A social conversation intelligent connection and activation method and system based on user profile

Abstract

The application discloses a kind of social conversation intelligent connection and activation method and system based on user profile, it is related to social network technical field.The method includes: in response to the connection request, the multidimensional profile of first user and second user is obtained;The quality of the multidimensional profile of first user is evaluated, and when meeting the preset condition, the feature analysis is carried out on the profile of both sides to identify the topic entry point;The topic entry point is input into AI model to generate personalized ice-breaking speech for the recipient and push to the first user;After conversation establishment, start layered time wheel to monitor the initiative message count of conversation, and mark the conversation as interrupted when not meeting the effective conversation standard;Activation judgment is executed on interrupted conversation, and activation speech is generated and pushed when meeting the activation condition.The application realizes the whole process intelligent management of social conversation from connection to activation, improves the matching efficiency and user experience.

Description

Technical Field

[0001] This invention relates to the field of social network technology, specifically to a method and system for intelligent connection establishment and activation of social conversations based on user data. Background Technology

[0002] With the rapid development of internet technology, social networking applications have become an important medium for people's daily communication and expanding interpersonal relationships. Stranger social networking and dating platforms use recommendation algorithms to match users with potential partners. After establishing initial contact through the platform, users need to deepen their understanding through conversational exchanges. However, in the actual use of existing social applications, the following common problems exist.

[0003] First, breaking the ice is difficult. After a successful match, users often don't know how to start the initial conversation. Traditional social applications often use a fixed template library for matchmaking messages, lacking the ability to generate personalized messages based on specific user characteristics and matching points between the two parties. This results in monotonous and unattractive icebreakers, weak user willingness to respond, and low conversion rates after the match is established.

[0004] Secondly, the conversation interruption rate is high and there is a lack of proactive maintenance. A large number of social conversations are quickly interrupted in the early stages due to a lack of effective incentives. Existing technologies lack precise monitoring and proactive intervention mechanisms for conversation status, and cannot identify the risk of conversation interruption and activate it in a timely manner. They can only passively wait for users to resume the conversation on their own, resulting in the termination of a large number of potential social relationships in the early stages of their formation.

[0005] Therefore, there is an urgent need in this field for a technical solution that can fully utilize multi-dimensional user data, generate personalized icebreakers during the initial connection, proactively monitor the conversation status after connection is established, and intelligently activate the system when the conversation is interrupted, so as to form a complete closed loop from distribution to connection establishment to conversation maintenance, and systematically improve the efficiency of social matching and user experience. Summary of the Invention

[0006] To address the problems of low icebreaking efficiency, high session interruption rate, and lack of proactive maintenance in existing technologies, this invention provides a method and system for intelligent connection establishment and activation of social sessions based on user data, thereby resolving the aforementioned technical deficiencies.

[0007] This invention proposes a method for intelligent connection establishment and activation in social conversations based on user data, comprising the following steps: S1. In response to the connection request, obtain multi-dimensional information of the first user and multi-dimensional information of the second user; S2. Conduct a quality assessment of the multi-dimensional data of the first user. When the quality assessment results meet the preset conditions, conduct feature analysis on the multi-dimensional data of the first user and the multi-dimensional data of the second user respectively, and identify the topic entry point based on the feature analysis results. S3. Input the topic entry point into the AI ​​model, and the AI ​​model will generate a personalized icebreaker script for the recipient based on the topic entry point, and push the personalized icebreaker script to the first user's client. S4. After detecting the establishment of a session between the first user and the second user, start session status monitoring. The session status monitoring uses a hierarchical time wheel as the scheduling data structure. Set up a monitoring task in the hierarchical time wheel. The monitoring task is used to trigger a check of the active message count of the session when the preset delay arrives. Determine whether the session has reached the valid session standard based on the active message count. If the valid session standard has not been reached, mark the session as an interrupted session. S5. Obtain the session identifier of the interrupted session and perform activation determination based on the session identifier; when all conditions in the activation condition set are met, mark the interrupted session as a session to be activated, generate the activation message and push it to the client of the first user or the client of the second user.

[0008] This invention also proposes a social conversation intelligent connection building and activation system based on user data, for implementing the method described above. The system includes: The data acquisition module is configured to respond to the connection establishment request and acquire multi-dimensional data of the first user and multi-dimensional data of the second user. The quality assessment and routing module is configured to perform quality assessment on the multi-dimensional data of the first user. When the quality assessment results meet the preset conditions, feature analysis is performed on the multi-dimensional data of the first user and the multi-dimensional data of the second user respectively, and the topic entry point is identified based on the feature analysis results. The AI ​​dialogue generation module is configured to input the topic entry point into the AI ​​model, and the AI ​​model will generate a personalized icebreaker dialogue for the recipient based on the topic entry point, and push the personalized icebreaker dialogue to the client of the first user. The session status monitoring module is configured to start session status monitoring after detecting the establishment of a session between the first user and the second user. The session status monitoring uses a hierarchical time wheel as the scheduling data structure. A monitoring task is set in the hierarchical time wheel. The monitoring task is used to trigger a check of the active message count of the session when a preset delay is reached. The active message count is used to determine whether the session has reached the valid session standard. If the valid session standard has not been reached, the session is marked as an interrupted session. The intelligent activation decision module is configured to obtain the session identifier of the interrupted session and perform activation determination based on the session identifier; when all conditions in the activation condition set are met, the interrupted session is marked as a session to be activated, an activation script is generated and pushed to the client of the first user or the client of the second user.

[0009] The present invention also proposes a terminal device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the computer program, it implements the steps of the above-described method for intelligent connection establishment and activation of social conversations based on user data.

[0010] The present invention also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method for intelligent connection establishment and activation of social conversations based on user data.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) By introducing a quality assessment and conditional routing mechanism, the present invention first performs a quality assessment on the multi-dimensional data of the first user after obtaining the user's multi-dimensional data. Only when the quality assessment result meets the preset conditions will the subsequent AI feature analysis and speech generation process be started, thus avoiding the waste of AI computing resources on low-quality users with missing data, and realizing the reasonable allocation of computing resources and the improvement of overall connection efficiency. The multi-dimensional data is normalized and weighted to obtain a quality score. The quality score is compared with a preset threshold to determine whether the preset conditions are met, so that the routing decision has a clear quantitative basis.

[0012] (2) This invention constructs a fusion scheme of text feature vectors and structured feature vectors of multi-dimensional data, performs cross-matching of the data from both parties, calculates topic potential scores to identify the optimal topic entry point, and then calls an AI model to generate personalized icebreakers. This scheme can make full use of the multi-dimensional data from both parties to mine matching points and generate targeted messages, thereby effectively improving users' willingness to respond and the success rate of establishing connections.

[0013] (3) By adopting a hierarchical time wheel as the scheduling data structure for session state monitoring, this invention achieves efficient scheduling of monitoring tasks in high-concurrency social scenarios, effectively supporting the parallel monitoring needs of large-scale sessions. At the same time, when the session active message count does not reach the valid session standard, the session state is migrated to a cooling state, and the active message count continues to be monitored within the cooling observation window. The session is only marked as interrupted when the cooling observation window expires and no new messages are detected, thus avoiding misjudgments caused by network latency or the user's brief absence.

