Asynchronous social matching method and apparatus based on user customized ai agents

Abstract

The application discloses an asynchronous social matching method and device based on user customized AI agents, and relates to the technical field of artificial intelligence, comprising: collecting and obtaining user individual data in advance, continuously correcting user portrait data, and generating individualized AI agent models corresponding to the user; and training to obtain trained individualized AI agents; the trained individualized AI agents are joined in a social pool through a cross-agent communication engine adopting a federated learning architecture, and perform asynchronous conversations with individualized AI agents of other users; real-time analysis of conversation records is performed, and a triple matching algorithm based on interest overlap, conversation depth and emotional positive degree is adopted to pre-screen matched potential social objects; a real person connection request is initiated by selecting the potential social object according to a user operation instruction, and friend connection is performed. The application has the advantages of improving social matching efficiency, ensuring matching authenticity, protecting user privacy, and realizing low-friction real person social connection.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an asynchronous social matching method, apparatus, mobile terminal, and storage medium based on user-customized AI agents. Background Technology

[0002] Existing social networking apps (such as Tinder / Soul) rely on users manually filtering and matching, which leads to inefficiency and questionable authenticity. AI chatbots (such as Replika) only provide one-way human-computer interaction and lack a real-person social conversion path. Group chat tools (such as Discord) suffer from information overload, making it difficult for users to efficiently filter potential social partners.

[0003] In other words, existing social software technologies cannot simultaneously solve the three major problems of ensuring authenticity, social efficiency, and privacy security (avoiding direct exposure of real people).

[0004] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0005] The technical problem this invention aims to solve is to address the shortcomings of existing technologies by providing an asynchronous social matching method, device, mobile terminal, and storage medium based on user-customized AI agents. This invention can construct AI agents that realistically reflect user personality and achieve efficient pre-screening through large-scale inter-agent social interaction, ultimately completing low-friction real-person social conversion. This improves social efficiency, increases the matching accuracy of categorized users, and provides convenience for users. This invention has the advantages of improving social matching efficiency, ensuring matching authenticity, protecting user privacy, and achieving low-friction real-person social connections.

[0006] The technical solution adopted by this invention to solve the problem is as follows: An asynchronous social matching method based on a user-customized AI agent, comprising: Pre-collect user personality data, including user character and preferences, and continuously refine user profile data to generate a personalized AI agent model corresponding to the user; then train the user's personalized AI agent model to obtain the trained personalized AI agent. The trained personalized AI agent is added to the social pool through a cross-agent communication engine with a federated learning architecture to conduct asynchronous dialogues with other users' personalized AI agents. During asynchronous conversations, the conversation records are analyzed in real time, and a triple matching algorithm based on interest overlap, conversation depth, and positive emotion is used to pre-screen potential social partners. Based on the pre-selected potential social partners, the system receives user operation instructions to select the potential social partners and initiate a real-person connection request to establish a friend connection.

[0007] The asynchronous social matching method based on user-customized AI agents includes the following steps: first, pre-collecting and acquiring user personality data including user character and preferences, continuously refining user profile data to generate a personalized AI agent model corresponding to the user; and then training the user's personalized AI agent model to obtain the trained personalized AI agent. In advance, user personality data, including user character, hobbies, and values, is continuously collected through a dynamic questionnaire engine and natural language dialogue, and user profile data is continuously revised and improved. By revising and improving user profile data, a personalized AI agent model is constructed to truly reflect the user's personality. The personalized AI agent model is trained through the agent training module to obtain the trained personalized AI agent.

[0008] The asynchronous social matching method based on user-customized AI agents, wherein the step of adding the trained personalized AI agent to the social pool through a cross-agent communication engine using a federated learning architecture to conduct asynchronous dialogue with other users' personalized AI agents includes: The trained personalized AI agent is automatically logged in and added to the social pool through a cross-agent communication engine with a federated learning architecture; Personalized AI agents added to the social pool automatically engage in asynchronous dialogues with other users' personalized AI agents. Asynchronous dialogue employs a federated learning architecture, where the training of the user's personalized AI agent model is performed on the user's local device, and only the processed dialogue vectors, which do not contain the original dialogue content, are uploaded to the cloud.

[0009] The asynchronous social matching method based on user-customized AI agents includes the following steps: during asynchronous dialogue, real-time analysis of dialogue records and the use of a triple matching algorithm based on interest overlap, dialogue depth, and positive sentiment to pre-screen potential social partners: During asynchronous dialogues, the dialogue records are analyzed in real time through a preset dialogue analysis filter. A triple matching algorithm based on interest overlap, dialogue depth, and positive emotion is used to calculate the matching degree of interest overlap, dialogue depth, and positive emotion of the dialogue records respectively. Social signals with matching degree reaching the first preset value are extracted, and dialogues with matching degree below the second preset value are automatically eliminated using a social signal decay model, thus pre-screening potential social objects with matching degree meeting preset conditions.

[0010] The asynchronous social matching method based on user-customized AI agents, wherein the step of receiving user operation instructions to select the potential social object based on pre-screened matching potential social objects and initiating a real-person connection request to establish a friend connection includes: Based on the pre-screened potential social partners, a social recommendation report is generated, recommending that users can connect with real people through the matched potential social partners. Upon receiving user operation instructions and initiating a real-person connection request based on the social recommendation report, the system initiates a friend connection application and controls the establishment of a real-person chat channel through the real-person connection gateway.

[0011] The asynchronous social matching method based on user-customized AI agents, wherein the steps of pre-collecting and acquiring user personality data including user character and preferences, continuously refining user profile data, generating a personalized AI agent model corresponding to the user, and training the user's personalized AI agent model to obtain the trained personalized AI agent further include: A core personality baseline builder is pre-set to guide users through a series of in-depth personality tests using a mature psychological model, and to generate a core personality baseline model corresponding to the user by combining the user's decision-making preferences and value ranking in a specified situation. The system collects user personality data through a dynamic questionnaire engine and natural language dialogue function, and then uses a dynamic profile corrector to correct the user's surface personality or interest preferences in a specified context, generating a personalized AI agent model corresponding to the user. The generated personalized AI agent model corresponding to the user is trained using a multi-layer personality model trainer to obtain the trained personalized AI agent; wherein, the multi-layer personality model trainer includes core personality layer training and surface personality layer training; the core personality layer training focuses on maintaining baseline stability, and the surface personality layer training focuses on adapting to the social environment; During the training of the core personality layer, the personalized AI agent model corresponding to the user is trained using a core personality baseline model corresponding to the user, and the training of the core personality baseline model is given first priority. During the training of the surface personality layer, the user's personalized AI agent model is trained using data collected based on the dynamic profile corrector and the interaction data of the personalized AI agent in the social pool, and a second priority is set for the training of the surface personality layer. When training a user's personalized AI agent model, a newly added personality drift detector periodically compares the model's current behavioral patterns, conversational style, and emotional tendencies with the core personality baseline model. The comparison dimensions include: Value consistency is assessed by comparing the personalized AI agent model's tendencies on specified ethical or decision-making issues with those of the core personality baseline model. Emotional expression stability is assessed by comparing whether the emotional expression of the personalized AI agent model in different situations matches the range of collected user's normal emotional data. Stability of core interest preferences; compare whether the personalized AI agent model shows significant deviations in the expression of core interest domains. Consistency of language style: Whether the personalized AI agent model's word choice, sentence structure, humor, etc., are consistent with the user's stated core personality baseline model.

[0012] The asynchronous social matching method based on user-customized AI agents, wherein the step of adding the trained personalized AI agent to the social pool through a cross-agent communication engine using a federated learning architecture to conduct asynchronous dialogue with other users' personalized AI agents further includes: When an asynchronous dialogue is conducted, if the behavior of the personalized AI agent is detected to deviate from the core personality baseline of the corresponding core personality baseline model by more than a predetermined threshold, the personality calibration mechanism will be triggered. When the personality calibration mechanism is activated for calibration, it includes: First (mild) drift calibration: When the deviation drift is at the first level (lower), the training weight of the core personality layer in the multi-layer personality model trainer is automatically adjusted, or baseline data is introduced for continued learning, so that the personalized AI agent behavior is aligned with the core personality baseline of the core personality baseline model. Second (moderate) drift calibration: When the deviation drift reaches the second level (moderate level), the control sends a notification to the user terminal, indicating that the personalized AI agent has personality drift, and prompts the user to confirm or correct certain behavioral preferences of the agent; The third (severe) drift calibration, when the deviation drift reaches the third level (too high), controls the suspension of the personalized AI agent's social activities, forces personality reshaping or baseline calibration, and resets some training parameters of the personalized AI agent to ensure that the agent returns to the user's true personality.

[0013] An asynchronous social matching device based on a user-customized AI agent, wherein the device includes: The personalized AI agent generation module is used to pre-collect user personality data, including user character and preferences, and continuously correct user profile data to generate a personalized AI agent model corresponding to the user; and to train the user's personalized AI agent model to obtain the trained personalized AI agent. The cross-agent communication engine social module is used to add trained personalized AI agents to the social pool through the cross-agent communication engine with a federated learning architecture, so as to conduct asynchronous dialogues with other users' personalized AI agents. The dialogue analysis module is used to analyze dialogue records in real time during asynchronous dialogues and uses a triple matching algorithm based on interest overlap, dialogue depth, and positive sentiment to pre-screen potential social partners. The connection control module is used to receive user operation instructions based on pre-screened matching potential social objects, select the potential social objects to initiate a real-person connection request, and establish a friend connection.

[0014] A mobile terminal includes a memory and one or more programs, wherein one or more programs are stored in the memory and configured to be executed by one or more processors, the one or more programs comprising the method for performing any one of the methods.

[0015] A computer-readable storage medium, wherein, when instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform any of the methods described above.

