Social data label attribute generation method for personalized recommendation-oriented large language model

Abstract

The 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, which comprises the following steps: constructing a multi-modal social feature correlation matrix; generating an initial label set adapted to a social scene; realizing multi-dimensional quantification of label attributes; constructing a label hierarchical relationship through an improved semantic dependency tree, calculating a time decay coefficient based on a behavior occurrence time interval, and taking the user proportion of the same label in a core social circle as a group correlation value. Through the construction of the multi-modal social feature correlation matrix, the complexity of multi-modal data in the social scene can be adapted, and the generated label system is more suitable for the multi-element characteristics of social data. Through the improved semantic dependency tree, the label hierarchical relationship and the time decay coefficient are automatically constructed, and the depth analysis ability 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 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 interests.

[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 the hierarchical association of labels, nor set a time decay mechanism for labels, making it difficult to capture the dynamic changes of 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 the 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:

[0007] 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;

[0008] 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 within 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 a large language model fine-tuned on social corpus according to the template format, and generating an initial label set containing appearance probability description;

[0009] 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 group association value as the proportion of users with the same label in the core social circle, forming a label attribute set containing hierarchical attributes, time decay attributes, and group association attributes;

[0010] S400, dynamically adjusting label and attribute weights: set a real-time monitoring window, when a user generates a new social behavior (such as publishing new content, high-frequency interaction, joining a new group chat and speaking), trigger the initial weight assignment of the new label, regularly adjust the weight according to subsequent interaction data (such as other user interaction, secondary forwarding), and decay the weight of non-active labels (such as no associated behavior for more than a set period) according to the set period;

[0011] S500, output personalized recommendation adaptation result: according to the recommendation scenario (such as content recommendation / friend recommendation), preset the label attribute weight distribution (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.

[0012] 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:

[0013] 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;

[0014] 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");

[0015] Meanwhile, the pre-trained language model in the 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;

[0016] In addition, the pooling operation in the 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.

[0017] S110.2, behavior data feature extraction: quantize the likes, forwards, and comments into numerical values, count the frequency of each behavior according to the time window, and construct a behavior feature vector;

[0018] S110.3, relationship data feature extraction: constructing a social relationship graph with users as nodes and interaction frequency as edge weight, generating node embedding vectors as relationship feature vectors by using graph embedding algorithm;

[0019] S110.4, multimedia data feature extraction: extracting key features from pictures and videos to generate multimedia feature vectors;

[0020] S110.5, feature normalization processing: using normalization algorithm to process text feature vectors, behavior feature vectors, relationship feature vectors and multimedia feature vectors respectively, and mapping each modal feature to the same feature space.

[0021] As a further improvement of the technical solution, in the S100, the co-occurrence correlation degree between different modal data is calculated, and a weighted fusion feature set is generated, including the following steps:

[0022] S120.1, correlation degree calculation:

[0023] The cosine similarity of the text feature vector and the behavior feature vector is calculated as the text-behavior correlation degree :

[0024] ;

[0025] Where, denotes the vector dot product, denotes the L2 norm of the vector;

[0026] The mutual information entropy of the relationship feature vector and the multimedia feature vector is calculated as the relationship-multimedia correlation degree :

[0027] ;

[0028] Where, denotes the joint probability distribution; and are the marginal probability distributions, respectively; is the value set of the relationship feature vector ; is the value set of the multimedia feature vector ;

[0029] S120.2, weight matrix generation: based on the correlation degree of each modal data, a 4×4 modal weight matrix is constructed:

[0030] ;

[0031] wherein, represents the association degree of text and relation features, and the calculation method is the same as ; represents the association degree of text and multimedia features, and the calculation method is the same as ; represents the association degree of behavior and relation features, and the calculation method is the same as ; represents the association degree of behavior and multimedia features, and the calculation method is the same as ; respectively, are the symmetric values of the corresponding association degrees.

[0032] Further, the association degrees of text and relation features , the association degrees of text and multimedia features , the association degrees of behavior and relation features , and the association degrees of behavior and multimedia features all adopt cosine similarity calculation, and the calculation logic is consistent with the calculation of the text-behavior association degree . For example, taking the association degree of text and multimedia features as an example, the calculation formula is as follows:

[0033] ;

[0034] wherein, represents the vector dot product, represents the L2 norm of the vector.

[0035] In addition, since the association between modal features has symmetry (such as the association degree of text and relation is equal to the association degree of relation and text), therefore:

[0036] (the relation-text association degree is equal to the text-relation association degree);

[0037] (the relation-behavior association degree is equal to the behavior-relation association degree);

[0038] (the multimedia-relation association degree is equal to the relation-multimedia association degree);

[0039] (the multimedia-behavior association degree is equal to the behavior-multimedia association degree);

[0040] The above symmetric association degrees do not need to be calculated additionally, and the values of the corresponding association degrees can be directly reused.

[0041] S120.3, feature fusion:

[0042] The normalized text feature vector , the behavior feature vector , the relationship feature vector , the multimedia feature vector The weighted matrix The fusion feature vector is generated by weighted sum :

[0043] ;

[0044] Wherein, is the weight coefficient, and ;

[0045] Further, the modal weight coefficient is calculated as follows:

[0046] First, the sum of the correlation degrees of each modal is calculated:

[0047] ;

[0048] ;

[0049] ;

[0050] ;

[0051] Then, the normalization processing is performed:

[0052] ;

[0053] ;

[0054] ;

[0055] ;

[0056] S120.4, feature verification:

[0057] The reconstruction error of the fusion feature vector and each original modal feature vector :

[0058] ;

[0059] Wherein, represents the Euclidean distance square of the fusion feature vector and the original modal feature vector; represents the error weight coefficient, and the initial value is 0.25;

[0060] If If the number of times of exceeding the preset threshold is greater than a preset threshold, the weight matrix is adjusted and the fusion feature vector is recalculated until the requirement is met.

[0061] As a further improvement of the technical solution, the step of dividing the core social circle based on the interaction intensity of the user social relationship in S200 includes the following steps:

[0062] S210.1, constructing a user social network graph, wherein the nodes represent users and the edges represent the interaction relationship between users, and the edge weight is the interaction frequency in a preset time window;

[0063] S210.2, calculating the interaction intensity index of each user based on the social network graph, and the interaction intensity index is obtained by integrating the interaction frequency, interaction timeliness and interaction diversity;

[0064] S210.3, sorting the interaction intensity indexes of all users in descending order, and selecting the top N% of users to form an initial core social circle;

[0065] S210.4, performing connectivity optimization on the initial core social circle, removing isolated users with sparse interactions with other members, and forming a final core social circle.