[0014] (4) This invention introduces a set of activation conditions, including activation count constraints, online status of both parties, receiving user's willingness to receive, and cumulative number of cards received on the same day. After a session is marked as interrupted, it automatically performs a multi-condition joint activation judgment and pushes the activation message only when all conditions are met, thus achieving precise session activation intervention. Unlike the existing technology that relies on user-initiated auxiliary methods, this invention achieves proactive and precise intervention of the system in interrupted sessions.

[0015] (5) This invention collects user feedback on personalized icebreaker and activation scripts, obtains positive signals such as clicking to send, replying within a preset time, and reaching the effective conversation standard, as well as negative signals such as not clicking after display, closing the page after display, and not replying within a preset time, constructs preference data pairs, and uses preference optimization methods to iteratively fine-tune the AI ​​model, thus constructing a complete feedback loop from data collection to model optimization, so that the quality of script generation can be continuously improved as the system runs.

[0016] (6) Steps S1 to S6 of this invention form a complete distribution-connection-activation technology closed loop through data flow, from data acquisition, quality assessment and routing, AI script generation, session establishment, session status monitoring, interruption marking to intelligent activation. Each step is mutually coordinated: the quality assessment result drives the routing decision, and the routing decision determines whether to start the AI ​​analysis and subsequent connection process; the session status monitoring identifies the interruption and triggers the activation judgment; the activation script generation in the activation process depends on the topic entry point identified in step S2; the user's feedback on the activation script is fed back to the AI ​​model for iterative optimization through step S6, enabling the system as a whole to have the ability to continuously improve itself. This system-level collaborative design allows the technical effects of each link to be superimposed, realizing intelligent management of the entire process from matching and distribution to session establishment and then to session maintenance. Attached Figure Description

[0017] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments, taken with reference to the accompanying drawings: Figure 1 This is a flowchart of a method for intelligent connection establishment and activation in social conversations based on user data; Figure 2 This is a structural diagram of a social conversation intelligent connection and activation system based on user data; Figure 3 This is a schematic diagram of the structure of a computer system suitable for implementing the embodiments of the present invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] Figure 1A flowchart illustrating a method for intelligent connection building and activation in social conversations based on user profiles is shown. Figure 1 As shown, the method includes the following steps S1 to S5.

[0021] S1. In response to the connection request, obtain multi-dimensional information of the first user and multi-dimensional information of the second user.

[0022] In social networking application platforms, the first user and the second user establish an initial connection through the platform's matchmaking mechanism. When the system receives a connection request from the first user for the second user, it responds by retrieving multi-dimensional information about both the first and second users from the platform's database. This multi-dimensional information includes textual, image-based, and structured data. Textual data includes self-introductions, tags, and ideal types. Image-based data includes lifestyle photos. Structured data includes age, height, wealth level, verification status, car and property ownership status, and number of activity updates.

[0023] S2. Conduct quality assessment on the multi-dimensional data of the first user. When the quality assessment results meet the preset conditions, perform feature analysis on the multi-dimensional data of the first user and the multi-dimensional data of the second user respectively. Identify the topic entry point based on the feature analysis results. Specifically, this includes the following sub-steps S211 to S214.

[0024] S211. Obtain the raw values ​​of each data dimension from the first user's multi-dimensional profile. Data dimensions include the number of photos, authentication status, number of tags, number of activity posts, number of words in the personal introduction, ideal type, car and property status, etc. For each data dimension, the system extracts the raw value of that dimension from the first user's multi-dimensional profile.

[0025] S212. Map the raw values ​​of each data dimension to a normalized score. For count dimensions, use a saturation function to map the raw values ​​to the interval 0 to 1, as shown in the following expression: ; in, It is a saturation function. These are the original values ​​for this data dimension. This sets a preset upper limit for each dimension to prevent excessive scoring in a single dimension. For Boolean dimensions, the normalized score is 0 or 1. For textual dimensions, scores are awarded based on information content. Missing dimensions are scored as 0. Through this mapping, the original value of each data dimension is converted into a normalized score within the range of 0 to 1.

[0026] S213. The normalized scores for all data dimensions are weighted and summed to obtain the first user's quality score. Quality Score The calculation formula is: ; in Traverse all data dimensions, For the first The original values ​​of each data dimension For the first Normalized score function for each data dimension For the first The weights of each data dimension satisfy the following: .

[0027] Weight Using historical matchmaking records as samples, the features are scores across various dimensions. The labels are peer response rate or 7-day effective session rate. A gradient boosting tree (GBDT) model is trained, and the normalized feature importance of each dimension is used as the basis for evaluation. To improve the quality rating It is strongly correlated with the actual connection-building effect. The weights are refitted weekly based on the latest data to adapt to changes in user behavior and content preferences.

[0028] S214, Quality Score With preset threshold Comparison. When quality scores... Less than the preset threshold If the first user is determined to have low-quality profile, the traditional matchmaking process will proceed without consuming AI-generated resources, and the process will end. When the quality score... Greater than or equal to the preset threshold Once the quality assessment results meet the preset conditions, the subsequent AI feature analysis and script generation process will be initiated.

[0029] Furthermore, once the quality score meets the preset conditions, the system also... With low threshold and high threshold The comparison results are used for three-level differentiated routing. When Less than If the data is deemed low-quality, it will be processed using the traditional matchmaking process, without consuming AI-generated resources. Less than or equal to and Less than At that time, the process of connecting users with data and information is followed, meaning the AI ​​model generates the script for the male user. Greater than or equal to At that time, the process of data interaction and matchmaking is prioritized, that is, the AI ​​model generates the female user's side of the conversation and presents it in the form of interactive cards. and Based on historical quality scores The quantiles were initialized and fine-tuned through A / B experiments to achieve the optimal overall connection efficiency.

[0030] The system of this invention also defines whiteboard data determination rules. A user is determined to have whiteboard data if they meet one of the following conditions: Condition 1: no profile picture; Condition 2: profile picture but no real-person verification; Condition 3: profile picture and real-person verification, and simultaneously meets the conditions of no personal photos and less than [amount missing] basic information. Item, no certification information, no personal introduction, no ideal type, no tags, no updates. This is the threshold for the completeness of basic data, ranging from 1 to 10 items, with a default value of 3 items. Whiteboard data allows users to execute traditional matching processes without consuming AI-generated resources.

[0031] In step S2, feature analysis is performed on the multi-dimensional data of the first user and the multi-dimensional data of the second user respectively, and the topic entry point is identified based on the feature analysis results. Specifically, this includes the following sub-steps S221 to S224.

[0032] S221. Extract the text feature vector of textual data and the structured feature vector of structured data from the multi-dimensional data of the first user and the multi-dimensional data of the second user, respectively.

[0033] For text-based materials, semantic vectors are extracted from text fields such as self-introduction, tags, and ideal type using the pre-trained language model BERT. Each text field is encoded separately and then attention-weighted pooling is used to generate text feature vectors. The dimension is 256.

[0034] For image-based data, a pre-trained visual model is used to extract semantic tags from everyday photos, including fitness, travel, food, etc. The extracted semantic tags are then mapped to text and processed in the same way as text-based data.

