Social recommendation ice-breaking method and system based on multi-modal intention recognition and behavior perception

Abstract

The application discloses a social recommendation ice-breaking method and system based on multi-modal intention recognition and behavior perception, and belongs to the technical field of artificial intelligence and social networks. The interactive process is started through two paths of user active input and behavior perception active triggering. The user intention is analyzed by combining a text intention recognition model and a picture intention recognition model with a dialogue memory context, and is divided into two processing paths of chatting intention and recommendation intention. By fusing multi-dimensional information such as user behavior data, preference analysis results, dialogue memory and user portrait, and combining a large language model, different scripts are generated according to different scenes: in the chatting scene, chatting reply scripts are generated according to the context; in the recommendation scene, complete combinations of chatting scripts, recommendation reasons and ice-breaking scripts are generated; potential dating demands are continuously monitored in the chatting process, and the recommendation process is naturally guided; after the interaction is completed, the dialogue memory is updated asynchronously to form a preference evolution closed loop, and the continuous evolution of recommendation accuracy and script quality is realized.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and social network technology, and in particular relates to a social recommendation icebreaker method and system based on multimodal intent recognition and behavior perception. Background Technology

[0002] In existing social applications, user matching and friend recommendations mainly rely on users manually filling in age range, region, height and other tag-based filtering conditions. Users' expressive ability is limited by the platform's preset filtering dimensions, making it difficult to convey vague and emotional preferences such as "gentle long-haired girl" or "like the type in a certain photo". Tag-based filtering cannot cover the aesthetic tendencies carried by natural language and images.

[0003] In terms of triggering mechanisms, most existing recommendation systems are passive, only displaying results when users actively enter the recommendation page. When user activity declines, such as when they do not check private messages or initiate conversations for a long time, the platform lacks effective means to identify when users "need help". Simple timed push strategies are crude and can easily cause user resentment.

[0004] In the post-matching interaction phase, many users give up on initiating a conversation even after finding a suitable match because they don't know what to say first. There is a lack of system support for generating personalized opening lines for specific recommended individuals between matching and conversation, which limits the conversion of matching into effective conversations.

[0005] In terms of preference learning, existing recommendation systems typically perform calculations based on static user profiles or single sessions. Each recommendation is independent of the others, making it impossible to accumulate changes in user friendship preferences across sessions, and the recommendation results are difficult to continuously optimize as the interaction process progresses.

[0006] In summary, existing social recommendation systems rely on static tagging and passive triggering, lacking a coherent processing chain from natural user preference expression, demand timing awareness, candidate matching to icebreaker copy generation. Furthermore, user preferences cannot accumulate and evolve across conversations. Consequently, they suffer from problems such as high barriers to preference expression, coarse and potentially disruptive proactive outreach, lack of icebreaker support after matching, and recommendation results failing to evolve with historical interactions. These are the technical problems that need to be solved in this field. Summary of the Invention

[0007] To address the aforementioned technical problems, the present invention aims to provide a social recommendation icebreaker method and system based on multimodal intent recognition and behavior perception. This method sets up two parallel paths on the triggering side: active input and behavior perception. The active path uses the receipt of valid input as the trigger condition. The behavior perception passive trigger path generates a passive trigger signal after the behavior collected by the embedded points hits the preset trigger behavior, and then generates the signal through interlocking verification of three-level frequency control, cooling constraints, and pre-admission. The two signals are combined into a global trigger criterion according to a logical OR relationship, enabling the system to proactively initiate guidance when user activity decreases, while reducing user disturbance through multiple constraints.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows: In a first aspect, the present invention provides a social recommendation icebreaker method based on multimodal intent recognition and behavior awareness, comprising the following steps: An active trigger signal is generated through an active input path, and an original passive trigger signal is generated through a behavior-aware passive trigger path. The original passive trigger signal is filtered by a triple interlocking system of multi-level frequency control, cooling window constraint and pre-access verification, and then integrated with the active trigger signal to output a global trigger criterion to trigger subsequent interaction processes. The input multimodal data is processed by modality splitting, and the preference features corresponding to each modality are extracted differentially. Short-term dialogue memory with time decay weight is loaded for fusion and disambiguation, and the intent category, preference feature vector and intent confidence are output. The process is divided according to the identified intent category. If the intent category is chat intent, user behavior data, preference analysis results, dialogue memory context, and user profile information are integrated to generate chat response text that conforms to the scenario constraints. During the dialogue, signals of seeking friendship are continuously monitored, and the user is guided naturally into the recommendation process according to confidence level, followed by the dialogue memory update step. If the intent category is recommendation intent, the following recommendation matching process is executed: The system loads three heterogeneous data sources in parallel: dialogue memory data, user profile data, and current preference features. After processing through three stages—dimensional alignment and standardization, priority weighted fusion, and conflict detection and resolution—enhanced query conditions are constructed, including hard constraints, strong preference conditions, weak preference conditions, and exclusion conditions. The entire pool of candidate users is subjected to three levels of processing: hard constraint coarse screening, multi-dimensional weighted fine ranking, and diversity rearrangement. The fine ranking stage adopts a weighted multiplication comprehensive scoring model, which uses the two-way matching degree score as a multiplication factor to participate in the final score calculation and outputs the final list of candidate target objects. Based on enhanced query conditions and candidate object features, a dimensional matching degree correlation analysis is performed to select the optimal matching dimension and generate three sets of differentiated copy: chat guidance copy, recommendation reason copy, and icebreaker copy. After content security verification, these are assembled into a recommendation card sequence and sent to the user. After the recommendation matching process is completed, the dialogue memory update step is executed: the user input, system response, entity label and user feedback signal of this interaction are asynchronously written into the memory module, and the short-term dialogue memory and long-term preference memory are updated respectively, forming a closed loop for continuous evolution of recommendation accuracy.

