A method, apparatus, device, and storage medium for improving connection establishment rate.

Abstract

This invention provides a method, apparatus, device, and storage medium for improving the connection establishment rate. It generates user feature vectors by fusing behavioral data and textual semantics through a pre-trained feature extraction model, constructs a multi-level user tag system using a tag generation model, and then extracts common features between the two parties using a tag clustering model. Subsequently, combining the target user's historical interaction data and tag system, a sentiment analysis model is used to obtain their current emotional state, which is then displayed as a pet emoji on the initiator's interface. Finally, the matching features, emotional state, and tag system are input into a copywriting generation model to generate a personalized greeting with appropriate emotion. The model parameters are dynamically optimized based on the response, thereby improving the relevance and persuasiveness of the information and ultimately increasing the connection establishment success rate.

Description

Technical Field

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

[0002] In today's social media platforms and instant messaging applications, the initial connection between users is a crucial step in social interaction, and the connection rate directly impacts the platform's user activity and retention rates. Traditional methods of establishing social connections mainly rely on users manually crafting greetings or using templated greetings provided by the platform, which have many limitations. When connecting with strangers, users often struggle to accurately grasp the other person's emotional state and interests, resulting in greetings that lack relevance and appeal, easily being ignored or rejected. Furthermore, while existing social platforms collect vast amounts of user behavior data, they haven't effectively utilized this data to optimize the connection process, particularly in terms of emotional matching and expression.

[0003] Existing technologies in social networking scenarios generally suffer from issues such as high repetition of greetings, lack of personalization, and obvious AI-generated traces. Users often receive monotonous messages, making it difficult to stimulate a desire to respond. Furthermore, existing systems cannot 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, often leading to awkward or ineffective initial interactions. Regarding AI-assisted copywriting, existing technologies generally produce content that feels mechanical, lacks authenticity and emotional warmth, and fails to simulate real interpersonal communication scenarios.

[0004] In view of the above, this application is hereby submitted. Summary of the Invention

[0005] This invention discloses a method, apparatus, device, and storage medium for improving the connection establishment rate, aiming to solve the problem of low connection establishment rate caused by the lack of personalized matching and emotional interaction in the existing user connection process.

[0006] The first embodiment of the present invention provides a method for improving the establishment rate, comprising:

[0007] Acquire the full-link behavior data of the initiating user and the target user in the target community, and input the full-link behavior data into a pre-trained feature extraction model to generate feature vectors of the initiating user and the target user.

[0008] The feature vectors of the initiating user and the target user are respectively input 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.

[0009] The tag system of the initiating user and the tag system of the target user are input into a pre-trained tag clustering model to extract and calculate the commonalities between the two and generate a matching feature set.

[0010] The historical interaction data of the target user is obtained, and the historical interaction data and the target user's tag system are input into a pre-trained sentiment analysis model to generate the target user's current sentiment state value.

[0011] The target user's current emotional state value is mapped to a pet emoji and displayed on the initiating user's chat interface. The matching feature set, the target user's current emotional state value, the initiating user, and the target user's multi-level user tag system are input into a pre-trained copywriting generation model to generate a greeting copywriting that is appropriate for the target user's emotions. The target user's response is obtained, and the parameters of the tag generation model, sentiment analysis model, and copywriting generation model are updated based on the response.

[0012] Preferably, the pre-trained feature extraction model is an LSTM-BERT fusion architecture, and the specific extraction process is as follows:

[0013] The LSTM time series model is used to analyze user behavior sequences and calculate the transition probability matrix P(s) of the behavior path. t+1 |s t To identify high-frequency behavioral patterns, where s t This represents the behavioral state at time t;

[0014] Based on the frequency of user interaction behaviors observed in the aforementioned high-frequency behavior patterns, 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 a 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;

[0015] The BERT model is used to process user text content and extract semantic features. The time-weighted interaction intensity features, the behavior 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 dimensionality of the fused feature vector and output the user's feature vector.

[0016] Preferably, the construction process of the multi-level user tagging system includes:

[0017] Based on user feature vectors, basic tags are extracted from basic attribute data through a rule engine. These basic tags include age group, region, and device type.

[0018] For the text semantic part in the user feature vector, the TF-IDF algorithm is used to calculate the keyword weights. The top m keywords with the highest weights are input into the LDA topic model, the number of topics is set to m, and user interest tags are generated.

[0019] Calculate the weight W for each interest tag. 兴趣 Its expression is W 兴趣 =0.6 × Content interaction frequency / Total number of interactions + 0.4 × Dwell time / Total online time;

[0020] The system acquires user behavior sequences and identifies changes in user behavior using the Flink streaming computing framework. When a specific behavior pattern is detected, behavior label generation is triggered.

[0021] The label weights are dynamically updated using exponential smoothing: W 新 =0.7×W 当日 +0.3×W 历史 W 当日 W is the weight calculated based on the behavior of the day. 历史 Historical cumulative weighting;

[0022] The basic tags, interest tags, and behavior tags are combined to form a multi-level tag system.

[0023] Preferably, the pre-trained label clustering model uses a multimodal fusion method to calculate the matching feature set, specifically including:

[0024] Based on the initiator's tag set and the target's tag set, calculate the behavioral statistical weight W. 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;

[0025] The interest tags of the initiator and the target are input into the BERT-wwm model to calculate the semantic similarity between the tags and obtain the semantic association weight W. semantic Its expression is:

[0026] W semantic =(cosinesimilarity+1) / 2, where cosinesimilarity is the cosine similarity, with a value range of [-1,1].

[0027] Construct a user social graph where nodes represent users and edges represent social relationships. Edge weights are determined by interaction frequency. The GraphSAGE algorithm is used to generate user embedding vectors. Based on these vectors, the PageRank value of each tag in the social network is calculated to obtain the social propagation weight Wsocial. A gating network is used to dynamically fuse the three weights, and the final matching score is calculated: FinalScore = α × W behavior +β×W semantic +γ×W social , where α, β, and γ are learning parameters, and the common points of FinalScore>z are extracted as the matching feature set.

[0028] Preferably, the analysis process of the pre-trained sentiment analysis model includes:

[0029] Obtain the most recent b historical interaction records of the target user, including posted content, comments, and likes. For each interaction record, use the BERT sentiment classifier to calculate the initial sentiment value sentimentscore, which ranges from [-1, 1].

[0030] Set interaction weights based on interaction type, which includes weights for publishing original content, commenting on others' content, and liking.

[0031] Calculate the time decay factor for each record: timedecay=e -0.05×Δt Where Δt is the number of hours since the current time, the weighted sentiment score for a single record is calculated as: ei = sentimentscore × interactionweight × timedecay

[0032] Personality tags are extracted from the target user's tag set, and a personality factor is calculated. This personality factor reflects the user's basic emotional tendency. The current emotional state value of the target user is then comprehensively calculated.