[0035] For structured data, numerical features such as age, height, and wealth level are discretized by binning and then embedded and encoded. Boolean features such as authentication status and car / property ownership are directly embedded and encoded. The embedded numerical features and embedded Boolean features are concatenated and mapped to a structured feature vector through a fully connected layer. The dimension is 128.

[0036] S222. Merge the text feature vector and structured feature vector of each user to generate a unified feature vector for the corresponding user.

[0037] Text feature vectors With structured feature vectors The input is a gated fusion network, which adaptively weights and fuses the text feature vector and the structured feature vector through gate weights, outputting a unified feature vector. The fusion expression is: ; ; in For gating weights, For the sigmoid function, This is the weight matrix. For bias terms, This represents the concatenation operation of two vectors. To unify the feature vector, the dimension is 256.

[0038] S223. Based on the unified feature vector of the first user and the unified feature vector of the second user, perform cross-matching to calculate the matching degree between the first user and the second user in each data dimension.

[0039] Obtain the feature vectors of both men and women and Next, the matching degree between the first user and the second user on the same data dimension is calculated. The matching degree is the cosine similarity between the unified feature vector of the first user and the unified feature vector of the second user on that dimension: ; in Represents cosine similarity. and Iterate through each data dimension of male and female users separately. Indicates the first user in the... The data dimension and the second user in the first Matching degree across data dimensions.

[0040] S224. Based on the matching degree, extract candidate dimension pairs from all data dimensions, calculate the topic potential score for each candidate dimension pair, and select one or more candidate dimension pairs with the highest topic potential score as topic entry points.

[0041] From the matching degree matrix The data dimension pairs with the highest matching scores are extracted as candidate dimension pairs, and the topic potential score is calculated for each candidate dimension pair. Topic Potential Score The calculation formula is: ; in To determine the topic's novelty, this dimension is calculated in reverse based on the frequency of its appearance in historical rhetoric; that is, the higher the frequency of this dimension in historical rhetoric, the lower the novelty. To assess interactive potential, statistics are based on historical response rates for similar topics. , , The weighting coefficients are determined by regression analysis of historical data.

[0042] All candidate dimension pairs are sorted from highest to lowest according to their topic potential scores, and one or more candidate dimension pairs with the highest topic potential scores in the sorting results are selected as the topic entry points.

[0043] After identifying the topic entry point, the system executes routing decisions based on the matchmaking type and traffic allocation configuration. In the red envelope matchmaking scenario, the traffic allocation ratio X for data interaction matchmaking determines the proportion of traffic allocated to data interaction matchmaking, with a default value of 30%. In the message matchmaking scenario, the traffic allocation ratio Y for data message matchmaking determines the proportion of traffic allocated to data message matchmaking, with a default value of 50%. The system allocates traffic to non-whiteboard data users according to the A / B test configuration. In the red envelope matchmaking scenario, female users who hit the test group are matched with data interaction matchmaking; in the message matchmaking scenario, male users who hit the test group are matched with data message matchmaking; and users who do not hit the test group or are whiteboard data users execute the traditional matchmaking process.

[0044] Parameters X and Y can also be dynamically determined through online optimization. The system models each matchmaking type, including data interaction matchmaking, data message matchmaking, and traditional matchmaking, as one arm of a multi-armed slot machine, with the return rate or effective session rate as the reward, and uses Thompson sampling for online optimization. A Beta distribution Beta(αi, βi) is maintained for each arm i, where αi is the number of historical successes plus 1, and βi is the number of historical failures plus 1. Sampling is performed on each arm during each traffic split. The system selects the arm with the largest θi for execution and updates the corresponding arm's αi and βi upon receiving connection feedback. As samples accumulate, the frequency of each arm being selected automatically converges to a dynamic proportion reflecting its true effect. The system further introduces contextual features, including gender, age group, activity level, and data quality score Q, and uses contextual arbitrage tools such as LinUCB to determine the optimal proportion for different groups to learn differentiated functions. To ensure online stability, a default proportion is used during the cold start phase, i.e., X=30% and Y=50%, and the system switches to arbitrage decision-making after the accumulated samples reach a threshold. The proportion change in a single cycle does not exceed 10%. The system retains ε exploration flow, ε is 5%, to continuously observe the changes in the effect of each arm. When the system determines that the current connection request should execute the AI ​​connection type, it proceeds to step S3.

[0045] S3. Input the topic entry point into the AI ​​model, and the AI ​​model will generate a personalized icebreaker script for the recipient based on the topic entry point. The personalized icebreaker script will be pushed to the client of the first user. Specifically, it includes the following sub-steps S31 to S34.

[0046] S31. Generate a prompt template based on the recipient's role information.

[0047] The recipient is the user who receives the personalized icebreaker. When the connection request is initiated by the first user to the second user, the recipient is the second user. Differentiated prompt templates are generated based on the recipient's role. When the recipient is a female user, the prompt template emphasizes a curious and appreciative style; when the recipient is a male user, the prompt template emphasizes a follow-up and extension style.

[0048] S32. Retrieve reference phrases from the historical phrase database based on the topic entry point.

[0049] Based on the topic entry point identified in step S2, reference dialogues with a scenario similarity to the topic entry point that meets a preset threshold are retrieved from the historical high-response dialogue database. The historical dialogue database stores high-quality dialogue samples with high instant response rates in the past, and each dialogue is labeled with the corresponding topic dimension, scenario type, and response rate.

[0050] S33. The generated prompt template, the first user's data summary, the second user's data summary, the topic entry point, and the reference dialogue are concatenated in order to form a generation input sequence. The generated input sequence is input into the AI ​​model, which generates multiple candidate icebreaker dialogues. Several (e.g., 2) candidate icebreaker dialogues are generated each time, and each candidate icebreaker dialogue contains 1 to 2 sentences.

[0051] S34. Each candidate icebreaker script is comprehensively scored, and one or more candidate icebreaker scripts with the highest comprehensive scores are selected as the personalized icebreaker scripts to be issued.

[0052] The overall score is calculated as follows: For each candidate icebreaker, a relevance score, naturalness score, and safety score are calculated separately. These scores are then weighted and summed to obtain the overall score for each candidate icebreaker. The relevance score represents the semantic relevance of the candidate icebreaker to the information provided by both parties. The naturalness score measures whether the candidate icebreaker conforms to everyday social language habits; it is calculated using the inverse normalization of perplexity, with lower perplexity indicating greater conformity to natural language expression habits. The safety score determines whether the candidate icebreaker violates sensitive word or prohibited expression filtering rules. The candidate icebreakers with the highest overall scores are selected as the personalized icebreaker options for distribution.

[0053] During the script generation process, this invention also introduces a diversity guarantee mechanism to prevent the repeated delivery of similar scripts to the same user. A script deduplication mechanism based on similarity fingerprints calculates a fingerprint for each generated script and compares it with the fingerprint cache of recently delivered scripts to the target user; the Hamming distance between the two must be greater than a threshold. , The default value is 3; otherwise, it is considered a duplicate script and regeneration is triggered. The topic rotation mechanism records the topic dimensions recently used by the target user. When a certain topic has been used more than twice in the user's recent scripts, its topic potential score weight is automatically reduced, guiding the system to select other topic dimensions. The temperature parameter dynamic adjustment mechanism dynamically increases the temperature parameter of the generation model based on the cumulative number of scripts received by the target user, increasing the randomness of generation.