[0009] As a further aspect of the present invention, the active input path receives three types of multimodal input: text, images, and voice. Text carries fuzzy dating preferences, images convey appearance aesthetic preferences, and voice is transcribed into text by an ASR engine and then incorporated into a unified processing pipeline. An active trigger signal is generated upon detecting any valid user input.

[0010] As a further aspect of the present invention, the behavior-aware passive triggering path collects user application behavior through client-side tracking points, and generates an original passive triggering signal when any one of the following six types of triggering behaviors is met: homepage browsing pause time exceeds 8 seconds; no private messages are viewed in the foreground for 30 seconds; no chat is initiated for 30 consecutive seconds; no opening remarks are sent in a new conversation for 120 seconds; three user private messages are viewed consecutively and read without replying; no input is made for 10 seconds after entering the assistant page.

[0011] As a further aspect of the present invention, the multi-level frequency control adopts a top-down three-level frequency control mechanism, where high-priority rules directly reject low-priority triggering requests, including: Global daily frequency control: The maximum number of passive triggers per day is 10; Session lifecycle frequency control: The maximum number of times this function can be triggered within the lifecycle of a single application is 5. Event-level frequency control: Count the trigger frequency of each type of behavior at the hourly and daily levels, and block it if it exceeds the corresponding threshold; The total frequency control factor is taken as the minimum value of the three-level frequency control factors.

[0012] As a further aspect of the present invention, the cooling window constraint is set to a uniform cooling cycle of 2 minutes, and the cooling factor is calculated using an exponential decay formula:

[0013] In the formula, This is the attenuation coefficient, with a value of 0.5. ; To standardize the cooling cycle; The current moment; The time since the last trigger is used; if the interval between the last trigger and the last trigger is sufficient, the passage is allowed; if the interval is insufficient, the passage is blocked by exponential decay.

[0014] As a further aspect of the present invention, the pre-access verification is passively triggered when the following three conditions are met simultaneously: The message interval between the user and the assistant is greater than or equal to the cooldown period; The message interval between the assistant and the user is greater than or equal to the cooldown period; The users are existing users who are interacting with the platform, not new users of the new features. The global trigger criterion is the maximum value of the active trigger signal and the filtered passive trigger signal. If the criterion is 1, the intent recognition process is started; otherwise, the system remains silent.

[0015] As a further aspect of the present invention, modal splitting processing is performed on the input multimodal data, including the following steps: Text input is fed into the LLM intent recognition model, which splits intent categories through cue engineering and few-shot learning. The image input is fed into the VLM visual model, which extracts three types of visual features: body shape and clothing, scene atmosphere, and temperament and style, and determines them as recommendation intent by default. After voice input is transcribed into text, the text processing link is reused. All modal features are uniformly mapped to the same semantic vector space.

[0016] As a further aspect of the present invention, the most recent 5-10 rounds of short-term dialogue memory are loaded for fusion and disambiguation, and the memory weights are calculated using an exponential time decay formula:

[0017] In the formula, The decay rate factor is 0.1d. -1 ; This refers to the current system time. For the first The timing of the dialogue rounds; the weight of long-term dialogues decreases while the weight of recent demands increases; the historical memory vector is weighted and aggregated, and then superimposed with the current input semantic features to resolve semantic ambiguity from three dimensions: topic continuity, emotional fluctuation, and preference accumulation.