[0066] As a further improvement of the technical solution, in S200, the common features of the core social circle are extracted to construct a dynamic prompt template, and the fusion feature set is input into a large language model fine-tuned by social corpus according to the template format, including the following steps:

[0067] S220.1, core social circle common feature extraction:

[0068] Topic distribution feature extraction: performing word frequency statistics on the publishing content and interaction text of the core social circle members, screening keywords with an appearance frequency higher than a preset threshold, generating a topic network containing topic words and association strength through co-occurrence analysis, and mapping the topic network to a topic distribution feature vector , wherein the topic word corresponds to dimension, and the association strength is used as the weight value of the corresponding dimension;

[0069] Interaction mode feature extraction: based on the behavior data of the core social circle members, the occurrence frequency and behavior sequence relationship of comment, forward and like behaviors are counted, the conversion probability between different behaviors is calculated, and an interaction mode feature vector is generated ;

[0070] Time feature extraction: segmenting the active time of the core social circle members, identifying high-frequency active periods and behavior-intensive cycles, and forming a time feature distribution table;

[0071] S220.2, dynamic prompt template construction:

[0072] Design a basic prompt framework, including task instruction section (such as explicitly requiring to generate user interest labels and occurrence probability), feature input section (reserved common feature embedding position), output format section (specify label quantity and probability representation method);

[0073] Convert the extracted topic distribution features, interaction mode features, and time features into natural language descriptions, sort and embed them into the feature input section of the basic prompt framework according to their importance, and form an initial prompt template;

[0074] Adjust the template length dynamically according to the size of the core social circle, and use feature sampling mechanism to retain key information when the number of members exceeds the preset value;

[0075] S220.3, text conversion of the fusion feature set:

[0076] Perform semantic analysis on the weighted fusion feature set generated by S100, and extract the core identifiers of each modal feature (such as keywords in text features and high-frequency behavior types in behavior features);

[0077] Convert the feature core identifiers into structured description text through a pre-trained text generation model, which includes feature categories, weight proportions, and relationship explanations;

[0078] Perform length adaptation processing on the description text to ensure its capacity matches the feature input section of the dynamic prompt template;

[0079] S220.4, prompt text integration and model input:

[0080] Insert the text-based fusion feature set into the specified position of the dynamic prompt template according to the preset format, and enhance the coherence of feature description through semantic connectors;

[0081] Perform compliance check on the integrated prompt text, remove redundant information that exceeds the input length limit of large language models, and retain core feature descriptions;

[0082] Input the optimized prompt text into a large language model fine-tuned on social data, and set the generation parameters;

[0083] S220.5, initial label set generation and processing:

[0084] Receive the candidate labels and corresponding occurrence probabilities output by the large language model, and generate a preliminary label set through threshold filtering (retain labels with probability higher than the set value);

[0085] Perform deduplication processing on the preliminary label set, merge labels with the same or similar semantics (based on a pre-built synonym dictionary), and recalculate the occurrence probability of the merged labels (take the mean or weighted value);

[0086] Output the final initial label set containing label names and corresponding occurrence probabilities.

[0087] As a further improvement of the technical solution, the label hierarchy relationship in S300 is constructed by improving the semantic dependency tree, including the following steps:

[0088] S310.1, Label semantic quantification:

[0089] After deduplication of the initial label set, each label is converted into a semantic vector using a pre-trained language model. The semantic similarity between any two labels is calculated:

[0090] ;

[0091] Where, is the vector dot product, is the L2 norm of the vector, ;

[0092] S310.2, Basic hierarchical clustering:

[0093] Use hierarchical agglomerative clustering to construct the initial hierarchy, initialize each label as an independent node; merge the two nodes with the highest similarity in each iteration, the merging condition is where is a pre-set similarity threshold, and , generate parent node , whose parent node semantic vector is the weighted average of the child node vectors:

[0094] ;

[0095] Where, , are the weights of the child nodes, taking the value of the occurrence probability of the corresponding label in the initial label set; is the semantic vector of the parent node ;

[0096] S310.3, Hierarchy logical verification:

[0097] Generate a description text for the parent node through a large language model, and calculate the semantic inclusion degree between the parent node and the child node:

[0098] ;

[0099] Where,​​​​​ For multilayer perceptrons, This is a vector concatenation operation. ;

[0100] like If so, remove the hierarchical association and re-cluster; where, This is the inclusion threshold;

[0101] S310.4, Dynamic Hierarchical Update:

[0102] 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 ;

[0103] S310.5, Definition of Tag Hierarchy Relationship Symbols:

[0104] 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 .

[0105] 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:

[0106] S320.1 Calculation of aging attenuation coefficient:

[0107] 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:

[0108] Recent Range (Including day 0 and day 7, corresponding to behavior within the last 7 days);

[0109] medium-term range (Excluding the 7th day but including the 30th day, corresponding to actions taken between the 7th and 30th days prior);

[0110] Long-term interval (not including the 30th day, corresponding to the behavior before 30 days);

[0111] Statistical target label behavior frequency in each interval , and calculate the time decay coefficient :

[0112] ;

[0113] Wherein, is the preset basic decay coefficient, which is determined based on the statistical law of social platform user behavior decay;

[0114] S320.2, group correlation value calculation:

[0115] Based on the core social circle determined in S200, the number of users carrying the target label is counted and the total number of users , calculate the initial group correlation value :

[0116] ;

[0117] When , introduce similar social circle data for correction: wherein, is a preset threshold value for judging whether the number of users in the core social circle is sufficient to support the calculation of the group correlation value;

[0118] The similar social circle is determined by calculating the cosine similarity of the target social circle and other social circles in the topic distribution feature vector and the interaction mode feature vector , the formula is:

[0119] ;

[0120] Wherein, is the target social circle, is the candidate social circle, denotes the L2 norm of the vector; is the social circle similarity;

[0121] Select the top social circles with the highest similarity as similar social circles;

[0122] Calculate the average label correlation value of the similar social circle, and correct the initial correlation value through the weighted formula:

[0123] ;

[0124] Wherein, is a weight coefficient, and , wherein is the average number of users in the similar social circle; is the final group correlation value;

[0125] S320.3, attribute set integration:

[0126] The label hierarchical relationship constructed in S310 , the timeliness decay coefficient calculated in S320.1 , the group correlation value calculated in S320.2 are integrated into a triple label attribute set , which is used for subsequent personalized recommendation model training.

[0127] As a further improvement of the technical solution, in the S400, a real-time monitoring window is set, and when a new social behavior is generated by a user, the initial weight assignment of a new label is triggered, including the following steps:

[0128] S410.1, binding construction of the monitoring window:

[0129] Based on the time interval of the timeliness decay in S300, a corresponding monitoring window is constructed:

[0130] The short-term monitoring window covers 1-7 days and is bound to the recent interval in S300;

[0131] The medium-term monitoring window covers 8-30 days and is bound to the medium-term interval in S300;

[0132] The window starts at the current time of the system and ends when the initial weight calculation is triggered;

[0133] S410.2, definition of the trigger condition of the new social behavior:

[0134] According to the behavior data type in S100, the following three types of trigger behaviors are defined:

[0135] The topic label carried by the content published by the user has a semantic similarity ≥ the preset similarity threshold in S300 with the existing label in the label hierarchical relationship in S300 ;

[0136] The cumulative number of comments, forwards, and likes in a unit of time ≥ the average total interaction frequency of the core social circle members in S200 in the same unit of time;

[0137] The number of speeches of the user within 72 hours after joining a new social circle ≥ the average speech frequency of the social circle, and the common label is associated;

[0138] S410.3, establishment of the assignment rule of the initial weight:

[0139] According to the behavior type, the pre-sequenced feature assignment is associated:

[0140] Content publishing trigger: weight association S300 group association value ;