[0054] The system pushes a personalized icebreaker script to the first user's client. The first user can view the script options in the client interface and click to send it directly to the second user.

[0055] S4. After detecting the establishment of a session between the first user and the second user, start session status monitoring. The session status monitoring adopts a hierarchical time wheel as the scheduling data structure. Set a monitoring task in the hierarchical time wheel. The monitoring task is used to trigger a check on the active message count of the session when a preset delay is reached. The preset delay is recorded as the intervention time window H. Determine whether the session has reached the valid session standard based on the active message count. If the valid session standard is not reached, mark the session as an interrupted session.

[0056] A conversation is established when the first user clicks the personalized icebreaker option and sends it, and the second user replies. The standard for establishing a conversation is that both the male and female users send at least one proactive message. Proactive messages are those entered by the user themselves or sent by clicking on a card, excluding messages sent by the system.

[0057] Intervention time window The method for determining this is as follows: An event is defined as the transition from a conversation establishing a connection (i.e., each person sends one initiating message) to a valid conversation (i.e., each person sends one initiating message). The survival time of an active message is the time interval from connection establishment to achieving a valid session. Messages that have not achieved a valid session by the end of the observation period are censored. The survival function S(t) for messages that have not achieved a valid session is estimated based on the Kaplan-Meier method: ; in For a moment The size of the risk set, i.e., at time [time]. The number of valid sessions has not yet been reached. For a moment Number of valid sessions achieved. Calculate the risk rate. : ; when The moment when the rate of spontaneous conversion of a conversation into a valid conversation begins to decline significantly, i.e., the rate at which this decline begins to decrease significantly. This is the optimal time for intervention, as the marginal benefit of activation is maximized at this point. As a baseline intervention time window, and in Fine-tuning was performed within the range using AB experiments. The activation time is 5 minutes, and the activation success rate is selected. The value is used as the final configuration. The default value is 15 minutes, and the range is 5 to 60 minutes.

[0058] After a session is established, the system initiates session status monitoring. Session status monitoring uses a hierarchical time wheel as the scheduling data structure. The hierarchical time wheel includes multiple time granularity levels: second-level wheel, minute-level wheel, and hour-level wheel. The number of slots in the second-level wheel is... Time granularity per tank Seconds, covering a 60-second cycle, are used to precisely trigger short-term check events. Minute-level wheel groove count. Time granularity per tank Minutes, covering a 60-minute cycle, are used by the main scheduler for session interruption checks. Hourly wheel slot count. Time granularity per tank The time granularity level covers a 24-hour period and is used for long-term monitoring tasks spanning multiple hours. Tasks are migrated from higher to lower levels through cascading degradation at different time granularity levels.

[0059] Configure the monitoring task in the hierarchical time wheel. The system creates the monitoring task when the session is established. ,in A unique identifier for the session. To monitor the expiration time of the task. Equal to the current moment With intervention time window The sum of .

[0060] The process of inserting a monitoring task is as follows: Calculate the remaining delay time of the task to be inserted. .like The total coverage time does not exceed the second level. That is, 60 seconds, then calculate the target slot index. ,in For the slot index where the current pointer of the hierarchical time wheel is located, append the task to the task list of that slot in the second-level wheel. The coverage time exceeds the second-level wheel but does not exceed the minute-level wheel coverage time. That is, if it's 60 minutes, then calculate the target slot for the minute-level wheel in the same way and insert it. If More than the duration of minute-level wheel coverage but less than the duration of hour-level wheel coverage. If the task is inserted within a 24-hour period, it will be inserted into the corresponding slot of the hourly wheel. This hierarchical decision ensures that the time cost of task insertion does not increase with the number of tasks.

[0061] The time wheel operates at a fixed frequency The second-level wheel pointer advances, checking all tasks in the current slot and triggering their execution each time it advances. When the second-level wheel pointer advances, the monitoring task corresponding to the current slot is triggered. When the minute-level wheel pointer advances one step, the remaining delay of the task in that slot is recalculated and inserted into the second-level wheel. When the hour-level wheel pointer advances one step, the remaining delay of the task in that slot is recalculated and inserted into the minute-level wheel. When a higher-level slot expires, the task within it is demoted to the corresponding slot of the next lower level, refining the process layer by layer until it is triggered.

[0062] When a monitoring task is triggered, the system retrieves the active message count for the session. The active message count includes the number of active messages sent by the first user and the number sent by the second user. The system maintains a first user message counter and a second user message counter for each session, updating them in real-time via an event-driven approach. When a user sends an active message, the message service generates an event containing the session identifier, sender role, timestamp, and message type. The monitoring service subscribes to this event and only counts events with the message type "active message," excluding non-active messages such as those sent on behalf of the system or displayed as cards.

[0063] The system determines whether a session meets the criteria for a valid session based on the number of proactive messages sent. A valid session is defined as both the number of proactive messages sent by the first user and the number of proactive messages sent by the second user reaching a preset threshold. This preset threshold is determined using the marginal utility inflection point method: statistically analyzing historical session depths, specifically the number of proactive messages sent by men and women. sentence, The corresponding 7-day relationship retention rate This refers to the probability that both parties will send messages again within 7 days, and it is used to calculate the marginal increment of retention rate with respect to session depth. ,when The first time it drops to less than 1% The value is the effective session threshold. The default value is 5 sentences. Further support is provided for user segmentation, including age group, gender, and activity level, to perform the marginal utility inflection point analysis described above, and to set differentiated parameters. value.

[0064] If the session meets the valid session criteria, that is, the first user's active message count reaches... And the second user's active message count reached If the monitoring ends normally, the session status is marked as valid, and the pending tasks for that session in the time wheel are canceled.

[0065] If a session fails to meet the valid session criteria, i.e., the active message count does not reach the target when the monitoring task expires, the session will be transitioned to a cooling-off state. Once in the cooling-off state, the system will display the session in the cooling-off observation window. Continue monitoring the active message count of the session; the default cooldown observation window is 3 minutes. Insert a new task into the time wheel when entering the cooldown state. Upon expiration, an interruption determination is triggered. If a new active message is detected within the cooling-off observation window (i.e., the first or second user sends a new message within the cooling-off observation window), the session state is transitioned back from the cooling-off state to the active state, and continues to operate according to the original hierarchical time-round scheduling mechanism. If the cooling-off observation window expires and no new active message is detected, the session is marked as an interrupted session.

[0066] The system also uses ordered sets as an auxiliary index for session states, binning by date and state. Each bin corresponds to an independent ordered set, forming a logical partition like "Session Monitoring:{Date}:{State}". The timestamp of a session entering the current state is used as the sorting score, and the session's unique identifier is used as a set member. When a state changes, the session is removed from the ordered set corresponding to the old state and added to the ordered set corresponding to the new state. This index structure supports batch retrieval of a list of sessions in a specific state based on a timestamp range, for periodic inspection and batch processing.

