Method and device for improving link establishment rate, equipment and storage medium

Abstract

The invention provides a method, a device, equipment and a storage medium for improving a link establishment rate. Behavior data and text semantics are fused through a pre-trained feature extraction model to generate user feature vectors, a multi-level user tag system is constructed by using a tag generation model, and common features of both parties are extracted through a tag clustering model. And then, in combination with historical interaction data and a label system of the target user, obtaining a current emotion state by using an emotion analysis model, and displaying the emotion on an initiator interface in a pet expression form. And finally, inputting the matching features, the emotional state and the tag system into a copywriting generation model, generating an emotion-adaptive personalized recruitment, and dynamically optimizing each model parameter according to the response, thereby improving the correlation and the infectivity of the information, and further improving the success rate of establishing the recruitment.

Description

Technical Field

[0001] The present invention relates to the field of data processing of social platforms, and in particular to a method, apparatus, device and storage medium for improving connection rate. Background Art

[0002] On today's social platforms and instant messaging applications, the initial connection between users has become a critical part of social interaction, and the connection rate directly impacts user activity and retention on social platforms. Traditional social connection methods rely primarily on users manually writing greetings or using templated greetings provided by the platform, which have numerous limitations. When connecting with strangers, users often struggle to accurately grasp their emotional state and interests, resulting in greetings that lack specificity and appeal, and are easily ignored or rejected. Furthermore, while existing social platforms collect a wealth of user behavior data, they fail to effectively leverage this data to optimize the user connection process, particularly in terms of emotional matching and expression.

[0003] Existing technologies in social networking scenarios generally suffer from problems such as high repetition of greetings, lack of personalization, and obvious traces of AI generation. The messages users receive are often stereotyped, making it difficult to stimulate a desire to reply. In addition, existing systems are unable to dynamically adjust communication strategies based on the other party's real-time emotional state and interest tags, nor can they help users understand the other party's emotional state through visualization, making initial communication often awkward or ineffective. In terms of AI-assisted copy generation, the content generated by existing technologies generally has problems such as a strong mechanical feel, lack of authenticity and emotional warmth, and difficulty in simulating real interpersonal communication scenarios.

[0004] In view of this, this application is filed. Summary of the Invention

[0005] The present invention discloses a method, device, equipment and storage medium for improving the connection rate, aiming to solve the problem that the existing user connection rate is low due to the lack of personalized matching and emotional interaction.

[0006] A first embodiment of the present invention provides a method for improving a connection rate, comprising: Obtaining full-link behavior data of the initiator user and the target user in the target community, inputting the full-link behavior data into a pre-trained feature extraction model to generate a feature vector of the initiator user and a feature vector of the target user; Inputting the feature vector of the initiator user and the feature vector of the target user into a pre-trained label generation model respectively to generate a multi-level user label system for the initiator user and a multi-level user label system for the target user; The tag system of the initiating user and the tag system of the target user are input into the pre-trained tag clustering model to extract and calculate the common points between the two parties and generate a matching feature set; Obtaining historical interaction data of the target user, inputting the historical interaction data and the target user's label system into a pre-trained sentiment analysis model to generate a current emotional state value of the target user; The current emotional state value of the target user is mapped into a pet expression and displayed on the chat interface of the initiating user, and the matching feature set, the current emotional state value of the target user, the initiating user, and the multi-level user label system of the target user are input into a pre-trained copy generation model to generate a greeting copy that is adapted to the target user's emotion, wherein the target user's response is obtained, and the parameters of the label generation model, the sentiment analysis model, and the copy generation model are updated based on the response.

[0007] Preferably, the pre-trained feature extraction model uses an LSTM-BERT fusion architecture, and the specific extraction process is: The LSTM time series model is used to analyze the user behavior sequence and calculate the transition probability matrix P(s) of the behavior path. t+1 |s t ) to identify high-frequency behavior patterns, where s t represents the behavior state at time t; Based on the high-frequency behavior pattern, the frequency of user interaction behavior is counted and the interaction intensity index I=0.4×N is calculated. 评论 +0.3×N 转发 +0.2×N 私信 +0.1×N 打赏 , and introduce the time decay factor γ=e -0.02×Δt Perform dynamic weight adjustment to obtain I weighted =I×γ, where Δt is the time difference from the present; The BERT model is used to process user text content and extract semantic features. The time-weighted interaction intensity features, the behavioral pattern features extracted by LSTM, and the BERT semantic features are concatenated to form a fused feature vector. Principal component analysis (PCA) is used to reduce the dimension of the fused feature vector and output the user's feature vector.

[0008] Preferably, the process of constructing the multi-level user tag system includes: Based on the user feature vector, a rule engine is used to extract basic tags from basic attribute data. The basic tags include age group, region, and device type. For the semantic part of the text in the user feature vector, the TF-IDF algorithm is used to calculate the keyword weight. The top m keywords with the highest weight are input into the LDA topic model, the number of topics is set to m, and the user interest tags are generated. Calculate the weight W of each interest tag 兴趣 , whose expression is W 兴趣 =0.6×content interaction frequency / total number of interactions + 0.4×stay duration / total online time; Obtain user behavior sequences and identify changes in user behavior through the Flink streaming computing framework. When specific behavior patterns are detected, behavior label generation is triggered. Use exponential smoothing method to dynamically update the label weight: W 新 =0.7×W 当日 +0.3×W 历史 , where W 当日 is the weight calculated based on the behavior of the day, W 历史 is the historical cumulative weight; Combine basic tags, interest tags, and behavior tags to form a multi-level tag system.

[0009] Preferably, the pre-trained label clustering model uses a multimodal fusion method to calculate the matching feature set, specifically including: Calculate the behavior statistical weight W based on the initiator tag set and the target tag set behavior , its expression is: W behavior =Σ(behavior intensity i × time decay i) / user activity baseline, where user activity baseline = total number of user behaviors in the past y days / y; Input the interest tags of the initiator and the target into the BERT-wwm model, calculate the semantic similarity between the tags, and obtain the semantic association weight W semantic , whose expression is: W semantic =(cosinesimilarity+1) / 2, where cosinesimilarity is the cosine similarity, ranging from [-1,1]; Construct a user social graph, where nodes are users and edges are social relationships. Edge weights are determined by interaction frequency. GraphSAGE algorithm is used to generate user embedding vectors. Based on the user embedding vectors, the PageRank value of the tag in the social network is calculated to obtain the social communication weight Wsocial. The gated network is used to dynamically fuse the three-way weights to calculate the final matching score: FinalScore = α × W behavior +β×W semantic +γ×W social , where α, β, and γ are learning parameters, and the common points with FinalScore>z are extracted as the matching feature set.

[0010] Preferably, the pre-trained sentiment analysis model analysis process includes: Obtain the target user's most recent b historical interaction records, including published content, comment content, and like records. Use the BERT sentiment classifier to calculate the initial sentiment value sentimentscore for each interaction record, with a value range of [-1, 1]; Set interaction weight based on the interaction type, which includes posting original content, commenting on other people's content, and like weight; Calculate the time decay coefficient for each record: timedecay=e -0.05×Δt , where Δt is the number of hours between the record and the current time. Calculate the weighted sentiment value of a single record: ei=sentimentscore×interactionweight×timedecay Extract personality tags from the target tag set and calculate the personality factor, where the personality factor is used to reflect the user's basic emotional tendency, and comprehensively calculate the current emotional state value of the target user: Estate=0.6×(Σei / 30)+0.4×personalityfactor, normalize Estate to the [0,1] interval as the final emotional state value output.