[0018] As a further aspect of the present invention, the input context generated by the large language model is constructed by integrating information from four dimensions: User behavior data includes triggering method, triggering event type, time interval since last interaction, and cumulative number of interactions on the day. Preference analysis results: Preference keywords extracted from each dimension of intent identification and their corresponding confidence levels; Dialogue memory context: the K most recent dialogue records, each with a time decay weight; User profile: basic demographic attributes, social style tags, interest tags, and social behavior tags; A dynamic attention mechanism is used to calculate the weights of each dimension, and the weights are dynamically adjusted according to the scenario: the weight of behavioral data increases in passively triggered scenarios, the weight of dialogue memory increases in emotional expression scenarios, and the weight of preference analysis increases in explicit preference expression scenarios.

[0019] As a further aspect of the present invention, the continuous monitoring of dating request signals specifically includes: Extract four types of signal features from the current message and conversation context: loneliness, friendship intention, companionship need, and explicit recommendation request; and calculate the confidence level of friendship request. Based on the confidence threshold, the guidance strategy is divided into three levels: Confidence level below the low threshold: Pure companionship chat, without recommendation guidance; Confidence level between low and high thresholds: Social topics are subtly incorporated into responses to provide guidance; Confidence level higher than or equal to the high threshold: Explicitly provide recommendations and direct guidance; Once the user replies with confirmation after being directly guided, the recommendation and matching process officially begins.

[0020] As a further aspect of the present invention, the three heterogeneous data streams are specifically as follows: Dialogue memory data is divided into short-term memory and long-term memory layers. Short-term memory stores the complete context of the most recent 5 to 10 rounds of dialogue and recommendation feedback signals. Long-term memory stores cross-conversation preference frequency statistics, positive feedback inductive features, and negative feedback inductive features, using exponential time decay weighting. User profile data includes basic demographic attributes, authentication information, social behavior profiles, and system behavior tags, which are used to provide context for self-awareness. Current preference features: explicit and implicit preferences obtained from the intent recognition output have the highest priority in the fusion process.

[0021] As a further aspect of the present invention, the priority weighted fusion is performed according to the priority of current preference > short-term memory > long-term memory > user profile inference, with corresponding weight coefficients of 0.5, 0.25, 0.15, and 0.10, respectively; the conflict detection and resolution adopts the principle of proximity in time: current preference covers short-term memory, and short-term memory covers long-term memory; the feature markers of explicit negative feedback from users are used as exclusion conditions.

[0022] As a further aspect of the present invention, the enhanced query conditions also include preference confidence metadata, which marks the source and confidence rating of each dimension of the conditions for differentiated weighting during the recommendation and ranking stage; When there are too few effective dimensions, an active completion strategy is initiated, which generates inferred preferences based on the statistical preferences of users in the same group, and fills the blank dimensions with the lowest confidence level.

[0023] As a further aspect of the present invention, the hard-constrained coarse sieve specifically performs four-stage filtration: Precise matching filtering with hard constraints; Exclusion rules filter feature types that have been explicitly rejected. Basic compliance filtering: Excludes banned, deactivated, inactive, and blocked / blacklisted accounts; 72-hour history deduplication filter: excludes items that have been recommended within the last 72 hours; After initial screening, the candidate set was reduced from millions to tens of millions to thousands.

[0024] As a further aspect of the present invention, the multi-dimensional weighted ranking and fusion of four-dimensional scores are as follows: Preference matching score: Gaussian decay function is used for continuous dimensions, hierarchical matching is used for discrete dimensions, and Jaccard coefficient is used for labeled dimensions; weighted according to preference confidence. Object quality score: calculated based on a combination of data completeness, authentication level, recent activity level, and historical social behavior; Two-way matching score: Based on the set of users who have accepted and rejected candidates in the past, calculate the cosine similarity between the current user and the users who have historical preferences, which represents the probability of acceptance by the other party; Timeliness and freshness score: Newly registered users and users with newly updated information receive bonus points, and long-tail untouched objects receive discovery bonus points.

[0025] As a further aspect of the present invention, the formula for the comprehensive scoring model is:

[0026] In the formula, Candidates Preference matching score; Candidates Object quality rating; Candidates Timeliness and freshness rating; Candidates Two-way matching score; , , The weighting coefficients are 0.6, 0.25, and 0.15, respectively. After sorting the candidates in descending order of their comprehensive scores, the top 50 to 100 candidates are selected for the re-sorting stage.