[0067] The system of this invention also establishes a disaster recovery and consistency guarantee mechanism. During each state transition, the new state is synchronously written to the ordered set index and the state table of the persistent storage layer, such as a relational database, ensuring that all pending tasks in the time wheel can be reconstructed from the persistent data after an abnormal system restart. The state transition operation uses a Compare-and-Swap (CAS) mechanism to ensure idempotency. Before executing the transition, it verifies whether the current state of the session is the expected source state; if they do not match, the operation is abandoned. A fixed order is adopted: first updating the persistent index, then operating the time wheel. If the time wheel operation fails, the state in the index can be reconciled through a periodic reconciliation task, which performs a full scan every 5 minutes to detect inconsistencies and automatically repair them, re-registering tasks that exist in the index but are missing from the time wheel back to the time wheel.

[0068] This invention defines a six-state finite state machine to model the lifecycle of each session. The six states include: active, active, cooldown, interrupted, activated, and closed. State transitions strictly follow predefined conditions. Once a session is established, it enters an active state; Once the criteria for a valid conversation are met, meaning both the man and woman have initiated N messages each, the conversation transitions from an active state to a valid state, and the monitoring ends. If a time wheel trigger check is performed in the active state but the valid session criteria are not met, the session will be moved to the cooldown state. If either party in the cooling state sends a new initiative message, it will transition back to the active state. If the cooling observation window for the cooling state expires and no new messages are received, the system will migrate to the interrupt state. In the interrupted state, activation is determined, and if the conditions are met, an activation card is pushed and the process moves to the activated state. Once the number of messages between the two parties in the activated state reaches the valid session standard, the session will be moved to the valid state. Timeout after activation in the already activated state The standard for a valid conversation has not yet been met. The default time is 60 minutes. Once the process is switched to the closed state, it will terminate.

[0069] A single session enters the active state at most once during its entire lifecycle. This is ensured through irreversible state machine transitions, preventing excessive disruption to the user. After entering the interrupted state, even if the user resumes the conversation, the state is only marked as active once the message count reaches the target, without reverting to the active state, thus avoiding repeated entry into the activation determination logic.

[0070] S5. Obtain the session identifier of the interrupted session and perform activation determination based on the session identifier; when all conditions in the activation condition set are met, mark the interrupted session as a session to be activated, generate the activation message and push it to the client of the first user or the client of the second user.

[0071] When the session state machine transitions from the cooldown state to the interrupt state, the intelligent activation decision module is triggered. The system obtains the session identifier of the interrupted session and performs activation determination based on the session identifier. The activation determination executes multiple determination conditions sequentially according to a preset priority, employing a short-circuit evaluation strategy, that is, if any condition is not met, subsequent determinations are immediately terminated.

[0072] The activation condition set includes the following conditions: First, the number of activations for a session has not exceeded the preset limit. The system queries the historical activation count for the session. If the limit has been reached (a single session can only be activated once), the result is to skip the activation and no activation operation is performed. This constraint ensures the idempotency of the activation behavior.

[0073] Second, both the first and second users are online. The real-time online status of both users is queried. The online status determination uses a dual-confirmation mechanism: the first confirmation is a coarse screening, reading the user's online status cache, maintained by the IM long-connection service, with a delay of approximately 3 to 5 seconds, to initially filter online sessions between the two parties. The second confirmation is a fine screening; for sessions that pass the coarse screening, there is a delay before actual push notifications. The system queries the real-time online status interface again after a few seconds. Both parties are considered online only if they are still online. If either party is offline, the result is considered delayed, and the session is added to the delayed retry queue for reassessment on the next retry.

[0074] When activation is delayed due to user offline status, the system re-adds the activation task for that session to the time wheel scheduling queue and dynamically adjusts the retry interval using an exponential backoff strategy. The formula for calculating the retry interval is: ; in The waiting interval for the nth retry. The basic interval is set to 5 minutes. The maximum interval is set to 30 minutes. This is the current retry count, starting from 0, with a maximum of 4 retries. That is, the first retry waits 5 minutes, the second retries wait 10 minutes, the third retries wait 20 minutes, and the fourth retries wait 30 minutes. Activation is abandoned after the maximum number of retries is exceeded. Each retry involves the system re-executing the complete multi-condition joint judgment process.

[0075] Third, the receiving user's willingness to receive is not blocked. The receiving user is either the first user or the second user. Check whether the receiving user has blocked the other party or enabled Do Not Disturb mode. If so, the result is to skip, respecting the user's wishes and not performing activation.

[0076] Fourth, the cumulative number of activation cards received by the user that day has not reached the daily limit. The daily limit for a single user to receive activation cards is... The default is 3 times, implemented by maintaining a date-separated counter for each user, which automatically expires and resets at midnight each day. If the user has already reached the daily limit for the number of activation cards received, the process will be skipped to avoid excessive disruption.

[0077] When all the above conditions are met, the decision is to activate the session. The system will mark the interrupted session as a session to be activated, generate an activation message, and push it to the receiving user.

[0078] The activation script is generated using a hierarchical context encoding method. The system acquires the scenario-level information of the current session, including the connection type, trigger stage, and time attributes. The trigger stage includes initial connection establishment, interrupted activation, and response. The time attributes include weekdays or weekends, mornings, afternoons, and evenings. It acquires the profile-level information of the first and second users, including key matching points, identified topic entry points, and highlight tags in their profiles. It acquires the dialogue-level information of the current session, which is a sequence of historical messages generated in the current session. A sliding window strategy is used to extract the most recent K rounds of dialogue, where K defaults to 5 (the most recent 10 messages). Each message retains the speaker's role, message text, and sending time. The scenario-level, profile-level, and dialogue-level information are concatenated in the order from scenario-level to profile-level to dialogue-level to form a conditional input sequence, which is then used to generate the activation script using an AI model.

[0079] The context-aware response module employs differentiated generation strategies based on different triggering stages. In the initial connection stage, an icebreaker strategy is used, aiming to pique the other party's curiosity and willingness to respond. The generated style leans towards relaxed, appreciative, and questioning expressions, with constraints including avoiding sensitive topics such as age and income, refraining from overly intimate terms of address, and keeping the message length between 15 and 40 characters. In the interruption-activated stage, an activation strategy is used, aiming to rekindle the other party's interest in the conversation after an interruption. The generated style leans towards a natural transition, introducing new topics different from the previous ones, and using low-threshold response formats, including multiple-choice and yes / no questions, to reduce the other party's response cost. Constraints include prohibiting statements implying the other party's lack of response, such as "Why didn't you reply?", prohibiting urging tones, and requiring the introduction of at least one new topic. The response phase employs a continuation strategy, with the goal of extending the current topic and advancing the relationship between the two parties. The generated style leans towards responsiveness and extension, and may include appropriate self-disclosure to enhance intimacy. Constraints include semantic coherence with the last message above, not repeating content already said by the other party, and not deviating from the main theme of the current topic.