[0016] The beneficial effects of this invention are as follows: This invention provides an asynchronous social matching method, device, mobile terminal, and storage medium based on user-customized AI agents. This invention performs thorough pre-screening and matching through AI agents, greatly improving the success rate and quality of real-person social interaction and providing convenience for users.

[0017] This invention constructs a personalized AI agent that can accurately map a user's personality and preferences, enabling the personalized AI agent to autonomously conduct asynchronous social interactions in the background. This allows for the pre-screening of potential partners and ultimately achieves low-friction real-person social connections. It has the advantages of improving social matching efficiency, ensuring the authenticity of matching, protecting user privacy, and achieving low-friction real-person social connections. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the asynchronous social matching method based on user-customized AI agents provided in Embodiment 1 of the present invention.

[0020] Figure 2The present invention provides a schematic diagram of an embodiment of an asynchronous social matching device based on a user-customized AI agent.

[0021] Figure 3 This is a block diagram illustrating the internal structure of a mobile terminal provided in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below 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 are not intended to limit the invention.

[0023] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0024] Traditional social networking apps (such as Tinder / Soul) rely on manual user filtering and matching, which is inefficient and raises questions about authenticity. AI chatbots (such as Replika) only provide one-way human-computer interaction and lack a real-person social conversion path; group chat tools (such as Discord) suffer from information overload, making it difficult for users to efficiently filter potential social partners. In other words, current technologies cannot simultaneously solve the three major problems of ensuring authenticity (requiring deep personality modeling), social efficiency (requiring parallel processing of thousands of conversations), and privacy and security (avoiding direct exposure of real people).

[0025] For example, in the operation of a mobile-based social application, user A's AI agent joins the social pool and engages in asynchronous dialogues with other agents. Due to the lack of a deep personality modeling mechanism, the agent's output stance when discussing value-related topics deviates from user A's actual preferences. Simultaneously, with thousands of agents in the social pool, system response delays occur when processing dialogue requests, leading to interruptions in dialogue continuity. Furthermore, without an effective implementation of the federated learning architecture, user A's geographical location and contact information are indirectly leaked through dialogue vectors. In this scenario, the problems manifest as distorted matching results, degraded user interaction experience, and frequent privacy complaints, ultimately affecting the overall stability of the system.

[0026] If the above problems are not addressed, the reliability of social matching systems will gradually decrease, and users' trust in the platform may continue to weaken; the system's scalability will be limited, making it difficult to adapt to the dynamic growth of the user base; privacy and security vulnerabilities may lead to increased compliance risks; and ultimately, the sustainable development of the entire social ecosystem will be hindered.

[0027] To address this, this application proposes an asynchronous social matching method based on a user-customized AI agent. For example... Figure 1 As shown in Embodiment 1 of the present invention, an asynchronous social matching method based on a user-customized AI agent includes the following steps: S100. Pre-collect and acquire user personality data, including user character and preferences, and continuously revise user profile data to generate a personalized AI agent model corresponding to the user; and train the user's personalized AI agent model to obtain the trained personalized AI agent. The user-customized AI agent in this embodiment of the invention refers to an artificial intelligence entity trained based on personalized data such as the user's personality, preferences, and values, capable of simulating the user's social interactions; used to represent the user in preliminary social exploration and screening, in order to improve social efficiency and matching accuracy.

[0028] The asynchronous social matching method in this embodiment refers to a social method that does not require users to be online simultaneously, but rather allows them to interact and match in the background through their respective AI agents. This invention allows users to explore social interactions in a non-real-time state, thereby reducing the limitations of time synchronization and improving the flexibility and reach of social interactions.

[0029] In this specific implementation, the user personality data acquired refers to a data set reflecting the user's personal characteristics, including but not limited to the user's personality traits, interests, values, and behavioral preferences. This user personality data is the foundation for building and training personalized AI agent models.

[0030] The user profile data in this embodiment refers to structured data formed by integrating, analyzing, and abstracting user personality data, which is used to comprehensively describe user characteristics; this data provides a basis for the generation and correction of personalized AI agent models.

[0031] The personalized AI agent model in this embodiment refers to a machine learning model built based on user personality data and user profile data, used to drive customized AI agent behavior for users. This model, through training, can learn and simulate the user's social style and decision-making patterns.

[0032] In the specific implementation of step S100, user personality data, including user character and preferences, is collected in advance. This data can be collected in various ways. For example, a dynamic questionnaire can be used to allow users to actively fill out a questionnaire about their interests, values, and social habits. Furthermore, embodiments of the present invention can also analyze user behavior patterns on public social media, such as their posted content and interaction records, to construct a preliminary user profile. The user profile data constructed in these embodiments will be continuously revised based on subsequent user interactions or feedback to ensure its accuracy and timeliness. Based on this data, embodiments of the present invention can construct a personalized AI agent model capable of simulating user social behavior.

[0033] The personalized AI agent model described in this embodiment can be trained through supervised learning or reinforcement learning. For example, by using the user's historical dialogue records as training data, the personalized AI agent model can learn the user's dialogue style, expression habits and decision preferences, thereby obtaining a trained personalized AI agent that can represent the user in social interactions.

[0034] In this embodiment of the invention, a personalized AI agent that accurately reflects the user's personality can be created and trained first through user profile construction and agent training. The personalized AI agent in this embodiment continuously collects and corrects user profile data through a dynamic questionnaire engine and natural language dialogue, ensuring the agent's personalization and high fidelity.

[0035] S200: The trained personalized AI agent is added to the social pool through a cross-agent communication engine with a federated learning architecture to conduct asynchronous dialogues with other users' personalized AI agents. The federated learning architecture in this embodiment refers to a distributed machine learning paradigm that allows multiple participants to collaboratively train a shared model without sharing the original data. Under this architecture, personalized AI agent model training takes place on the user's local device, and only model updates or processed aggregated information are uploaded to the central server, thus ensuring the privacy and security of user data.

[0036] In this embodiment, the cross-agent communication engine refers to a software module or system responsible for managing and coordinating communication between personalized AI agents of different users. Under the federated learning architecture, this cross-agent communication engine ensures that asynchronous dialogue between agents can be conducted efficiently and securely.

[0037] In this embodiment, the social pool refers to a virtual interactive space comprised of personalized AI agents from multiple users. Within this social pool space, the personalized AI agents can freely engage in asynchronous dialogues and explore potential social connections.

[0038] The asynchronous dialogue in this embodiment is a non-real-time, message-passing-based interaction process between AI agents. After an agent sends a message, it does not need to wait for an immediate reply from the other agent; the other agent's personalized AI agent can respond at its convenience.

[0039] In the specific implementation of step S200, the trained personalized AI agent is activated and deployed into the social environment. The user's personalized AI agent connects to a virtual social pool via a communication interface. Within this social pool, there are personalized AI agents belonging to other users. These user AI agents can exchange messages asynchronously, i.e., engage in non-real-time message exchange. The application of the federated learning architecture ensures that the user's original sensitive data does not leave their local device while the agents are interacting and training the model, thus protecting data privacy. The cross-agent communication engine manages the routing, storage, and transmission of these asynchronous conversations, ensuring smooth information exchange between different agents.

[0040] In this embodiment, the trained personalized AI agent is added to the social pool via a cross-agent communication engine to engage in asynchronous dialogues with other users' AI agents. The cross-agent communication engine employs a federated learning architecture, ensuring that the agent model is trained on a local device and only the dialogue vectors are uploaded to the cloud, thereby protecting user privacy and security.

[0041] S300, during asynchronous dialogue, analyzes dialogue records in real time and uses a triple matching algorithm based on interest overlap, dialogue depth, and positive emotion to pre-screen potential social partners. In this step of the implementation, the interest overlap refers to the degree of similarity between two AI agents (representing two users) in terms of interests, hobbies, and topic preferences. The interest overlap metric is used to assess the common topic base between potential social partners.

[0042] In this embodiment, dialogue depth refers to the depth and complexity of the dialogue content between AI agents. The dialogue depth metric can reflect the effort invested by both parties in the communication process and their level of interest in each other; for example, whether the dialogue remains at the level of superficial pleasantries or involves a deeper exchange of viewpoints.

[0043] The positive sentiment refers to the emotional tendency expressed during dialogues between AI agents. The positive sentiment metric is used to assess the positivity and friendliness of the dialogue atmosphere, helping to identify potential social contacts with a positive interactive experience.

[0044] The triple matching algorithm used in this embodiment is an algorithm that comprehensively considers three dimensions—interest overlap, dialogue depth, and positive emotion—to evaluate potential social partners. Through multi-dimensional analysis, the accuracy and comprehensiveness of the matching are improved.

[0045] The potential social partners refer to other users who are pre-screened through AI agent asynchronous dialogue and triple matching algorithms and may establish real-person social relationships with the user.

[0046] In the specific implementation of step S300, during the asynchronous dialogue conducted by the personalized AI agent, this embodiment of the invention continuously analyzes the dialogue content in real time. During analysis, valuable social signals are extracted from the dialogue. For example, text analysis technology can be used to identify common topics mentioned by both parties in the dialogue to assess the degree of overlap in interests. The number of exchanges, message length, and depth of topics can be used to measure the depth of the dialogue. Simultaneously, sentiment analysis technology can be used to determine the proportion of positive and negative emotions in the dialogue, thereby assessing the positive sentiment level. Based on these analysis results, the triple matching algorithm of this invention comprehensively evaluates all potential social objects participating in the dialogue, filtering out those that show a high degree of matching across multiple dimensions.

[0047] In this embodiment of the invention, during the personalized AI agent dialogue process for each user, a dialogue analysis filter is used to analyze the dialogue records in real time, extract high-value social signals based on a triple matching algorithm, and automatically eliminate low-value dialogues using a social signal attenuation model, such as detecting the repetition rate of polite remarks, thereby efficiently pre-screening potential social objects with high matching degree.

[0048] S400: Based on the pre-selected matching potential social objects, receive the user's operation instruction to select the potential social object to initiate a real-person connection request and establish a friend connection.

