresponding to actions taken 30 days prior); Statistical analysis of the frequency of actions performed by the target label within each interval. And calculate the aging attenuation coefficient. : ; in, The preset base attenuation coefficient is determined based on the statistical law of attenuation of user behavior on social platforms; In this step, This is determined based on statistical patterns of user behavior decay on social platforms, for example: (Recently) (Mid-term) (In the long term), it can be adjusted based on platform user activity, such as high-activity users. It can be increased to 0.9.

[0054] S320.2, Calculation of Group Association Value: Based on the core social circles identified in S200, the number of users carrying the target tags was counted. With total number of users Calculate the initial group association value : ; when At that time, similar social circle data was introduced for correction: among which, The preset threshold is used to determine whether the number of users in the core social circle is sufficient to support the calculation of the group association value; Similar social circles are calculated by comparing the topic distribution feature vectors of the target social circle with those of other social circles. and interaction pattern feature vector The cosine similarity on the x-axis is determined by the formula: ; in, For the target social circle, For candidate social circles, The L2 norm of a vector; Social circle similarity; Select the first with the highest similarity Each social circle is considered as a similar social circle; Calculate the average tag association value of similar social circles And the initial correlation values ​​are corrected using a weighted formula: ; in, These are the weighting coefficients, and In the formula The average number of users in similar social circles; This represents the final group association value. S320.3, Attribute Set Integration: The tag hierarchy constructed in S310 The aging decay coefficient calculated by S320.1 Population association value calculated by S320.2 Integrate into a set of triplet tag attributes This is used for subsequent training of personalized recommendation models.

[0055] S400, Dynamically Adjust Tag and Attribute Weights: Set up a real-time monitoring window. When a user engages in new social behaviors (such as posting new content, high-frequency interaction, joining a new group chat and speaking), the initial weight of the newly added tag is triggered. The weight is adjusted periodically based on subsequent interaction data (such as interaction with other users, secondary forwarding). Inactive tags (such as those with no related behavior for more than a set period) have their weights decayed according to the set period. In this step, a real-time monitoring window is set up in S400. When a user engages in new social behavior, the initial weight assignment of the newly added tag is triggered, including the following steps: S410.1, Binding and Construction of Monitoring Window: Based on the time interval of aging decay in S300, a corresponding monitoring window is constructed: The short-term monitoring window covers 1-7 days and is linked to the recent range of the S300. The mid-term monitoring window covers 8-30 days and is linked to the mid-term interval of the S300. The window starts at the current system time and triggers the initial weight calculation when it ends; As a further explanation of this step, the binding of the monitoring window in this embodiment is directly associated with the time interval of time decay in S300: the short-term monitoring window covers 1-7 days and is bound to the recent interval [0,7], and the medium-term monitoring window covers 8-30 days and is bound to the medium-term interval (7,30]. All windows are based on the current system time. Starting from, respectively at +7 days +30 days will end, at which point the initial weight calculation for newly added tags will be automatically triggered. During the window's operation, user behavior data is stored in real time in the format of "behavior type-occurrence time-associated tag", such as "posted content-20231005-outdoor hiking" and "comment-20231006-marathon", providing basic data support for subsequent weight calculations.

[0056] S410.2 Define the triggering conditions for new social behaviors: Based on the behavior data type of S100, the following three types of triggering behaviors are defined: The semantic similarity between the topic tags carried in user-posted content and existing tags in the S300 tag hierarchy is greater than or equal to the preset similarity threshold in S300. ; The cumulative number of comments, reposts, and likes per unit of time is greater than or equal to the average total interaction frequency of core social circle members in S200 within the same unit of time. After a user joins a new social circle, the number of times they post within 72 hours is greater than or equal to the average posting frequency of that social circle, and their common tags are associated with it; As a further explanation of this step, when defining the triggering conditions for new social behaviors in S410.2, based on the behavior data type in S100, the three types of triggering behaviors are associated with the parameters of the preceding steps: The semantic similarity between the topic tags carried in the user's published content and the existing tags in the S300 tag hierarchy is greater than or equal to the similarity threshold set by S300 (this threshold is the core criterion for S300 tag clustering). The weighted cumulative number of comments, reposts, and likes within a unit of time (calculated as “comment×2+repost×3+like×1”) ≥ the average total frequency of interaction of core social circle members in S200 within the same unit of time (this frequency is the statistical mean of the interaction behavior of core social circle members in S200 within a unit of time). Within 72 hours of joining a new social circle, the number of times a user posts is greater than or equal to the average posting frequency of that circle over the past 30 days (total posts ÷ number of members), and the posts are associated with common tags determined by the circle using the S200 common feature extraction method.