[0011] Preferably, the process of generating the greeting copy by the pre-trained copy generation model is as follows: Select the tone parameters of the copy based on the current emotional state value Estate, and extract the top q common points with the highest weight from the matching feature set as the core elements of the topic; Build a topic agent system, extract keywords from the core elements of the topic through BERTopic technology, and calculate the topic weight based on real-time hot spots: whot=0.6×(trend24h / max(trend))+0.4×(localheat / avg(heat)) Among them, trend24h is the 24-hour trend value, localheat is the regional heat; Calculate the comprehensive score of each candidate topic: Topicscore = 0.5 × commonweight + 0.3 × emotionmatch + 0.2 × freshness, where commonweight is the commonality weight, emotionmatch is the match between the topic and the emotional state value, and freshness is the freshness of the topic; The topic with the highest Topicscore is selected. Combining the tone parameters of the copy and the tag sets of both parties, a copy generation prompt is constructed. The prompt is input into the pre-trained copy generation model to obtain the greeting copy. The SimHash algorithm is used to detect the similarity between the generated copy and the historical copy. If the similarity is greater than the preset value, the copy is regenerated.

[0012] Preferably, the step of obtaining the target user's response and updating the parameters of the tag generation model, sentiment analysis model, and copywriting generation model based on the response is as follows: Obtaining response data of the target user, calculating a comprehensive response feedback score based on the response data of the target user, and updating a weight fusion parameter of the label generation model using a gradient descent method based on the comprehensive response feedback score; When it is determined that the target user's emotional state value deviates from the predicted emotional state value by more than a preset value, the attention weight matrix of the pre-trained sentiment analysis model is updated; Positive samples and negative samples are selected based on the response rate, and the LoRA parameters of the copywriting generation model are updated according to the positive and negative sample contrast learning losses.

[0013] A second embodiment of the present invention provides a device for improving a connection rate, comprising: A feature vector generation unit is used to obtain the full-link behavior data of the initiator user and the target user in the target community, input the full-link behavior data into the pre-trained feature extraction model, and generate the feature vector of the initiator user and the feature vector of the target user; a label system generating unit, configured to input the feature vector of the initiating user and the feature vector of the target user into a pre-trained label generation model, respectively, to generate a multi-level user label system for the initiating user and a multi-level user label system for the target user; A matching feature set generation unit is used to input the tag system of the initiating user and the tag system of the target user into a pre-trained tag clustering model, extract and calculate the common points between the two parties, and then generate a matching feature set; An emotional state value generating unit is configured to obtain historical interaction data of a target user, input the historical interaction data and the target user's label system into a pre-trained sentiment analysis model, and generate a current emotional state value of the target user; A feedback adjustment unit is used to map the current emotional state value of the target user into a pet expression and display it on the chat interface of the initiating user, input the matching feature set, the current emotional state value of the target user, the initiating user, and the multi-level user label system of the target user into a pre-trained copy generation model, and generate a greeting copy that is adapted to the target user's emotion, wherein the target user's response is obtained, and the parameters of the label generation model, the sentiment analysis model, and the copy generation model are updated based on the response.

[0014] A third embodiment of the present invention provides a device for improving the connection rate, including a memory and a processor, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement a method for improving the connection rate as described in any one of the above items.

[0015] A fourth embodiment of the present invention provides a computer-readable storage medium, characterized in that it stores a computer program, and the computer program can be executed by a processor of the device where the computer-readable storage medium is located to implement a method for improving the connection rate as described in any one of the above items.

[0016] Based on the method, device, equipment and storage medium for improving the connection rate provided by the present invention, a user feature vector is generated by fusing behavioral data and text semantics through a pre-trained feature extraction model, a multi-level user tag system is constructed using a tag generation model, and then the common features of both parties are extracted through a tag clustering model. Subsequently, the target user's historical interaction data and tag system are combined with the sentiment analysis model to obtain their current emotional state, and the emotions are displayed in the form of pet expressions on the initiator's interface. Finally, the matching features, emotional state and tag system are input into the text generation model to generate emotionally adapted personalized greetings, and the parameters of each model are dynamically optimized according to the response, thereby improving the relevance and appeal of the information, and thus improving the success rate of connection. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of a method for improving the connection rate provided by the first embodiment of the present invention; Figure 2 2 is a schematic diagram of a module of a device for improving connection rate provided by the second embodiment of the present invention. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0019] In order to better understand the technical solution of the present invention, the embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0020] The present invention discloses a method, device, equipment and storage medium for improving the connection rate, aiming to solve the problem that the existing user connection rate is low due to the lack of personalized matching and emotional interaction.

[0021] A first embodiment of the present invention provides a method for improving a connection rate. The method may be performed by a device for improving a connection rate (hereinafter referred to as a device or system), and in particular, by one or more processors within the device or system to implement at least the following steps: S101, obtaining full-link behavior data of an initiator user and a target user in a target community, inputting the full-link behavior data into a pre-trained feature extraction model, and generating a feature vector of the initiator user and a feature vector of the target user; In this embodiment, the system first needs to obtain the full-link behavioral data of the initiator user and the target user in the target community. Full-link behavioral data refers to the complete behavioral trajectory of users on the social platform, covering the entire process of users from registration, browsing, interaction to consumption. These data include static information such as basic user attributes such as age, height, education, and dynamic interactive behavior data such as clicks, stay time, comments, forwarding, private messages, rewards and other operation records, as well as semantic information such as text content and message copy posted by users. The system collects these multi-dimensional data in real time through the data acquisition interface and performs pre-processing, including data cleaning, deduplication, format standardization and other operations to ensure the quality of input data.

[0022] After acquiring the raw data, the system feeds this full-link behavioral data into a pre-trained feature extraction model for processing. This feature extraction model utilizes a fusion LSTM-BERT architecture, primarily considering the temporal nature of user behavior and the semantic characteristics of text content, capable of extracting user features from different dimensions. LSTM processes temporally dependent behavioral sequence data, capturing the evolving patterns and long-term dependencies of user behavior, while BERT focuses on understanding the deep semantics of user-generated content. The fusion of these two enables the model to comprehensively characterize user profiles.

[0023] In the specific feature extraction process, the system first uses the LSTM time series model to analyze the user's behavior sequence. The user's behavior on the platform has obvious sequential characteristics. For example, the user may first browse the content, then click on the part of interest, and then comment or forward. This behavior sequence contains the user's interest evolution and decision-making process. The LSTM model can effectively learn this temporal dependency through its unique gating mechanism, including input gate, forget gate and output gate. The model represents the user's behavior sequence as a state sequence and calculates the transition probability matrix P(s t+1 |s t ), which describes the transition from the current behavior state s t Transfer to the next behavioral state s t+1 By analyzing this transition probability matrix, the system can identify users' high-frequency behavior patterns, such as typical behavior links like "browsing videos, liking, and commenting," to reflect users' usage habits and interests.

[0024] Based on the identified high-frequency behavior patterns, the system further counts the frequency of various user interaction behaviors, including core social behaviors such as comments, forwarding, private messages, and rewards. Different types of interactive behaviors reflect different levels of user engagement, so differentiated weight coefficients are designed to calculate the comprehensive interaction intensity index. The specific calculation formula is I=0.4×N 评论 +0.3×N 转发 +0.2×N 私信 +0.1×N 打赏 Among them, commenting behavior has the highest weight of 0.4, because commenting requires users to invest more thinking and time to express their opinions, which best reflects the user's deep participation; forwarding behavior has a weight of 0.3, which represents the user's recognition of the content and willingness to spread it; private messaging has a weight of 0.2, which reflects the user's social initiative; rewarding has a weight of 0.1, although it involves economic expenditure but the frequency is relatively low.