[0080] To ensure semantic coherence between the generated responses and the conversation history, the system performs a contextual coherence check on each candidate response. (Coherence score) The calculation formula is: ; semantic relevance The semantic vector cosine similarity between the candidate response and the most recent message from the other party in the dialogue layer, with a value ranging from 0 to 1, where a higher value indicates greater relevance; topic continuity. Determine whether the topic dimension involved in the candidate response overlaps with the topic dimensions of the two most recent rounds of dialogue. If there is overlap, assign a value of 1; otherwise, assign a value of 0.3. Novelty The minimum semantic distance between a candidate response and all previously received utterances for that user is normalized and mapped to a range of 0 to 1. Weights , , Dynamically adjusted based on the trigger phase: Initial connection establishment phase is set as follows: , , Interruption activation phase settings , , The response phase is set up as follows: , , .

[0081] Final ranking score of candidate responses The calculation formula is: ; in The security score is calculated using sensitive word matching and a compliance classifier. The naturalness score is calculated by normalizing the inverse of the perplexity score; a lower perplexity score indicates a better fit with natural language expression habits. (δ, ε) These are the weighting coefficients for the coherence score, security score, and naturalness score, respectively. Preferably, the system selects... The top two candidates are selected as the final output options.

[0082] The system also maintains user-level topic usage records to achieve topic dimension rotation and prevent duplication. Each time a response is generated, the main topic dimension used in that response is marked, including fitness, travel, food, pets, etc., and recorded in the user's topic usage history with a timestamp. Each topic dimension enters a cooldown period after use. The default cooldown period is 72 hours. During this period, candidate responses for this topic dimension will have their penalty points automatically deducted from the ranking. The calculation formula is: ; in This refers to the time elapsed since the topic was first used. This represents the total duration of the topic cooling-off period. The penalty for a recently used topic is the highest, close to 1, gradually decreasing to 0 over time. When all available topic dimensions in both parties' information are within their cooling-off period, the system automatically reduces the penalty weight of the topic with the longest cooling-off period, selects the dimension with the lowest penalty for reuse, and adopts different angles in the response. For example, in the fitness topic, the first question could be about exercise frequency, and the second could be about fitness goals, achieving content differentiation under the same topic through angle rotation.

[0083] When multiple interrupted sessions simultaneously meet the activation condition set and compete for the same user's activation quota, the system calculates the comprehensive priority score for each interrupted session, sorts the sessions according to their comprehensive priority scores from highest to lowest, and prioritizes activating the session with the highest comprehensive priority score. (Comprehensive Priority Score) The calculation formula is: ; in As a freshness factor, , The decay function is the number of minutes since the interruption. It ensures that sessions that have just entered the interrupted state are processed first, because the shorter the interruption time, the higher the activation success rate. , where is the matching factor, and is the normalized value of the topic potential score in the cross-analysis of the two parties' data; This is the historical response rate factor, representing the statistical value of the probability that a user actually responded after receiving an activation card in the past. Weight , , Determined through regression fitting of historical activation success data.

[0084] The system will push the activation script to the receiving users. When the activation script is pushed to the first user, the first user is the receiving user; when the activation script is pushed to the second user, the second user is the receiving user. After the activation process is completed, if the female user replies, the male user will receive the message, and the system will again call the AI ​​model to generate a follow-up script for the male user based on the previous conversation, and push the information interaction card to the male user.

[0085] S6. Collect behavioral feedback from the first and second users regarding personalized icebreaker and activation techniques, and iteratively optimize the AI ​​model based on this feedback. Step S6 specifically includes the following sub-steps S61 to S64.

[0086] S61. The action of the first user or the second user clicking on the dialogue option and sending it is taken as the first positive signal; the action of the first user or the second user replying within a preset time is taken as the second positive signal; and the action of the conversation between the first user and the second user reaching the valid conversation standard is taken as the third positive signal.

[0087] The system defines evaluation metrics at three time levels. The immediate level metrics are collected within 0 to 5 minutes after a user receives the message, including the message click-through rate (the ratio of the number of times a user clicks the message option and sends it to the number of times the message is displayed), the immediate response rate (the probability that the other party replies within 3 minutes of receiving the message), and the first-round reply character count (the number of characters in the first reply message). The mid-term level metrics are collected within 15 to 60 minutes after the conversation is established, including conversation depth (the total number of messages initiated by both parties), effective conversation conversion rate (the probability that the conversation ultimately reaches the effective conversation standard after the message is sent), and conversation duration (the time span from the first message to the last message). The long-term level metrics are collected within 1 to 7 days after the conversation is established, including the 7-day relationship retention rate (the probability that both parties will initiate a message again within 7 days), and the user negative feedback rate (the proportion of users who report, block, or mark the message as uninterested).

[0088] Overall performance score of indicators at each level The calculation formula is: ; in , , These are the normalized composite scores for immediate, medium-term, and long-term indicators, respectively. The normalized composite scores are obtained by mapping the original values ​​of each indicator to the 0-1 range and then summing them with weights. Preferably, the weights... , , The long-term effect has the highest weight, reflecting the optimization goal of building real social relationships rather than making short-term data look good.

[0089] S62. The behavior of the first user or the second user sending content without clicking after displaying the conversation is taken as the first negative signal; the behavior of the first user or the second user closing the conversation page after displaying the conversation is taken as the second negative signal; and the behavior of the first user or the second user not replying within a preset time is taken as the third negative signal.

[0090] S63. Construct preference data pairs using the first positive signal, the second positive signal, the third positive signal, the first negative signal, the second negative signal, and the third negative signal.

[0091] The system assigns a strength weight to each type of feedback signal. A positive signal retained for 7 days has a strength of 1.0, a signal sent by the user has a strength of 0.6, and a signal closed after display has a strength of 0.3. For the same set of input conditions—i.e., the same summary of information from both parties, the same topic entry point, and the same context—if the system generates multiple candidate statements and the user selects one (by clicking send), the selected statement is marked as a preference. Items that are not selected are marked as non-preference items. This forms a set of preference data pairs.

[0092] S64. Input the preference data into the AI ​​model and use the preference optimization method to iteratively fine-tune the AI ​​model.

[0093] The objective function for AI model optimization is to maximize the log probability difference between preferred and unpreferred statements, and the Direct Preference Optimization (DPO) method is adopted: ; in This represents the loss function of the direct preference optimization method. The expectation operator indicates that the mathematical expectation of the expression within the parentheses is taken. For input conditions, For preferred rhetoric, For non-preferred rhetoric, For the current model, The reference model is the previous version. For temperature hyperparameters, This is the sigmoid function.

[0094] The model is fine-tuned every two weeks, using preference data pairs accumulated over the past two weeks, with a minimum of 5000 valid data pairs required. Fine-tuning employs the QLoRA quantized low-rank adaptation method, updating the gradient of only the low-rank adaptation matrix while keeping the main model parameters frozen.

[0095] The search database is dynamically updated weekly. The database calculates the immediate response rate of each sample dialogue in the database after it has been used as a reference sample within the past 30 days. If the response rate is lower than 70% of the global average, it is marked as an inefficient sample and removed from the database. From online dialogues of the past 7 days, high-quality dialogues with an immediate response rate higher than 130% of the global average and whose corresponding conversations meet the criteria for effective conversations are added to the database. Before addition, the semantic similarity between the dialogue and existing samples in the database is calculated. If the similarity exceeds 0.85 with any existing sample, the dialogue is not added to avoid redundancy. The total capacity of the search database is capped. The limit is 10,000 entries. When the limit is exceeded after supplementation, entries will be eliminated from the list based on the historical response rate from low to high until the limit is reached.