[0141] High-frequency interaction trigger: weight association S200 interaction behavior conversion probability

[0142] Social circle joining trigger: weight association similarity between new social circle and core social circle

[0143] As a further improvement of the technical solution, in the S400, the weight is adjusted regularly according to the subsequent interaction data, and the non-active label decays the weight according to the set period, including the following steps:

[0144] S420.1, binding weight adjustment period:

[0145] Bind the weight adjustment period with the time interval of the 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 interval ; the adjustment time is synchronized with the end time of the monitoring window of S410;

[0146] S420.2, hierarchical processing of subsequent interaction data:

[0147] According to the behavior data type of S100, the influence degree of interaction on weight is distinguished: the influence of comment behavior is associated with the semantic inclusion degree of label in S300 ; the influence of forwarding behavior is associated with the behavior conversion probability of interaction mode in S200; the influence of like behavior is associated with the final group association value in S300 ;

[0148] S420.3, calculate the current label weight Label current weight is calculated by historical weight decay + new interaction gain: historical weight decays according to the time decay coefficient of S300 (Decay is more significant);New interaction gain is calculated based on the hierarchical influence degree of S420.2 (the higher the influence degree, the greater the gain);

[0149] S420.4, execute non-active label decay mechanism:

[0150] Non-active label determination: labels without S420.2 effective interaction data in the last 2 adjustment periods

[0151] Decay method: use the same interval weighting logic as S300 time decay (short-term period decay is faster than medium-term period);

[0152] Elimination rule: when the weight decays below the probability threshold of the initial label set in S200, it is marked as a label to be eliminated and no longer participates in the recommendation calculation of S500.

[0153] As a further improvement of the technical solution, in the S500, the output personalized recommendation adaptation result comprises the following steps:

[0154] S510.1, Scene-based label attribute weight preset Based on the recommended scene type, the label attribute set generated in S300 Preset weight allocation rule:

[0155] Content recommendation scenario: Time decay coefficient Weight coefficient of Final group correlation value Weight coefficient of Label hierarchy Weight coefficient of Satisfies ;

[0156] Friend recommendation scenario: Weight coefficient satisfies ;

[0157] Wherein, ; The weight value can be fine-tuned based on the label weight dynamically adjusted in S400, and the fine-tuning logic is positively correlated with the label activity;

[0158] S510.2, Calculate the matching score of the recommended object and the label , the formula is:

[0159] ;

[0160] Wherein, is the time decay coefficient corresponding to the recommended object, which reuses the calculation logic of in S320.1; is the group correlation value corresponding to the recommended object, which reuses the calculation logic of G in S320.2; is the hierarchical matching degree between the recommended object and the label, which is determined based on the semantic containment degree in S310; ;

[0161] S510.3, Result sorting and core label extraction:

[0162] Arrange the recommended objects in descending order of matching score , forming a recommendation sequence;

[0163] The core contribution label is the label corresponding to the attribute with the highest weight coefficient in the score composition (such as the content recommendation scenario At the highest, take and corresponding attribute parameter value.

[0164] Compared with the prior art, the present application has the beneficial effects:

[0165] 1. The present application can adapt to the complexity of multi-modal data in social scenarios by constructing a multi-modal social feature correlation matrix, integrating user text data, behavior data, relationship data and multimedia data and mapping them to the same feature space, so that the generated label system is more in line with the multi-dimensional features of social data.

[0166] 2. The present application generates initial labels based on a large language model fine-tuned on social corpus, and automatically constructs label hierarchical relationships and calculates time decay coefficients through improved semantic dependency trees, reducing the dependence on artificial rules and improving the depth analysis ability of social semantics, which helps to capture the dynamic changes of user interest.

[0167] 3. The present application introduces a group correlation value as one of the label attributes, and predefines the label attribute weight distribution according to the recommended scene, which makes up for the limitations of single classification dimension, can support the differentiated needs of different scenes such as content recommendation and friend recommendation, and enhances the adaptability of recommendation.

[0168] 4. The present application sets a real-time monitoring window to dynamically adjust the label and attribute weight, so that the label weight can be updated in real time with new social behavior and interaction data of the user, avoiding the lag of static weight, and making the label more reflect the current interest state of the user. BRIEF DESCRIPTION OF DRAWINGS

[0169] Figure 1 The present application is a method for the present application. DETAILED DESCRIPTION

[0170] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application. Embodiment 1

[0171] As Figure 1 shown, the present embodiment provides a large language model social data label attribute generation method for personalized recommendation, including the following steps:

[0172] S100, construct a multi-modal social feature correlation matrix: feature extraction is performed on the text data (such as user post content, comments) of the user, the behavior data (such as likes, forwards, interaction frequency), the relationship data (such as friend links, group chat participation) and the multimedia data (such as picture semantic tags, video key frame text description), and the co-occurrence correlation degree between different modal data is calculated through feature normalization processing mapping to the same feature space, and a weighted fusion feature set is generated;

[0173] In this step, in S100, 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:

[0174] S110.1, text data feature extraction: the user post content and comments are segmented, a pre-trained language model is used to extract semantic vectors, and a text feature vector is generated through a pooling operation;

[0175] As a further description of this step, the segmentation processing in this embodiment uses a general Chinese segmentation tool (such as jieba), and the user's post and comment are segmented, and stop words (such as filtering meaningless words such as "de") are filtered;

[0176] At the same time, 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;

[0177] In addition, the pooling operation in this embodiment uses mean pooling to take the average of the token-level semantic vector to generate a unified dimension text feature vector (such as 768 dimensions) as the basis for subsequent fusion.

[0178] S110.2, behavior data feature extraction: quantize the like, forward and comment behaviors into numerical values, count the behavior frequency according to the time window, and construct a behavior feature vector;

[0179] As a further description of this step, the behavior quantization rule in this embodiment is as follows: like is recorded as 1, forward is recorded as 2, and comment is recorded as 3 (comment length exceeding 20 words adds 1);

[0180] The time window is set to 30 days, and 5 time sub-windows are divided according to the week, and the cumulative value of the behavior in each sub-window is counted (such as 3 likes and 1 forward in the first week, corresponding to the sub-window value 3x1+1x2=5);

[0181] The behavior feature vector dimension is consistent with the number of time sub-windows (such as 5 dimensions), and the vector value is the cumulative behavior quantization result of each sub-window.

[0182] S110.3, relationship data feature extraction: construct a social relationship graph with users as nodes and interaction frequency as edge weight, generate node embedding vectors as relationship feature vectors using graph embedding algorithm;

[0183] As a further description of this step, the social relationship graph in this embodiment is constructed as follows: the target user is the center node, the direct friends are the first-level nodes, and the users with interaction frequency ≥ 5 times in the group chat are the second-level nodes; the edge weight is the total interaction frequency in the last 90 days (such as 20 interactions between friends, the weight is 20);

[0184] At the same time, the graph embedding algorithm uses Node2Vec, sets the walk length to 10 and the window size to 5, and generates 128-dimensional node embedding vectors as relationship feature vectors. The vector values reflect the position and association strength of the user in the social network.