[0033] Estate = 0.6 × (Σei / 30) + 0.4 × personality factor, normalize Estate to the interval [0,1] as the final emotional state value output.

[0034] Preferably, the process by which the pre-trained copywriting generation model generates greeting copy is as follows:

[0035] Based on the current emotional state value Estate, select the tone parameter of the copywriting, and extract the top q common points with the highest weight from the matching feature set as the core elements of the topic;

[0036] Build a topic agent system that extracts keywords from core topic elements using BERTopic technology and calculates topic weight by combining real-time trending topics.

[0037] whot=0.6×(trend24h / max(trend))+0.4×(localheat / avg(heat))

[0038] Among them, trend24h is the 24-hour trend value, and localheat is the regional heat.

[0039] Calculate the overall score for each candidate topic: Topicscore = 0.5 × commonweight + 0.3 × emotionmatch + 0.2 × freshness, where commonweight is the weight of common points, emotionmatch is the degree of matching between the topic and the emotional state value, and freshness is the freshness of the topic;

[0040] The topic with the highest Topicscore is selected, and combined with the copywriting tone parameters and the tag sets of both parties, copywriting generation prompts are constructed. The prompts are then input into a pre-trained copywriting generation model to obtain 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 a preset value, the copy is regenerated.

[0041] 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 specifically involves:

[0042] Obtain the response data of the target user, calculate the comprehensive response feedback score based on the response data of the target user, and update the weight fusion parameters of the label generation model based on the comprehensive response feedback score using the gradient descent method;

[0043] When it is determined that the deviation between the target user's emotional state value and the predicted emotional state value is greater than a preset value, the attention weight matrix of the pre-trained sentiment analysis model is updated.

[0044] Positive and negative samples are selected based on the response rate, and the LoRA parameters of the copywriting generation model are updated based on the learning loss of the positive and negative samples.

[0045] A second embodiment of the present invention provides an apparatus for improving the establishment rate, comprising:

[0046] The feature vector generation unit is used to acquire the full-link behavior data of the initiating user and the target user in the target community, and input the full-link behavior data into the pre-trained feature extraction model to generate the feature vectors of the initiating user and the target user.

[0047] The tag system generation unit is used to input the feature vectors of the initiating user and the target user into the pre-trained tag generation model to generate a multi-level user tag system for the initiating user and a multi-level user tag system for the target user.

[0048] The 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 the pre-trained tag clustering model, extract and calculate the common points between the two parties, and then generate a matching feature set.

[0049] The emotion state value generation unit is used to acquire the target user's historical interaction data, input the historical interaction data and the target user's tag system into a pre-trained emotion analysis model, and generate the target user's current emotion state value.

[0050] The feedback adjustment unit is used to map the current emotional state value of the target user into a pet emoji and display it on the chat interface of the initiating user. The matching feature set, the current emotional state value of the target user, the initiating user, and the multi-level user tag system of the target user are input into a pre-trained copywriting generation model to generate a greeting copywriting that is adapted to the target user's emotion. The unit also obtains the target user's response and updates the parameters of the tag generation model, the sentiment analysis model, and the copywriting generation model based on the response.

[0051] The third embodiment of the present invention provides a device for improving the connection establishment rate, including a memory and a processor. The memory stores a computer program, which can be executed by the processor to implement a method for improving the connection establishment rate as described in any of the above embodiments.

[0052] The fourth embodiment of the present invention provides a computer-readable storage medium, characterized in that it stores a computer program, which can be executed by a processor of the device in which the computer-readable storage medium is located, to implement a method for improving connection rate as described in any of the above claims.

[0053] Based on the method, apparatus, device, and storage medium provided by this invention for improving the connection establishment rate, a pre-trained feature extraction model is used to fuse behavioral data and text semantics to generate user feature vectors. A multi-level user tag system is constructed using a tag generation model, and then a tag clustering model is used to extract common features between the two parties. Subsequently, combining the target user's historical interaction data and tag system, a sentiment analysis model is used to obtain their current emotional state, which is then displayed on the initiator's interface in the form of a pet emoji. Finally, the matching features, emotional state, and tag system are input into a copywriting generation model to generate a personalized greeting with appropriate emotion. The model parameters are dynamically optimized based on the response, thereby improving the relevance and persuasiveness of the information and thus increasing the connection establishment success rate. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating a method for improving the connection rate according to the first embodiment of the present invention;

[0055] Figure 2 This is a schematic diagram of a module of a device for improving the establishment rate provided in the second embodiment of the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] To better understand the technical solution of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0058] This invention discloses a method, apparatus, device, and storage medium for improving the connection establishment rate, aiming to solve the problem of low connection establishment rate caused by the lack of personalized matching and emotional interaction in the existing user connection process.

[0059] The first embodiment of the present invention provides a method for improving the connection establishment rate, which can be executed by a device (hereinafter referred to as the device or system) for improving the connection establishment rate, and in particular, by one or more processors within the device or system, to at least implement the following steps:

[0060] S101, Obtain the full-link behavior data of the initiating user and the target user in the target community, and input the full-link behavior data into the pre-trained feature extraction model to generate the feature vector of the initiating user and the feature vector of the target user.

[0061] In this embodiment, the system first needs to acquire the full-link behavioral data of both the initiating and target users in the target community. Full-link behavioral data refers to the complete behavioral trajectory of users on the social platform, covering the entire process from registration, browsing, interaction to consumption. This data includes basic user attributes such as age, height, and education level (static information), as well as dynamic interactive behavior data such as clicks, dwell time, comments, reposts, private messages, and tipping records. It also includes semantic information such as text content and message copy posted by users. The system collects this multi-dimensional data in real time through a data acquisition interface and performs preprocessing, including data cleaning, deduplication, and format standardization, to ensure the quality of the input data.

[0062] After acquiring the raw data, the system inputs this end-to-end behavioral data into a pre-trained feature extraction model for processing. This feature extraction model employs an LSTM-BERT fusion architecture, primarily considering the temporal characteristics of user behavior and the semantic characteristics of text content, enabling the extraction of user features from different dimensions. LSTM handles temporally dependent behavioral sequence data, capturing the evolutionary patterns and long-term dependencies of user behavior, while BERT focuses on understanding the deep semantic information of user-generated content. The fusion of the two allows the model to comprehensively characterize the user profile.