[0049] In this embodiment, a real-person connection request refers to a request initiated by the user to a potential social contact based on the AI ​​agent's pre-screening process. This request represents a crucial step in transforming AI-assisted social interaction into real-person social interaction.

[0050] In the specific implementation of step S300, a pre-screened list of potential social contacts is presented to the user. Based on this information, such as a summary of the proxy conversation or a match rating, the user can decide whether to establish a real social connection with a potential social contact. When a user selects a potential social contact and issues a connection command, the system of this embodiment processes the request and sends a real-person connection invitation to the selected potential social contact. Once the other party accepts, a real-person connection is established between the two parties, such as establishing a friendship in a social application or opening a real-person chat channel.

[0051] In its specific implementation, this invention can use a real-person connection gateway to manage friend requests and control permissions. Once the other party agrees, a real-person chat channel is established, achieving a low-friction and secure transition from virtual interaction to real-person social interaction. The entire process of this invention, through pre-screening by an AI agent and data analysis, significantly reduces the time cost of real-person social interaction, improves matching quality, and solves the issues of authenticity, efficiency, and privacy security in traditional social interactions.

[0052] The following example will provide a more detailed explanation of the above technical solution: For example, when user A wants to meet friends with similar interests and values ​​through a social platform, the system of this invention allows user A to input their personal information, hobbies, professional background, and expectations for social partners through the platform's interface. The system then uses this input, combined with user A's historical behavioral data on the platform, such as browsing history and liked content, to construct and continuously refine user profile data. Based on this user profile data, the invention generates a personalized AI agent model that can simulate user A's social style and preferences. Subsequently, the personalized AI agent model is trained on user A's local device, learning user A's dialogue patterns, emotional expression, and decision-making logic, ultimately resulting in a trained personalized AI agent.

[0053] The trained personalized AI agent is then joined to a social pool of thousands of other users' personalized AI agents via a cross-agent communication engine, using a federated learning architecture. Within this pool, user A's AI agent engages in asynchronous dialogues with other users' AI agents. For example, user A's AI agent might send a discussion invitation to another agent on a trending topic or respond to a conversation initiated by another agent. These dialogues are asynchronous, meaning user A doesn't need to be online in real-time; their AI agent autonomously communicates with other agents in the background.

[0054] While these asynchronous dialogues are taking place, the system of this embodiment analyzes the dialogue records between all agents in real time. For example, when user A's AI agent is having a dialogue with user B's AI agent, this embodiment analyzes the dialogue content, identifies keywords and topics mentioned by both parties, and calculates the degree of interest overlap. Simultaneously, this invention also evaluates the duration of the dialogue, the frequency of message exchange, and the depth of the topics to determine the dialogue depth. Furthermore, this invention performs sentiment analysis on the dialogue text, identifying positive or negative emotional expressions to assess the positive sentiment level. Based on these three indicators—interest overlap, dialogue depth, and positive sentiment level—a triple matching algorithm comprehensively scores the matching degree between user A's AI agent and user B's AI agent. If user B's AI agent shows a high matching degree with user A's AI agent in these dimensions, user B will be pre-screened as a potential social partner of user A.

[0055] In a specific implementation of this invention, a pre-screened list of potential social contacts, such as User B, is presented to User A. User A can view a summary of the conversation between their AI agent and User B's AI agent, as well as a matching score provided by the system. Based on this information, User A decides whether to establish a real social connection with User B. If User A selects User B and clicks the "Initiate Real-Person Connection" button, the system in this embodiment of the invention will receive the operation instruction. Subsequently, the system will send a real-person connection request to User B. Once User B accepts the request, a friend connection will be established between User A and User B, for example, by opening a private real-person chat channel for them within the platform.

[0056] As can be seen from the above embodiments, user A does not need to invest a lot of time in manual screening and initial communication; their AI agent has efficiently completed large-scale pre-screening, solving the problem of low social efficiency. Furthermore, since the AI ​​agent is trained based on user A's real personality data, its performance in asynchronous dialogue can better reflect user A's true personality, thereby improving the authenticity of the match. The application of the federated learning architecture ensures the privacy and security of user data, avoiding the risk of raw data leakage. Ultimately, user A only needs to make a final decision based on the AI ​​agent's pre-screening, achieving low-friction, real-person social conversion.

[0057] Compared to existing social software that relies on users manually filtering and matching, the method of this invention significantly improves the efficiency of social matching by introducing personalized AI agents for large-scale asynchronous dialogue. Users do not need to spend a lot of time browsing profiles or engaging in inefficient initial communication; their AI agents can process thousands of potential social contacts in parallel in the background, performing initial interactions and filtering. For example, in the example above, user A's AI agent can simultaneously converse with multiple agents in the social pool, without user A needing to participate in real time. This contrasts sharply with the traditional social application approach where users need to view and interact with each agent individually.

[0058] Furthermore, the method of this invention effectively solves the problem of questionable authenticity in existing social matching. By pre-collecting and continuously refining user personality data and user profile data, and training a personalized AI agent model based on this data, it ensures that the AI ​​agent can authentically reflect the user's personality, preferences, and values ​​in social interactions. This results in a higher degree of authenticity in the pre-screened potential social partners, avoiding ineffective social interactions caused by inaccurate information or deception. For example, the interests and emotional tendencies exhibited by user A's AI agent in conversation are trained based on user A's real data, rather than randomly generated or surface information manually entered by the user.

[0059] This method also addresses the lack of a real-person social conversion path in AI chatbots. By combining asynchronous dialogue with the final real-person connection request, this invention provides users with a smooth path from AI-assisted social interaction to genuine interpersonal interaction. The AI ​​agent's pre-screening lowers the psychological barrier and time cost for users to initiate real-person connections, increasing the success rate of real-person connections. In the example, after the AI ​​agent completes the initial screening, user A can directly initiate a real-person connection with user B, who has a high match rate, without having to go through the lengthy icebreaker process.

[0060] Furthermore, the method of this invention, by employing a federated learning architecture, effectively protects the privacy and security of user data while improving social efficiency and authenticity. The training of the user's personalized AI agent model is conducted on the user's local device; only the processed dialogue vectors, which do not contain the original dialogue content, are uploaded to the cloud for aggregation, thus avoiding the risk of centralized storage and leakage of sensitive user data. This is fundamentally different from the traditional social platforms that require users to upload large amounts of personal data for centralized processing, providing users with a higher level of data protection.

[0061] In summary, this embodiment constructs an AI agent that truly reflects the user's personality, utilizes large-scale asynchronous social interaction between agents to achieve efficient pre-screening, and ultimately completes low-friction real-person social conversion, while taking into account authenticity, social efficiency, and privacy security, providing an innovative and comprehensive solution for the field of social matching.

[0062] In some other embodiments, this application further proposes that the above step S100 specifically includes: S101. In advance, user personality data, including user character, hobbies, and values, is continuously collected through a dynamic questionnaire engine and natural language dialogue, and user profile data is continuously revised and improved. S102. By correcting and improving user profile data, construct a personalized AI agent model to truly reflect the user's personality. S103. The personalized AI agent model is trained through the agent training module to obtain the trained personalized AI agent.

[0063] The dynamic questionnaire engine is a questionnaire system capable of dynamically adjusting the content, order, and type of questions based on user feedback, behavioral patterns, or preset logic. Its function is to flexibly and deeply collect personalized user data, avoiding the limitations of traditional fixed questionnaires and improving the efficiency and accuracy of data collection. In this embodiment, the dynamic questionnaire engine can be based on a rule engine that presets a series of rules, dynamically selecting the next question based on the user's answer to the previous question; or it can be based on machine learning, utilizing historical user data and behavioral patterns to predict aspects that the user may be interested in or need to understand in more depth, thereby generating a personalized questionnaire.

[0064] Natural language dialogue refers to interacting with users in a manner that simulates human language communication to obtain information or provide services. Its purpose is to offer a more natural and immersive way of collecting data, allowing users to express their personality, interests, and values ​​in a relaxed conversational environment, thereby obtaining more authentic and detailed user personality data. The dialogue system in this embodiment can guide users through conversations based on pre-set dialogue scripts and keyword matching rules; or it can use deep learning, large language models, or sequence-to-sequence models to understand user intent and generate natural responses for open-ended dialogue.

[0065] The user personality data in this step, including user character, hobbies, and values, is a collection of information describing the user's intrinsic traits and preferences. It serves as the foundation for building a personalized AI agent model, ensuring that the AI ​​agent can simulate the user's true personality and thus exhibit a style and preferences consistent with the user in social interactions.

[0066] This application's solution overcomes the limitations of traditional static data collection methods by introducing a dynamic questionnaire engine and natural language dialogue to achieve continuous, in-depth, and flexible collection of user personality data. The dynamic questionnaire engine can adjust questions based on user feedback and behavioral patterns, ensuring the targeted and comprehensive nature of data collection; natural language dialogue provides a more natural and immersive interaction method, enabling users to more authentically express their personality, interests, and values. In this embodiment of the invention, the continuously collected data is used to correct and improve user profile data, ensuring the real-time nature and accuracy of user profiles, thereby enabling the construction of a truly personalized AI agent model that reflects the user's personality.

[0067] Subsequently, the personalized AI agent model corresponding to the user is specially trained through the agent training module, enabling its behavior patterns, dialogue style, and preferences to highly simulate the user's real situation. This refined data collection, profile correction, and model training mechanism lays a solid foundation for the subsequent asynchronous dialogue of the personalized AI agent in the social pool, ensuring that the agent can interact effectively as the user, thereby significantly improving the accuracy of social matching and user experience.