[0185] S110.4, multimedia data feature extraction: extract key features from pictures and videos to generate multimedia feature vectors;

[0186] As a further description of this step, the picture feature extraction and video feature extraction in this embodiment specifically include:

[0187] Picture feature extraction: a pre-trained image feature extraction model (such as ResNet50) is used to encode the features of the pictures posted by the user, outputting 2048-dimensional original features, which are reduced to 256-dimensional features through principal component analysis;

[0188] Video feature extraction: a key frame is extracted every 10 seconds, processed according to the picture feature extraction method, and the mean value of the multi-frame features is taken to generate a 256-dimensional video feature vector, which is consistent in dimension with the picture feature vector.

[0189] S110.5, feature normalization processing: use normalization algorithm to process text feature vectors, behavior feature vectors, relationship feature vectors, and multimedia feature vectors respectively, and map each modal feature to the same feature space.

[0190] As a further description of this step, the Min-Max normalization algorithm is used in this embodiment to process text, behavior, relationship, and multimedia feature vectors respectively, and map each vector value to the [0, 1] interval; the mapping formula is:

[0191] ;

[0192] wherein, is the original vector value, , the minimum and maximum values of the modal feature respectively, to ensure that different modal features are comparable within the same numerical range.

[0193] In this step, in S100, the co-occurrence correlation degree between different modal data is calculated, and a weighted fusion feature set is generated, including the following steps:

[0194] S120.1, correlation degree calculation:

[0195] Calculate the text feature vector and the cosine similarity of the behavior feature vector as the text-behavior correlation degree :

[0196] ;

[0197] wherein, denotes the vector dot product, denotes the L2 norm of the vector;

[0198] Calculate the relationship feature vector and the mutual information entropy of the multimedia feature vector as the relationship-multimedia correlation degree :

[0199] ;

[0200] wherein, denotes the joint probability distribution; and are the marginal probability distributions, respectively; is the value set of the relationship feature vector ; is the value set of the multimedia feature vector ;

[0201] For example, taking a user as an example, the text feature vector (768 dimensions) describes the "outdoor hiking" related content, the behavior feature vector (5 dimensions) shows that the user frequently likes "hiking equipment" posts in the past 30 days, and the cosine similarity between the two is calculated, indicating that the text and behavior are highly correlated;

[0202] The relationship feature vector (128 dimensions) of the "hiking group" that the user belongs to reflects that the interaction within the group is close, and the picture feature vector (256 dimensions) published by the user is mostly in the hiking scene, and the mutual information entropy is calculated, indicating that the relationship network and the multimedia content are moderately correlated.

[0203] S120.2, weight matrix generation: based on the correlation degree of each modal data, a 4x4 modal weight matrix is constructed:

[0204] ;

[0205] wherein, represents the relevance of text and relationship features, and the calculation method is the same as ; represents the relevance of text and multimedia features, and the calculation method is the same as ; represents the relevance of behavior and relationship features, and the calculation method is the same as ; represents the relevance of behavior and multimedia features, and the calculation method is the same as ; respectively, are the symmetric values of the corresponding relevance;

[0206] S120.3, feature fusion:

[0207] The normalized text feature vector , the behavior feature vector , the relationship feature vector , the multimedia feature vector are weighted and summed according to the weight matrix to generate a fusion feature vector :

[0208] ;

[0209] wherein, is a weight coefficient, and ;

[0210] Further, the modal weight coefficient is calculated as follows:

[0211] First, the sum of the relevance of each modal is calculated:

[0212] ;

[0213] ;

[0214] ;

[0215] ;

[0216] Then, normalization processing is performed:

[0217] ;

[0218] ;

[0219] ;

[0220] ;

[0221] S120.4, feature verification:

[0222] Calculate the fusion feature vector Reconstruction error with each original modal feature vector :

[0223] ;

[0224] Wherein, represents the Euclidean distance square of the fusion feature vector and the original modal feature vector; represents the error weight coefficient, and the initial value is 0.25;

[0225] If exceeds the preset threshold, adjust the weight matrix And recalculate the fusion feature vector until satisfy the requirements.

[0226] For example, taking a social platform user "Zhang San" as an example, the complete implementation process of constructing a multi-modal social feature correlation matrix is as follows:

[0227] First, collect Zhang San's text data (post "weekend hiking" in the past 30 days, comments), behavior data (like "hiking route" post 12 times, forward 5 times), relationship data (join "outdoor enthusiasts" group, interact with 10 people in the group more than 30 times), multimedia data (post 3 hiking photos);

[0228] Then, extract each modal feature and normalize according to S110.1-S110.5 to get 、 、 、 ;

[0229] Then, calculate the correlation degree of each modal according to S120.1 to generate the weight matrix ;

[0230] Next, calculate the weight coefficient and fuse the features according to S120.3 to get ;

[0231] Finally, verify the reconstruction error according to S120.4, and confirm that satisfy the threshold, output the fusion feature set with weight, which is used for subsequent label attribute generation.

[0232] It needs to be supplemented that, in order to calculate the relationship feature and the multimedia feature The mutual information entropy needs to be discretized first. The embodiment adopts the equal-width binning method as the default scheme:

[0233] The 128-dimensional relationship features output by Node2Vec The 64-dimensional multimedia features after ResNet50+PCA dimensionality reduction The global minimum value min and maximum value max of each dimension are counted respectively.

[0234] Divide equal-width intervals (interval width ), and map the continuous feature value to the corresponding interval (for example interval 3); count the co-occurrence frequency of and as the basis for calculating the joint probability .

[0235] If the feature distribution has a long-tail characteristic (such as the interaction frequency class feature), the equal-frequency binning method can be switched (to ensure that the sample size of each interval is comparable), which can be dynamically adjusted according to the feature dimension (such as 5≤ ≤10) to adapt to the feature distribution of different social scenarios.

[0236] S200, generating an initial label set adapted to a social scenario: dividing a core social circle based on the interaction intensity of user social relationships (such as a user set whose interaction frequency and reply rate reach a set threshold within the last 90 days), extracting common features of the core social circle (such as interaction frequency and common topic proportion) to construct a dynamic prompt template, inputting the fusion feature set into a large language model fine-tuned on social corpus according to the template format, and generating an initial label set containing appearance probability description;

[0237] In this step, the core social circle based on the interaction intensity of user social relationships in S200 includes the following steps:

[0238] S210.1, constructing a user social network graph, wherein the nodes represent users and the edges represent the interaction relationship between users, and the edge weight is the interaction frequency within a preset time window;

[0239] As a further description of this step, the construction of the user social network graph in this embodiment includes the following steps:

[0240] First, define the network nodes and edges: take the target user as the center node, other users as associated nodes, and the edge represents the interaction relationship between users;

[0241] Then, specify the interaction behavior range: comments, reposts, and private messages are recognized as valid interactions (simple likes are not counted), and a single valid interaction is recorded as 1;

[0242] Then, set the time window: the time window is the recent 90 days (which can be adjusted according to the platform interaction frequency), denoted as ;

[0243] Next, calculate the edge weight: the edge weight of the user and the interaction weight of the user is calculated according to the following formula: , wherein is whether there is an effective interaction on the th day (1 if there is, otherwise 0);

[0244] Finally, generate a network graph: label the user as the node, and label the calculated value as the edge.