[0027] As a further aspect of the present invention, the diversity rearrangement performs three types of constraints: Geographic diversity constraint: Candidates from the same city shall not exceed 60%; Style diversity constraint: It must include at least two visual style types; Interaction state diversity constraint: It includes both completely new, untouched objects and objects with whom there has been some initial interaction; After rearrangement, the final 3 to 5 candidate target objects are output.

[0028] As a further aspect of the present invention, the generation constraints for the three sets of differentiated copywriting are as follows: Chat prompts should connect to recent conversation topics, convey a sense of companionship, and naturally lead to recommended results. Recommendation copy: Based on the Top-K optimal matching dimension, highlighting the commonalities or complementarities between the two parties, including specific details; Icebreaker copywriting: Focus on the candidate's interests, match the user's social style, and create a hook for conversation by combining common ground between the two parties.

[0029] As a further aspect of the present invention, the quality score of the icebreaker copywriting is calculated by weighting three dimensions:

[0030] In the formula, The relevance of the icebreaker copy to the characteristics of the target audience; To ensure the icebreaker copy matches the user's communication style; To ensure the responsiveness of icebreaker copywriting; , , These are the weight coefficients for the corresponding dimensions.

[0031] As a further solution of the present invention, a fallback mechanism is set up: when the generated text is empty, a fallback text is selected from the pre-configured fallback pool; the fallback pool is shuffled by Fisher-Yates and then allocated in order to ensure that the text is not repeated in scenarios with multiple candidate objects; when the content security verification fails, the fallback text is sent in the actively triggered scenario, and is not sent in the passively triggered scenario.

[0032] As a further aspect of the present invention, the recommended card sequence arrangement rule is as follows: Image 1: Chat prompts, recommended card information, and keyword highlighting; assistant avatar displayed. Cards 2 through N-1: Purely recommended card information, assistant avatar hidden; Card N: Recommended card information and interactive operation buttons; assistant avatar hidden. Multiple cards are sent at fixed intervals with progressively increasing timestamps, and each icebreaker message is assigned a unique identifier for persistent storage.

[0033] As a further aspect of the present invention, the long-term preference memory update includes statistical preference dimension occurrence frequency, summarizing positive feedback features and summarizing negative feedback features, and the updated memory data is used for intent recognition, feature fusion and recommendation matching steps.

[0034] Secondly, the present invention provides a social recommendation icebreaker system based on multimodal intent recognition and behavior awareness, comprising: Intelligent triggering module: used to generate trigger signals through active input paths and behavior-aware passive triggering paths. After the passive triggering signal is filtered by multi-level frequency control, cooling window constraints and pre-entry verification, it is integrated with the active triggering signal to output global triggering criteria. Multimodal intent recognition module: This module is used to split multimodal inputs, extract the preference features of each modality, and combine them with short dialogue memory with time decay for disambiguation, outputting intent category, preference feature vector and confidence score. Intent Triage and Chat Interaction Module: This module is used to triage and process messages based on intent categories. It integrates multi-dimensional information to generate chat reply text under each chat intent and continuously monitors signals of friendship requests, guiding users to naturally transition into the recommendation process in a tiered manner. The memory loading and feature fusion module is used to load heterogeneous data from three sources—dialogue memory, user profile, and current preferences—in parallel under the recommendation intent. After dimensional alignment, weighted fusion, and conflict resolution, it constructs enhanced query conditions. Multi-level funnel recommendation module: It is used to perform hard constraint coarse screening, multi-dimensional weighted fine ranking and diversity rearrangement on the candidate user pool in sequence under the recommendation intention. It uses a weighted multiplication model to calculate the comprehensive score by using the bidirectional matching degree as the multiplication factor and outputs a list of candidate target objects. Recommendation copy generation module: used to generate chat prompts, recommendation reasons and icebreakers based on preference and feature correlation analysis under the recommendation intent, and then assemble them into a recommendation card sequence for sending after security verification; Dialogue Memory Module: Used to asynchronously store interaction data, update short-term dialogue memory and long-term preference memory respectively, forming a closed loop of preference evolution.

[0035] As a further aspect of the present invention, the intelligent triggering module monitors six types of user behaviors through behavior perception passive triggering path monitoring: homepage browsing pauses for more than 8 seconds, no private messages viewed in the foreground for 30 seconds, no chat initiated for 30 consecutive seconds, no opening remarks sent in a new conversation for 120 seconds, no reply to 3 consecutive read private messages, and no input on the assistant page for 10 seconds; multi-level frequency control includes three levels of frequency control: global daily level, conversation cycle, and event granularity, with high-priority rules directly rejecting low-priority requests.