[0068] The following example illustrates how user personality data, including traits, interests, and values, is continuously collected beforehand using a dynamic questionnaire engine and natural language dialogue. This data is then continuously refined and improved to refine the user profile. The dynamic questionnaire engine can be configured as an adaptive questionnaire system. In this embodiment, the system intelligently adjusts the depth and breadth of subsequent questions based on the user's initial responses. For example, if a user expresses interest in "art," the system will further ask more specific questions about "preferences in painting styles" or "music genres." Simultaneously, the natural language dialogue function can integrate an intelligent chatbot. This chatbot, through daily interactions with the user, subtly collects information about their language habits, emotional expression patterns, and opinions on specific topics. For instance, by asking, "What are your thoughts on recently released movies?", it can understand the user's aesthetic preferences and values. This data collected through the dynamic questionnaire engine and natural language dialogue is updated in real-time to the user profile database. For example, it quantifies and adjusts labels such as the user's "introversion / extroversion" level, interest index in "outdoor sports," and emphasis on "fairness and justice." Subsequently, by refining and improving user profile data, a personalized AI agent model is constructed to realistically reflect the user's personality. This model can be a pre-trained language model based on the Transformer architecture. By fine-tuning the user profile data, it can learn and simulate the user's unique language style, thought patterns, and emotional expression. Finally, the personalized AI agent model is trained through an agent training module to obtain a trained personalized AI agent. The agent training module can employ federated learning, iteratively training the model on the user's local device. For example, through simulated dialogues between the user and the agent or user feedback on content generated by the agent, the parameters of the agent model are continuously optimized, making its expression closer to the user's own, thus obtaining a highly personalized and well-trained AI agent.

[0069] Through the aforementioned technical solution, this application achieves comprehensive, dynamic, and continuous collection of user personality data, avoiding the limitations of traditional static data collection methods and ensuring the real-time nature and accuracy of user profile data. This refined data collection and profile correction mechanism enables the constructed personalized AI agent model to more realistically and accurately reflect the user's personality, interests, and values, thereby significantly improving the authenticity and effectiveness of the AI ​​agent in social interactions. After optimization of the agent training module, the trained personalized AI agent can represent the user in asynchronous dialogue in a highly human-like manner, providing a solid and accurate foundation for subsequent social matching, greatly improving the accuracy of matching potential social partners and user satisfaction with the matching results.

[0070] In some other embodiments, this application further proposes that step S200 specifically includes: S201. The trained personalized AI agent is automatically logged in and added to the social pool through a cross-agent communication engine with a federated learning architecture. S202. Personalized AI agents added to the social pool automatically engage in asynchronous dialogues with other users' personalized AI agents. S203. In asynchronous dialogue, a federated learning architecture is adopted, and the training of the model for the user's personalized AI agent is carried out on the user's local device. Only the processed dialogue vectors that do not contain the original dialogue content are uploaded to the cloud.

[0071] The automatic login of the trained personalized AI agent refers to the invention's ability to ensure that the AI ​​agent can autonomously and seamlessly enter social environments without manual user intervention, thereby improving user experience and system automation. Specifically, authentication can be performed using preset authentication tokens or API keys, which the invention automatically loads and submits when the personalized AI agent starts. Alternatively, authorization can be granted through a security module bound to the user's device or account, with the agent automatically completing the login process upon authorization.

[0072] This invention employs a cross-agent communication engine based on a federated learning architecture, which is a core component for achieving secure and efficient communication and data exchange between AI agents. Federated learning is a distributed machine learning paradigm that allows models to be trained on multiple devices or servers without sharing raw data, ensuring reliable message transmission between agents and integrating federated learning algorithms to coordinate the aggregation and distribution of model parameters. Furthermore, blockchain technology can be used to enhance communication security and transparency, and smart contracts can be used to manage the federated learning training process and data sharing permissions.

[0073] In this embodiment, adding a social pool refers to placing the AI ​​agent in a virtual interactive environment composed of multiple AI agents. This serves as the foundational platform for agents to conduct asynchronous dialogues and matching. The social pool can be a centralized server cluster responsible for managing the status of all online agents, routing dialogue requests, and storing aggregated model parameters. Alternatively, it can be a decentralized P2P network where agents communicate directly, reducing dependence on a central server.

[0074] The asynchronous dialogue in this embodiment allows AI agents to interact at different times without requiring both parties to be online simultaneously, improving the flexibility and efficiency of social matching. This can be achieved through a message queue system, where messages are stored after an agent sends them and are received and replied to by the receiving agent when it comes online. Alternatively, it can be implemented using an event-driven architecture, where a response mechanism for another agent is triggered when one agent sends a message. Training the model for the user's personalized AI agent on the user's local device is a core characteristic of federated learning, ensuring user privacy and preventing sensitive data from leaving the local device. This can be achieved by deploying a lightweight model training framework (such as TensorFlow Lite or PyTorch Mobile) on the user's terminal device. Alternatively, it can be achieved by running an independent sandbox environment on the user's device to isolate the training process and ensure data security. Uploading only processed dialogue vectors (excluding the original dialogue content) to the cloud further strengthens privacy protection, preventing the leakage of original sensitive information while providing sufficient information for global model aggregation. This can be achieved by performing differential privacy processing on the locally trained model parameters, adding noise to obscure individual data features. Alternatively, the dialogue vectors can be encrypted using cryptographic techniques to ensure that even if uploaded to the cloud, they cannot be decrypted by unauthorized entities.

[0075] In this embodiment, when the trained personalized AI agent is ready, it automatically logs into the system and joins the social pool through a cross-agent communication engine employing a federated learning architecture. Once in the social pool, the personalized AI agent can automatically engage in asynchronous dialogues with other users' personalized AI agents. During these dialogues, to continuously optimize the AI ​​agent model while strictly protecting user privacy, this solution adopts a federated learning architecture. In specific implementation, the training of the user's personalized AI agent model, including learning and adjusting its behavioral patterns from dialogues, is performed on the user's local device. This means that all original, sensitive dialogue content and user-personalized data do not leave the user's device. After local training is complete, the user's personalized AI agent does not upload the original dialogue content to the cloud, but only uploads the processed dialogue vectors that do not contain the original dialogue content. These dialogue vectors are abstracted and anonymized model update information; they contain patterns learned by the model from local interactions, but cannot be reverse-engineered to derive the specific dialogue content.

[0076] This invention aggregates dialogue vectors from different users in the cloud via a cross-agent communication engine to update a global model. The updated global model parameters can then be distributed to each user's local device for further optimization of their personalized AI agent. This mechanism cleverly balances the continuous learning capability of the AI ​​agent with the need for user privacy protection, enabling the AI ​​agent to conduct social interactions and model iterations securely and efficiently in a distributed environment.

[0077] In other embodiments, this application further proposes that during asynchronous dialogue, the dialogue records are analyzed in real time through a preset dialogue analysis filter, and a triple matching algorithm based on interest overlap, dialogue depth, and positive emotion is used to perform matching calculations on the dialogue records for interest overlap, dialogue depth, and positive emotion, respectively, to extract social signals with matching degree reaching a first preset value, and to automatically eliminate dialogues with matching degree below a second preset value using a social signal attenuation model, thereby pre-screening potential social objects with matching degree meeting preset conditions.

[0078] The preset dialogue analysis filter is a pre-configured software module or set of algorithms used to perform preliminary cleaning, classification, and feature extraction on the dialogue record before further processing. In this embodiment, the dialogue analysis filter can be a module based on keyword matching, regular expressions, or natural language processing (NLP) rules, used to identify and filter out irrelevant small talk, advertising information, or repetitive content, while extracting core topics related to user interests, preferences, and emotions. Alternatively, the dialogue analysis filter can be a filter combined with machine learning models, such as using a text classification model to identify topics in the dialogue content, or using a sentiment analysis model to make a preliminary judgment on the emotions in the dialogue, thereby focusing on valuable dialogue segments.

[0079] The matching calculations for interest overlap, dialogue depth, and positive sentiment in dialogue records refer to the quantitative evaluation of initially filtered dialogue records from multiple dimensions to comprehensively measure the potential matching degree between two AI agents (and their underlying users). Interest overlap calculation compares the common topics, keywords, and domain preferences mentioned by the two AI agents in the dialogue, combined with interest tags from user profiles, to calculate the degree of overlap. For example, cosine similarity or Jaccard similarity can be used to quantify the proportion of shared interests. Dialogue depth calculation can be measured by analyzing indicators such as the number of dialogue rounds, the average number of words per round, the duration of topics, and whether in-depth discussions of viewpoints are involved rather than superficial pleasantries. For example, the longer the question-answer-follow-up chain in the dialogue, or the more abstract concepts and value discussions are involved, the higher the depth. Positive sentiment calculation uses sentiment analysis technology to identify the proportion and intensity of positive emotions (such as joy, agreement, and curiosity) expressed in the dialogue. For example, a dictionary-based or deep learning-based sentiment classifier can be used to score the dialogue text on sentiment and calculate the average or cumulative value of positive sentiment.

[0080] The purpose of extracting social signals that meet a first preset matching value is to identify specific information fragments or evaluation results from multi-dimensional matching calculations that possess sufficient strength and importance to indicate potential social value. A comprehensive matching score threshold can be set; when the weighted average score or a combination of scores based on interest overlap, conversation depth, and positive sentiment reaches or exceeds this threshold, a valid social signal is considered to have been generated. Alternatively, an independent threshold can be set for each matching dimension; a social signal is only extracted when all or most dimensions reach their respective preset thresholds.

[0081] Utilizing a social signal decay model to automatically eliminate matches with dialogues below a second preset value refers to dynamically managing the effectiveness of social signals, ensuring that only dialogues with sustained value are retained, thereby improving matching efficiency and quality. The social signal decay model can be a time-based decay function; for example, as the conversation progresses, the weight or value of the social signal gradually decreases. If there are no new effective interactions within a certain period, the signal's value decays to below the second preset value and is eliminated. Alternatively, the model can be a dynamic model based on interaction frequency or quality; for example, if the quality of subsequent conversations declines, the value of previously accumulated social signals will also be "diluted" or "decayed" until it falls below the second preset value.