[0245] S210.2, calculate the interaction intensity index of each user based on the social network graph, and the interaction intensity index is obtained by integrating the interaction frequency, interaction timeliness and interaction diversity;

[0246] As a further description of this step, the calculation of the interaction intensity index in this embodiment includes the following steps:

[0247] First, determine the weighting coefficient: set , respectively as the weight coefficient of the interaction frequency, timeliness and diversity, the sum of the three is 1, based on the importance distribution of social interaction characteristics);

[0248] Subsequently, calculate the interaction frequency normalization value: take the maximum edge weight of the target user and all associated users, calculate ;

[0249] Then, calculate the timeliness factor: is the number of days from the last interaction to the current day, (the timeliness decay coefficient), and the calculation formula is: ;

[0250] Next, calculate the interaction diversity: (the total number of effective interaction types, i.e. comments, forwards and private messages), is the number of interaction types of the user , calculate ;

[0251] Finally, integrate the indexes: calculate the interaction intensity :

[0252] .

[0253] S210.3, sort the interaction intensity indicators of all users in descending order, and select the top N% of users to form an initial core social circle;

[0254] As a further description of this step, the selection of the initial core social circle in this embodiment includes the following steps:

[0255] First, sort the interaction intensity: sort the interaction intensity indicators of all associated users in descending order;

[0256] Then, set the selection ratio N: the total number of associated users N = 20 when N = 15 (N is the selection ratio, which is dynamically adjusted according to the density of the social network);

[0257] Then, determine the number of selections: calculate the initial number of members according to the top N% of the total number of associated users (for example, if there are 80 associated users and N = 20, the number of selected users is

[0258] Finally, select the initial circle: select the top N% of users after sorting to form the initial core social circle.

[0259] S210.4, perform connectivity optimization on the initial core social circle, remove isolated users with sparse interactions with other members, and form a final core social circle.

[0260] As a further description of this step, the connectivity optimization of the initial core social circle in this embodiment includes the following steps:

[0261] First, calculate the average interaction frequency: the average interaction frequency between members of the initial core social circle ; wherein, is the average interaction frequency between members of the initial circle; is the sum of edge weights between all members of the initial circle; is the total number of edges between members of the initial circle (i.e., the number of actual interaction relationships between members);

[0262] Then, set the isolation threshold: the isolation threshold is (based on network connectivity experience value, used to determine isolated users);

[0263] Then, identify isolated users: check each member of the initial circle, if the interaction frequency between a user and all other members is less than the isolation threshold, it is determined to be an isolated user;

[0264] Finally, generate the final circle: remove all isolated users, and the remaining members form the final core social circle.

[0265] ​​In this step, in S200, the common features of the core social circle are extracted to construct a dynamic prompt template, and the fusion feature set is input into the large language model fine-tuned by social corpus in the template format, including the following steps:

[0266] S220.1, core social circle common feature extraction:

[0267] Topic distribution feature extraction: word frequency statistics are performed on the publishing content and interactive text of the core social circle members, keywords with an occurrence frequency higher than a preset threshold are screened, a topic network containing topic words and association strength is generated through co-occurrence analysis, and the topic network is mapped into a topic distribution feature vector , wherein the topic word corresponds to the dimension of , and the association strength is used as the weight value of the corresponding dimension;

[0268] Interaction mode feature extraction: based on the behavior data of the core social circle members, the occurrence frequency and behavior sequence relationship of comment, forward, and like behaviors are counted, the conversion probability between different behaviors is calculated, and an interaction mode feature vector is generated ;

[0269] Time feature extraction: the active time of the core social circle members is segmented and counted, high-frequency active time periods and behavior-intensive cycles are identified, and a time feature distribution table is formed;

[0270] As a further description of this step, the topic distribution feature extraction in this embodiment includes the following steps:

[0271] First, pre-process the text: remove duplicates, filter out emojis and links, use a segmentation tool (such as jieba) to segment the text and remove stop words (such as “de” and “le”);

[0272] Then, screen the topic words: count the word frequency and keep the words with an occurrence frequency greater than 5% (preset proportion) of the total number of texts as topic words;

[0273] Then, calculate the co-occurrence strength: count the number of times the topic words co-occur ( , the number of times the topic words and the topic words appear in the same text at the same time), and take ( , the maximum value of the co-occurrence strength of the topic word a and all other topic words);

[0274] Next, generate a feature vector: the dimension of the topic distribution feature vector corresponds to the number of topic words, and the dimension value , wherein: is the topic word The sum of the co-occurrence strengths of all other subject words;

[0275] Finally, output as the topic distribution feature.

[0276] Further, the interaction mode feature extraction in the embodiment includes the following steps:

[0277] First, count the behavior frequency: record the total number of comments and forwarding behaviors of the core circle members, and calculate the behavior proportion (such as comment proportion = comment number / total behavior number);

[0278] Then, count the conversion number: count the conversion number of continuous behaviors within 7 days (such as the number of times of commenting first and then forwarding);

[0279] Then, calculate the conversion probability:

[0280] ;

[0281] Wherein: is the probability of behavior conversion behavior ; is the total number of behaviors occurring within 7 days ; represents the total number of behaviors occurring within the statistical period;

[0282] Next, construct a feature vector: the dimension of the interaction mode feature vector is “behavior proportion + conversion probability” (such as containing 2 behaviors of comments and forwarding, the dimension is 2 behavior proportions + 2 conversion probabilities = 4 dimensions);

[0283] Finally, output as the interaction mode feature.

[0284] S220.2, dynamic prompt template construction:

[0285] Design a basic prompt framework, including a task instruction segment (such as explicitly requiring to generate user interest labels and occurrence probability), a feature input segment (reserved common feature embedding position), and an output format segment (specify the number of labels and probability representation method);

[0286] Convert the extracted topic distribution feature, interaction mode feature, and time feature into natural language description, sort and embed the feature input segment of the basic prompt framework according to the feature importance, and form an initial prompt template;

[0287] According to the size of the core social circle, dynamically adjust the length of the template, when the number of members exceeds the preset value, use feature sampling mechanism to retain key information;

[0288] As a further illustration of this step, the dynamic prompt template construction in this embodiment includes the following specific steps:

[0289] First, design the basic framework: include the task instruction section (clearly generate interest tags and probability range, format requirements), feature input section (reserve "[topic feature]", "[interaction mode]", "[time feature]" placeholders), output format section (provide examples);

[0290] Subsequently, convert the features into natural language: for example, describe the topic feature as "core circle high-frequency discussion XX, co-occurrence intensity XX", describe the interaction mode as "mainly XX behavior (XX proportion), XX conversion probability XX", and describe the time feature as "high-frequency active period is XX";

[0291] Then, embed according to importance: in the order of "topic feature > interaction mode > time feature" (based on the direct impact on tag generation), embed the natural language features into the corresponding placeholders;

[0292] Next, adjust the template length: if the core circle members are > 50, take the top 50% of the theme words by word frequency, retain behavior types with a proportion > 10%, and ensure that the total length of the template is 500 words;

[0293] Finally, generate the dynamic prompt template.