[0036] As a further aspect of the present invention, the multimodal intent recognition module includes a text intent recognition unit, a visual intent recognition unit, and a memory fusion disambiguation unit; the visual intent recognition unit inputs an image, extracts three types of visual features—body shape and clothing, scene atmosphere, and temperament / style—and determines them as recommendation intents by default; the memory fusion disambiguation unit loads 5-10 rounds of short-term dialogue memory, uses exponential time decay weighted aggregation, and resolves semantic ambiguity in a single round. As a further embodiment of the present invention, the intent triage and chat interaction module includes a multi-dimensional context fusion unit, a chat text generation unit, and a friend-making request monitoring and guidance unit; the friend-making request monitoring and guidance unit extracts four types of signals: loneliness, friend-making intention, companionship need, and explicit request, calculates the confidence level, and then executes a three-level guidance strategy of pure chat, topic hints, and direct recommendation.

[0037] As a further aspect of the present invention, the multi-level funnel recommendation module includes a coarse screening unit, a fine ranking unit, and a diversity rearrangement unit; the fine ranking unit calculates scores for four dimensions: preference matching degree, object quality, bidirectional matching degree, timeliness, and freshness, and calculates a comprehensive score through a weighted multiplication model, with the bidirectional matching degree score participating in the calculation as a multiplication factor.

[0038] As a further embodiment of the present invention, the recommended text generation module includes a matching degree correlation analysis unit, three sets of text generation units, a security verification unit, and a card assembly unit; the card assembly unit arranges the recommended card sequence according to the rule that the first card has a guiding text and a highlight, the middle cards are pure cards, and the last card has an operation button.

[0039] Compared with existing technologies, the multimodal intent recognition and behavior perception-based social recommendation icebreaking method and system provided in this application have the following beneficial effects: This invention extracts intent through three modalities—text, image, and voice—combined with time-decayed dialogue memory for joint intent recognition. Users can directly express nuanced preferences using natural language or images, which helps lower the threshold for preference expression and reduce preference loss due to insufficient tag dimensions. It identifies six types of implicit request behaviors through passive triggering paths based on behavioral perception, and uses a three-level frequency control system (global daily level, application session level, and segmented event level), exponential decay cooling, and pre-entry interlocking with existing users before triggering. This allows for the identification of when users need help while constraining the frequency of contact, achieving a balance between increasing activity and avoiding disturbance. Finally, it generates recommendation reasons with specific matching details and interests surrounding the target through dimensional correlation analysis between users and candidate objects. The invention employs icebreaker text that balances user social style and reproducibility, pre-filled into input boxes to reduce the burden on users to organize opening remarks and provide topics for follow-up between matched individuals. Through exponential decay weighting of short-term and long-term memories, conflict resolution based on temporal proximity, and asynchronous memory updates after interaction, recent preferences receive higher weight, and negative feedback features are promptly eliminated, allowing for continuous adjustment of recommendation results and text perspectives as user preferences evolve. Furthermore, the invention introduces a two-way matching degree based on the historical acceptance and rejection of candidate individuals, using a multiplicative factor in the comprehensive scoring. Simultaneously, it uses three diversity constraints—region, style, and interaction status—for re-ranking, balancing the possibility of mutual acceptance with the exploration space of the recommendation list. Attached Figure Description

[0040] Figure 1 This is a flowchart of a social recommendation icebreaker method based on multimodal intent recognition and behavior perception provided by the present invention.

[0041] Figure 2 This is a flowchart of intent triage and chat interaction in a social recommendation icebreaker method for multimodal intent recognition and behavior perception provided by the present invention.

[0042] Figure 3 This is a flowchart of the recommendation scenario copy generation process in a social recommendation icebreaker method based on multimodal intent recognition and behavior perception provided by the present invention.

[0043] Figure 4 This is a structural block diagram of a social recommendation icebreaker system for multimodal intent recognition and behavior perception provided by the present invention. Detailed Implementation

[0044] The present invention will now be described in further detail with reference to the accompanying drawings and specific 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.

[0045] To address the shortcomings of existing social recommendation systems that rely on static tagging and passive triggering, lacking a seamless processing chain from natural user preference expression, demand timing awareness, candidate matching to icebreaker copy generation, and the inability of user preferences to accumulate and evolve across conversations, this invention aims to provide a social recommendation icebreaker method and system based on multimodal intent recognition and behavior awareness. This method sets up two parallel paths on the triggering side: active input and behavior awareness. The active path uses the receipt of valid input as the trigger condition. The passive triggering path, after the behavior collected by the tracking points matches the preset trigger behavior, generates a passive trigger signal through a three-level frequency control, cooling constraints, and pre-entry interlocking verification. The two signals are logically ORed and merged into a global trigger criterion, enabling the system to proactively initiate guidance when user activity decreases, while minimizing user disturbance through multiple constraints.