[0063] In the specific feature extraction process, the system first uses an LSTM time-series model to analyze user behavior sequences. User behavior on the platform exhibits clear sequential characteristics; for example, a user might first browse content, then click on a section of interest, and subsequently comment or share. This behavioral sequence encapsulates the user's interest evolution and decision-making process. The LSTM model, through its unique gating mechanism, including input gates, forget gates, and output gates, can effectively learn these temporal dependencies. The model represents the user's behavior sequence as a state sequence and calculates the transition probability matrix P(s) of the behavior path. t+1 |s t This matrix describes the behavior from the current state s. t Transition to the next behavioral state t+1 The probability distribution is determined by analyzing this transition probability matrix. The system can identify high-frequency user behavior patterns, such as typical behavioral chains like "browsing videos, liking, and commenting," to reflect user habits and interests.

[0064] Based on the identified high-frequency behavioral patterns, the system further analyzes the frequency of various user interactions, including core social behaviors such as commenting, forwarding, private messaging, and tipping. Different types of interactions reflect different levels of user engagement; therefore, differentiated weighting coefficients are designed to calculate a 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 received the highest weight of 0.4, because commenting requires users to invest more thought and time in expressing their opinions, and best reflects the user's deep participation; forwarding behavior had a weight of 0.3, representing the user's recognition of the content and willingness to spread it; private messaging had a weight of 0.2, reflecting the user's social initiative; and tipping had a weight of 0.1, although it involves economic expenditure, the frequency is relatively low.

[0065] Considering the time-sensitive nature of user behavior—that recent behavior reflects a user's current interest state better than historical behavior—the system introduces a time decay factor for dynamic weight adjustment. The time decay factor γ = e -0.02×Δt An exponential decay model is adopted, where Δt represents the number of days between the occurrence of the behavior and the current time. This gives higher weight to more recent behaviors, while the influence of older behaviors gradually decreases. The time-weighted interaction strength I is obtained by multiplying the original interaction strength index I by the time decay factor. weighted =I×γ, this weighted value more accurately reflects the user's current activity level and interests.

[0066] In processing user-generated text content, the system employs the BERT model to extract deep semantic features. As a pre-trained language model based on the Transformer architecture, BERT, through its bidirectional attention mechanism, fully understands the contextual relationships within the text. The system inputs text data such as user-posted updates, comments, and private messages into the BERT model. After processing by multiple layers of Transformer encoders, a 768-dimensional semantic feature vector is finally extracted from the CLS marker positions. This vector encapsulates multi-dimensional semantic information about the user's text content, including its theme, sentiment, and style.

[0067] After extracting features from each dimension, the system will use the time-weighted interaction intensity feature I. weighted The behavioral pattern features extracted by LSTM and the semantic features extracted by BERT are concatenated. This multi-source feature fusion strategy can integrate user behavior patterns, interaction intensity, and content preferences to form a more comprehensive user representation. The concatenated feature vector typically has a dimension of over 800. While such high-dimensional features are information-rich, they also bring computational complexity and the risk of overfitting.

[0068] To address the challenges posed by high-dimensional features, the system employs Principal Component Analysis (PCA) to reduce the dimensionality of the fused feature vectors. PCA projects the original high-dimensional features onto the direction of maximum variance through a linear transformation, reducing feature dimensionality while preserving key data information. In this embodiment, the system sets a dimensionality reduction criterion of retaining over 95% of the information. By calculating the eigenvalues ​​and eigenvectors of the feature covariance matrix, the top 200 principal components are selected as the feature representation after dimensionality reduction. This ensures both the expressive power of the features and keeps computational complexity within a reasonable range, ultimately outputting a 200-dimensional user feature vector as the comprehensive feature representation for that user.

[0069] S102, input the feature vector of the initiating user and the feature vector of the target user into the 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.

[0070] In this embodiment, the first step in tag generation is to extract basic tags, which reflect the user's fundamental attributes and objective characteristics. The system parses the user's feature vector using a rule engine, a reasoning system based on a predefined rule set that automatically extracts corresponding tags based on the basic attribute information encoded in the feature vector. In the first 50 dimensions of the feature vector, the system encodes the user's basic attribute information. The rule engine extracts age group tags through threshold judgment and interval mapping. For example, when the age feature value falls within the 18-25 range, it generates tags like "post-95s" or "Generation Z," while the 26-35 range generates tags like "newcomer to the workforce" or "youth group." Geographical tags are extracted based on user registration information and IP address resolution. The system maintains a geographic coding table containing provincial, municipal, and district-level administrative divisions, generating precise geographic tags such as "city-district" through table lookup mapping. Device type tags are obtained by parsing the user's device information, including operating system type (e.g., iOS or Android) and fine-grained information such as device model.

[0071] The system needs to mine users' interests and preferences from their content interaction behavior. The user feature vector contains semantic features extracted by BERT, which encode the topic information of user-generated and interacted content. The system first decodes these semantic features to reconstruct the text content interacted with by the user, 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 how frequently a word appears in user content, while inverse document frequency reduces the weight of common words, highlighting the discriminative power of feature words. The system sorts the TF-IDF values ​​of all words and selects the top m keywords with the highest weights as input for topic modeling. In this embodiment, m is set to 20, a number that ensures coverage of the user's main interests without introducing too much noise.

[0072] After selecting keywords, the system inputs them into the LDA (Latent Dirichlet Allocation) topic model for topic discovery. LDA is a probabilistic topic model that assumes each document is a mixture of multiple topics, and each topic is a probability distribution of words. The system sets the number of topics K to 20, a parameter determined based on extensive experiments and business experience, which can effectively balance the fine granularity and interpretability of topics. The LDA model iteratively updates the topic-vocabulary distribution and document-topic distribution through 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 corresponding interest tags based on topics with a probability value greater than 0.1 in the topic distribution, such as "Technology & Digital," "Food Exploration," and "Fitness & Sports," which intuitively reflect the user's areas of interest.

[0073] To more accurately quantify the importance of each interest tag, the system is configured with a weighting formula that comprehensively considers interaction depth and time investment. The weight W of each interest tag... 兴趣 Through formula W 兴趣The formula is calculated as follows: 0.6 × Content Interaction Frequency / Total Interactions + 0.4 × Dwell Time / Total Online Time. The design philosophy behind this formula is that the frequency of a user's interaction with a certain type of content reflects the breadth of their interests, while dwell time reflects the depth of those interests. Content interaction frequency tracks the number of times a user likes, comments, or shares content within that interest category, and is normalized by dividing by the total number of interactions to ensure comparability between users with different activity levels. Dwell time records the cumulative browsing time for that type of content, and is also normalized by dividing by the total online time. The two metrics are assigned weights of 0.6 and 0.4 respectively, a slight bias towards interaction frequency, as proactive interaction better reflects a user's true interests than passive browsing.