[0294] S220.3, text conversion of the fused feature set:

[0295] Perform semantic analysis on the weighted fused feature set generated by S100 to extract the core identifiers of each modality feature (such as keywords in text features, high-frequency behavior types in behavior features);

[0296] Convert the feature core identifiers into structured description text through a pre-trained text generation model, and the description text includes feature categories, weight proportions, and correlation relationship descriptions;

[0297] It can be understood that this embodiment converts the description text through a pre-trained text generation model (such as T5-small), for example: "User text features mainly include'marathon' (30%), high-frequency behaviors include liking fitness content (25%), and the two are highly correlated; closely interact with 10 fitness enthusiasts (20%). "The text length is adapted to be within 150 words to ensure compatibility with the template.

[0298] Perform length adaptation processing on the description text to ensure that it matches the feature input section capacity of the dynamic prompt template;

[0299] S220.4, prompt text integration and model input:

[0300] Insert the text-based fusion feature set into the designated position of the dynamic prompt template in the preset format, and enhance the coherence of the feature description through semantic linking words.

[0301] Perform compliance checks on the integrated prompt text, remove redundant information that exceeds the input length limit of the large language model, and retain the core feature description.

[0302] Input the optimized prompt text into the large language model fine-tuned on social corpus, and set the generation parameters; the social corpus is selected from public interactive text on social platforms (such as comments and private messages), and is processed by de-duplication, user identification desensitization, and low-quality content filtering (such as filtering invalid text with too few characters).

[0303] S220.5, initial label set generation and processing:

[0304] Receive the candidate labels and corresponding appearance probabilities output by the large language model, and generate a preliminary label set by threshold filtering (retain labels with a probability higher than a set value);

[0305] Perform de-duplication processing on the preliminary label set, merge labels with the same or similar semantics (based on a pre-built synonym dictionary), and recalculate the appearance probability of the merged labels (take the average or weighted value);

[0306] Output the final initial label set containing label names and corresponding appearance probabilities.

[0307] As a further explanation of this step, the initial label set generation and processing in this embodiment includes the following steps:

[0308] First, filter the labels: retain labels with a probability > 0.5 (filtering threshold, based on feature correlation) in the model output to form a preliminary label set;

[0309] Then, de-duplicate and merge: based on a pre-built synonym dictionary (including public word libraries and social field supplementary near-synonyms), merge labels with similar semantics;

[0310] Then, recalculate the probability: take the average of the original probabilities of the merged labels (if the probability difference 0.2);

[0311] Finally, output the final initial label set, with an example format of: "Label 1 (probability); Label 2 (probability);...".

[0312] S300, multi-dimensional quantification of label attributes: build label hierarchical relationships through improved semantic dependency trees (e.g., "outdoor hiking" belongs to "sports"), calculate time decay coefficients based on the time interval of behavior occurrence (recent behavior coefficients are 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;

[0313] In this step, the construction of label hierarchical relationships in S300 through improved semantic dependency trees includes the following steps:

[0314] S310.1, label semantic quantification:

[0315] After removing duplicates from the initial label set, a pre-trained language model is used to convert each label into a semantic vector of dimension , and the semantic similarity between any two labels is calculated :

[0316] ;

[0317] where is the dot product of the vectors, and are the L2 norms of the vectors, respectively;

[0318] As a further explanation of this step, the label semantic quantification in this embodiment includes the following steps:

[0319] First, perform a de-duplication process on the initial label set: based on a social field synonym library (e.g., containing near-synonym mappings such as "hiking" and "hiking", "fitness" and "bodybuilding"), merge semantically repeated labels, and retain unique labels (e.g., merge "outdoor hiking" and "wilderness hiking" into "outdoor hiking");

[0320] Subsequently, select a pre-trained language model: use the Sentence-BERT model suitable for short text semantic encoding (as this model performs stably in short phrase semantic similarity calculation and has uniform output dimensions), and convert each label into a fixed-dimensional semantic vector;

[0321] Then, calculate the similarity of all label pairs according to the semantic similarity formula, and output a structured similarity matrix (e.g., 5 labels correspond to a 5x5 matrix, and the matrix elements are the similarity values of each pair of labels);

[0322] Finally, store the similarity matrix in CSV format for subsequent clustering processing, where the row and column indices are the label names, and the cell values are the corresponding similarity.

[0323] S310.2, Basic Hierarchical Clustering:

[0324] 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.

[0325] ;

[0326] in, , The weight of the child node is the probability of the corresponding label appearing in the initial label set. parent node semantic vector;

[0327] As a further explanation of this step, the execution of the basic hierarchical clustering in this embodiment includes the following specific steps:

[0328] 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).

[0329] 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.

[0330] Then, according to the clustering rules disclosed in the claims, the iterative merging unit is: each time, ≥ are selected 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).

[0331] 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").

[0332] S310.3, Hierarchical Logic Verification:

[0333] Generating a parent node by a large language model , calculating a semantic inclusion degree of the parent node and the child node :

[0334] ;

[0335] wherein, is a multi-layer perceptron, is a vector concatenation operation, ;

[0336] If , remove the hierarchical association and re-cluster; wherein, is an inclusion degree threshold value;

[0337] Further, the 3MLP model for calculating the semantic inclusion degree in the embodiment is designed as follows:

[0338] Input layer: the dimension is equal to the dimension of the multi-modal fusion feature (such as 128 dimensions, directly connected to the fused feature vector, reflecting the compatibility of the method), and the fused feature vector is directly accessed;

[0339] Hidden layer: 64 neurons are set, and a ReLU activation function (to alleviate the gradient disappearance problem, which is a well-known activation function in the art) is used;

[0340] Output layer: 2 neurons are set, and a Softmax activation is used to output the probability distribution of “existence / nonexistence of semantic inclusion relationship”.

[0341] The model training uses a cross-entropy loss function.

[0342] S310.4, dynamic hierarchical update:

[0343] Set an update period , when the semantic vector of a newly added label has a maximum similarity with an existing hierarchical node, trigger hierarchical reconstruction, and repeat steps S310.2-S310.3 to update the semantic dependency tree; wherein, is an update threshold value, and ;

[0344] As a further description of this step, the dynamic hierarchical update in the embodiment includes the following specific steps:

[0345] First, set an update period : based on the frequency of adding labels on a social platform, set to 7 days (i.e., check every 7 days whether there are newly added labels);

[0346] Subsequently, a new label is received: the new label is derived from the label generation result triggered by the new behavior of the user (for example, the user recently frequently mentions "trail running", and "trail running" is a new label);

[0347] Then, the maximum similarity of the new label and all nodes in the existing hierarchy is calculated, and an update threshold is set (for example = 0.6, then = 0.5, and the merging threshold of the new label is lowered);

[0348] Next, if the maximum similarity is greater than or equal to , the new label is included in the clustering pool, and the clustering and checking process of S310.2-S310.3 is repeated to update the hierarchical tree;

[0349] Finally, an update log is recorded, including the name of the new label, the merged node, and the comparison of the hierarchical structure before and after the update.

[0350] S310.5, definition of label hierarchical relationship symbol:

[0351] The semantic dependency tree finally generated by S310.1-S310.4, and the parent-child node hierarchical structure and semantic inclusion degree contained therein are uniformly defined as label hierarchical relationship .