[0046] See Figures 1 to 3 As shown, this invention provides a social recommendation icebreaker method based on multimodal intent recognition and behavior awareness, comprising the following steps: S10. Generate an active trigger signal through an active input path and generate an original passive trigger signal through a behavior-aware passive trigger path. After being filtered by a triple interlock of multi-level frequency control, cooling window constraint and pre-access verification, the original passive trigger signal is integrated with the active trigger signal to output a global trigger criterion and trigger the interaction process.

[0047] In this step, a dual-path collaborative triggering architecture of active and passive approaches is built to address the shortcomings of traditional social recommendation systems, such as passive triggering and low user reach. The system initiates the interaction process based on two paths: active user input and behavior perception. Specifically, the active input path generates an active trigger signal when it receives valid user input, while the behavior perception passive triggering path generates a passive trigger signal when the user behavior collected by the client meets the preset trigger behavior conditions and passes the interlocking verification of frequency control constraints, cooling constraints, and pre-admission constraints. The system then outputs a global trigger criterion according to the logical OR relationship between the active and passive trigger signals.

[0048] The active input path accepts three types of input: text, images, and voice. Text conveys vague friendship preferences, images express aesthetic preferences, and voice input, after being transcribed into text via speech recognition, is processed in the same pipeline as the text input. A minimal triggering rule is set, initiating the process upon detecting any valid user input, without complex calculations. This active input path lowers the threshold for expressing user preferences, breaks through the limitations of platform-based tagging and filtering, and preserves users' natural expression needs. The preset triggering behavior conditions include at least one of the following six types of behavior: remaining on the homepage for more than a first duration, remaining on the foreground for a second duration without checking private messages, not initiating a chat for a third consecutive duration, not sending an opening message for a fourth duration after creating a new conversation, continuously viewing a preset number of users' private messages and not replying to any of them, and no input for a fifth duration after entering the assistant page.

[0049] By identifying implicit social needs against fixed thresholds, six types of triggering behaviors are pre-defined: homepage browsing stagnation exceeding 8 seconds, no checking of private messages for 30 seconds, no initiation of a chat for 30 consecutive seconds, no opening remarks for a new conversation for 120 seconds, reading but not replying to 3 consecutive user private messages, and no input for 10 seconds on the assistant page. Meeting any one of these behavioral conditions generates an initial passive trigger signal. To avoid disturbing users, a three-tiered frequency control, cooling-off rate limiting, and pre-access control are implemented, with all rules interlocked before a guided dialogue can be triggered.

[0050] The behavior-aware passive triggering path relies on real-time collection of user application behavior through client-side tracking points, identifying implicit social requests against fixed thresholds. The interlocking verification of this behavior-aware passive triggering path includes: frequency control based on three decreasing priority levels: global daily level, application session level, and sub-event level, with higher priority level constraints serving as the basis for rejecting lower priority level trigger requests; a cooling constraint that exponentially decays the trigger signal based on the difference between the time interval since the last trigger and the cooling period; and pre-entry constraints that verify the bidirectional message interval between the user and the assistant and the user's existing interaction status, allowing only users with existing interactions to be passively awakened.

[0051] Global daily frequency control counts the number of passive triggers per day and compares it with the daily upper limit. Application session-level frequency control counts the number of triggers within the lifecycle of a single application and compares it with the session upper limit. Event-level frequency control counts the hourly and daily trigger frequencies for each type of behavior. The minimum value of the three-level frequency control results is taken as the total frequency control factor. The cooldown period is two minutes, and the cooldown decay rate factor is 0.5 per minute. When the current time is within the cooldown period from the last trigger time, the cooldown constraint is released; otherwise, it is blocked by exponential decay. The global trigger criterion is the maximum value between the active trigger signal and the passive trigger signal.

[0052] In this embodiment, the multi-level frequency control adopts a top-down three-level frequency control mechanism, where high-priority rules directly reject low-priority triggering requests, including: Global daily frequency control: The maximum number of passive triggers per day is 10; Session lifecycle frequency control: The maximum number of times this function can be triggered within the lifecycle of a single application is 5. Event-level frequency control: Count the trigger frequency of each type of behavior at the hourly and daily levels, and block it if it exceeds the corresponding threshold; The total frequency control factor is taken as the minimum value of the three-level frequency control factors.