[0025] Taking into account the timeliness of user behavior, that is, recent behavior can better reflect the user's current interest status than historical behavior, the system introduces a time decay factor for dynamic weight adjustment. Time decay factor γ = e -0.02×Δt An exponential decay method is used, where Δt represents the number of days between the time of the behavior and the current time. This results in more recent behaviors receiving higher weights, while the influence of older behaviors gradually decreases. By multiplying the original interaction intensity index I by the time decay factor, we get the time-weighted interaction intensity I. weighted =I×γ, this weighted value more accurately reflects the user's current activity status and interest tendencies.

[0026] When processing user-generated text content, the system uses the BERT model to extract deep semantic features. As a pre-trained language model based on the Transformer architecture, BERT fully understands the context of text through a bidirectional attention mechanism. The system feeds user posts, comments, private messages, and other text data into the BERT model. After processing through a multi-layer Transformer encoder, it ultimately extracts a 768-dimensional semantic feature vector from the CLS marker position. This vector embodies multi-dimensional semantic information, including the theme, sentiment, and style of the user's text content.

[0027] After completing the feature extraction of each dimension, the system will use the time-weighted interaction intensity feature I weighted LSTM-derived behavioral pattern features and BERT-derived semantic features are combined for feature splicing. This multi-source feature fusion strategy integrates user behavior patterns, interaction intensity, and content preferences to form a more comprehensive user representation. The resulting spliced ​​feature vector typically exceeds 800 dimensions. While this high-dimensional feature vector is rich in information, it also carries the risk of computational complexity and overfitting.

[0028] In order to solve the problems caused by high-dimensional features, the system uses principal component analysis (PCA) technology to reduce the dimensionality of the fused feature vector. PCA projects the original high-dimensional features to the direction with the largest variance through linear transformation, reducing the feature dimension while retaining the main information of the data. In this embodiment, the system sets the information volume of more than 95% as the dimensionality reduction standard, and selects the first 200 principal components as the feature representation after dimensionality reduction by calculating the eigenvalues ​​and eigenvectors of the feature covariance matrix. This not only ensures the expressive power of the features, but also controls the computational complexity within a reasonable range, and finally outputs a 200-dimensional user feature vector as the comprehensive feature representation of the user.

[0029] S102: Input the feature vector of the initiating user and the feature vector of the target user into a pre-trained label generation model to generate a multi-level user label system for the initiating user and a multi-level user label system for the target user; In this embodiment, the first step in tag generation is to extract basic tags, which reflect the user's basic attributes and objective characteristics. The system parses the user's feature vector using a rules engine. A rules engine is an inference system based on a predefined set of rules that automatically extracts corresponding tags based on the basic attribute information encoded in the feature vector. The system encodes the user's basic attribute information in the first 50 dimensions of the feature vector. The rules engine extracts age group tags through threshold judgment and interval mapping. For example, when the age feature value falls between 18 and 25, it generates the tags "Post-95s" or "Generation Z," while the tags "Newcomer" or "Youth Group" are generated for the age feature value between 26 and 35. Regional tags are extracted based on user registration information and IP address resolution. The system maintains a regional code table that includes three levels of administrative divisions: province, city, and district. Through table lookup and mapping, precise regional tags such as "X City - X District" are generated. Device type tags are obtained by parsing the user's device information, including operating system type (such as iOS or Android) and fine-grained information such as device model.

[0030] The system needs to mine users' interests and preferences from their content interaction behaviors. The user feature vector contains semantic features extracted by BERT, which encode the topic information of user-generated content and interactive content. The system first decodes these semantic features to restore the text content with which the user has interacted, and then uses the TF-IDF algorithm to calculate the importance weight of each word. The TF-IDF algorithm measures the importance of a word to the user by calculating the product of term frequency and inverse document frequency. Term frequency reflects the frequency of a word in user content, while inverse document frequency reduces the weight of common words and highlights the distinctiveness of feature words. The system sorts the TF-IDF values ​​of all words and selects the top m keywords with the highest weight as the input for topic modeling. In this embodiment, m is set to 20. This number can ensure that the main points of interest of the user are covered without introducing too much noise.

[0031] After selecting keywords, the system inputs them into the Latent Dirichlet Allocation (LDA) topic model for topic discovery. LDA is a probabilistic topic model that assumes each document is a mixture of multiple topics, each of which is a probabilistic distribution of words. The system sets the number of topics, K, to 20. This parameter was determined based on extensive experiments and business experience to achieve an optimal balance between topic granularity and interpretability. The LDA model iteratively updates the topic-word distribution and document-topic distribution using Gibbs sampling or variational inference, ultimately generating a topic distribution vector for each user. Each dimension of the vector represents the user's preference for a particular topic. The system generates interest tags for topics with a probability greater than 0.1 in the topic distribution, such as "Technology and Digital," "Food Exploration," and "Fitness and Sports." These tags intuitively reflect the user's areas of interest.

[0032] In order to more accurately quantify the importance of each interest tag, the system configures a weight calculation formula that comprehensively considers the depth of interaction and time investment. The weight of the interest tag W 兴趣 By formula W 兴趣 =0.6×content interaction frequency / total number of interactions+0.4×stay duration / total online time. The design concept of this formula is that the frequency of user interaction with a certain type of content reflects the breadth of their interest, while the stay duration reflects the depth of interest. The content interaction frequency counts the number of times users like, comment, forward, and other behaviors on the content under this interest category, and normalizes it by dividing it by the total number of interactions to ensure comparability between users with different levels of activity. Stay duration records the user's cumulative browsing time on this type of content, and is also normalized by dividing it by the total online time. The two indicators are assigned weight coefficients of 0.6 and 0.4 respectively. This weight distribution is slightly biased towards the frequency of interaction, because active interaction behavior can better reflect the user's true interest than passive browsing.

[0033] Behavior labels are generated using real-time streaming computing, capturing users' immediate behavioral characteristics and state changes. The system continuously monitors user behavior streams using the Flink streaming computing framework. Multiple behavioral pattern detectors are defined within Flink, each responsible for identifying specific behavioral patterns. For example, if a user views more than 10 livestream rooms within an hour, the "livestream active user" label is generated; if a user has not interacted for three consecutive days, the "low activity" label is generated; and if a user is active between midnight and 3:00 AM, the "night owl" label is generated. These behavioral patterns are defined based on statistical analysis of massive amounts of user behavior data and business experience, effectively identifying user behavioral characteristics and usage habits. Flink combines event-time windows with processing-time windows to ensure the real-time and accurate generation of behavior labels.

[0034] The dynamic update mechanism of tag weights ensures that the tag system can adapt to changes in user interests. User interests and behaviors are not static and may change due to factors such as seasons, hot events, and personal status. Therefore, the system uses exponential smoothing to continuously update tag weights. Update formula W 新 =0.7×W 当日 +0.3×W 历史 This reflects a balanced consideration of recent behavior and historical accumulation. 当日 The weight is calculated in real time based on the user's daily behavior, reflecting the user's latest interest trends. Giving a higher weight of 0.7 ensures the system's sensitivity to changes in user interests. 当日 It is the cumulative value of all previous historical weights, representing the user's long-term stable interests and preferences. The weight of 0.3 ensures the stability of the tag system and avoids drastic changes in tags due to short-term behavioral fluctuations. After completing the generation and weight calculation of various tags, the system organically combines basic tags, interest tags, and behavior tags to build a complete multi-level tag system.