[0074] The generation of behavioral tags employs real-time streaming computation, capturing users' immediate behavioral characteristics and state changes. The system continuously monitors user behavior data using the Flink streaming computation framework. Multiple behavior pattern detectors are defined in Flink, each responsible for identifying specific behavioral patterns. For example, when a user is detected browsing more than 10 live streams consecutively within one hour, the "Active Live Stream User" tag is generated; when a user has not interacted for three consecutive days, a "Low Activity" tag is generated; and when a user is active between midnight and 3 AM, a "Night Owl" tag is generated. These behavioral patterns are defined based on statistical analysis of massive amounts of user behavior data and summaries of business experience, effectively identifying user behavioral characteristics and usage habits. Flink ensures the real-time nature and accuracy of behavioral tag generation through a combination of event time windows and processing time windows.

[0075] The dynamic update mechanism for tag weights ensures that the tag system can adapt to changes in user interests. User interests and behaviors are not static and may shift due to factors such as seasons, trending 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 behavior throughout the day, reflecting the user's latest interest trends. The relatively high weight of 0.7 ensures the system's sensitivity to changes in user interests. 当日 This is the cumulative value of all previous historical weights, representing the user's long-term stable interests and preferences. A weight of 0.3 ensures the stability of the tag system and avoids drastic changes in tags due to short-term behavioral fluctuations. After generating and calculating the weights of various tags, the system organically combines basic tags, interest tags, and behavioral tags to build a complete multi-level tag system.

[0076] S103, input the tag system of the initiating user and the tag system of the target user into the pre-trained tag clustering model, extract and calculate the common points between the two parties, and then generate a matching feature set;

[0077] The label clustering model first calculates the behavioral statistical weights W. behavior This weight reflects the similarity between two users in their behavioral patterns. The system extracts behavior-related tags from both users' tag sets, including behavioral feature tags such as active time periods, interaction frequency, and content preferences. For each shared behavioral tag, the system calculates its behavior intensity value, which comprehensively considers the frequency, duration, and depth of engagement of the behavior. For example, if two users both have the tag "video creator," the system will quantify the behavior intensity by counting the number of videos they have published, the number of interactions they have received, and the average duration of their videos. Simultaneously, the system introduces a time decay mechanism to adjust the weight of behaviors at different times, with recent behaviors receiving higher weights while the influence of older behaviors gradually weakens, making the matching results more reflective of the user's current state.

[0078] The formula for calculating behavioral statistical weights, Wbehavior = Σ(behavior intensity i × time decay i) / user activity baseline, fully considers the differences in user activity levels. The numerator is the weighted sum of the intensities of all common behavioral tags, with each behavior intensity multiplied by its corresponding time decay factor to ensure timeliness. The denominator, user activity baseline, is a normalization factor calculated by dividing the user's total number of behaviors over the past y days by y. In this embodiment, y is set to 30 days, meaning the user's average daily number of behaviors over the past month is calculated. This normalization process eliminates the bias caused by differences in user activity levels, making the behavioral weights of highly active and inactive users comparable. For example, an active user who generates 100 behaviors per day and a regular user who only generates 10 behaviors per day will receive similar weight values ​​for their similar behavioral patterns through normalization of the activity baseline, avoiding interference from activity level itself on the matching results.

[0079] The semantic association weight Wsemantic focuses on the semantic similarity of user interest tags. The system inputs the interest tag texts of the initiator and target into the BERT-wwm (Whole Word Masking) model. BERT-wwm is an improved version of the BERT model, employing a whole-word masking strategy for Chinese processing, which can better understand the complete semantics of Chinese words. For each pair of interest tags, the model first processes the tag text through a word segmenter, adding special markers [CLS] and [SEP], and then inputs it into the multi-layer Transformer encoder of BERT-wwm. After processing through 12 or 24 layers of 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.

[0080] After obtaining the semantic vectors of the tags, 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, with a value ranging from -1 to 1, where 1 indicates identical vectors, 0 indicates orthogonal and unrelated vectors, and -1 indicates completely opposite vectors. To convert the similarity into a more intuitive weight value, the system uses the formula Wsemantic = (cosinesimilarity + 1) / 2 for a linear transformation, mapping the value range to 0 to 1. This semantic similarity calculation can discover tag matches that appear different but are semantically related. For example, "basketball" and "NBA" are literally different but have high semantic similarity; "food" and "restaurant exploration" also have a strong semantic connection. Through the deep semantic understanding of the BERT-wwm model, the system can uncover more potential common interests, improving the accuracy and coverage of matching.

[0081] The calculation of the social propagation weight Wsocial requires the construction and analysis of users' social relationship networks. The system first constructs a large-scale user social graph, a directed weighted graph structure where nodes represent users on the platform, and edges represent social relationships between users, such as following or friendship. The weight of each edge is 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 interactions are assigned different weight coefficients; for example, private messages have a higher weight than simple likes because the former represents a deeper level of social connection. The system maintains real-time updates to this social graph; when users generate new social behaviors, the corresponding edge weights are dynamically adjusted.

[0082] Based on the constructed social graph, the system employs 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 the structural information of the graph 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 to a two-layer structure. The first layer aggregates and compresses the 768-dimensional feature vector of the node 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 the features of these neighboring nodes, concatenates it with the node's own features, and then passes it through a fully connected layer and an activation function to obtain a new feature representation of the node.

[0083] Based on the user embedding vectors generated by GraphSAGE, the system further calculates the propagation influence of tags in social networks. The PageRank algorithm is adopted here, treating tags as signals propagating in social networks and iteratively calculating the global influence score of each tag. The core assumption of PageRank is that if a tag is owned by a user with high influence, or shared by many users, then the tag itself has high influence. The system initializes the PageRank value of each tag to 1.0, and then updates it using the iterative 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 obtains a converged PageRank value, which serves as the social propagation weight Wsocial for the tag. This weight reflects the propagation potential and influence range of the tag in the social network. Shared tags with high social propagation weights mean that two users are likely in similar social circles, with more common topics and social connections.

[0084] After obtaining the weights for the three dimensions, the system uses a gating network to dynamically fuse these weights and generate the final matching score. A gating network is a variant of an attention mechanism that adaptively adjusts the importance of different weights based on input features. The network input is a feature vector containing three weight values, user feature statistics, and scene context. Passed through a two-layer fully connected network, it outputs three gating coefficients α, β, and γ. These coefficients are normalized using softmax to ensure their sum equals 1. During model training, these parameters undergo supervised learning using numerous successful and failed user connection cases, and are updated using the cross-entropy loss function and the Adam optimizer. After training, the model learns the optimal combination strategy for each dimension's weights in different scenarios. For example, in interest-based social scenarios, the semantic association weight β is relatively high; in familiar social scenarios, the social propagation weight γ dominates; and in stranger social scenarios, the behavioral statistics weight α is more important.