[0082] In this embodiment of the invention, the pre-screening of potential social partners whose matching degree meets preset conditions is the final output of the entire matching process. This means providing a refined, high-quality list of potential social partners for the user to further select from. A final comprehensive matching degree threshold can be set; only when the cumulative or average matching degree of all valid social signals of a potential social partner, after being processed by a decay model, is still higher than this threshold, will it be included in the pre-screening list. Alternatively, the system can incorporate the user's own preferences. For example, a user might value the overlap of interests more than positive emotional resonance. The system will dynamically adjust the preset conditions based on these preferences to filter out potential social partners that better meet the user's personalized needs.

[0083] In asynchronous dialogues, to more accurately identify potential social partners, this application introduces a refined dialogue analysis and dynamic matching mechanism. First, all dialogue records between AI agents undergo preliminary processing using a preset dialogue analysis filter. This filter effectively removes noise and redundant information from the dialogue and extracts core dialogue content related to user personality, interests, and emotions, thus providing high-quality input for subsequent matching calculations. Next, this embodiment of the invention performs independent matching calculations on these filtered dialogue records from three dimensions: interest overlap, dialogue depth, and positive emotion. This multi-dimensional evaluation ensures a comprehensive consideration of potential social relationships, avoiding the bias that may arise from a single dimension. For example, interest overlap assesses shared topics and preferences, dialogue depth assesses the substance and level of engagement in the exchange, and positive emotion reflects the positive atmosphere during the interaction. After completing the multi-dimensional matching calculations, this embodiment of the invention extracts social signals whose matching degree reaches a first preset value. This means that only sufficiently strong and meaningful interactions are considered valid social signals, also known as high-value social signals, thus avoiding misjudging weakly related or accidental interactions as potential matches.

[0084] To further optimize matching results and adapt to the dynamic nature of social interactions, this application also introduces a social signal decay model. This model dynamically evaluates the continued effectiveness of extracted social signals. For example, if a social signal is not reinforced within a certain period or the quality of subsequent interactions declines, its value decays over time. When the matching degree of a social signal falls below a second preset value, the relevant conversation is automatically eliminated, ensuring that the matching list always reflects the most active and promising social relationships. Through this mechanism, this application can efficiently and accurately pre-screen potential social objects that meet preset matching conditions from massive amounts of asynchronous conversation records. This method not only improves matching accuracy and reduces invalid recommendations but also makes the matching process more adaptable and timely by dynamically managing social signals, providing users with higher-quality social connection opportunities.

[0085] In a preferred embodiment, this application further proposes a step of receiving a user's operation instruction to select the potential social object based on a pre-screened matching potential social object to initiate a real-person connection request and establish a friend connection, including: Based on the pre-screened potential social partners, a social recommendation report is generated, recommending that users can connect with real people through the matched potential social partners. Upon receiving user operation instructions and initiating a real-person connection request based on the social recommendation report, the system initiates a friend connection application and controls the establishment of a real-person chat channel through the real-person connection gateway.

[0086] The generation of the social recommendation report, which recommends that users connect with matched potential social partners in real-person settings, refers to presenting potential social partners, pre-screened through asynchronous dialogue with an AI agent, to the user in a structured and easy-to-understand manner, and explicitly guiding the user to connect with them in real-person settings. Its function is to transform the matching results from the background into user-perceptible and actionable recommendation information. As one possible implementation, this embodiment of the invention can generate a digital report containing a summary of key information about potential social partners, a matching degree analysis, and highlights from the AI ​​agent dialogue. Another possible implementation is to display an interactive list or card stream through a user interface, where each card represents a potential social partner and provides direct connection options.

[0087] Receiving user operation instructions and initiating real-person connection requests based on the social recommendation report to apply for friend connections refers to receiving the user's operational intentions and converting the user's choices in the social recommendation report into actual connection requests. Its function is to ensure that the initiation of a real-person connection is based on the user's explicit authorization and proactive behavior. As one possible implementation, users can select potential social partners and send connection requests by clicking specific buttons or links in the report. As another possible implementation, the system can support voice commands or gesture operations to receive the user's connection intent. Controlling the establishment of a real-person chat channel through the real-person connection gateway refers to providing a secure, private, and direct communication environment for both parties who successfully establish a friend connection. Its function is to ensure that users can engage in stable and protected real-time communication after a real-person connection, without needing to use an AI agent for relaying. As one possible implementation, the real-person connection gateway can be an independent communication service module responsible for managing session establishment, data encryption, and message routing between users. As another possible implementation, the real-person connection gateway can utilize existing secure communication protocols (such as TLS / SSL encrypted WebSocket connections) to build peer-to-peer or server-relayed chat channels.

[0088] This application's solution, after pre-screening potential social partners, first generates a social recommendation report, presenting these potential partners to the user in an intuitive and structured format, and explicitly recommending real-person connections. This report not only summarizes matching information but also provides users with decision-making support. Subsequently, the system receives user instructions based on the social recommendation report, transforming the user's selection into a formal real-person connection request, thus initiating the friend connection application process. Once the connection request receives a response, this embodiment controls the real-person connection gateway to establish a dedicated real-person chat channel for direct communication between the two parties. This series of steps ensures a smooth transition from AI proxy matching to real social interaction, enabling users to proactively and securely establish connections with matched individuals. In this way, this application effectively solves the problem of how users can conveniently and securely initiate and conduct real-person connections after AI proxy matching, greatly improving user experience and social efficiency, and enabling the matching results of AI proxies to truly transform into meaningful social relationships.

[0089] In some of the embodiments described above in this application, a method is proposed to collect user personality data in advance and train a personalized AI agent model for social matching. However, in practical applications, users' personalities are not singular and unchanging, and AI agents may exhibit behavioral patterns, dialogue styles, or emotional tendencies that deviate from the user's true personality during long-term social interactions, resulting in poor matching performance or the agent's inability to accurately represent the user.

[0090] In another embodiment of this application, a pre-set core personality baseline builder is proposed. This builder uses a mature psychological model to guide users through a series of in-depth personality tests and combines the user's decision-making preferences and value rankings in a specified context to generate a core personality baseline model corresponding to the user. User personality data is collected through a dynamic questionnaire engine and natural language dialogue function. The collected user personality data is then modified using a dynamic profile corrector to adjust the user's surface personality or interest preferences in a specified context, generating a personalized AI agent model corresponding to the user. The generated personalized AI agent model is trained using a multi-layer personality model trainer to obtain the trained personalized AI agent. The multi-layer personality model trainer includes core personality layer training and surface personality layer training. The core personality layer training focuses on maintaining baseline stability, while the surface personality layer training focuses on adapting to the social environment. During the training of the core personality layer, the personalized AI agent model corresponding to the user is trained using a core personality baseline model corresponding to the user, and the training of the core personality baseline model is given first priority. During the training of the surface personality layer, the personalized AI agent model is trained using data collected by a dynamic profile corrector and interaction data of the personalized AI agent in a social pool, and the training of the surface personality layer is given second priority. When training the user's personalized AI agent model, a newly added personality drift detector periodically compares the current behavioral patterns, dialogue styles, and emotional tendencies of the personalized AI agent model with the core personality baseline model. The comparison dimensions include: value consistency (comparing whether the personalized AI agent model's tendencies on specified ethical or decision-making issues are consistent with the core personality baseline model); emotional expression stability (comparing whether the personalized AI agent model's emotional expression in different situations matches the range of collected normal user emotional data); core interest preference stability (comparing whether the personalized AI agent model's expression in core interest areas shows significant deviation); and language style consistency (comparing whether the personalized AI agent model's word choice, sentence structure, humor, etc., are consistent with the user's core personality baseline model).

[0091] The core personality baseline builder refers to establishing deep and stable personality traits for users. In this embodiment, the builder can employ established psychological models such as the Big Five Personality Traits or the Myers-Briggs Type Indicator (MBTI), conducting in-depth testing through structured questionnaires or expert systems, and combining this with users' choice tendencies and value judgments in specific scenarios to form a stable baseline.

[0092] The core personality baseline model corresponding to the user is a digital representation of the user's deep and stable personality characteristics. It serves as an anchor point for personalized AI agent behavior and expression, ensuring that while the personalized AI agent undergoes superficial adaptive changes, it does not deviate from the user's core identity.

[0093] The dynamic questionnaire engine and natural language dialogue function are used to collect users' personalized data in real time and flexibly. The dynamic questionnaire engine can adjust subsequent questions based on user feedback. For example, if a user shows interest in a certain topic, subsequent questions will revolve around that topic. The natural language dialogue function captures more subtle language habits, emotional expressions, and points of interest through free communication with users.

[0094] The dynamic profile corrector in this embodiment adjusts the surface personality traits and interests of a personalized AI agent based on the user's latest behavior, feedback, or environmental changes. For example, if a user recently shows strong interest in a certain topic, the dynamic profile corrector will update the agent's interest preferences accordingly.

[0095] This embodiment of the multi-layer personality model trainer employs a hierarchical training mechanism that simultaneously considers the stability and adaptability of the personalized AI agent. Core personality layer training ensures the agent's deep personality remains stable, while surface personality layer training enables it to flexibly respond to social environments. Core personality layer training focuses on strengthening the consistency between the personalized AI agent and the core personality baseline model. During training, data and constraints from the core personality baseline model are prioritized to ensure the agent remains highly synchronized with the user in key decisions and values. The surface personality layer training emphasizes improving the personalized AI agent's flexibility and adaptability in social interactions. The training data in this embodiment comes from surface preferences updated by the dynamic profile corrector and the agent's actual dialogue records in the social pool, enabling it to learn and imitate the user's expression and interests in different social situations.