[0035] S103: Input the tag system of the initiating user and the tag system of the target user into a pre-trained tag clustering model, extract and calculate the common points between the two parties, and then generate a matching feature set; The label clustering model first calculates the behavior statistical weight W behavior, this weight reflects the similarity between the two users' behavior patterns. The system extracts behavior-related tags from the tag sets of both parties, including behavioral feature tags such as active time periods, interaction frequency, and content preferences. For each common behavior tag, the system calculates its behavior intensity value. The calculation of behavior intensity takes into account the frequency, duration, and depth of participation of the behavior. For example, if two users have the tag "video creator", the system will count indicators such as the number of videos posted by each user, the number of interactions obtained, and the average length of the video to quantify the intensity of the behavior. At the same time, the system introduces a time decay mechanism to adjust the weights of behaviors in different periods. Recent behaviors receive higher weights, while the influence of behaviors with a long history gradually weakens, so that the matching results can better reflect the user's current status.

[0036] The calculation formula for the behavioral statistical weight is Wbehavior=Σ(behavior intensity i×time decay i) / user active baseline, which fully takes into account the differences in user activity. The numerator is the sum of the weighted intensities of all common behavioral tags, and each behavioral intensity is multiplied by the corresponding time decay factor to ensure timeliness. The user active baseline in the denominator is a normalization factor, which is calculated by dividing the total number of behaviors of the user in the past y days by y. In this embodiment, y is set to 30 days, that is, the average daily number of behaviors of the user in the past month is calculated. This normalization process eliminates the deviation caused by the difference in activity of different users, making the behavioral weights of highly active users and low-active users comparable. For example, an active user who generates 100 behaviors per day and an ordinary user who only generates 10 behaviors per day, through the normalization of the active baseline, their behavioral patterns of the same proportion will obtain similar weight values, avoiding the interference of the activity itself on the matching results.

[0037] The calculation of the semantic relevance weight Wsemantic focuses on the semantic similarity of user interest tags. The system inputs the interest tag text of the initiator and target into the BERT-wwm (Whole Word Masking) model. BERT-wwm is an improved version of the BERT model that uses a whole-word masking strategy for Chinese language processing, which can better understand the complete semantics of Chinese vocabulary. For each pair of interest tags, the model first processes the tag text through a word segmenter, adds special tags [CLS] and [SEP], and then inputs it into the BERT-wwm multi-layer Transformer encoder. After processing through a 12-layer or 24-layer self-attention mechanism and feedforward network, the system extracts a 768-dimensional sentence vector from the [CLS] position as the semantic representation of the tag.

[0038] After obtaining the semantic vectors of the labels, the system calculates the cosine similarity between the two vectors. Cosine similarity measures the degree of similarity by calculating the cosine of the angle between the two vectors. Cosine similarity ranges from -1 to 1, where 1 indicates exact duplicates, 0 indicates orthogonal indifference, and -1 indicates complete opposites. To convert similarity into a more intuitive weight, the system uses a linear transformation using the formula Wsemantic = (cosinesimilarity + 1) / 2, mapping the value range to between 0 and 1. This semantic similarity calculation can identify matches between seemingly dissimilar but semantically related labels. For example, "basketball" and "NBA," while literal differences, have high semantic similarity; "food" and "exploring restaurants" also have a strong correlation in semantic space. Leveraging the deep semantic understanding of the BERT-wwm model, the system can uncover more potential commonalities of interest, improving matching accuracy and coverage.

[0039] Calculating the social communication weight, Wsocial, requires constructing and analyzing the user's social network. The system first builds a large-scale user social graph, a directed, weighted graph structure in which nodes represent users on the platform and edges represent social relationships between users, such as following and friendship. Edge weights are determined by the frequency of interaction between users. The calculation of interaction frequency comprehensively considers various interactive behaviors, including mutual comments, likes, private messages, and mentions. Different types of interaction are assigned different weights. For example, private messages carry a higher weight than simple likes because they represent deeper social connections. The system maintains real-time updates to this social graph, and when users engage in new social behaviors, the corresponding edge weights are dynamically adjusted.

[0040] Based on the constructed social graph, the system uses the GraphSAGE (Graph Sample and Aggregate) algorithm to generate user embedding vectors. GraphSAGE is an inductive graph neural network algorithm that learns node representations by aggregating the features of neighboring nodes. The core idea of ​​the algorithm is to gradually integrate graph structural information and node features into the node embedding representation through multiple layers of neighbor sampling and feature aggregation. In this embodiment, the GraphSAGE model is set as a two-layer structure. The first layer aggregates and compresses the node's 768-dimensional feature vector to 256 dimensions, and the second layer further compresses it to 128 dimensions to generate the final user embedding vector. The aggregation function uses a mean aggregator. For each node, the algorithm first samples its K neighboring nodes (K is set to 10), then calculates the mean of these neighboring node features, concatenates them with the node's own features, and passes them through a fully connected layer and an activation function to obtain the node's new feature representation.

[0041] Based on the user embedding vectors generated by GraphSAGE, the system further calculates the tag's influence within the social network. This algorithm employs the PageRank algorithm, treating tags as signals that propagate within the social network. An iterative calculation derives a global influence score for each tag. The core assumption of PageRank is that a tag possessed by influential users or shared by many users is inherently influential. The system initializes the PageRank value of each tag to 1.0 and then iteratively updates it using the formula PR(tag) = (1-d) / N + d×Σ(PR(user) / C(user)), where d is the damping coefficient (set to 0.85), N is the total number of tags, PR(user) is the PageRank value of the user who owns the tag, and C(user) is the number of tags owned by the user. After 100 iterations, the system converges to a PageRank value, which serves as the tag's social diffusion weight, Wsocial. This weight reflects the tag's potential for diffusion and reach within the social network. Shared tags with high social diffusion weights indicate that two users are likely to be in similar social circles, sharing more common topics and social connections.

[0042] After obtaining the weights for the three dimensions, the system uses a gating network to dynamically combine these weights to generate the final match score. A gating network is a variant of the attention mechanism that adaptively adjusts the importance of different weights based on input features. The network input is a feature vector containing the three weight values, user feature statistics, and scene context. Through a two-layer fully connected network, it outputs three gating coefficients α, β, and γ. These coefficients are softmax-normalized to ensure their sum is 1. During the model training phase, these parameters are supervised and learned using a large number of successful and failed user connection cases. The parameters are updated using the cross-entropy loss function and the Adam optimizer. After training, the model learns the optimal combination strategy for the weights of each dimension in different scenarios. For example, in interest-based social scenarios, the semantic association weight β is relatively high; in acquaintance social scenarios, the social communication weight γ is dominant; and in stranger social scenarios, the behavioral statistics weight α is more important.

[0043] The final match score is calculated using the formula: FinalScore = α × Wbehavior + β × Wsemantic + γ × Wsocial. This score comprehensively reflects the overall degree of match between two users across three dimensions: behavioral patterns, interests, and social connections. The system sets a threshold z (in this example, z = 0.6) and extracts all shared tags with a FinalScore greater than z as a matching feature set. This set includes not only the tags themselves but also metadata such as each tag's match score and key contributing dimensions. The matching feature set is sorted from highest to lowest by score, with the highest-priority shared points being prioritized in subsequent copy generation. For example, if the shared interest "photography" has a match score of 0.85, with the primary score coming from semantic association and behavioral statistics, the system will use this as the core topic element, generating a personalized opening line such as "I see you also like photography. Have you taken any interesting photos recently?"