[0085] The final match score is calculated using the formula FinalScore = α × Wbehavior + β × Wsemantic + γ × Wsocial. This score comprehensively reflects the overall matching degree between two users across three dimensions: behavioral patterns, interests, and social relationships. The system sets a threshold z (z = 0.6 in this embodiment) and extracts all common tags with a FinalScore greater than z as a matching feature set. This set includes not only the tags themselves but also meta-information such as the match score and main contribution dimensions for each tag. The matching feature set is sorted from highest to lowest score, and the highest priority commonalities will be used extensively in subsequent copywriting generation. For example, if the match score for the common interest "photography" reaches 0.85, and the main scores come from semantic association and behavioral statistics, the system will use it as a core topic element to generate a personalized opening line such as, "I see you also like photography. Have you taken any interesting photos lately?"

[0086] S104, Obtain the historical interaction data of the target user, input the historical interaction data and the target user's tag system into the pre-trained sentiment analysis model, and generate the target user's current sentiment state value;

[0087] The system first extracts historical interaction data of the target user from the database. This data records various expressions and interactive behaviors of the user on the platform. The acquisition of historical interaction data follows a reverse chronological order, prioritizing the extraction of the most recent interaction records, as recent behavior better reflects the user's current emotional state. The system is set to retrieve the most recent b historical interaction records; in this embodiment, b is set to 30. This number provides sufficient samples for sentiment analysis without introducing overly distant or potentially irrelevant emotional information due to excessive data. These 30 records cover various user interaction types, including user-generated original content such as updates, articles, and videos; user comments and replies to others' content; and lightweight interactive behaviors such as likes and favorites. When extracting this data, the system records detailed information for each record, including metadata such as interaction type, time of occurrence, content text, and interacting parties.

[0088] Historical interaction data and the target user's tagging system are input into a pre-trained sentiment analysis model, which uses deep learning technology to understand and quantify users' emotional expressions. At the core of the sentiment analysis model is a BERT sentiment classifier trained on large-scale sentiment-annotated data, capable of understanding the sentiment tendency and intensity in text. For each interaction record, the system first extracts its text content. For posted and commented content, the raw text is used directly; for non-textual behaviors such as likes, the system extracts the text of the liked content as the analysis object. The text is preprocessed before being input into the BERT model. Preprocessing includes word segmentation, adding special tags, truncating, or padding to a fixed length. The BERT model uses its 12-layer Transformer encoder to deeply encode the text, capturing complex semantic relationships between words and sentiment expression patterns.

[0089] The BERT sentiment classifier outputs an initial sentiment score (sentimentscore) for each record. This is a continuous value in the range [-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 satisfaction. This continuous value design can more finely characterize the intensity changes of emotions than simple sentiment classification (positive / negative / neutral). For example, "I'm so happy today!" might get a high score of 0.9, "It's okay" might get a slightly positive score of 0.3, and "I'm feeling a little down" might get a negative score of -0.4. During training, the model uses a sentiment corpus containing millions of labeled data points, covering various emotional expressions and language styles, ensuring the accuracy and robustness of sentiment recognition.

[0090] Different types of interactions reflect emotions with varying degrees of reliability; therefore, the system assigns different weights to each interaction type. Publishing original content is given the highest weight of 1.0 because users typically invest more thought and emotion when actively creating and sharing content, resulting in the most authentic and intense emotional expression. Users spending time writing text, recording a video, or carefully editing a set of images demonstrates their desire to express and share, containing the richest emotional information. Commenting on others' content is weighted at 0.8. While comments also contain user opinions and emotional expressions, compared to original content, comments are often reactions to others' content and may be influenced by the original content's emotional tone, thus having slightly lower reliability. Liking behavior is weighted at 0.5. Liking is a lightweight interaction; users may like it out of politeness, habit, or casual browsing, and it may not fully represent their true emotional state. However, a large number of likes can still reflect a user's emotional inclinations.

[0091] The time factor plays a crucial role in sentiment analysis. Users' emotional states are dynamic, and interactions more recent in time better reflect their current emotions. The system uses a time decay coefficient, timedecay=e -0.05×Δt This time-related impact is quantified, where Δt is the number of hours since the interaction. The exponential decay function design assigns a weight close to 1 to recently occurring interactions (Δt close to 0), while the weight decreases exponentially over time. The decay coefficient of 0.05 has been validated through extensive experimentation. It reduces the weight of interactions from 24 hours ago to approximately 0.3, and interactions from a week ago to near 0. This decay rate ensures the dominance of recent emotions without completely ignoring earlier emotional context. For example, a "feeling great" post from one hour ago receives a time weight of 0.95, while similar content posted at the same time yesterday only receives a weight of 0.3.

[0092] Taking into account three factors—sentiments score, interaction weight, and time decay—the system calculates the weighted sentiment score (ei) for each record as: ei = sentimentscore × interactionweight × timedecay. This comprehensive indicator fully reflects the contribution of the interaction record to the user's current emotional state. For example, a newly published original post, "Finally finished the project, feels great!", has a sentimentscore of 0.8, an interaction weight of 1.0, and a timedecay close to 1, resulting in an ei of approximately 0.8. However, a like posted 3 days ago, even with a high sentiment score, has a low interaction weight of only 0.5 and a time decay of 0.22, resulting in an ei of only 0.088, which has a minimal impact on the current mood.

[0093] 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 tagging system. These tags are stable characteristics derived from long-term behavioral analysis, such as "optimistic and cheerful," "introverted and quiet," and "rational and calm." Based on these personality tags, the system calculates the personality factor, an indicator reflecting the user's basic emotional tendency. The calculation of the personality factor uses the Big Five personality model from psychology as its theoretical basis, mapping personality tags to five dimensions: openness, conscientiousness, extraversion, agreeableness, and neuroticism. Then, a weighted combination is used to obtain a comprehensive emotional tendency score. For example, a user tagged "extroverted and lively" will have a personality factor that leans towards a positive value, perhaps around 0.6; while a "introverted and cautious" user might have a personality factor around 0.4. This factor represents the user's baseline emotional level under the influence of no special events, a relatively stable emotional baseline.

[0094] The final emotional state value is calculated using the formula Estate = 0.6 × (Σei / 30) + 0.4 × personalityfactor. The first part of the formula is the average of 30 weighted emotional values, reflecting the emotional state derived from recent behavior analysis; the weight of 0.6 reflects the emphasis on real-time behavior. The second part is the contribution of the personality factor; the weight of 0.4 ensures the stability of the emotional analysis and avoids biases in emotional judgment caused by individual extreme behaviors. This weighted combination method allows the emotional analysis to capture user emotional fluctuations without deviating too far from their basic personality traits. For example, a user with a normally cheerful personality (personalityfactor = 0.7), even if they have recently expressed some negative emotions (average ei = 0.2), will still have an Estate value of 0.6 × 0.2 + 0.4 × 0.7 = 0.4, which is neutral to slightly positive, a more reasonable conclusion than one based solely on recent behavior. The calculated Estate value may initially not be within the standard [0,1] range; the system normalizes it to map it into this range.