[0096] Weights in the topic potential scoring formula , , Continuous updates are achieved through an online learning mechanism. Each generated script is considered a delivery, with the feedback (whether the other party responds) serving as a label. The input features are three sub-scores of the topic entry point used: relevance, novelty, and interaction potential. The weights are incrementally updated using online logistic regression. ; ; ; in , , The weighting coefficients before the update. , , The updated weighting coefficients, The learning rate is set to 0.01. Let cross-entropy be the loss function. , , The loss function pairs are respectively , , The partial derivatives of the weights. A weight update is performed for each feedback signal received. The cumulative change in each weight within a single day cannot exceed 10% of the initial value; if it does, the change is truncated to the boundary value.

[0097] The deployment of the updated model adopts a phased, gradual rollout strategy. During the offline evaluation phase, historical preference datasets are used to assess the new model's offline metrics, requiring the new model to maintain a preference prediction accuracy no lower than the old model. A small-scale rollout phase is conducted with 5% of users for 3 days, monitoring instant response rate and effective session conversion rate. The new model group is required to achieve at least 95% of these two metrics compared to the control group. If any core metric is significantly lower than the control group, statistical testing will be conducted. If the value is less than 0.05, an automatic rollback will occur. The expanded gray-scale phase will last for 3 days with 20% of users, requiring core metrics to be no less than 98% of the control group. The monitoring phase after full release will last for 7 days, continuously monitoring all levels of metrics, focusing on whether there is an abnormal increase in the user negative feedback rate. If the negative feedback rate exceeds 150% of the historical baseline, an automatic alarm will be triggered and a rollback will be prepared.

[0098] Thus, a complete closed-loop technology system for distribution, connection, and activation has been formed, encompassing data acquisition, quality assessment and routing, AI-generated scripts, session establishment, session status monitoring, interruption marking, intelligent activation, and feedback optimization. Each module is interconnected through data and control flows, enabling intelligent management of the entire process from matching and distribution to session establishment and maintenance.

[0099] The method of the present invention will be illustrated by an example using a specific application scenario below.

[0100] In a social networking app for strangers, User A is a female user, and User B is a male user. The system recommends User A and User B to be connected via a red envelope (cash gift). The system extracts User A's profile information and evaluates its quality. User A has a verified profile picture, 5 basic profile items, 3 lifestyle photos, vehicle ownership verification, and 10 activity posts, thus being determined to be a user with a non-blank profile. User A is matched with the A / B experimental group, and the system executes the profile interaction and connection process.

[0101] The system extracts multi-dimensional data from users A and B and performs cross-analysis. The analysis reveals that user B's activity posts include fitness photos, and user A's tags include "fitness enthusiast," thus identifying fitness as a matching point. The AI ​​model generates a female user-side message based on the data from both users: Option 1 is "Your fitness photos show great physique; I'm a fitness newbie and would like to ask for advice"; Option 2 is "Your cooking in your activity posts is amazing; I wonder if I'll ever get to try it." The system pushes an interactive data card to user A, who clicks option 1 to send it.

[0102] After receiving the message, the system generates a response for the male user: Option 1 is "Just practice more, I can teach you"; Option 2 is "You have a good figure, you probably exercise regularly." User B clicks Option 1 to reply, and the conversation begins.

[0103] The system initiated session status monitoring. After 15 minutes, a check revealed that User A and User B had each sent only two initiating messages, failing to meet the criteria for a valid session. The system determined that both users were online and sent an activation card to User A. The card titled "Disconnected? Try the following materials to rekindle your connection:" and provided two activation script options. User A clicked the activation script and sent it. After User B received the message, the system generated a follow-up script, and the session was successfully activated.

[0104] Further reference Figure 2 As an implementation of the above method, this invention also proposes a social conversation intelligent connection and activation system based on user data, which can be applied to various electronic devices. This user data-based social conversation intelligent connection and activation system 200 includes the following modules: Data acquisition module 210 is configured to acquire multi-dimensional data of the first user and multi-dimensional data of the second user in response to connection establishment request; The quality assessment and routing module 220 is configured to perform quality assessment on the multi-dimensional data of the first user. When the quality assessment result meets the preset conditions, feature analysis is performed on the multi-dimensional data of the first user and the multi-dimensional data of the second user respectively, and the topic entry point is identified based on the feature analysis result. The AI ​​dialogue generation module 230 is configured to input the topic entry point into the AI ​​model, and the AI ​​model will generate a personalized icebreaker dialogue for the recipient based on the topic entry point, and push the personalized icebreaker dialogue to the client of the first user. The session status monitoring module 240 is configured to start session status monitoring after detecting the establishment of a session between the first user and the second user. The session status monitoring adopts a hierarchical time wheel as the scheduling data structure. A monitoring task is set in the hierarchical time wheel. The monitoring task is used to trigger a check of the active message count of the session when a preset delay is reached. The active message count is used to determine whether the session has reached the valid session standard. If the valid session standard has not been reached, the session is marked as an interrupted session. The intelligent activation decision module 250 is configured to obtain the session identifier of the interrupted session, perform activation determination based on the session identifier, and when all conditions in the activation condition set are met, mark the interrupted session as a session to be activated, generate an activation script and push it to the client of the first user or the client of the second user.

[0105] The following is for reference. Figure 3 It shows a schematic diagram of the structure of a computer system 300 suitable for implementing terminal devices or servers of the present invention. Figure 3 The terminal device or server shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of the present invention.

[0106] like Figure 3As shown, the computer system 300 includes a central processing unit (CPU) 301, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 302 or programs loaded from storage section 308 into random access memory (RAM) 303. The RAM 303 also stores various programs and data required for the operation of the computer system 300. The CPU 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0107] The following components are connected to I / O interface 305: an input section 306 including a keyboard, mouse, etc.; an output section 307 including a liquid crystal display (LCD) and speakers, etc.; a storage section 308 including a hard disk, etc.; and a communication section 309 including a network interface card such as a LAN card and a modem, etc. The communication section 309 performs communication processing via a network such as the Internet. A drive 310 is also connected to I / O interface 305 as needed. A removable medium 311, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 310 as needed so that computer programs read from it can be installed into storage section 308 as needed.

[0108] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 309, and / or installed from removable medium 311. When the computer program is executed by central processing unit (CPU) 301, it performs the functions defined in the methods of this invention. It should be noted that the computer-readable medium described in this invention can be a computer-readable signal medium or a computer-readable medium or any combination thereof. The computer-readable medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this invention, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on a computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0109] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0110] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0111] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention is not limited to the specific combination of the above-described technical features, but also includes other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in this invention.