[0044] S104, obtaining historical interaction data of the target user, inputting the historical interaction data and the target user's label system into a pre-trained sentiment analysis model, and generating a current emotional state value of the target user; The system first extracts the historical interaction data of the target user from the database. This data records the user's various expressions and interactive behaviors on the platform. The acquisition of historical interaction data follows the principle of reverse chronological order, and the most recent interaction records are extracted first, because recent behaviors can better reflect the user's current emotional state. The system is set to obtain the most recent b historical interaction records. In this embodiment, b is set to 30. This number can provide enough samples for sentiment analysis without introducing too old and possibly irrelevant emotional information due to excessive data. These 30 records cover a variety of user interaction types, including original content actively posted by users, such as dynamics, articles, videos, etc., user comments and replies to other people's content, and lightweight interactive behaviors such as likes and favorites. When extracting this data, the system will record detailed information for each record, including metadata such as interaction type, occurrence time, content text, and interaction object.

[0045] The historical interaction data and the target user's label system are input into a pre-trained sentiment analysis model, which uses deep learning technology to understand and quantify users' emotional expressions. The core of the sentiment analysis model is a BERT sentiment classifier trained with large-scale sentiment annotation data. This classifier can understand the emotional tendency and intensity in the text. For each interaction record, the system first extracts its text content. For posted content and comment content, the original text is used directly; for non-text actions such as likes, the system extracts the text of the liked content as the analysis object. The text is preprocessed and input into the BERT model. The preprocessing includes operations such as word segmentation, adding special tags, truncation or padding to a fixed length, etc. The BERT model deeply encodes the text through its 12-layer Transformer encoder, capturing the complex semantic relationships between words and emotional expression patterns.

[0046] The BERT sentiment classifier outputs an initial sentiment score, sentimentscore, for each record. This is a continuous value in the interval [-1, 1], where -1 represents extremely negative emotions such as anger, sadness, and disappointment, 0 represents a neutral or calm emotional state, and 1 represents extremely positive emotions such as happiness, excitement, and contentment. This continuous value design captures the intensity of emotions more nuancedly than simple sentiment classification (positive / negative / neutral). For example, "I'm so happy today!" might receive a high score of 0.9, "Not bad" might receive a slightly positive score of 0.3, and "I'm feeling a little down" might receive a negative score of -0.4. The model is trained using a sentiment corpus containing millions of annotated data points, covering a wide range of emotional expressions and language styles, ensuring accurate and robust sentiment recognition.

[0047] Different types of interactions reflect emotions to varying degrees of reliability, so the system assigns different weights (interactionweight) to each type of interaction. Original content is assigned the highest weight of 1.0 because users actively create and share content, often putting more thought and emotion into it, and the emotions they express are the most authentic and intense. When users take the time to write a paragraph, record a video, or meticulously edit a collection of images, these actions inherently reflect their desire to express and share, and they contain the richest emotional information. Commenting on others' content is assigned a weight of 0.8. While commenting also reflects a user's opinion and emotions, compared to original content, comments are often reactions to others' content and may be influenced by the original content's emotions, making them slightly less reliable. Likes are assigned a weight of 0.5. Likes are lightweight interactions, often expressed out of courtesy, habit, or casual browsing. They may not fully reflect their true emotional state, but a large number of likes, when accumulated, can still reflect a user's emotional tendencies.

[0048] Time factor plays a key role in sentiment analysis. The user's emotional state changes dynamically. The closer the interaction behavior is to the current time, the more it can reflect the user's current mood. The system uses the time decay coefficient timedecay=e -0.05×Δt To quantify this temporal influence, we use the function Δt, where Δt is the number of hours since the interaction was recorded. This exponential decay function is designed so that recent interactions (Δt close to 0) receive a weight close to 1, while the weight decreases exponentially over time. The choice of a decay coefficient of 0.05, validated through extensive experimentation, reduces the weight of interactions from 24 hours ago to approximately 0.3, and the weight of interactions from a week ago to nearly 0. This decay rate ensures the dominance of recent sentiment while not completely ignoring earlier emotional context. For example, a user posting "I'm in a great mood" one hour ago would receive a temporal weight of 0.95, while a similar post posted at the same time yesterday would only receive a weight of 0.3.

[0049] Taking into account sentiment, interaction weight, and time decay, the system calculates a weighted sentiment value (ei) for each record: sentimentcore × interactionweight × timedecay. This comprehensive metric comprehensively reflects the contribution of an interaction record to the user's current emotional state. For example, a newly published original content, "Finally finished the project, feeling great!", has a sentimentcore of 0.8, an interactionweight of 1.0, and a timedecay close to 1, resulting in an ei of approximately 0.8. On the other hand, a like record from three days ago, even if the content has a high sentiment value, may only have an interaction weight of 0.5 and a time decay of 0.22, resulting in an ei of only 0.088, significantly less impacting the current mood.

[0050] A user's emotional state is influenced not only by short-term behavior but also by their long-term personality traits. The system extracts personality-related tags from the target user's tag system. These tags represent stable characteristics derived through long-term behavioral analysis, such as "optimistic and cheerful," "introverted and quiet," and "rational and calm." Based on these personality tags, the system calculates a personality factor, an indicator of the user's basic emotional tendencies. The calculation of the personality factor uses the Big Five personality model from psychology as its theoretical foundation, mapping personality tags to the five dimensions of openness, conscientiousness, extraversion, agreeableness, and neuroticism. A weighted combination then generates a comprehensive emotional tendency score. For example, a user tagged as "extroverted and lively" will likely have a personality factor that is skewed towards the positive, perhaps around 0.6; while a user tagged as "introverted and cautious" may have a personality factor around 0.4. This factor represents the user's baseline emotional level when unaffected by specific events and serves as a relatively stable emotional baseline.

[0051] The final emotional state value is calculated using the formula: Estate = 0.6 × (Σei / 30) + 0.4 × personality factor. The first component of the formula is the average of 30 weighted sentiment values, reflecting the emotional state derived from recent behavior analysis. The weight of 0.6 emphasizes real-time behavior. The second component is the contribution of personality factors. The weight of 0.4 ensures the stability of emotional analysis and avoids bias in emotional judgment caused by individual extreme behaviors. This weighted combination allows emotional analysis to capture user emotional fluctuations while remaining consistent with their underlying personality traits. For example, a user with a generally cheerful personality (personality factor = 0.7) who has recently expressed some negative emotions (average ei = 0.2) will still have a final Estate of 0.6 × 0.2 + 0.4 × 0.7 = 0.4, placing them in a neutral to positive state. This is more reasonable than a conclusion based solely on recent behavior. The calculated Estate value may not initially fall within the standard [0, 1] range; the system normalizes it to this range.

[0052] S105, mapping the current emotional state value of the target user into a pet expression, and displaying it on the chat interface of the initiating user, inputting the matching feature set, the current emotional state value of the target user, the initiating user, and the multi-level user label system of the target user into a pre-trained copy generation model, and generating a greeting copy that is adapted to the target user's emotion, wherein the target user's response is obtained, and the parameters of the label generation model, the sentiment analysis model, and the copy generation model are updated based on the response.

[0053] Visualizing emotional states can significantly enhance the user experience. The system converts abstract emotional values ​​into concrete pet expression animations, allowing users to understand the other person's emotional state at a glance. The system has preset pet images with different personality traits. The lively cat, with its large, lively eyes and raised tail, is well-suited to expressing cheerfulness and extroversion; the gentle rabbit, with its soft ears and gentle expression, excels at conveying peaceful and friendly emotions; the loyal dog, with its warm and friendly nature, displays a tail-wagging frequency and amplitude that accurately reflects the intensity of the emotion; the mysterious fox, with its elegant posture and penetrating eyes, is well-suited to expressing complex and subtle emotions; and the lazy panda, with its adorable demeanor, conveys a relaxed and comfortable atmosphere. Based on the personality tags in the target user's tag system, the system uses a mapping algorithm to select the most matching pet image as their emotional avatar.