[0095] S105, map the current emotional state value of the target user to a pet emoji 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 tag system of the target user into the pre-trained copywriting generation model to generate a greeting copywriting that is adapted to the target user's emotion. In this process, the response of the target user is obtained and the parameters of the tag generation model, the sentiment analysis model, and the copywriting generation model are updated based on the response.

[0096] Visualizing emotional states greatly enhances the user experience. The system transforms abstract emotional values ​​into concrete pet facial animations, allowing users to understand the other person's emotional state at a glance. The system pre-defines pet images with different personality traits. Lively cats have bright, expressive eyes and upturned tails, suitable for portraying cheerful and outgoing emotions; gentle rabbits have soft ears and mild expressions, adept at conveying peaceful and friendly feelings; loyal dogs exhibit enthusiastic and friendly traits, with the frequency and amplitude of their tail wagging accurately reflecting emotional intensity; mysterious foxes have elegant postures and deep eyes, suitable for expressing complex and subtle emotions; and lazy pandas convey a relaxed and comfortable atmosphere with their endearing appearance. Based on the target user's personality tags within their tagging system, the system uses a mapping algorithm to select the most suitable pet image as their emotional representation.

[0097] The dynamic mapping of pet expressions follows a sophisticated emotional range segmentation strategy. When the Estate value is greater than 0.7, the system determines that the user is in a positive and happy emotional state, and the pet will display a series of happy expressions, including a smile with eyes curving into crescent shapes, a wide laugh with upturned corners of the mouth, and actions such as jumping excitedly or spinning in place. These animations are not simply looped, but dynamically adjusted according to the intensity of the emotion. The closer Estate is to 1, the higher the amplitude and frequency of the animation. For example, a cat's tail will wag more joyfully, and a rabbit's ears will stand up excitedly and twitch slightly. When 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, curiously tilting its head to observe, and subtle movements such as slight wagging of the tail or ears. These movements convey a peaceful and open attitude towards communication. When Estate is below 0.3, pets will exhibit depressed mood characteristics, with their ears drooping naturally, their eyes looking slightly melancholy, and they may curl up, slowly lie down, or sigh softly. This visual presentation can evoke empathy from the user who initiated the interaction, prompting them to adopt a gentler and more considerate communication style.

[0098] The smoothness of the animation is achieved through high-precision interpolation technology. The system uses a Bézier curve interpolation algorithm to generate smooth transition animations 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 Bézier 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 happy emotions uses an ease-out curve, with a fast initial speed followed by a gradual slowdown, reflecting the natural release of emotions; while the animation of melancholy emotions uses an ease-in curve, with a slow start followed by a slight acceleration, reflecting the feeling of suppressed emotions. The animation is displayed as a floating layer in the initiating user's chat interface, with a duration of 2-3 seconds and looping, which is neither too abrupt and affects the chat experience, but also continuously reminds the user to pay attention to the other party's emotional state.

[0099] The system inputs the matching feature set, the target user's emotional state value, and the labeling system of both parties into a pre-trained copywriting generation model. The first step in copywriting generation is determining the appropriate tone, which directly affects the style and word choice of greetings. Based on different ranges of Estate values, the system has formulated differentiated tone strategies. When Estate is greater than 0.7, the system chooses a light and lively tone, with the emotional intensity parameter set to 0.8. This high-intensity emotional expression is conveyed through the use of exclamation marks, emoticons, and playful wording, such as "Wow, I see you also like 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 the emotional intensity parameter set to 0.5. The language expression is more neutral and peaceful, avoiding being overly enthusiastic or cold, such as "Hello, I noticed we both follow the technology field. Do you have any new discoveries you'd like to share?" When Estate falls below 0.3, the system switches to a caring and considerate tone, lowering the emotional intensity parameter to 0.3, using gentler and more cautious language, and focusing on emotional support. For example, "I saw your recent posts. If you'd like, could we talk about your life lately? Sometimes talking can make you feel better."

[0100] Topic selection is a key factor in the success of topic construction. The system extracts the top q most weighted commonalities 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 targeted without being too simplistic. The system constructs an intelligent topic agent system that integrates multiple topic discovery and optimization techniques. First, it performs in-depth analysis of the core elements of the topic using 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 commonality tags into text descriptions, inputs them into the BERTopic model, and extracts the most representative set of keywords through steps such as BERT encoding, UMAP dimensionality reduction, and HDBSCAN clustering.

[0101] To make topics more timely and engaging, the system incorporates a real-time trending topic analysis mechanism. The topic weight is calculated using the formula whot=0.6×(trend24h / max(trend))+0.4×(localheat / avg(heat)), which comprehensively considers overall online popularity and regional relevance. trend24h is obtained through access to the trending search APIs of mainstream platforms, reflecting changes in topic discussion popularity over the past 24 hours. The system maintains a real-time updated trending topic database, refreshed hourly, recording metrics such as search volume, discussion volume, and interaction volume for each topic. localheat, based on the user's geographic location information, calculates the topic's popularity within the user's city or region. This regionalized trending topic analysis makes topics more relevant to the user's living environment. For example, if both parties are in the same city and both are interested in food, the system might generate a topic based on local trending topics such as "a new restaurant opened at Universal Studios in a certain city."

[0102] The final score for each candidate topic is determined through a comprehensive evaluation. The formula, Topicscore = 0.5 × commonweight + 0.3 × emotionmatch + 0.2 × freshness, balances multiple dimensions. Commonweight comes directly from the weight values ​​in the matching feature set, representing the degree of topic relevance. Emotionmatch is calculated by determining the consistency between the topic's emotional tone and the user's current mood; for example, overly cheerful topics will receive lower scores when the user is feeling down. Freshness assesses the topic's novelty, prioritizing recently emerging topics or topics that users have recently followed but haven't yet discussed in depth. The system generates 5-10 candidate topics and selects the highest-scoring one as the core content of the copy.

[0103] 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 a low-rank adapter and quantizing it. The system organizes the selected topic, copy tone parameters, and tag information of both parties into structured prompts. The design of the prompts follows a framework of "role setting - background information - task requirements - output constraints." For example, a prompt might be: "You are a friendly social assistant. User A (tags: photography enthusiast, travel expert) wants to connect with User B (tags: photographer, landscape enthusiast). User B is currently in a positive mood. Please generate a 50-100 word greeting, focusing on 'recent travel photography experiences,' with a relaxed, friendly, natural, and sincere tone." After the model generates the initial copy based on this prompt, the system performs multiple rounds of optimization, including length adjustment, sensitive word filtering, and personalized element injection.