[0057] S410.3 Establish the rules for assigning initial weights: Assign values ​​to preceding features based on behavior type: Content publishing trigger: Weight associated with the group association value of S300 ; High-frequency interaction trigger: The conversion probability of interaction behavior associated with S200; Social circle joining trigger: Weighted correlation of the similarity between the new social circle and the core social circle.

[0058] As a further explanation of this step, in step S410.3, which establishes the initial weight assignment rules, the specific method for associating preceding features based on the triggering behavior type is as follows: When content is published, the initial weight is calculated using the following formula: ,in (The correlation coefficient between content publishing and group association is set based on the characteristics of social scenarios.) This is the final population association value output in S300; When high-frequency interaction is triggered, the initial weight is calculated according to the following formula: ,in (The interaction conversion rate reflects the correlation between interaction behavior and interest.) The conversion probability of the interactive behavior is calculated in S200; When joining a social circle is triggered, the initial weight is calculated using the following formula: ,in (The influence coefficient of social circle similarity on tag relevance, highlighting the correlation strength of community attributes). The similarity between the new social circle and the core social circle (calculated according to the social circle similarity formula in S320.2).

[0059] Furthermore, in this embodiment, the triggering condition for high-frequency interaction is defined as follows: the frequency of user interaction exceeds 1.5 times the average frequency of interaction in the core social circle (based on the dynamic adjustment of the core circle to avoid the limitations of fixed threshold scenarios), and the average frequency of the core social circle is output through step S200.

[0060] In this step, within S400, the weights are periodically adjusted based on subsequent interaction data, and the weights of inactive tags decay according to a set period, including the following steps: S420.1, Binding Weight Adjustment Cycle: The weight adjustment period is linked to the time interval of time decay in S300: the short-term adjustment period is 7 days, matching the recent interval. The medium-term adjustment period is 30 days, matching the medium-term range. Adjust the time to synchronize with the end time of the S410 monitoring window; S420.2, Layered processing of subsequent interactive data: The impact of interactions on weights is differentiated based on the type of behavioral data in S100: the impact of comment behavior is correlated with the semantic content of tags in S300. The impact of forwarding behavior on the conversion probability of interaction patterns in S200; the impact of liking behavior on the final group correlation value in S300. ; S420.3 Calculate the current tag weight. The current tag weight is calculated by adding historical weight decay to the new interaction gain: historical weight is calculated using the time decay coefficient of S300. attenuation( The smaller the value, the more significant the attenuation); the new interactive gain is calculated based on the cumulative effect of the hierarchical influence in S420.2 (the higher the influence, the greater the gain). S420.4, Implement inactive tag decay mechanism: Inactive tag determination: Tags that have no valid interaction data with S420.2 within two consecutive adjustment periods; Attenuation method: Adopting the same interval weighted logic as the S300 time-related attenuation (short-term cycle attenuation is faster than medium-term cycle); Elimination rule: When the weight decays to below the probability threshold of the initial label set in S200, it is marked as a label to be eliminated and will no longer participate in the recommendation calculation of S500.