[0053] In the multi-level frequency control mechanism, the frequency control priority decreases from top to bottom: P0 global daily constraint > P1 application session constraint > P2 subdivided event constraint. High-priority rules directly reject low-priority triggered requests. Three sets of basic judgment formulas are configured, and rate limiting is achieved based on threshold Boolean judgment, without redundant weighted calculations. Specifically, this includes: (1) Global daily frequency control: Count the number of passive triggers per day. , daily upper limit =10, frequency control factor.

[0054] ; In the formula, This is the global daily frequency control factor; 1 indicates permission to pass, and 0 indicates interception. This represents the cumulative number of times the event was passively triggered that day. This is the maximum number of times a passive event can be triggered per day.

[0055] (2) Session lifecycle frequency control: Count the number of times a single application lifecycle is triggered. upper limit =5, frequency control factor.

[0056] ; In the formula, This is the session period frequency control factor; 1 indicates permission to proceed, and 0 indicates obstruction. This refers to the cumulative number of times a single application is passively triggered within its lifecycle. This is the maximum number of times a passive event can be triggered in a single session.

[0057] (3) Event-level frequency control: Count the hourly and daily trigger frequency of each type of behavior, and directly block those exceeding the corresponding threshold, integrating the total frequency control factors. In the formula, To integrate the total frequency control factor after three levels of frequency control, the minimum value of the three levels is taken; This is the global daily frequency control factor; For session period frequency control factor; For the first Event granularity frequency control factor for class behaviors.

[0058] To avoid user annoyance caused by continuous push notifications in a short period of time, a uniform cooldown period is set. In step S10, a unique exponential decay formula is configured to balance the smoothness of rate limiting with the mathematical consistency of the entire document. In this embodiment, the cooling window constraint is set to a uniform cooling cycle of 2 minutes, and the cooling factor is calculated using an exponential decay formula:

[0059] In the formula, This is the cooling window decay factor; the greater the interval, the closer the value is to 1, and the closer the interval, the more exponentially it decays. The cooling attenuation coefficient is set to 0.5. ; To ensure a consistent cooling cycle, this embodiment uses a value of 2 minutes. This refers to the current system time. The time since the last trigger is used; if there is enough time since the last trigger, the request is allowed; if the time is insufficient, the request is blocked exponentially to avoid annoying pop-up windows.

[0060] In this embodiment, the pre-access verification is passively triggered when the following three conditions are met simultaneously: The message interval between the user and the assistant is greater than or equal to the cooldown period; The message interval between the assistant and the user is greater than or equal to the cooldown period; The users are existing users who are interacting with the platform, not new users of the new features. The summary admission determination formula for the three admission conditions of passively triggered pre-verification is as follows: ; In the formula, This is a prerequisite admission factor; 1 indicates that the admission conditions are met, and 0 indicates that they are not met. The moment the user sends the last message to the assistant; The moment the assistant sends the last message to the user; To standardize the cooling cycle; This is a user status identifier; 1 represents an existing interactive user, and 0 represents a new user for the new feature.

[0061] The system verifies the interval between two-way messages between the user and the assistant, filters new users, and only allows passive activation to existing users, balancing reach effectiveness and user experience; it also integrates dual-path output to determine global trigger criteria. If the result is 1, the intent recognition process is initiated; otherwise, the system remains silent, completing step S10, full-link triggering scheduling. This is a global trigger criterion; 1 initiates the subsequent process, 0 maintains system silence. The user actively inputs a trigger signal; 1 indicates triggering, and 0 indicates no triggering. The original passive trigger signal is 1; 1 indicates action triggering, and 0 indicates no triggering. The global trigger criterion is the maximum value of the active trigger signal and the filtered passive trigger signal. If the criterion is 1, the intent recognition process is started; otherwise, the system remains silent.

[0062] S20. Multimodal Intent Recognition: Modal splitting is performed on the input multimodal data, and the preference features corresponding to each modality are extracted differentially; short-term dialogue memory with time decay weight is loaded for fusion and disambiguation, and the intent category, preference feature vector and intent confidence are output.

[0063] This step integrates multimodal input with historical dialogue context to resolve semantic ambiguity and accurately separate chat and recommendation intents. In this embodiment, modality splitting processing is performed on the input multimodal data, including the following steps: Text input is fed into the LLM intent recognition model, which uses prompting engineering and few-shot learning to segment intent categories; image input is fed into the VLM visual model, which extracts three types of visual features: body shape and clothing, scene atmosphere, and temperament / style, which are defaulted to recommendation intent; speech input is transcribed into text and reused in the text processing chain; features from all modalities are uniformly mapped to the same semantic vector space. To ensure comparability of modal features, they are uniformly mapped to the same semantic vector space, eliminating heterogeneous data barriers.