[0054] The dynamic mapping of pet expressions follows a sophisticated emotional range division strategy. When the Estate value is greater than 0.7, the system determines that the user is in a positive and joyful emotional state, and the pet displays a series of happy expressions, including a crescent-shaped smile with eyes curved into a crescent, a wide grin with the corners of the mouth raised, and excited jumping or spinning in circles. These animations are not simply looped but dynamically adjusted based on the intensity of the emotion. The closer the Estate is to 1, the higher the amplitude and frequency of the animation. For example, a cat's tail wags more cheerfully, and a rabbit's ears prick up and twitch slightly in excitement. When the Estate is in the middle range of 0.3 to 0.7, the pet displays a calm and gentle expression, including slow and rhythmic blinking, a curious tilt of the head, and subtle movements such as a slight wiggle of the tail or ears. These movements convey a calm and open attitude towards communication. When Estate is lower than 0.3, the pet will show low emotional characteristics, with ears drooping naturally and slightly melancholic eyes. It may curl up, lie down slowly, or sigh softly. This visual presentation can trigger empathy from the initiating user and prompt them to adopt a more gentle and considerate communication method.

[0055] The smoothness of the animation is achieved through high-precision interpolation technology. The system uses a Bezier curve interpolation algorithm to generate smooth transitions between keyframes, ensuring a rendering effect of 60 frames per second. Each expression animation sequence contains 8-12 keyframes, defining the position, angle, and deformation parameters of various parts of the pet. The control points of the Bezier curve are dynamically adjusted according to the intensity of the emotion, so that the easing effect of the animation matches the emotional state. For example, the animation of happiness uses an ease-out curve, with a fast initial speed and a gradual slowing down, reflecting the natural release of emotion; while the animation of melancholy uses an ease-in curve, with a slow start and a slight acceleration, reflecting the suppressed feeling of emotion. The animation is displayed as a floating layer on the initiating user's chat interface, with a duration of 2-3 seconds and looping. This is neither too abrupt nor disruptive to the chat experience, but it also provides a continuous reminder for the user to pay attention to the other person's emotional state.

[0056] The system inputs the matching feature set, the target user's emotional state, and the labeling systems of both parties into a pre-trained copywriting generation model. The first step in copywriting generation is determining the appropriate tone, which directly influences the style and word choice of the greeting. Based on different ranges of Estate values, the system develops differentiated tone strategies. When Estate is greater than 0.7, the system selects a light-hearted and lively tone, with an emotional intensity parameter set to 0.8. This high-intensity emotional expression is conveyed through the use of exclamation points, emojis, and playful phrases, for example, "Wow, I see you're into photography! Have you taken any amazing photos lately?" When Estate is between 0.3 and 0.7, a friendly and gentle tone is adopted, with an emotional intensity parameter of 0.5. The language is more neutral and calm, avoiding overly enthusiastic or cold expressions, such as "Hello, I noticed we both focus on technology. Any recent discoveries you'd like to share?" When the Estate parameter falls below 0.3, the system switches to a caring and considerate tone, with the emotional intensity parameter dropping to 0.3. The language used becomes more gentle and cautious, focusing on emotional support. For example, "I saw your recent post. If you're willing, could you tell me about your recent life? Sometimes talking can make you feel better." Topic selection is a key factor in determining the success of association building. The system extracts the top q common points with the highest weight from the matching feature set as the core elements of the topic. In this embodiment, q is set to 3 to ensure that the topic is both targeted and not too single. The system builds an intelligent topic agent system that integrates a variety of topic discovery and optimization technologies. First, the core elements of the topic are deeply analyzed through BERTopic technology. BERTopic is a topic modeling method based on BERT and clustering algorithms, which can accurately extract topic keywords from short texts. The system converts the common point labels into text descriptions, inputs them into the BERTopic model, and extracts the most representative keyword set through steps such as BERT encoding, UMAP dimensionality reduction, and HDBSCAN clustering.

[0057] To make topics more timely and engaging, the system introduces a real-time hotspot analysis mechanism. The formula for calculating topic weight, whot = 0.6 × (trend24h / max(trend)) + 0.4 × (localheat / avg(heat)), comprehensively considers both overall network popularity and regional relevance. Trend24h, obtained by accessing the hot search APIs of mainstream platforms, reflects the changes in the discussion popularity of a topic over the past 24 hours. The system maintains a real-time hotspot database, which is refreshed hourly, recording metrics such as search volume, discussion volume, and interaction volume for each topic. Localheat, based on the user's geographic location, calculates the popularity of a topic in the user's city or region. This localized hotspot analysis makes topics more relevant to the user's living environment. For example, if both parties are in a certain city and share a passion for food, the system might generate a topic based on local hotspots such as "Newly opened restaurant at Universal Studios in a certain city."

[0058] The final score for each candidate topic is determined through a comprehensive evaluation: Topicscore = 0.5 × commonweight + 0.3 × emotionmatch + 0.2 × freshness. This formula balances multiple considerations. Commonweight is derived directly from the weights in the matching feature set and represents the degree of topic match. Emotionmatch is calculated by calculating the consistency between the topic's emotional tone and the user's current mood. For example, when a user is feeling down, an overly cheerful topic will receive a lower match score. Freshness assesses the topic's freshness, prioritizing recent topics or topics that users have recently engaged with but haven't discussed in depth. The system generates 5-10 candidate topics, and the ones with the highest scores are selected as the core content of the copy.

[0059] The actual copy generation utilizes a large language model fine-tuned with QLoRA. QLoRA (Quantized Low-Rank Adaptation) is a parameter-efficient fine-tuning method that significantly reduces computational resource requirements while maintaining model performance by adding low-rank adapters and quantizing the pre-trained model. The system organizes the selected topic, copy tone parameters, and tag information from both parties into structured prompts. The prompts are designed following the framework of "role setting - background information - task requirements - output constraints." For example, a prompt might read: "You are a friendly social assistant. User A (labels: photography enthusiast, travel expert) would like to connect with User B (labels: photographer, landscape enthusiast). User B is currently in a positive mood. Please generate a 50-100-word greeting, focusing on 'recent travel photography experiences,' in a relaxed, friendly, and natural tone." After the model generates the initial copy based on this prompt, the system undergoes multiple rounds of optimization, including length adjustment, sensitive word filtering, and the incorporation of personalized elements.

[0060] To ensure the diversity and novelty of the copy, the system uses the SimHash algorithm to detect the similarity between the newly generated copy and the historical copy library. SimHash is a locally sensitive hashing algorithm that can quickly calculate the fingerprint of the text and evaluate the similarity. The system divides the copy into words and calculates the hash value of each word. The SimHash fingerprint of the copy is obtained by weighted summation, and then the Hamming distance between the new copy and the historical copy fingerprint is calculated. If the similarity exceeds the preset threshold of 0.7, the system will adjust the generation parameters such as temperature coefficient, top-p value, etc., and regenerate the copy, trying up to 5 times. This mechanism effectively prevents users from receiving stereotyped greetings, and each connection can provide a unique communication experience.