[0061] As a further explanation of this step, the implementation logic for periodically adjusting weights and decaying inactive tags based on subsequent interaction data in S420 is as follows: When binding the weight adjustment cycle in S420.1, the cycle is bound to the time interval of time decay in S300. The short-term adjustment cycle is 7 days (matching the recent interval), and the medium-term adjustment cycle is 30 days (matching the medium-term interval). The adjustment time is synchronized with the end time of the monitoring window in S410 to ensure that the weight adjustment is consistent with the timeliness of user interests. When processing subsequent interaction data in S420.2, the degree of influence is distinguished according to the behavioral data type of S100: The impact of commenting behavior The calculation formula is: ,in (The reinforcement coefficient of the comment on the semantics of the tag) The semantic inclusion degree of the tag in S300; Impact of forwarding behavior The calculation formula is: ,in (The spread impact coefficient of forwarding) The probability of behavioral conversion in the interaction mode of S200; The impact of liking The calculation formula is: ,in (The weak influence coefficient of likes) This is the final group association value in S300.

[0062] Furthermore, in S420.3, the current label weight is calculated. When, calculate according to the following formula: ; in The label weights from the previous period. This is the aging attenuation coefficient of S300 (the smaller the value, the more significant the attenuation). This is to adjust the total impact of all new interactions within the adjustment period.

[0063] Furthermore, when the inactive tag decay mechanism is implemented in S420.4, an inactive tag refers to a tag that has no effective interaction data in S420.2 within two consecutive adjustment periods. The decay method adopts the same interval weighted logic as the time-based decay in S300 (the decay is faster in the short-term period than in the medium-term period). When the weight decays to below the probability threshold of the initial tag set in S200, it is marked as a tag to be eliminated and no longer participates in the recommendation calculation of S500.

[0064] S500 Output personalized recommendation adaptation results: Based on the recommendation scenario (such as content recommendation / friend recommendation), preset tag attribute weight allocation (timeliness attribute weight is higher in content recommendation, and group association attribute weight is higher in friend recommendation), calculate the matching score between the object to be recommended and the tag, and output the results and core contribution tags in sorted by score.

[0065] In this step, the personalized recommendation adaptation results are output in the S500, including the following steps: S510.1, Contextualized Tag Attribute Weight Preset: Based on the recommended scenario type, the tag attribute set generated by S300 is pre-set. Preset weight allocation rules: Content recommendation scenario: Time-delay coefficient Weighting coefficients Final group association value Weighting coefficients Tag hierarchy Weighting coefficients satisfy ; Friend recommendation scenario: weighting coefficients satisfy ; in, The weight values ​​can be fine-tuned based on the dynamically adjusted tag weights of S400, and the fine-tuning logic is positively correlated with the tag activity level. As a further explanation of this step, Furthermore, the size relationships are distinguished based on scene characteristics: Content recommendation scenario: Priority should be given to matching users' recent interests, therefore the weighting coefficients meet the following requirements. Example value , , (Set based on the strong influence of content timeliness on user attention); Friend recommendation scenario: Priority should be given to matching commonalities within the social circle, therefore the weighting coefficients must meet the following requirements. Example value , , (The core role of social matching is set based on group association).

[0066] Furthermore, the weight fine-tuning logic is positively correlated with tag activity: if a tag's dynamically adjusted weight in S400 is higher than the average of similar tags (e.g., "outdoor hiking" has a weight of 0.7, higher than the average of 0.5), then the weight coefficient of its corresponding attribute can be increased by 5%-10% (e.g., in content recommendations). (Adjust from 0.5 to 0.525), and vice versa, to ensure that the influence of active tags is strengthened.

[0067] S510.2 Matching Score Calculation: The matching score between the recommended object and the tag. The formula is: ; in, The time-related decay coefficient for the object to be recommended is reused from S320.1. The calculation logic is based on the time series of the associated behaviors of the object to be recommended, dividing it into recent (0-7 days), medium (7-30 days), and long (>30 days) intervals, counting the frequency of behaviors in each interval, and then calculating the weighted average of the intervals. For the group association value corresponding to the object to be recommended, the calculation logic of G in S320.2 is reused, that is, the percentage of users who pay attention to the object in the core social circle is counted. If the number of users in the social circle is small, similar social circle data is introduced for correction (the correction method is the same as S320.2). The hierarchical matching degree between the object to be recommended and the tag is based on the semantic inclusion degree in S310. Sure( ); S510.3, Result Sorting and Core Label Extraction: By match score The objects to be recommended are sorted in descending order to form a recommendation sequence; The core contribution tag is the tag corresponding to the attribute with the highest weight coefficient in the score composition (such as in content recommendation scenarios). At its highest point, take and (Associated tags), and mark the attribute parameter values ​​corresponding to the tags.