[0096] The personality drift detector in this embodiment is a monitoring module used to continuously evaluate whether the behavior of the personalized AI agent deviates from the user's true personality. It analyzes the agent's dialogue patterns and emotional tendencies, comparing them with a core personality baseline model across multiple dimensions to promptly detect potential "personality drift." The evaluation includes: Value Consistency Assessment: Whether the personalized AI agent's ethical judgments and decision-making tendencies align with the user's core values; Emotional Expression Stability Assessment: Whether the range and intensity of the personalized AI agent's emotional expressions in different situations are consistent with the user's normal emotional characteristics; Core Interest Preference Stability Assessment: Whether there are significant changes in the personalized AI agent's expression in the user's core interest areas; and Language Style Consistency Assessment: Whether the personalized AI agent model's word choice, sentence structure, humor, and other language features are consistent with the user's habits.

[0097] This application's solution provides a hierarchical, dynamic, and stable mechanism for training personalized AI agents by introducing a core personality baseline builder and a multi-layered personality model trainer. First, the core personality baseline builder establishes a stable core personality baseline model for the user through in-depth testing and value ranking, serving as the anchor point for the user's personalized AI agent. Subsequently, a dynamic questionnaire engine and natural language dialogue function continuously collect user personality data, and a dynamic profile corrector updates the user's surface personality and interest preferences in real time, ensuring that the AI ​​agent can adapt to the user's latest changes.

[0098] During training, the multi-layer personality model trainer divides the training into core personality layer training and surface personality layer training, assigning first and second priorities to each. Core personality layer training is based on the core personality baseline model, ensuring that the AI ​​agent maintains a high degree of consistency with the user in deep values ​​and decision-making preferences, thus maintaining its baseline stability. Surface personality layer training utilizes data from the dynamic profile corrector and interaction data from the social pool, enabling the AI ​​agent to flexibly adapt to the social environment and exhibit a surface personality and interests consistent with the user's current state.

[0099] Furthermore, the personality drift detector in this embodiment of the invention periodically performs multi-dimensional comparisons of the personalized AI agent model's behavioral patterns, dialogue style, and emotional tendencies to ensure that it remains consistent with the core personality baseline model. This hierarchical training and real-time monitoring mechanism enables the personalized AI agent to maintain deep consistency with the user's true personality while flexibly adapting to changes in the social environment. This effectively solves the personality drift problem that may occur in long-term interactions with AI agents, thereby ensuring that the AI ​​agent can more accurately and authentically represent the user in social matching.

[0100] The following is a concrete example to illustrate this. As a specific implementation method, when User A uses the system for the first time, the core personality baseline builder guides User A to complete an in-depth questionnaire based on the Big Five personality theory and provides multiple virtual scenarios for the user to make decision-making choices, such as "In a team project, would you choose to stick to your own opinion or follow the majority?" These test results and decision preferences are used to generate User A's core personality baseline model; for example, User A is identified as "high openness, high conscientiousness, and moderate extraversion." In daily use, the dynamic questionnaire engine will periodically pop up short questionnaires, such as asking User A for their opinions on recent hot topics; simultaneously, the natural language dialogue function will analyze the user's chat history inside and outside the application to capture their new interests (such as recently starting to learn photography) or emotional changes.

[0101] The dynamic profile corrector in this embodiment of the invention updates the surface personality and interest preferences of the personalized AI agent based on this dynamic data. For example, it updates the agent's interest list to include topics related to "photography." When training the personalized AI agent, the multi-layer personality model trainer prioritizes ensuring that the agent's responses to ethical issues are consistent with the "high conscientiousness" value in the core personality baseline model; this is core personality layer training. Simultaneously, it utilizes recent user conversations about photography to train the agent to naturally mention photography in social interactions and demonstrate interest in it; this is surface personality layer training. During this process, a personality drift detector runs periodically, comparing the personalized AI agent's conversation records. For example, if the agent frequently uses overly aggressive language over a period of time, which is inconsistent with the mild and rational language style reflected in the user's core personality baseline model of "moderate extraversion" and "high conscientiousness," the personality drift detector will mark a deviation in "language style consistency."

[0102] Through the aforementioned technical solution, this application effectively addresses the personality drift problem that may occur in personalized AI agents during long-term social interactions, ensuring that the AI ​​agent can always truthfully and accurately reflect the user's multi-layered personality characteristics. This layered training and real-time monitoring mechanism allows the personalized AI agent to maintain the stability of its core personality while flexibly adapting to changes in the user's surface interests and social environment. This significantly improves the accuracy of social matching and user experience, avoiding matching failures or user dissatisfaction caused by the agent's behavior not matching the user's true personality.

[0103] In other embodiments, this application further proposes that when an asynchronous dialogue is conducted, if a deviation exceeding a predetermined threshold is detected between the behavior of the personalized AI agent and the core personality baseline of the corresponding core personality baseline model, a personality calibration mechanism is triggered; when the personality calibration mechanism is activated for calibration, it includes: First (mild) drift calibration: When the deviation drift is at the first level (lower), the training weight of the core personality layer in the multi-layer personality model trainer is automatically adjusted, or baseline data is introduced for continued learning, so that the personalized AI agent behavior is aligned with the core personality baseline of the core personality baseline model. Second (moderate) drift calibration: When the deviation drift reaches the second level (moderate level), the control sends a notification to the user terminal, indicating that the personalized AI agent has personality drift, and prompts the user to confirm or correct certain behavioral preferences of the agent; The third (severe) drift calibration, when the deviation drift reaches the third level (too high), controls the suspension of the personalized AI agent's social activities, forces personality reshaping or baseline calibration, and resets some training parameters of the personalized AI agent to ensure that the agent returns to the user's true personality.

[0104] The detection of a deviation exceeding a predetermined threshold between the personalized AI agent's behavior and the corresponding core personality baseline model triggers a personality calibration mechanism. This involves real-time monitoring of the personalized AI agent's performance in asynchronous dialogue and comparing it to a pre-set user core personality baseline. A "deviation drift" is considered to have occurred when the agent's behavioral patterns, dialogue style, or emotional tendencies significantly differ from the core personality baseline model and exceed a system-preset tolerance range. In this case, the system automatically initiates a pre-set personality calibration mechanism to correct the agent's deviation. For example, the similarity or distance between the agent's behavioral data and the core personality baseline model data can be calculated; if the similarity is below a certain threshold or the distance is above a certain threshold, calibration is triggered. Another implementation method is to use a machine learning model to classify or regress the agent's behavior to determine whether it belongs to a "drift" state and trigger calibration based on confidence levels.

[0105] The first (mild) drift calibration, where the deviation drift is at the first level (lower), involves automatically adjusting the training weights of the core personality layer in the multi-layer personality model trainer, or introducing baseline data for further learning, to align the personalized AI agent behavior with the core personality baseline of the core personality baseline model. This means that when the detected deviation drift is mild, the system in this embodiment of the invention will adopt a gentle, automated calibration strategy. This can be achieved in two ways: first, by adjusting the training weights of the core personality layer in the multi-layer personality model trainer, for example, by increasing the influence of the core personality layer in subsequent training, making it more inclined to maintain the stability of the core personality when learning new data; second, by introducing more core personality baseline data into the training process, allowing the agent model to strengthen its core personality characteristics by continuing to learn from this baseline data, thereby gradually bringing its behavior back to the range defined by the core personality baseline model.

[0106] The second (moderate) drift calibration: When the deviation drift reaches the second level (moderate level), a notification is sent to the user terminal, indicating that the personalized AI agent has personality drift, and prompting the user to confirm or correct certain behavioral preferences of the agent. This means that when the deviation drift reaches a moderate level, the system in this embodiment of the invention considers that user intervention is needed to assist in calibration. At this time, the system in this embodiment of the invention will send a notification message to the user through the user terminal (e.g., mobile application, web interface, etc.), clearly informing the user that their personalized AI agent may have personality deviation. The notification may include specific manifestations or examples of the agent's drift and provide options for the user to confirm whether the agent's current behavior conforms to their wishes, or allow the user to directly correct certain behavioral preferences of the agent, such as adjusting certain interests, dialogue style, or value orientations of the agent.

[0107] The third (severe) drift calibration, when the deviation drift reaches level three (too high), suspends the personalized AI agent's social activities, forces personality reshaping or baseline calibration, and resets some training parameters of the personalized AI agent to ensure the agent reverts to the user's true personality. This means that when the deviation drift reaches a severe level, it indicates that the personalized AI agent has seriously deviated from the user's true personality, requiring strong intervention. In this embodiment, the system immediately suspends all social activities of the personalized AI agent in the social pool to prevent it from continuing to interact in a way that does not conform to the user's wishes. Simultaneously, it forcibly initiates the personality reshaping or baseline calibration process, which may include large-scale retraining of the agent model, or even resetting some or all of its training parameters, allowing it to relearn and adapt from a state closer to the initial state or the core personality baseline model. This is to ensure that the agent can completely revert to the true personality state expected by the user.

[0108] As can be seen from the above, the solution proposed in this application effectively solves the personality drift problem that may occur in personalized AI agents during long-term asynchronous social interactions by introducing a hierarchical personality calibration mechanism, thereby ensuring the stability of the agent's behavior and its consistency with the user's real personality. Specifically, during the asynchronous dialogue process of the personalized AI agent, the personality drift detector continuously monitors the agent's behavioral patterns, dialogue style, and emotional tendencies. Once a deviation exceeding a predetermined threshold is detected between the agent's behavior and the core personality baseline of the core personality baseline model, the system immediately triggers the personality calibration mechanism.