Claims

1. A method for intelligent connection establishment and activation in social conversations based on user data, characterized in that, Includes the following steps: S1. In response to the connection request, obtain multi-dimensional information of the first user and multi-dimensional information of the second user; S2. Perform quality assessment on the multi-dimensional data of the first user. When the quality assessment result meets the preset conditions, perform feature analysis on the multi-dimensional data of the first user and the multi-dimensional data of the second user respectively, and identify the topic entry point based on the feature analysis result. S3. Input the topic entry point into the AI ​​model, and the AI ​​model generates a personalized icebreaker script for the recipient based on the topic entry point. The personalized icebreaker script is then pushed to the client of the first user. S4. After detecting the establishment of a session between the first user and the second user, start session status monitoring. The session status monitoring adopts a hierarchical time wheel as the scheduling data structure. Set a monitoring task in the hierarchical time wheel. The monitoring task is used to trigger a check of the active message count of the session when a preset delay is reached. Determine whether the session has reached the valid session standard based on the active message count. If the valid session standard is not reached, mark the session as an interrupted session. S5. Obtain the session identifier of the interrupted session, and perform an activation determination based on the session identifier; When all conditions in the activation condition set are met, the interrupted session is marked as a session to be activated, an activation message is generated and pushed to the client of the first user or the client of the second user.

2. The method for intelligent connection establishment and activation of social conversations based on user data according to claim 1, characterized in that, In step S2, the quality of the multi-dimensional data of the first user is assessed, which specifically includes the following sub-steps: S211. Obtain the original values ​​of each data dimension in the multi-dimensional data of the first user; S212. Map the original value of each of the data dimensions to a normalized score; S213. The normalized scores of all the data dimensions are weighted and summed to obtain the quality score of the first user. S214. Compare the quality score with a preset threshold. When the quality score is greater than or equal to the preset threshold, determine that the quality assessment result meets the preset condition.

3. The method for intelligent connection establishment and activation of social conversations based on user data according to claim 1, characterized in that, In step S2, feature analysis is performed on the multi-dimensional data of the first user and the multi-dimensional data of the second user, respectively. The topic entry point is identified based on the feature analysis results. This specifically includes the following sub-steps: S221. Extract the text feature vector of textual data and the structured feature vector of structured data from the multi-dimensional data of the first user and the multi-dimensional data of the second user, respectively. S222. Merge the text feature vector of each user with the structured feature vector to generate a unified feature vector for the corresponding user; S223. Based on the unified feature vector of the first user and the unified feature vector of the second user, perform cross-matching to calculate the matching degree between the first user and the second user in each data dimension. S224. Based on the matching degree, extract candidate dimension pairs from all data dimensions, calculate the topic potential score for each candidate dimension pair, and select one or more candidate dimension pairs with the highest topic potential score as the topic entry point.

4. The method for intelligent connection establishment and activation of social conversations based on user data according to claim 1, characterized in that, In step S3, the topic entry point is input into the AI ​​model, and the AI ​​model generates a personalized icebreaker script for the recipient based on the topic entry point. This specifically includes the following sub-steps: S31. Generate a prompt template based on the recipient's role information; S32. Retrieve reference scripts from the historical script database based on the topic entry point; S33. The first user's data summary, the second user's data summary, the topic entry point, the generated prompt template, and the reference dialogue are concatenated into a generated input sequence. The generated input sequence is then input into the AI ​​model, which generates multiple candidate icebreaker dialogues. S34. Each candidate icebreaker is comprehensively scored, and one or more candidate icebreakers with the highest comprehensive score are selected as personalized icebreakers to be issued.

5. The method for intelligent connection establishment and activation of social conversations based on user data according to claim 1, characterized in that, In step S4, the hierarchical time wheel includes multiple time granularity levels, and each time granularity level achieves the gradual migration of tasks from higher levels to lower levels through cascading degradation. When the number of active messages in the session does not reach the valid session standard, the session state will be transitioned to a cooling-off state. Continue monitoring the active message count of the session within the cooling observation window; If a new active message is detected within the cooling observation window, the session state is transitioned back from the cooling state to the active state. If the cooling observation window expires and no new active message is detected, the session is marked as interrupted.

6. The method for intelligent connection establishment and activation of social conversations based on user data according to claim 1, characterized in that, In step S5, the set of activation conditions includes the following condition: the number of times the session is activated has not exceeded a preset limit; Both the first user and the second user are online. The recipient's willingness to receive the activation message was not blocked; the cumulative number of activation cards received by the recipient on that day did not reach the daily limit; When all conditions in the activation condition set are met and multiple interrupted sessions compete for the same user's activation quota, the comprehensive priority score of each interrupted session is calculated, and the multiple interrupted sessions are sorted from high to low according to the comprehensive priority score. The interrupted session with the highest comprehensive priority score is activated first.

7. The method for intelligent connection establishment and activation of social conversations based on user data according to claim 1, characterized in that, It also includes step S6, collecting behavioral feedback from the first user and the second user regarding the personalized icebreaker and activation techniques, and iteratively optimizing the AI ​​model based on the behavioral feedback, specifically including the following sub-steps: S61. The behavior of the first user or the second user clicking on the dialogue option and sending it is taken as a first positive signal; the behavior of the first user or the second user replying within a preset time is taken as a second positive signal; and the behavior of the conversation between the first user and the second user reaching the valid conversation standard is taken as a third positive signal. S62. The behavior of the first user or the second user sending content by entering it without clicking after displaying the conversation is taken as a first negative signal; the behavior of the first user or the second user closing the conversation page after displaying the conversation is taken as a second negative signal; and the behavior of the first user or the second user not replying within a preset time is taken as a third negative signal. S63. Construct a preference data pair using the first positive signal, the second positive signal, the third positive signal, the first negative signal, the second negative signal, and the third negative signal; S64. Input the preference data into the AI ​​model and use the preference optimization method to iteratively fine-tune the AI ​​model.

8. A social conversation intelligent connection establishment and activation system based on user data, used to implement the method as described in any one of claims 1 to 7, characterized in that, The system includes: The data acquisition module is configured to respond to the connection establishment request and acquire multi-dimensional data of the first user and multi-dimensional data of the second user. The quality assessment and routing module is configured to perform quality assessment on the multi-dimensional data of the first user, and when the quality assessment result meets the preset conditions, perform feature analysis on the multi-dimensional data of the first user and the multi-dimensional data of the second user respectively, and identify the topic entry point based on the feature analysis result; The AI ​​dialogue generation module is configured to input the topic entry point into the AI ​​model, and the AI ​​model generates a personalized icebreaker dialogue for the recipient based on the topic entry point, and pushes the personalized icebreaker dialogue to the client of the first user. The session status monitoring module is configured to initiate session status monitoring upon detecting the establishment of a session between the first user and the second user. The session status monitoring uses a hierarchical time wheel as the scheduling data structure. A monitoring task is set in the hierarchical time wheel, which is used to trigger a check of the active message count of the session when a preset delay is reached. The active message count is used to determine whether the session has reached the valid session standard. If the valid session standard is not reached, the session is marked as an interrupted session. The intelligent activation decision module is configured to obtain the session identifier of the interrupted session, perform activation determination based on the session identifier, and when all conditions in the activation condition set are met, mark the interrupted session as a session to be activated, generate an activation message and push it to the client of the first user or the client of the second user.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent connection and activation method for social conversations based on user data as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent connection and activation method for social conversations based on user data as described in any one of claims 1 to 7.

Images