[0068] As a further explanation of this step, the result sorting and core tag extraction in this embodiment specifically include: Results sorted by match score The objects to be recommended are sorted in descending order to form a recommendation sequence, with objects with higher scores being recommended first. Core contribution tag extraction: Obtain the tag corresponding to the attribute with the highest weight coefficient in the component (e.g., in content recommendation scenarios). At its highest point, take and (related tags), and annotate the attribute parameter values ​​corresponding to the tag (e.g. or (Specific calculation results).

[0069] Those skilled in the art will understand that the process of implementing all or part of the steps of the above embodiments can be carried out by hardware or by a program instructing the relevant hardware.

[0070] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for generating social data tag attributes from a large language model for personalized recommendation, characterized in that: Includes the following steps: S100. Construct a multimodal social feature association matrix: Extract features from users' text data, behavioral data, relational data, and multimedia data, map them to the same feature space through feature normalization, calculate the co-occurrence association degree between different modal data, and generate a weighted fusion feature set. S200. Generate an initial tag set adapted for social scenarios: Divide the core social circle based on the interaction intensity of users' social relationships, extract the common features of the core social circle to construct a dynamic prompt template, input the fused feature set into the large language model fine-tuned by the social corpus according to the template format, and generate an initial tag set containing occurrence probability descriptions. S300, Achieve multi-dimensional quantification of tag attributes: Construct a hierarchical relationship of tags through an improved semantic dependency tree, calculate the time decay coefficient based on the time interval of behavior occurrence, and statistically analyze the proportion of users with the same tags in the core social circle as the group association value to form a tag attribute set containing hierarchical attributes, time-effect attributes, and group association attributes; S400, Dynamically Adjust Tag and Attribute Weights: Set up a real-time monitoring window. When a user engages in new social behavior, the initial weight of the newly added tag is assigned. The weight is adjusted periodically based on subsequent interaction data. The weight of inactive tags decays according to a set period. S500 Output personalized recommendation adaptation results: Based on the preset tag attribute weight allocation for the recommendation scenario, calculate the matching score between the object to be recommended and the tag, and output the results and core contribution tags in sorted order by score.

2. The method for generating social data tag attributes for large language models oriented towards personalized recommendation according to claim 1, characterized in that, In step S100, features are extracted from the user's text data, behavioral data, relational data, and multimedia data, and mapped to the same feature space through feature normalization processing, including the following steps: S110.1 Text Data Feature Extraction: Segment user-posted content and comments into words, extract semantic vectors using a pre-trained language model, and generate text feature vectors through pooling operations; S110.2 Behavioral data feature extraction: Quantify likes, reposts, and comments into numerical values, count the frequency of each behavior according to time windows, and construct a behavioral feature vector; S110.3, Relationship Data Feature Extraction: Construct a social relationship graph with users as nodes and interaction frequency as edge weights, and use a graph embedding algorithm to generate node embedding vectors as relationship feature vectors; S110.4 Multimedia Data Feature Extraction: Extract key features from images and videos to generate multimedia feature vectors; S110.5 Feature Normalization Processing: A normalization algorithm is used to process the text feature vector, behavior feature vector, relation feature vector, and multimedia feature vector respectively, mapping each modality feature to the same feature space.