[0104] To ensure the diversity and novelty of the copywriting, the system uses the SimHash algorithm to detect the similarity between newly generated copy and the historical copy library. SimHash is a locality-sensitive hashing algorithm that can quickly calculate the fingerprint of text and evaluate similarity. After segmenting the copy into words, the system calculates the hash value of each word, obtains the SimHash fingerprint of the copy through weighted summation, and then calculates the Hamming distance between the fingerprints of the new copy and the historical copy. If the similarity exceeds a preset threshold of 0.7, the system will adjust the generation parameters, such as temperature coefficient and top-p value, and regenerate the copy, attempting a maximum of 5 times. This mechanism effectively avoids users receiving monotonous greetings, ensuring a unique communication experience each time a connection is established.

[0105] The system's intelligence is reflected in its continuous learning and optimization capabilities. When a user responds to a greeting, the system records detailed response data, including whether a reply was received, the reply time, the length of the reply, the emotional tone, and the number of subsequent dialogue rounds. Based on this response data, the system calculates a comprehensive response feedback score R, which considers both immediate feedback and long-term effects. Immediate feedback primarily assesses whether a reply was received and the speed of the reply; a quick reply within 5 minutes receives the highest score, decreasing with time. Long-term effects evaluate the continuity and depth of the dialogue, including indicators such as the number of dialogue rounds, average reply length, and positive emotional tone.

[0106] Based on the response feedback scores, the system updates the parameters of the three core models accordingly. For the label clustering model, the system uses gradient descent to update the weight fusion parameters α, β, and γ, with a learning rate set to 0.01. By calculating the partial derivatives of the loss function with respect to each parameter, the system adjusts the parameter values ​​in a direction that improves matching accuracy. When the weight contribution of a certain dimension is particularly prominent in successfully established cases, the system appropriately increases the weight coefficient of that dimension. For example, if the success rate of establishing connections through interest similarity matching is found to be particularly high, the semantic association weight β will be gradually increased.

[0107] The update to the sentiment analysis model primarily aims to improve prediction accuracy. The system analyzes the actual sentiment of responses to determine the user's true emotional state. When this actual value deviates from the previously predicted Estate value by more than 0.3, a model update mechanism is triggered. The update mainly adjusts the attention weight matrix in the BERT model, calculates the gradient of the prediction error with respect to the attention parameters using backpropagation, and updates the parameters using the Adam optimizer. This update enables the model to better understand the emotional expression characteristics of specific user groups, improving the accuracy of subsequent sentiment analysis.

[0108] The copywriting generation model is optimized using a contrastive learning strategy. The system labels copywriting with a response rate higher than 0.7 as positive samples and those lower than 0.3 as negative samples, constructing a training set for contrastive learning. By calculating the representational differences between positive and negative samples in the semantic space, the LoRA parameters of the model are optimized, making the generated copywriting more closely resemble the features of positive samples. The 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 validated through A / B testing. Full deployment is only implemented when the connection establishment success rate improves by more than 5%, ensuring a steady improvement in system performance.

[0109] Please see Figure 2 The second embodiment of the present invention provides an apparatus for improving the establishment rate, comprising:

[0110] Feature vector generation unit 201 is used to obtain the full-link behavior data of the initiating user and the target user in the target community. The full-link behavior data is input into the pre-trained feature extraction model to generate the feature vectors of the initiating user and the target user.

[0111] The tag system generation unit 202 is used to input the feature vector of the initiating user and the feature vector of the target user into the pre-trained tag generation model to generate a multi-level user tag system for the initiating user and a multi-level user tag system for the target user.

[0112] 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 the pre-trained tag clustering model, extract and calculate the common points between the two parties, and then generate a matching feature set.

[0113] The emotion state value generation unit 204 is used to acquire the historical interaction data of the target user, input the historical interaction data and the target user's tag system into a pre-trained emotion analysis model, and generate the target user's current emotion state value.

[0114] Feedback adjustment unit 205 is used to map the current emotional state value of the target user to a pet emoji and display it on the chat interface of the initiating user. It inputs the matching feature set, the current emotional state value of the target user, the initiating user, and the multi-level user tag system of the target user into a pre-trained copywriting generation model to generate a greeting copywriting that is adapted to the target user's emotion. The model obtains the target user's response and updates the parameters of the tag generation model, the sentiment analysis model, and the copywriting generation model based on the response.

[0115] The third embodiment of the present invention provides a device for improving the connection establishment rate, including a memory and a processor. The memory stores a computer program, which can be executed by the processor to implement a method for improving the connection establishment rate as described in any of the above embodiments.

[0116] The fourth embodiment of the present invention provides a computer-readable storage medium, characterized in that it stores a computer program, which can be executed by a processor of the device in which the computer-readable storage medium is located, to implement a method for improving connection rate as described in any of the above claims.

[0117] Based on the method, apparatus, device, and storage medium provided by this invention for improving the connection establishment rate, 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 common features between the two parties are extracted using a tag clustering model. Subsequently, combining the target user's historical interaction data and tag system, the current emotional state is obtained using a sentiment analysis model, and the emotion is displayed on the initiator's interface in the form of a pet emoji. Finally, the matching features, emotional state, and tag system are input into a copywriting generation model to generate a personalized greeting with appropriate emotion. The parameters of each model are dynamically optimized based on the response, thereby improving the relevance and persuasiveness of the information and thus increasing the connection establishment success rate.

[0118] Exemplary examples show that the computer program described in the third and fourth embodiments of the present invention can be divided into one or more modules, which 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 capable of performing specific functions, which describe the execution process of the computer program in a device for implementing an improved connection rate. For example, the apparatus described in the second embodiment of the present invention.

[0119] The processor referred to can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the method for improving connectivity, connecting various parts of the implementation of the method for improving connectivity through various interfaces and lines.

[0120] The memory can be used to store the computer program and / or modules. The processor, by running or executing the computer program and / or modules stored in the memory, and by calling data stored in the memory, implements various functions of a method to improve connection rate. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, text conversion function, etc.), etc.; the data storage area may store data created according to 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 may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital card (SD card), flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0121] If the implemented 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, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0122] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0123] 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 variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for improving the establishment rate of connections, characterized in that, include: Acquire the full-link behavior data of the initiating user and the target user in the target community, and input the full-link behavior data into a pre-trained feature extraction model to generate feature vectors of the initiating user and the target user. The feature vectors of the initiating user and the target user are respectively input 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 tag system of the initiating user and the tag system of the target user are input into a pre-trained tag clustering model to extract and calculate the commonalities between the two and generate a matching feature set. The historical interaction data of the target user is obtained, and the historical interaction data and the target user's tag system are input into a pre-trained sentiment analysis model to generate the target user's current sentiment state value. The target user's current emotional state value is mapped to a pet emoji and displayed on the initiating user's chat interface. The matching feature set, the target user's current emotional state value, the initiating user, and the target user's multi-level user tag system are input into a pre-trained copywriting generation model to generate a greeting copywriting that is appropriate for the target user's emotions. The target user's response is obtained, and the parameters of the tag generation model, sentiment analysis model, and copywriting generation model are updated based on the response.