[0061] The system's intelligence is reflected in its ability to continuously learn and optimize. When the target user responds to a greeting, the system records the response data in detail, including whether a reply was received, the time of response, the length of the response, the sentiment, the number of subsequent conversation turns, and other multi-dimensional information. Based on this response data, the system calculates a comprehensive response feedback score R, which takes into account both immediate feedback and long-term effects. Immediate feedback primarily considers whether a reply was received and the speed of the response. Quick replies within 5 minutes receive the highest score, and the score decreases as time passes. Long-term effects assess the continuity and depth of the conversation, including indicators such as the number of conversation turns, average reply length, and emotional positivity.

[0062] Based on the response feedback scores, the system performs targeted parameter updates on the three core models. For the label clustering model, the system uses the gradient descent method to update the weight fusion parameters α, β, and γ, with the learning rate set to 0.01. By calculating the partial derivatives of the loss function with respect to each parameter, the parameter values ​​are adjusted in the direction that can improve matching accuracy. When the weight contribution of a certain dimension in a successful association case is particularly prominent, the system will appropriately increase the weight coefficient of that dimension. For example, if the association success rate through interest similarity matching is found to be particularly high, the semantic association weight β will gradually increase.

[0063] The update to the sentiment analysis model primarily aims to improve prediction accuracy. The system determines the user's true emotional state by analyzing the actual sentiment of the reply. When this actual value deviates by more than 0.3 from the previously predicted "Estate" value, a model update is triggered. This update primarily adjusts the attention weight matrix in the BERT model. Backpropagation is used to calculate the gradient of the prediction error with respect to the attention parameter, and the Adam optimizer is used to update the parameters. This update enables the model to better understand the emotional expressions of specific user groups, improving the accuracy of subsequent sentiment analysis.

[0064] The copy generation model is optimized using a contrastive learning strategy. The system labels copies with a response rate above 0.7 as positive samples and those below 0.3 as negative samples, constructing a training set for contrastive learning. By calculating the difference in representation between positive and negative samples in semantic space, the model's LoRA parameters are optimized, making the copy generated by the model more similar to the characteristics of positive samples. LoRA parameters are updated by minimizing the contrastive loss function, which is designed to maximize the similarity of positive sample pairs and minimize the similarity of negative sample pairs. The temperature parameter τ is set to 0.07 to control the difficulty of sample differentiation. A batch update is triggered every 100 accumulated response records. The updated model is verified through A / B testing. Full deployment is only carried out when the success rate of connection establishment exceeds 5%, ensuring steady improvement in system performance.

[0065] See also Figure 2 A second embodiment of the present invention provides a device for improving the connection rate, including: A feature vector generation unit 201 is used to obtain the full-link behavior data of the initiator user and the target user in the target community, input the full-link behavior data into the pre-trained feature extraction model, and generate the feature vector of the initiator user and the feature vector of the target user; The label system generating unit 202 is configured to input the feature vector of the initiating user and the feature vector of the target user into a pre-trained label generation model to generate a multi-level user label system for the initiating user and a multi-level user label system for the target user; The matching feature set generation unit 203 is used to input the tag system of the initiating user and the tag system of the target user into a pre-trained tag clustering model, extract and calculate the common points between the two parties, and then generate a matching feature set; The emotional state value generating unit 204 is configured to obtain historical interaction data of the target user, input the historical interaction data and the target user's label system into a pre-trained sentiment analysis model, and generate a current emotional state value of the target user; The feedback adjustment unit 205 is used to map the current emotional state value of the target user into a pet expression and display it on the chat interface of the initiating user, input the matching feature set, the current emotional state value of the target user, the initiating user, and the multi-level user label system of the target user into the pre-trained copy generation model, and generate a greeting text that is adapted to the target user's emotions, wherein the target user's response is obtained and the parameters of the label generation model, the sentiment analysis model, and the copy generation model are updated based on the response.

[0066] A third embodiment of the present invention provides a device for improving the connection rate, including a memory and a processor, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement a method for improving the connection rate as described in any one of the above items.

[0067] A fourth embodiment of the present invention provides a computer-readable storage medium, characterized in that it stores a computer program, and the computer program can be executed by a processor of the device where the computer-readable storage medium is located to implement a method for improving the connection rate as described in any one of the above items.

[0068] Based on the method, device, equipment and storage medium for improving the connection rate provided by the present invention, a pre-trained feature extraction model is used to fuse behavioral data and text semantics to generate a user feature vector, a label generation model is used to build a multi-level user label system, and then a label clustering model is used to extract common features between the two parties. Subsequently, the target user's historical interaction data and label system are combined with a sentiment analysis model to obtain their current emotional state, and the emotion is displayed in the form of a pet emoji on the initiator's interface. Finally, the matching features, emotional state and label system are input into the copy generation model to generate emotionally adapted personalized greetings. The model parameters are dynamically optimized based on the response, thereby improving the relevance and appeal of the information, and thus increasing the success rate of connection.

[0069] Exemplarily, the computer program described in the third and fourth embodiments of the present invention can be divided into one or more modules, and the one or more modules are stored in the memory and executed by the processor to complete the present invention. The one or more modules can be a series of computer program instruction segments that can perform specific functions, and the instruction segments are used to describe the execution process of the computer program in the device for implementing a method for improving the connection rate. For example, the device described in the second embodiment of the present invention.

[0070] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the method for improving the connection rate, and utilizes various interfaces and lines to connect the various parts of the method for improving the connection rate.

[0071] The memory can be used to store the computer programs and / or modules. The processor implements the various functions of a method for improving connection rates by running or executing the computer programs and / or modules stored in the memory and calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store an operating system and at least one application required for a function (such as a sound playback function, a text conversion function, etc.); the data storage area may store data generated based on the use of the mobile phone (such as audio data, text message data, etc.). In addition, the memory may include high-speed random access memory and non-volatile memory, such as a hard disk, internal memory, a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0072] If the implementation module is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention can also implement all or part of the process steps in the above-mentioned method embodiments by using a computer program to instruct the relevant hardware. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of each of the above-mentioned method embodiments. The computer program includes computer program code, which can be in source code form, object code form, executable file, or some intermediate form. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal, and software distribution medium. It should be noted that the content of the computer-readable medium can be appropriately increased or decreased based on the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media does not include electric carrier signals and telecommunication signals.

[0073] It should be noted that the device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed across multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment. In addition, in the drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. A person of ordinary skill in the art can understand and implement the present invention without inventive effort.

[0074] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for improving the connection rate, characterized in that: include: Obtaining full-link behavior data of the initiator user and the target user in the target community, inputting the full-link behavior data into a pre-trained feature extraction model to generate a feature vector of the initiator user and a feature vector of the target user; Inputting the feature vector of the initiator user and the feature vector of the target user into a pre-trained label generation model respectively to generate a multi-level user label system for the initiator user and a multi-level user label system for the target user; The tag system of the initiating user and the tag system of the target user are input into the pre-trained tag clustering model to extract and calculate the common points between the two parties and generate a matching feature set; Obtaining historical interaction data of the target user, inputting the historical interaction data and the target user's label system into a pre-trained sentiment analysis model to generate a current emotional state value of the target user; The current emotional state value of the target user is mapped into a pet expression and displayed on the chat interface of the initiating user, and the matching feature set, the current emotional state value of the target user, the initiating user, and the multi-level user label system of the target user are input into a pre-trained copy generation model to generate a greeting copy that is adapted to the target user's emotion, wherein the target user's response is obtained, and the parameters of the label generation model, the sentiment analysis model, and the copy generation model are updated based on the response.