3. The method for generating social data tag attributes for large language models oriented towards personalized recommendation according to claim 2, characterized in that, In step S100, the co-occurrence correlation degree between different modal data is calculated to generate a weighted fusion feature set, including the following steps: S120.1, Relevance Calculation: Calculate text feature vectors With behavioral feature vector Cosine similarity as a text-behavior correlation : ； in, Represents the vector dot product. The L2 norm of a vector; Calculate relation feature vectors With multimedia feature vectors Mutual information entropy, as a relation-multimedia correlation degree : ； in, Represents the joint probability distribution; and These are the marginal probability distributions; It is a relational feature vector The set of possible values; It is a multimedia feature vector The set of possible values; S120.2, Weight Matrix Generation: Based on the correlation degree of each modality data, construct a 4×4 modality weight matrix. : ； in, Indicates the degree of correlation between text and relational features; Indicates the degree of correlation between text and multimedia features; Indicates the degree of correlation between behavior and relational characteristics; Indicates the degree of correlation between behavior and multimedia features; These are the symmetric values ​​corresponding to the degree of correlation; S120.3, Feature Fusion: Normalized text feature vectors behavioral feature vector relational feature vectors Multimedia feature vectors By weight matrix Perform weighted summation to generate a fused feature vector. : ； in, All are weighting coefficients, and ; S120.4, Feature Verification: Calculate the fusion feature vector Reconstruction error of each original modal eigenvector : ； in, This represents the squared Euclidean distance between the fused feature vector and the original modality feature vector; All represent error weighting coefficients; like If the threshold is exceeded, the weight matrix will be adjusted. And recalculate the fused feature vector until... The requirements are met.

4. The method for generating social data tag attributes for large language models oriented towards personalized recommendation according to claim 3, characterized in that, The process of dividing the core social circle based on the interaction intensity of users' social relationships in S200 includes the following steps: S210.1 Construct a user social network graph, where nodes represent users, edges represent the interaction relationships between users, and the edge weight is the interaction frequency within a preset time window; S210.2 Calculate the interaction intensity index for each user based on the social network graph, and the interaction intensity index is obtained by integrating interaction frequency, interaction timeliness and interaction diversity; S210.3 Rank all users’ interaction intensity in descending order and select the top N% of users to form the initial core social circle; S210.4 Perform connectivity optimization on the initial core social circle, remove isolated users with sparse interactions with other members, and form the final core social circle.

5. The method for generating social data tag attributes for large language models for personalized recommendation according to claim 4, characterized in that, In step S200, common features of the core social circle are extracted to construct a dynamic prompt template. The fused feature set is then input into a large language model fine-tuned from the social corpus according to the template format. This includes the following steps: S220.1 Extraction of common features of the core social circle: Topic distribution feature extraction: Frequency analysis is performed on the content posted and interactive texts of core social circle members. Keywords with a frequency exceeding a preset threshold are selected. Co-occurrence analysis is used to generate a topic network containing topic words and association strengths. This topic network is then mapped to a topic distribution feature vector. ; Interaction pattern feature extraction: Based on the behavioral data of core social circle members, the frequency of comments, reposts, and likes is statistically analyzed, and the relationship between behavioral sequences is calculated. The conversion probability between different behaviors is then calculated to generate an interaction pattern feature vector. ; Time feature extraction: The active time of core social circle members is segmented and statistically analyzed to identify high-frequency active periods and intensive behavior cycles, forming a time feature distribution table; S220.2 Dynamic prompt template construction: Design a basic prompting framework, including a task instruction segment, a feature input segment, and an output format segment; The extracted topic distribution features, interaction pattern features, and time features are converted into natural language descriptions, and then embedded into the feature input segments of the basic prompt framework according to the importance of the features to form an initial prompt template. The template length is dynamically adjusted based on the size of the core social circle. When the number of members exceeds the preset value, a feature sampling mechanism is used to retain key information. S220.3, Textualization of Feature Sets: Semantic parsing is performed on the weighted fusion feature set generated by S100 to extract the core identifiers of each modality feature; The core feature identifiers are converted into structured descriptive text by a pre-trained text generation model. The descriptive text includes feature categories, weight percentages, and descriptions of relationships. The description text is length-adapted to ensure that it matches the capacity of the feature input segment of the dynamic prompt template; S220.4, Prompt Text Integration and Model Input: The textualized fusion feature set is inserted into the designated position of the dynamic prompt template according to the preset format, and the coherence of the feature description is enhanced by semantic connectors; The integrated prompt text undergoes a compliance check, and redundant information exceeding the input length limit of the large language model is removed, while core feature descriptions are retained. Input the optimized prompt text into a large language model that has been fine-tuned with social corpus, and set the generation parameters; S220.5 Initial Tag Set Generation and Processing: Receive candidate labels and their corresponding probabilities from the output of the large language model, and generate a preliminary label set by filtering through a threshold. Perform deduplication on the initial tag set, merge tags with the same or similar semantics, and recalculate the occurrence probability of the merged tags; The output contains the final initial tag set, which includes the tag names and their corresponding probabilities of occurrence.