2. The method for improving the establishment rate according to claim 1, characterized in that, The pre-trained feature extraction model uses an LSTM-BERT fusion architecture, and the specific extraction process is as follows: The LSTM time series model is used to analyze user behavior sequences and calculate the transition probability matrix P(s) of the behavior path. t+1 |s t To identify high-frequency behavioral patterns, where s t This represents the behavioral state at time t; Based on the frequency of user interaction behaviors observed in the aforementioned high-frequency behavior patterns, 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 a 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 behavior 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 dimensionality of the fused feature vector and output the user's feature vector.

3. The method for improving the establishment rate according to claim 2, characterized in that, The construction process of the multi-level user tagging system includes: Based on user feature vectors, basic tags are extracted from basic attribute data through a rule engine. These basic tags include age group, region, and device type. For the text semantic part in the user feature vector, the TF-IDF algorithm is used to calculate the keyword weights. The top m keywords with the highest weights are input into the LDA topic model, the number of topics is set to m, and user interest tags are generated. Calculate the weight W for each interest tag. 兴趣 Its expression is W 兴趣 =0.6 × Content interaction frequency / Total number of interactions + 0.4 × Dwell time / Total online time; The system acquires user behavior sequences and identifies changes in user behavior using the Flink streaming computing framework. When a specific behavior pattern is detected, behavior label generation is triggered. The label weights are dynamically updated using exponential smoothing: W 新 =0.7×W 当日 +0.3×W 历史 W 当日 W is the weight calculated based on the behavior of the day. 历史 Historical cumulative weighting; The basic tags, interest tags, and behavior tags are combined to form a multi-level tag system.

4. The method for improving the establishment 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: Based on the initiator's tag set and the target's tag set, calculate the behavioral statistical weight W. 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; The interest tags of the initiator and the target are input into the BERT-wwm model to calculate the semantic similarity between the tags and obtain the semantic association weight W. semantic Its expression is: W semantic =(cosinesimilarity+1) / 2, where cosinesimilarity is the cosine similarity, with a value range of [-1,1]. Construct a user social graph where nodes represent users and edges represent social relationships. Edge weights are determined by interaction frequency. The GraphSAGE algorithm is used to generate user embedding vectors. Based on these vectors, the PageRank value of each tag in the social network is calculated to obtain the social propagation weight Wsocial. A gating network is used to dynamically fuse the three weights, and the final matching score is calculated: FinalScore = α × W behavior +β×W semantic +γ×W social , where α, β, and γ are learning parameters, and the common points of FinalScore>z are extracted as the matching feature set.

5. The method for improving the establishment rate according to claim 1, characterized in that, The analysis process of the pre-trained sentiment analysis model includes: Obtain the most recent b historical interaction records of the target user, including posted content, comments, and likes. For each interaction record, use the BERT sentiment classifier to calculate the initial sentiment value sentimentscore, which ranges from [-1, 1]. Set interaction weights based on interaction type, which includes weights for publishing original content, commenting on others' content, and liking. Calculate the time decay factor for each record: timedecay=e -0.05×Δt Where Δt is the number of hours since the current time, the weighted sentiment score for a single record is calculated as: ei = sentimentscore × interactionweight × timedecay Personality tags are extracted from the target user's tag set, and a personality factor is calculated. This personality factor reflects the user's basic emotional tendency. The current emotional state value of the target user is then comprehensively calculated. Estate = 0.6 × (Σei / 30) + 0.4 × personality factor, normalize Estate to the interval [0,1] as the final emotional state value output.

6. The method for improving the establishment rate according to claim 5, characterized in that, The process by which the pre-trained copywriting generation model generates greeting copy is as follows: Based on the current emotional state value Estate, select the tone parameter of the copywriting, 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 that extracts keywords from core topic elements using BERTopic technology and calculates topic weight by combining real-time trending topics. whot=0.6×(trend24h / max(trend))+0.4×(localheat / avg(heat)) Among them, trend24h is the 24-hour trend value, and localheat is the regional heat. Calculate the overall score for each candidate topic: Topicscore = 0.5 × commonweight + 0.3 × emotionmatch + 0.2 × freshness, where commonweight is the weight of common points, emotionmatch is the degree of matching between the topic and the emotional state value, and freshness is the freshness of the topic; The topic with the highest Topicscore is selected, and combined with the copywriting tone parameters and the tag sets of both parties, copywriting generation prompts are constructed. The prompts are then input into a pre-trained copywriting generation model to obtain 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 a preset value, the copy is regenerated.

7. The method for improving the establishment 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 specifically involves: Obtain the response data of the target user, calculate the comprehensive response feedback score based on the response data of the target user, and update the weight fusion parameters of the label generation model based on the comprehensive response feedback score using the gradient descent method; When it is determined that the deviation between the target user's emotional state value and the predicted emotional state value is greater than a preset value, the attention weight matrix of the pre-trained sentiment analysis model is updated. Positive and negative samples are selected based on the response rate, and the LoRA parameters of the copywriting generation model are updated based on the learning loss of the positive and negative samples.

8. A device for improving the connection rate, characterized in that, include: The feature vector generation unit is used to acquire the full-link behavior data of the initiating user and the target user in the target community, and input the full-link behavior data into the pre-trained feature extraction model to generate the feature vectors of the initiating user and the target user. The tag system generation unit is used to input the feature vectors of the initiating user and the target user into the pre-trained tag generation model to generate a multi-level user tag system for the initiating user and a multi-level user tag system for the target user. The 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 the pre-trained tag clustering model, extract and calculate the common points between the two parties, and then generate a matching feature set. The emotion state value generation unit is used to acquire the target user's historical interaction data, input the historical interaction data and the target user's tag system into a pre-trained emotion analysis model, and generate the target user's current emotion state value. The feedback adjustment unit is used to map the current emotional state value of the target user into a pet emoji and display it on the chat interface of the initiating user. The matching feature set, the current emotional state value of the target user, the initiating user, and the multi-level user tag system of the target user are input into a pre-trained copywriting generation model to generate a greeting copywriting that is adapted to the target user's emotion. The unit also obtains the target user's response and updates the parameters of the tag generation model, the sentiment analysis model, and the copywriting generation model based on the response.

9. A device for improving the connection rate, characterized in that, The device includes a memory and a processor, wherein the memory stores a computer program that 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, The device contains a computer program that can be executed by a processor of the device in which the computer-readable storage medium is located, to implement a method for improving connection rate as described in any one of claims 1 to 7.

Images