[0352] In this step, in S300, the time decay coefficient is calculated based on the time interval of the behavior, and the proportion of users with the same label in the core social circle is calculated as the group association value, including the following steps:

[0353] S320.1, time decay coefficient calculation:

[0354] Based on the user behavior data extracted in S100, the behavior time sequence corresponding to the target label is determined , where is the time of the th behavior, and ; the interval between the current time and each behavior time is divided into mutually exclusive time interval sets , where:

[0355] The recent interval (includes the 0th day and the 7th day, corresponding to behaviors within the last 7 days);

[0356] The medium-term interval (does not include the 7th day but includes the 30th day, corresponding to behaviors before 7-30 days);

[0357] The long-term interval (not including the 30th day, corresponding to the behavior before 30 days);

[0358] Statistics of the frequency of the behavior of the target label in each interval , and calculate the time decay coefficient :

[0359] ;

[0360] wherein, is a preset basic decay coefficient, determined based on the statistical law of social platform user behavior decay;

[0361] In this step, determined based on the statistical law of social platform user behavior decay, for example: (recently), (medium term), (long term), which can be adjusted according to the platform user activity, such as high active users can be increased to 0.9.

[0362] S320.2, group correlation value calculation:

[0363] Based on the core social circle determined in S200, the number of users carrying the target label and the total number of users , calculate the initial group correlation value :

[0364] ;

[0365] When , introduce similar social circle data for correction: wherein, is a preset threshold value for judging whether the number of users in the core social circle is sufficient to support the calculation of the group correlation value;

[0366] The similar social circle is determined by calculating the cosine similarity of the target social circle and other social circles in the topic distribution feature vector and the interaction mode feature vector , the formula is:

[0367] ;

[0368] wherein, is the target social circle, is the candidate social circle, denotes the L2 norm of the vector; is the social circle similarity;

[0369] Select the top social circles with the highest similarity as similar social circles;

[0370] Calculate the average tag correlation value of the similar social circle And correct the initial correlation value by the weighted formula:

[0371] ;

[0372] Wherein, is the weight coefficient, and , wherein is the average number of users in the similar social circle; is the final group correlation value;

[0373] S320.3, attribute set integration:

[0374] The label hierarchical relationship constructed in S310 , the time decay coefficient calculated in S320.1 , the group correlation value calculated in S320.2 are integrated into a triple label attribute set , which is used for subsequent personalized recommendation model training.

[0375] S400, dynamically adjust the label and attribute weight: set a real-time monitoring window, when a user generates a new social behavior (such as publishing new content, high-frequency interaction, joining a new group chat and speaking), trigger the initial weight assignment of the new label, adjust the weight according to the subsequent interaction data (such as other user interaction, secondary forwarding), and decay the weight according to the set period for non-active labels (such as no correlation behavior for more than a set period);

[0376] In this step, in S400, the initial weight assignment of the new label is triggered when the user generates a new social behavior, including the following steps:

[0377] S410.1, binding construction of monitoring window:

[0378] Based on the time interval of time decay in S300, the corresponding monitoring window is constructed:

[0379] The short-term monitoring window covers 1-7 days, which is bound to the recent interval of S300;

[0380] The medium-term monitoring window covers 8-30 days, which is bound to the medium-term interval of S300;

[0381] The window starts at the current system time, and ends when the initial weight calculation is triggered;

[0382] As a further illustration of this step, the binding of the monitoring window in this embodiment directly correlates the time interval of the aging decay in S300: the short-term monitoring window covers 1-7 days, binding with the recent interval [0, 7], the medium-term monitoring window covers 8-30 days, binding with the medium-term interval (7, 30]. The windows all take the current time of the system as the starting point, and end at +7 days, +30 days, respectively. The end time automatically triggers the initial weight calculation of the new label. During the running of the window, user behavior data is stored in real time in the format of "behavior type-occurrence time-associated label", such as "publish content-20231005-outdoor hiking" and "comment-20231006-marathon", etc., to provide basic data support for subsequent weight calculation.

[0383] S410.2, define the trigger condition of the new social behavior:

[0384] According to the behavior data type in S100, the following three types of trigger behaviors are defined:

[0385] The semantic similarity of the topic label carried by the user's published content with the existing label in the label hierarchical relationship in S300 is greater than or equal to the preset similarity threshold in S300 ;

[0386] The weighted cumulative number of comments, forwards, and likes in a unit of time (calculated as "comments x 2 + forwards x 3 + likes x 1") is greater than or equal to the average total interaction frequency of the core social circle members in the same unit of time in S200 (this frequency is the statistical mean of the unit time interaction behavior of the core social circle members in S200);

[0387] The number of speeches of the user within 72 hours after joining the new social circle is greater than or equal to the average speech frequency of the social circle, and the common label is associated;

[0388] As a further illustration of this step, in S410.2, the trigger condition of the new social behavior is defined, and according to the behavior data type in S100, the three types of trigger behaviors are associated with the parameters in the previous steps:

[0389] The semantic similarity of the topic label carried by the user's published content with the existing label in the label hierarchical relationship in S300 is greater than or equal to the preset similarity threshold in S300 (this threshold is the core determination basis for S300 label clustering);

[0390] The weighted cumulative number of comments, forwards, and likes in a unit of time (calculated as "comments x 2 + forwards x 3 + likes x 1") is greater than or equal to the average total interaction frequency of the core social circle members in the same unit of time in S200 (this frequency is the statistical mean of the unit time interaction behavior of the core social circle members in S200);

[0391] ​The number of speeches of a user within 72 hours after joining a new social circle is greater than or equal to the average speech frequency of the circle in the past 30 days (total number of speeches ÷ number of members), and the speeches are associated with the common label of the circle determined by the S200 common feature extraction method.

[0392] S410.3, establish the assignment rule of the initial weight:

[0393] Assign the pre-sequence feature according to the behavior type:

[0394] Content publishing trigger: weight association S300 group association value ;

[0395] High-frequency interaction trigger: weight association S200 interaction behavior conversion probability;

[0396] Social circle joining trigger: weight association similarity between new social circle and core social circle.

[0397] As a further description of this step, in the step of establishing the assignment rule of the initial weight in S410.3, the specific way of associating the pre-sequence feature according to the trigger behavior type is as follows:

[0398] When the content publishing trigger occurs, the initial weight is calculated according to the following formula: , wherein (correlation coefficient of content publishing and group association, based on social scene characteristics), is the final group association value output in S300;

[0399] When the high-frequency interaction trigger occurs, the initial weight is calculated according to the following formula: , wherein (reflection coefficient of interaction conversion to interest, reflecting the association between interaction behavior and interest), is the interaction behavior conversion probability calculated in S200;

[0400] When the social circle joining trigger occurs, the initial weight is calculated according to the following formula: , wherein (influence coefficient of social circle similarity to label relevance, highlighting the association strength of social group attribute), is the similarity between the new social circle and the core social circle (calculated according to the social circle similarity formula in S320.2).

[0401] Further, the trigger condition of high-frequency interaction in this embodiment is defined as: the user interaction frequency is more than 1.5 times the average interaction frequency of the core social circle (based on the dynamic adjustment of the core circle, to avoid the scene limitation of fixed threshold), and the average frequency of the core social circle is output through the S200 step.