6. The method for generating social data tag attributes for large language models for personalized recommendation according to claim 5, characterized in that, The steps in S300 for constructing the tag hierarchy relationship using the improved semantic dependency tree include the following: S310.1, Tag Semantic Quantization: After deduplicating the initial label set, a pre-trained language model is used to process each label. Convert to 3D semantic vector Calculate any two tags and semantic similarity : ； in, For vector dot product, Let L2 norm be the vector. ; S310.2, Basic Hierarchical Clustering: Hierarchical agglomerative clustering is used to construct the initial hierarchy, and each label is initialized as an independent node; in each iteration, the two nodes with the highest similarity are merged, and the merging condition is as follows: ,in Generate parent nodes to preset similarity thresholds. The semantic vector of its parent node is a weighted average of the vectors of its child nodes. ； in, , The weight of the child node; parent node semantic vector; S310.3, Hierarchical Logic Verification: Generate parent nodes using a large language model The descriptive text is used to calculate the semantic inclusion degree between parent and child nodes. : ； in, For multilayer perceptrons, This is a vector concatenation operation. ; like If so, remove the hierarchical association and re-cluster; where, This is the inclusion threshold; S310.4, Dynamic Hierarchical Update: Set update cycle When adding a new tag The maximum similarity between the semantic vector and the existing hierarchical nodes When this occurs, hierarchical restructuring is triggered, and steps S310.2-S310.3 are repeated to update the semantic dependency tree; among which, To update the threshold, and ; S310.5, Definition of Tag Hierarchy Relationship Symbols: The semantic dependency tree finally generated by S310.1-S310.4, and its parent-child node hierarchy and semantic containment degree are included. Unified definition of tag hierarchy relationship .

7. The method for generating social data tag attributes for large language models for personalized recommendation according to claim 6, characterized in that, In S300, the time decay coefficient is calculated based on the time interval between behavior occurrences, and the proportion of users with the same tag in the core social circle is used as the group association value. This includes the following steps: S320.1 Calculation of aging attenuation coefficient: Based on the user behavior data extracted in S100, the behavior time series corresponding to the target tag is determined. ,in For the first The time of the next action, among which ; set the current time Interval with each action time Divide the time intervals to form a set of mutually exclusive time intervals. ,in: Recent Range ; medium-term range ; Long-term range ; Statistical analysis of the frequency of actions performed by the target label within each interval. And calculate the aging attenuation coefficient. : ； in, The preset base attenuation coefficient; S320.2, Calculation of Group Association Value: Based on the core social circles identified in S200, the number of users carrying the target tags was counted. With total number of users Calculate the initial group association value : ； when At that time, similar social circle data was introduced for correction: among which, The preset threshold is used to determine whether the number of users in the core social circle is sufficient to support the calculation of the group association value; The similar social circles are determined by calculating the topic distribution feature vectors of the target social circle and other social circles. and interaction pattern feature vector The cosine similarity on the x-axis is determined by the formula: ； in, For the target social circle, For candidate social circles, The L2 norm of a vector; Social circle similarity; Select the first with the highest similarity Each social circle is considered as a similar social circle; Calculate the average tag association value of similar social circles And the initial correlation values ​​are corrected using a weighted formula: ； in, These are the weighting coefficients, and In the formula The average number of users in similar social circles; This represents the final group association value. S320.3, Attribute Set Integration: The tag hierarchy constructed in S310 The aging decay coefficient calculated by S320.1 Population association value calculated by S320.2 Integrate into a triplet tag attribute set This is used for subsequent training of personalized recommendation models.