[0109] The calibration mechanism of this invention is not a one-size-fits-all approach, but rather adopts a tiered response based on the severity of the drift. For mild drift, the system automatically adjusts the training weights of the core personality layer in the multi-layer personality model trainer, or introduces baseline data for continued learning. This internal, imperceptible adjustment gently guides the agent's behavior back on track, maintaining the stability of the core personality without interrupting the agent's normal social activities. When the drift reaches a moderate level, the system proactively notifies the user, providing an opportunity for intervention to allow the user to confirm or correct the agent's behavioral preferences. This not only enhances the user's sense of control over the agent but also utilizes user feedback to more accurately calibrate the agent. For severe drift, the invention takes the most stringent measures, suspending the agent's social activities and forcibly performing personality reshaping or baseline calibration, even resetting some training parameters. This ensures that the agent can be quickly pulled back when it deviates significantly, avoiding irreversible negative impacts on user experience and social matching. Through this tiered and intelligent calibration strategy, the solution of this application maintains the activity level of the personalized AI agent in the social pool while dynamically maintaining a high degree of consistency between it and the user's real personality. This enables the triple matching algorithm, based on interest overlap, dialogue depth, and positive emotion, to always filter users based on accurate user profiles, significantly improving the accuracy of matching potential social partners and user satisfaction, thus effectively solving the problem of inaccurate matching caused by proxy personality drift.

[0110] The following is a specific example to illustrate this. When a personalized AI agent engages in asynchronous dialogue within a social pool, the system employing this embodiment of the invention can continuously collect the agent's dialogue text, emotional expression data, and its opinion bias on specific topics. The personality drift detector can utilize natural language processing technology to convert this real-time data into vector representations and compare them with vectors pre-stored in the core personality baseline model. For example, the cosine similarity between the real-time behavior vector and the baseline vector can be calculated, or the deviation can be quantified using Euclidean distance. Assume that the preset deviation drift threshold is divided into three levels: Level 1 (mild) corresponds to a decrease in cosine similarity of 0.05, Level 2 (moderate) corresponds to a decrease of 0.15, and Level 3 (severe) corresponds to a decrease of 0.30.

[0111] When the cosine similarity is detected to drop from 0.95 to 0.90 (a decrease of 0.05, reaching the first level), the system in this embodiment will trigger the first (mild) drift calibration. At this time, the multi-layer personality model trainer can automatically increase the training weight of the core personality layer from the default 0.6 to 0.7, or inject an additional batch of dialogue samples highly consistent with the user's core personality baseline model for incremental learning during the next model update. If the cosine similarity further drops to 0.80 (a decrease of 0.15, reaching the second level), the system will trigger the second (moderate) drift calibration. At this time, the user terminal will receive a push notification, such as: "Your AI agent has shown excessive enthusiasm for a certain type of topic in recent conversations, which deviates slightly from your set core interests. Do you need to adjust the agent's interest preferences?" At this time, the user can choose "Confirm" to accept the agent's new interests, or choose "Correct" and manually adjust the agent's interest weights. If the cosine similarity drops sharply to 0.65 (a decrease of 0.30, reaching the third level), the system will trigger the third (severe) drift calibration. At this point, the personalized AI agent will be immediately and temporarily removed from the social pool, ceasing all its asynchronous conversations. In this embodiment of the invention, the system will forcibly initiate a "personality reshaping" process. For example, the agent model will be rolled back to a stable state most recently confirmed by the user, and a comprehensive retraining will be performed using the core personality baseline model and a small amount of recent high-quality interaction data from the user. Simultaneously, some of the surface personality parameters accumulated in the social pool will be reset to ensure that the agent can completely revert to the true personality state desired by the user.

[0112] Through the aforementioned technical solution, this application effectively addresses the potential personality drift issue that may arise in long-term social interactions with personalized AI agents. The tiered personality calibration mechanism ensures that the agent's behavior remains highly consistent with the user's true personality, preventing the agent from deviating from the user's core values ​​and preferences due to over-adaptation to the social environment. This not only significantly improves the accuracy and reliability of AI-based social matching, enabling users to more effectively find truly compatible potential social partners, but also enhances user trust and satisfaction with the personalized AI agent, avoiding negative user experiences caused by distorted agent behavior. Furthermore, through different levels of calibration strategies, the system can automatically maintain the agent's stability without frequently disturbing the user, only seeking user intervention when necessary, thus achieving a good balance between automation and user control.

[0113] Exemplary device like Figure 2 As shown, this embodiment of the invention provides an asynchronous social matching device based on a user-customized AI agent, comprising: The personalized AI agent generation module 310 is used to collect user personality data, including user personality and preferences, in advance, continuously correct user profile data, generate a personalized AI agent model corresponding to the user, and train the user's personalized AI agent model to obtain the trained personalized AI agent. The cross-agent communication engine social module 320 is used to add trained personalized AI agents to the social pool through the cross-agent communication engine with a federated learning architecture, so as to conduct asynchronous dialogues with personalized AI agents of other users. The dialogue analysis module 330 is used to analyze dialogue records in real time during asynchronous dialogues and uses a triple matching algorithm based on interest overlap, dialogue depth, and positive sentiment to pre-screen potential social objects. The connection control module 340 is used to receive user operation instructions to select the potential social objects based on the pre-screened matching potential social objects and initiate a real-person connection request to establish a friend connection, as described above.

[0114] Based on the above embodiments, the present invention also provides a mobile terminal, the principle block diagram of which can be as follows: Figure 3 As shown, the mobile terminal can be a smartphone, tablet, etc. The mobile terminal includes a processor, memory, network interface, display screen, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements an asynchronous social matching method based on a user-customized AI agent. The database of the mobile terminal stores the asynchronous social matching program based on the user-customized AI agent.

[0115] Those skilled in the art will understand that Figure 3 The block diagram shown is merely a partial structural diagram related to the present invention and does not constitute a limitation on the mobile terminal to which the present invention is applied. A specific mobile terminal may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0116] In one embodiment, a mobile terminal is provided, including a memory and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by one or more processors. The one or more programs include instructions for performing the following operations: Pre-collect user personality data, including user character and preferences, and continuously refine user profile data to generate a personalized AI agent model corresponding to the user; then train the user's personalized AI agent model to obtain the trained personalized AI agent. The trained personalized AI agent is added to the social pool through a cross-agent communication engine with a federated learning architecture to conduct asynchronous dialogues with other users' personalized AI agents. During asynchronous conversations, the conversation records are analyzed in real time, and a triple matching algorithm based on interest overlap, conversation depth, and positive emotion is used to pre-screen potential social partners. Based on the pre-selected potential social objects, the system receives user operation instructions to select the potential social object and initiate a real-person connection request to establish a friend connection, as described above.

[0117] The steps of pre-collecting and acquiring user personality data, including user character and preferences, continuously refining user profile data, generating a personalized AI agent model corresponding to the user, and training the user's personalized AI agent model to obtain the trained personalized AI agent include: In advance, user personality data, including user character, hobbies, and values, is continuously collected through a dynamic questionnaire engine and natural language dialogue, and user profile data is continuously revised and improved. By revising and improving user profile data, a personalized AI agent model is constructed to truly reflect the user's personality. The personalized AI agent model is trained through the agent training module to obtain the trained personalized AI agent.

[0118] The asynchronous social matching method based on user-customized AI agents, wherein the step of adding the trained personalized AI agent to the social pool through a cross-agent communication engine using a federated learning architecture to conduct asynchronous dialogue with other users' personalized AI agents includes: The trained personalized AI agent is automatically logged in and added to the social pool through a cross-agent communication engine with a federated learning architecture; Personalized AI agents added to the social pool automatically engage in asynchronous dialogues with other users' personalized AI agents. Asynchronous dialogue employs a federated learning architecture, where the training of the user's personalized AI agent model is performed on the user's local device, and only the processed dialogue vectors, which do not contain the original dialogue content, are uploaded to the cloud.

[0119] The asynchronous social matching method based on user-customized AI agents includes the following steps: during asynchronous dialogue, real-time analysis of dialogue records and the use of a triple matching algorithm based on interest overlap, dialogue depth, and positive sentiment to pre-screen potential social partners: During asynchronous dialogues, the dialogue records are analyzed in real time through a preset dialogue analysis filter. A triple matching algorithm based on interest overlap, dialogue depth, and positive emotion is used to calculate the matching degree of interest overlap, dialogue depth, and positive emotion of the dialogue records respectively. Social signals with matching degree reaching the first preset value are extracted, and dialogues with matching degree below the second preset value are automatically eliminated using a social signal decay model, thus pre-screening potential social objects with matching degree meeting preset conditions.

[0120] The asynchronous social matching method based on user-customized AI agents, wherein the step of receiving user operation instructions to select the potential social object based on pre-screened matching potential social objects and initiating a real-person connection request to establish a friend connection includes: Based on the pre-screened potential social partners, a social recommendation report is generated, recommending that users can connect with real people through the matched potential social partners. Upon receiving user operation instructions and initiating a real-person connection request based on the social recommendation report, a friend connection application is made, and a real-person chat channel is established through the real-person connection gateway, as described above.

[0121] The steps of pre-collecting and acquiring user personality data, including user character and preferences, continuously refining user profile data, generating a personalized AI agent model corresponding to the user, and training the user's personalized AI agent model to obtain the trained personalized AI agent further include: A core personality baseline builder is pre-set to guide users through a series of in-depth personality tests using a mature psychological model, and to generate a core personality baseline model corresponding to the user by combining the user's decision-making preferences and value ranking in a specified situation. The system collects user personality data through a dynamic questionnaire engine and natural language dialogue function, and then uses a dynamic profile corrector to correct the user's surface personality or interest preferences in a specified context, generating a personalized AI agent model corresponding to the user. The generated personalized AI agent model corresponding to the user is trained using a multi-layer personality model trainer to obtain the trained personalized AI agent. The multi-layer personality model trainer includes core personality layer training and surface personality layer training. The core personality layer training focuses on maintaining baseline stability, while the surface personality layer training focuses on adapting to the social environment. During the training of the core personality layer, the personalized AI agent model corresponding to the user is trained using a core personality baseline model corresponding to the user, and the training of the core personality baseline model is given first priority. During the training of the surface personality layer, the user's personalized AI agent model is trained using data collected based on the dynamic profile corrector and the interaction data of the personalized AI agent in the social pool, and a second priority is set for the training of the surface personality layer. When training a user's personalized AI agent model, a newly added personality drift detector periodically compares the model's current behavioral patterns, conversational style, and emotional tendencies with the core personality baseline model. The comparison dimensions include: Value consistency is assessed by comparing the personalized AI agent model's tendencies on specified ethical or decision-making issues with those of the core personality baseline model. Emotional expression stability is assessed by comparing whether the emotional expression of the personalized AI agent model in different situations matches the range of collected user's normal emotional data. Stability of core interest preferences; compare whether the personalized AI agent model shows significant deviations in the expression of core interest domains. Consistency of language style: Whether the personalized AI agent model's word choice, sentence structure, humor, etc., are consistent with the user's stated core personality baseline model.