[0402] In this step, in S400, the weight is adjusted regularly according to the subsequent interaction data, and the non-active label is decayed in a set period, including the following steps:

[0403] S420.1, binding weight adjustment period:

[0404] The weight adjustment period is bound to the time interval of the 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 interval ; the adjustment time is synchronized with the end time of the monitoring window in S410;

[0405] S420.2, hierarchical processing of subsequent interaction data:

[0406] According to the behavior data type in S100, the influence degree of interaction on weight is distinguished: the influence of comment behavior is related to the semantic inclusion degree of label in S300 ; the influence of forwarding behavior is related to the behavior conversion probability of interaction mode in S200; the influence of like behavior is related to the final group correlation value in S300 ;

[0407] S420.3, calculating the current label weight Label current weight is calculated by historical weight decay + new interaction gain: historical weight is decayed according to the time decay coefficient in S300 (Decay is more significant);New interaction gain is calculated based on the hierarchical influence degree in S420.2 (the higher the influence degree, the greater the gain);

[0408] S420.4, execute non-active label decay mechanism:

[0409] Non-active label determination: the label without S420.2 effective interaction data in the last 2 adjustment periods;

[0410] Decay method: adopt the interval weighting logic consistent with S300 time decay (short-term period decay is faster than medium-term period);

[0411] Elimination rule: when the weight decay is lower than the probability threshold of the initial label set in S200, mark it as a label to be eliminated, and no longer participate in the recommendation calculation in S500.

[0412] As a further description of this step, the implementation logic of adjusting weight regularly according to subsequent interaction data and non-active label decay in S420 is as follows:

[0413] In S420.1, the weight adjustment period is bound to the time interval of the 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 interval), the adjustment time is synchronized with the end time of the monitoring window in S410, and the weight adjustment is consistent with the time decay of user interest;

[0414] In S420.2, the influence degree is processed according to the behavior data type in S100:

[0415] The influence degree of comment behavior The calculation formula is: Wherein (the reinforcement coefficient of comment on label semantics), is the semantic inclusion degree of label in S300;

[0416] The influence degree of forwarding behavior The calculation formula is: Wherein (the propagation influence coefficient of forwarding), is the behavior conversion probability of interaction mode in S200;

[0417] The influence degree of like behavior The calculation formula is: Wherein (the weak influence coefficient of like), is the final group correlation value in S300.

[0418] Further, in S420.3, the current label weight is calculated according to the following formula:

[0419] ;

[0420] Wherein is the label weight of the last period, is the time decay coefficient of S300 (the smaller the value, the more significant the decay), is the total influence degree of all new interactions in the adjustment period.

[0421] Further, in S420.4, the non-active label decay mechanism is executed, the non-active label refers to the label without effective interaction data in S420.2 in the last two adjustment periods, the decay method adopts the interval weighting logic consistent with the time decay in S300 (the short-term period decays faster than the medium-term period), when the weight decays to less than the probability threshold of the initial label set in S200, it is marked as a label to be eliminated and no longer participates in the recommendation calculation in S500.

[0422] S500, outputting the personalized recommendation adaptation result: according to the preset label attribute weight distribution (the time effectiveness attribute weight proportion is higher in content recommendation, and the group correlation attribute weight proportion is higher in friend recommendation), calculating the matching score of the recommended object and the label, and outputting the result and the core contribution label in order of score.

[0423] In this step, in S500, outputting the personalized recommendation adaptation result includes the following steps:

[0424] S510.1, scene-based label attribute weight preset based on the recommendation scene type, the label attribute set generated in S300 Preset weight distribution rule:

[0425] Content recommendation scenario: time decay coefficient Weight coefficient of Final group correlation value Weight coefficient of Label hierarchy Weight coefficient of Satisfy ;

[0426] Friend recommendation scenario: weight coefficient satisfies ;

[0427] Among them, ; The weight value can be fine-tuned based on the label weight adjusted dynamically in S400, and the fine-tuning logic is positively correlated with the label activity level;

[0428] As a further description of this step, , and the size relationship is distinguished according to the scene characteristics:

[0429] Content recommendation scenario: it is necessary to match the user's recent interest first, so the weight coefficient satisfies , the example value is , , (based on the strong influence of content timeliness on user attention);

[0430] Friend recommendation scenario: it is necessary to match the commonality of social circle first, so the weight coefficient satisfies , the example value is , , (based on the core role of group correlation in social matching).

[0431] Further, the weight fine-tuning logic is positively correlated with the label activity level: if the weight of a certain label after dynamic adjustment in S400 is higher than the average weight of the same type of label (e.g., the weight of "outdoor hiking" is 0.7, which is higher than the average 0.5), the weight coefficient of the attribute corresponding to the label can be increased by 5-10% (e.g., in the content recommendation scenario , the weight is fine-tuned from 0.5 to 0.525), and vice versa, to ensure that the influence of active labels is strengthened.

[0432] S510.2, match score calculation of the recommended object and the label The formula is:

[0433] ;

[0434] Wherein, is the time decay coefficient corresponding to the recommended object, which reuses the calculation logic in S320.1, that is, based on the time sequence of the associated behavior of the recommended object, the recent (0-7 days), medium-term (7-30 days), and long-term (> 30 days) intervals are divided, the behavior frequency of each interval is counted, and the interval is weighted for calculation; is the group association value corresponding to the recommended object, which reuses the calculation logic of G in S320.2, that is, the proportion of users in the core social circle who follow the object is counted, and if the number of social circle users is small, similar social circle data is introduced for correction (the correction method is consistent with S320.2); is the hierarchical matching degree of the recommended object and the label, which is determined based on the semantic inclusion degree in S310; ; ;

[0435] S510.3, result sorting and core label extraction:

[0436] The recommended objects are arranged in descending order of match score to form a recommendation sequence;

[0437] The core contribution label is the label corresponding to the attribute with the highest weight coefficient in the score (e.g., in the content recommendation scenario , the label associated with is taken), and the attribute parameter value corresponding to the label is marked.

[0438] As a further description of this step, the result sorting and core label extraction in this embodiment specifically includes:

[0439] Result sorting: the recommended objects are arranged in descending order of match score to form a recommendation sequence, and the objects with higher scores are recommended first;

[0440] Core contribution label extraction: obtain the label corresponding to the attribute with the highest weight coefficient in the score composition (such as the content recommendation scenario The highest time, take the label associated with And mark the attribute parameter value (such as Or The specific calculation result).

[0441] Those skilled in the art can understand that the process of implementing all or part of the steps of the above embodiments can be completed by hardware, or by program to instruct related hardware to complete.

[0442] The basic principles, main features and advantages of the present application are shown and described above. Those skilled in the art should understand that the present application is not limited by the above embodiments, and the above embodiments and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application 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; 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 ; 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 oriented towards 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, 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.

7. The method for generating social data tag attributes for large language models for personalized recommendation according to claim 6, 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.

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, 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 interaction data in the three categories of S420.2 within two consecutive adjustment periods, including the semantic inclusion degree of tags in S300 related to comment behavior. Interaction data, forwarding behavior correlation with the conversion probability of interaction patterns in S200, and like behavior correlation with the final group correlation value in S300. Interactive data; 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.

9. The method for generating social data tag attributes for large language models for personalized recommendation according to claim 8, 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