8. The method for generating social data tag attributes for large language models for personalized recommendation according to claim 7, characterized in that, In S400, a real-time monitoring window is set up. When a user generates a new social behavior, the initial weight assignment of the newly added tag is triggered, including the following steps: S410.1, Binding and Construction of Monitoring Window: Based on the time interval of aging decay in S300, a corresponding monitoring window is constructed: The short-term monitoring window covers 1-7 days and is linked to the recent range of the S300. The mid-term monitoring window covers 8-30 days and is linked to the mid-term interval of the S300. The window starts at the current system time and triggers the initial weight calculation when it ends; S410.2 Define the triggering conditions for new social behaviors: Based on the behavior data type of S100, the following three types of triggering behaviors are defined: The semantic similarity between the topic tags carried in user-posted content and existing tags in the S300 tag hierarchy is greater than or equal to the preset similarity threshold in S300. ; The cumulative number of comments, reposts, and likes per unit of time is greater than or equal to the average total interaction frequency of core social circle members in S200 within the same unit of time. After a user joins a new social circle, the number of times they post within 72 hours is greater than or equal to the average posting frequency of that social circle, and their common tags are associated with it; S410.3 Establish the rules for assigning initial weights: Assign values ​​to preceding features based on behavior type: Content publishing trigger: Weight associated with the group association value of S300 ; High-frequency interaction trigger: The conversion probability of interaction behavior associated with S200; Social circle joining trigger: Weighted correlation of the similarity between the new social circle and the core social circle.

9. The method for generating social data tag attributes for large language models for personalized recommendation according to claim 8, characterized in that, In S400, the weights are adjusted periodically based on subsequent interaction data, and the weights of inactive tags decay according to a set period, including the following steps: S420.1, Binding Weight Adjustment Cycle: The weight adjustment period is linked to the time interval of time decay in S300: the short-term adjustment period is 7 days, matching the recent interval. The medium-term adjustment period is 30 days, matching the medium-term range. Adjust the time to synchronize with the end time of the S410 monitoring window; S420.2, Layered processing of subsequent interactive data: The impact of interactions on weights is differentiated based on the type of behavioral data in S100: the impact of comment behavior is correlated with the semantic content of tags in S300. The impact of forwarding behavior on the conversion probability of interaction patterns in S200; the impact of liking behavior on the final group correlation value in S300. ; S420.3 Calculate the current tag weight. The current tag weight is calculated by adding historical weight decay to the new interaction gain: historical weight is calculated using the time decay coefficient of S300. Attenuation; the new interactive gain is calculated based on the cumulative effect of the hierarchical influence of S420.2; S420.4, Implement inactive tag decay mechanism: Inactive tag determination: Tags that have no valid interaction data as described in S420.2 within two consecutive adjustment periods; Attenuation method: Adopts the same interval weighted logic as the S300 time-related attenuation; Elimination rule: When the weight decays to below the probability threshold of the initial label set in S200, it is marked as a label to be eliminated and will no longer participate in the recommendation calculation of S500.

10. The method for generating social data tag attributes for large language models oriented towards personalized recommendation according to claim 9, characterized in that, In S500, outputting personalized recommendation and adaptation results includes the following steps: S510.1, Contextualized Tag Attribute Weight Preset: Based on the recommended scenario type, the tag attribute set generated by S300 is pre-set. Preset weight allocation rules: Content recommendation scenario: Time-delay coefficient Weighting coefficients Final group association value Weighting coefficients Tag hierarchy Weighting coefficients satisfy ; Friend recommendation scenario: weighting coefficients satisfy ; in, ; S510.2 Matching Score Calculation: The matching score between the recommended object and the tag. The formula is: ； in, The time-decrease coefficient corresponding to the object to be recommended; The group association value corresponding to the object to be recommended; The hierarchical matching degree between the object to be recommended and the tag is based on the semantic inclusion degree in S310. Sure; S510.3, Result Sorting and Core Label Extraction: By match score The objects to be recommended are sorted in descending order to form a recommendation sequence; The core contribution label is the label corresponding to the attribute with the highest weight coefficient in the score composition, and the attribute parameter value corresponding to the label is marked.

Images