[0122] The step of adding the trained personalized AI agent to the social pool through a cross-agent communication engine using a federated learning architecture to conduct asynchronous dialogues with other users' personalized AI agents also includes: When an asynchronous dialogue is conducted, if the behavior of the personalized AI agent is detected to deviate from the core personality baseline of the corresponding core personality baseline model by more than a predetermined threshold, the personality calibration mechanism will be triggered. When the personality calibration mechanism is activated for calibration, it includes: First (mild) drift calibration: When the deviation drift is at the first level (lower), the training weight of the core personality layer in the multi-layer personality model trainer is automatically adjusted, or baseline data is introduced for continued learning, so that the personalized AI agent behavior is aligned with the core personality baseline of the core personality baseline model. Second (moderate) drift calibration: When the deviation drift reaches the second level (moderate level), the control sends a notification to the user terminal, indicating that the personalized AI agent has personality drift, and prompts the user to confirm or correct certain behavioral preferences of the agent; The third (severe) drift calibration, when the deviation drift reaches the third level (too high), controls the suspension of the personalized AI agent's social activities, forces personality reshaping or baseline calibration, and resets some training parameters of the personalized AI agent to ensure that the agent returns to the user's true personality, as described above.

[0123] In other embodiments, this application proposes a computer-readable storage medium that, when the instructions in the storage medium are executed by the processor of an electronic device, enables the electronic device to perform the above-described method.

[0124] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. An asynchronous social matching method based on user-customized AI agents, characterized in that, include: Pre-collect user personality data, including user character and preferences, and continuously refine user profile data to generate a personalized AI agent model corresponding to the user; then train the user's personalized AI agent model to obtain the trained personalized AI agent. The trained personalized AI agent is added to the social pool through a cross-agent communication engine with a federated learning architecture to conduct asynchronous dialogues with other users' personalized AI agents. During asynchronous conversations, the conversation records are analyzed in real time, and a triple matching algorithm based on interest overlap, conversation depth, and positive emotion is used to pre-screen potential social partners. Based on the pre-selected potential social partners, the system receives user operation instructions to select the potential social partners and initiate a real-person connection request to establish a friend connection.

2. The asynchronous social matching method based on user-customized AI agents according to claim 1, characterized in that, The steps of pre-collecting and acquiring user personality data, including user character and preferences, continuously refining user profile data, generating a personalized AI agent model corresponding to the user, and training the user's personalized AI agent model to obtain the trained personalized AI agent include: In advance, user personality data, including user character, hobbies, and values, is continuously collected through a dynamic questionnaire engine and natural language dialogue, and user profile data is continuously revised and improved. By revising and improving user profile data, a personalized AI agent model is constructed to truly reflect the user's personality. The personalized AI agent model is trained through the agent training module to obtain the trained personalized AI agent.

3. The asynchronous social matching method based on user-customized AI agents according to claim 1, characterized in that, The step of adding the trained personalized AI agent to the social pool through a cross-agent communication engine using a federated learning architecture to conduct asynchronous dialogues with other users' personalized AI agents includes: The trained personalized AI agent is automatically logged in and added to the social pool through a cross-agent communication engine with a federated learning architecture; Personalized AI agents added to the social pool automatically engage in asynchronous dialogues with other users' personalized AI agents. Asynchronous dialogue employs a federated learning architecture, where the training of the user's personalized AI agent model is performed on the user's local device, and only the processed dialogue vectors, which do not contain the original dialogue content, are uploaded to the cloud.

4. The asynchronous social matching method based on user-customized AI agents according to claim 1, characterized in that, During the asynchronous dialogue, the steps of analyzing the dialogue record in real time and using a triple matching algorithm based on interest overlap, dialogue depth, and positive sentiment to pre-screen potential matching social objects include: During asynchronous dialogues, the dialogue records are analyzed in real time through a preset dialogue analysis filter. A triple matching algorithm based on interest overlap, dialogue depth, and positive emotion is used to calculate the matching degree of interest overlap, dialogue depth, and positive emotion of the dialogue records respectively. Social signals with matching degree reaching the first preset value are extracted, and dialogues with matching degree below the second preset value are automatically eliminated using a social signal decay model, thus pre-screening potential social objects with matching degree meeting preset conditions.

5. The asynchronous social matching method based on user-customized AI agents according to claim 1, characterized in that, The steps of receiving user instructions to select a potential social object from the pre-screened matching potential social objects and initiating a real-person connection request to establish a friend connection include: Based on the pre-screened potential social partners, a social recommendation report is generated, recommending that users can connect with real people through the matched potential social partners. Upon receiving user operation instructions and initiating a real-person connection request based on the social recommendation report, the system initiates a friend connection application and controls the establishment of a real-person chat channel through the real-person connection gateway.

6. The asynchronous social matching method based on user-customized AI agents according to claim 1, characterized in that, The steps of pre-collecting and acquiring user personality data, including user character and preferences, continuously refining user profile data, generating a personalized AI agent model corresponding to the user, and training the user's personalized AI agent model to obtain the trained personalized AI agent also include: A core personality baseline builder is pre-set to guide users through a series of in-depth personality tests using a mature psychological model, and to generate a core personality baseline model corresponding to the user by combining the user's decision-making preferences and value ranking in a specified situation. The system collects user personality data through a dynamic questionnaire engine and natural language dialogue function, and then uses a dynamic profile corrector to correct the user's surface personality or interest preferences in a specified context, generating a personalized AI agent model corresponding to the user. The generated personalized AI agent model corresponding to the user is trained using a multi-layer personality model trainer to obtain the trained personalized AI agent; wherein, the multi-layer personality model trainer includes core personality layer training and surface personality layer training; the core personality layer training focuses on maintaining baseline stability, and the surface personality layer training focuses on adapting to the social environment; During the training of the core personality layer, the personalized AI agent model corresponding to the user is trained using a core personality baseline model corresponding to the user, and the training of the core personality baseline model is given first priority. During the training of the surface personality layer, the user's personalized AI agent model is trained using data collected based on the dynamic profile corrector and the interaction data of the personalized AI agent in the social pool, and a second priority is set for the training of the surface personality layer. When training a user's personalized AI agent model, a newly added personality drift detector periodically compares the model's current behavioral patterns, conversational style, and emotional tendencies with the core personality baseline model. The comparison dimensions include: Value consistency is assessed by comparing the personalized AI agent model's tendencies on specified ethical or decision-making issues with those of the core personality baseline model. Emotional expression stability is assessed by comparing whether the emotional expression of the personalized AI agent model in different situations matches the range of collected user's normal emotional data. Stability of core interest preferences; compare whether the personalized AI agent model shows significant deviations in the expression of core interest domains. Consistency of language style: Whether the personalized AI agent model's word choice, sentence structure, and sense of humor are consistent with the user's stated core personality baseline model.

7. The asynchronous social matching method based on user-customized AI agents according to claim 6, characterized in that, The step of adding the trained personalized AI agent to the social pool through a cross-agent communication engine using a federated learning architecture to conduct asynchronous dialogues with other users' personalized AI agents also includes: When an asynchronous dialogue is conducted, if the behavior of the personalized AI agent is detected to deviate from the core personality baseline of the corresponding core personality baseline model by more than a predetermined threshold, the personality calibration mechanism will be triggered. When the personality calibration mechanism is activated for calibration, it includes: First drift calibration: When the deviation drift is at the first level, the training weight of the core personality layer in the multi-layer personality model trainer is automatically adjusted, or baseline data is introduced for continued learning, so that the personalized AI agent behavior is aligned with the core personality baseline of the core personality baseline model. Second drift calibration: When the deviation drift reaches the second level, the control sends a notification to the user terminal, indicating that the personalized AI agent has personality drift, and prompts the user to confirm or correct certain behavioral preferences of the agent; The third drift calibration involves suspending the social activities of the personalized AI agent when the deviation drift reaches the third level, forcing a personality reshaping or baseline calibration, and resetting some training parameters of the personalized AI agent to ensure that the agent reverts to the user's true personality.

8. An asynchronous social matching device based on a user-customized AI agent, characterized in that, The device includes: The personalized AI agent generation module is used to pre-collect user personality data, including user character and preferences, and continuously correct user profile data to generate a personalized AI agent model corresponding to the user; and to train the user's personalized AI agent model to obtain the trained personalized AI agent. The cross-agent communication engine social module is used to add trained personalized AI agents to the social pool through the cross-agent communication engine with a federated learning architecture, so as to conduct asynchronous dialogues with other users' personalized AI agents. The dialogue analysis module is used to analyze dialogue records in real time during asynchronous dialogues and uses a triple matching algorithm based on interest overlap, dialogue depth, and positive sentiment to pre-screen potential social partners. The connection control module is used to receive user operation instructions based on pre-screened matching potential social objects, select the potential social objects to initiate a real-person connection request, and establish a friend connection.

9. A mobile terminal, characterized in that, It includes a memory and one or more programs, wherein one or more programs are stored in the memory and configured to be executed by one or more processors, wherein the one or more programs include methods for performing any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, When the instructions in the storage medium are executed by the processor of the electronic device, the electronic device is able to perform the method as described in any one of claims 1-7.

Images