2. A method for improving the connection rate according to claim 1, characterized in that: The pre-trained feature extraction model uses the LSTM-BERT fusion architecture. The specific extraction process is as follows: The LSTM time series model is used to analyze the user behavior sequence and calculate the transition probability matrix P(s) of the behavior path. t+1 |s t ) to identify high-frequency behavior patterns, where s t represents the behavior state at time t; Based on the high-frequency behavior pattern, the frequency of user interaction behavior is counted and the interaction intensity index I=0.4×N is calculated. 评论 +0.3×N 转发 +0.2×N 私信 +0.1×N 打赏 , and introduce the time decay factor γ=e -0.02×Δt Perform dynamic weight adjustment to obtain I weighted =I×γ, where Δt is the time difference from the present; The BERT model is used to process user text content and extract semantic features. The time-weighted interaction intensity features, the behavioral pattern features extracted by LSTM, and the BERT semantic features are concatenated to form a fused feature vector. Principal component analysis (PCA) is used to reduce the dimension of the fused feature vector and output the user's feature vector.

3. A method for improving the connection rate according to claim 2, characterized in that: The process of constructing the multi-level user tag system includes: Based on the user feature vector, a rule engine is used to extract basic tags from basic attribute data. The basic tags include age group, region, and device type. For the semantic part of the text in the user feature vector, the TF-IDF algorithm is used to calculate the keyword weight. The top m keywords with the highest weight are input into the LDA topic model, the number of topics is set to m, and the user interest tags are generated. Calculate the weight W of each interest tag 兴趣 , whose expression is W 兴趣 =0.6×content interaction frequency / total number of interactions + 0.4×stay duration / total online time; Obtain user behavior sequences and identify changes in user behavior through the Flink streaming computing framework. When specific behavior patterns are detected, behavior label generation is triggered. Use exponential smoothing method to dynamically update the label weight: W 新 =0.7×W 当日 +0.3×W 历史 , where W 当日 is the weight calculated based on the behavior of the day, W 历史 is the historical cumulative weight; Combine basic tags, interest tags, and behavior tags to form a multi-level tag system.

4. The method for improving the connection rate according to claim 3, characterized in that: The pre-trained label clustering model uses a multimodal fusion method to calculate the matching feature set, specifically including: Calculate the behavior statistical weight W based on the initiator tag set and the target tag set behavior , its expression is: W behavior =Σ(behavior intensity i × time decay i) / user activity baseline, where user activity baseline = total number of user behaviors in the past y days / y; Input the interest tags of the initiator and the target into the BERT-wwm model, calculate the semantic similarity between the tags, and obtain the semantic association weight W semantic , whose expression is: W semantic =(cosinesimilarity+1) / 2, where cosinesimilarity is the cosine similarity, ranging from [-1,1]; Construct a user social graph, where nodes are users and edges are social relationships. Edge weights are determined by interaction frequency. GraphSAGE algorithm is used to generate user embedding vectors. Based on the user embedding vectors, the PageRank value of the tag in the social network is calculated to obtain the social communication weight Wsocial. The gated network is used to dynamically fuse the three-way weights to calculate the final matching score: FinalScore = α × W behavior +β×W semantic +γ×W social , where α, β, and γ are learning parameters, and the common points with FinalScore>z are extracted as the matching feature set.

5. The method for improving the connection rate according to claim 1, characterized in that: The analysis process of the pre-trained sentiment analysis model includes: Obtain the target user's most recent b historical interaction records, including published content, comment content, and like records. Use the BERT sentiment classifier to calculate the initial sentiment value sentimentscore for each interaction record, with a value range of [-1, 1]; Set interaction weight based on the interaction type, which includes posting original content, commenting on other people's content, and like weight; Calculate the time decay coefficient for each record: timedecay=e -0.05×Δt , where Δt is the number of hours between the record and the current time. Calculate the weighted sentiment value of a single record: ei=sentimentscore×interactionweight×timedecay Extract personality tags from the target tag set and calculate the personality factor, where the personality factor is used to reflect the user's basic emotional tendency, and comprehensively calculate the current emotional state value of the target user: Estate=0.6×(Σei / 30)+0.4×personalityfactor, normalize Estate to the [0,1] interval as the final emotional state value output.

6. A method for improving the connection rate according to claim 5, characterized in that: The process of generating the greeting copy by the pre-trained copy generation model is as follows: Select the tone parameters of the copy based on the current emotional state value Estate, and extract the top q common points with the highest weight from the matching feature set as the core elements of the topic; Build a topic agent system, extract keywords from the core elements of the topic through BERTopic technology, and calculate the topic weight based on real-time hot spots: whot=0.6×(trend24h / max(trend))+0.4×(localheat / avg(heat)) Among them, trend24h is the 24-hour trend value, localheat is the regional heat; Calculate the comprehensive score of each candidate topic: Topicscore = 0.5 × commonweight + 0.3 × emotionmatch + 0.2 × freshness, where commonweight is the commonality weight, emotionmatch is the match between the topic and the emotional state value, and freshness is the freshness of the topic; The topic with the highest Topicscore is selected. Combining the tone parameters of the copy and the tag sets of both parties, a copy generation prompt is constructed. The prompt is input into the pre-trained copy generation model to obtain the greeting copy. The SimHash algorithm is used to detect the similarity between the generated copy and the historical copy. If the similarity is greater than the preset value, the copy is regenerated.

7. The method for improving the connection rate according to claim 6, characterized in that: The step of obtaining the target user's response and updating the parameters of the tag generation model, sentiment analysis model, and copywriting generation model based on the response is as follows: Obtaining response data of the target user, calculating a comprehensive response feedback score based on the response data of the target user, and updating a weight fusion parameter of the label generation model using a gradient descent method based on the comprehensive response feedback score; When it is determined that the target user's emotional state value deviates from the predicted emotional state value by more than a preset value, the attention weight matrix of the pre-trained sentiment analysis model is updated; Positive samples and negative samples are selected based on the response rate, and the LoRA parameters of the copywriting generation model are updated according to the positive and negative sample contrast learning losses.

8. A device for improving the connection rate, characterized in that: include: A feature vector generation unit is used to obtain the full-link behavior data of the initiator user and the target user in the target community, input the full-link behavior data into the pre-trained feature extraction model, and generate the feature vector of the initiator user and the feature vector of the target user; a label system generating unit, configured to input the feature vector of the initiating user and the feature vector of the target user into a pre-trained label generation model, respectively, to generate a multi-level user label system for the initiating user and a multi-level user label system for the target user; A matching feature set generation unit is used to input the tag system of the initiating user and the tag system of the target user into a pre-trained tag clustering model, extract and calculate the common points between the two parties, and then generate a matching feature set; An emotional state value generating unit is configured to obtain historical interaction data of a target user, input the historical interaction data and the target user's label system into a pre-trained sentiment analysis model, and generate a current emotional state value of the target user; A feedback adjustment unit is used to map the current emotional state value of the target user into a pet expression and display it on the chat interface of the initiating user, input the matching feature set, the current emotional state value of the target user, the initiating user, and the multi-level user label system of the target user into a pre-trained copy generation model, and generate a greeting copy that is adapted to the target user's emotion, wherein the target user's response is obtained, and the parameters of the label generation model, the sentiment analysis model, and the copy generation model are updated based on the response.

9. A device for improving the connection rate, characterized in that: It includes a memory and a processor, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement a method for improving the connection rate as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that A computer program is stored, and the computer program can be executed by a processor of the device where the computer-readable storage medium is located to implement a method for improving the connection rate as described in any one of claims 1 to 7.

Images