[0064] Single-turn dialogues can be semantically ambiguous, so historical dialogue memories are used to help correct intentions. In this embodiment, short-term dialogue memories from the most recent 5 to 10 rounds are loaded for fusion and disambiguation, and the memory weights are calculated using an exponential time decay formula:

[0065] In the formula, For the first The weighting of memory in turn-based dialogues; the more recent the time, the higher the weighting. This is the memory decay rate factor, with a value of 0.1d. -1 ; This refers to the current system time. For the first The timing of the dialogue rounds; the weight of long-term dialogues decreases while the weight of recent demands increases; the historical memory vector is weighted and aggregated, and then superimposed with the current input semantic features to resolve semantic ambiguity from three dimensions: topic continuity, emotional fluctuation, and preference accumulation.

[0066] By weighted aggregation of historical memory vectors and superimposed with current input semantic features for joint reasoning, ambiguity is resolved from three dimensions: topic continuity, emotional fluctuation, and preference accumulation, avoiding misjudgment of intent due to misinterpretation; at the same time, explicit and implicit friendship preferences of users are extracted and stored in short-term and long-term memory in real time, accumulating fragments of user profiles.

[0067] When outputting the intent category, preference feature vector, and intent confidence, the contextual features are fused to output the bi-class posterior probability, defining the intent determination rules: ; In the formula, The final determination of the intent category; The posterior probability of the recommendation intention; Let be the posterior probability of the chat intention.

[0068] The system splits recommendation and chat commands into two categories and connects them to downstream business processes. It synchronously calculates the confidence level of intent determination, using the highest posterior probability as the confidence score. If the score is lower than a preset threshold, the system proactively requests supplementary information to correct any ambiguity. The recognition process extracts preferences for dating entities in parallel, synchronizing with the memory module in real time to enable iterative preference iteration during the interaction process, providing structured input for multi-level funnel recommendation matching.

[0069] S30, Intent Triage and Chat Interaction: Triage is performed based on the identified intent category; if the intent category is chat intent, user behavior data, preference analysis results, dialogue memory context, and user profile information are integrated to generate chat response text that conforms to the scenario constraints, and the signal of seeking friendship is continuously monitored during the dialogue, and guided to the recommendation process naturally according to the confidence level, and then the dialogue memory update step S70 is executed; if the intent category is recommendation intent, the recommendation matching process of S40-S60 is executed, and then the dialogue memory update step S70 is executed.

[0070] In this step, the intent category determined by the dual-path intelligent trigger is divided for processing. Specifically, when the intent category is chat intent, user behavior data, preference analysis results, dialogue memory context and user profile are aggregated and input into a large language model to generate chat reply text; during the dialogue, potential friend-making requests are continuously monitored, and when implicit recommendation needs are identified, the user is naturally guided to the recommendation process; when the intent category is recommendation intent, the recommendation and text generation process of memory loading and feature fusion to dialogue memory update is executed.

[0071] Specifically, when performing traffic splitting based on the identified intent category, the intent category of the output is intelligently triggered based on the dual-path mechanism. Execution of traffic splitting. When the intent is chat, the system aggregates information from four dimensions to construct the input context generated by the LLM. In this embodiment, the fusion of information from four dimensions to construct the input context generated by the large language model specifically includes: (1) User behavior data: including triggering method (active / passive), triggering event type, time interval since last interaction, and cumulative number of interactions on the day. Among these, user behavior data reflects the user's current state and the urgency of their needs, and the behavioral feature vector is defined as follows: ; In the formula, For triggering method identification, 0 represents user-initiated input, and 1 represents system-triggered input. This is a passively triggered event type. , indicates the type of event triggered (when passively triggered); Indicates the time interval (in seconds) since the last interaction; This indicates the cumulative number of interactions on that day.

[0072] (2) Preference Analysis Results: Preference keywords and corresponding confidence levels extracted during the intent recognition process. Even if the intent is determined to be chat, preference information may still be included. The extracted preference keywords are represented as follows: ; In the formula, For the first Each preference dimension (age / region / interests, etc.); For the first The preference values ​​corresponding to each preference dimension; For the first Confidence level corresponding to each preference dimension; This represents the total number of extracted preference dimensions.

[0073] (3) Dialogue Memory Context: The K most recent dialogue records, each with a time decay weight; wherein, the K most recent dialogue records are loaded from the short-term memory module, and the dialogue memory context H is represented as: ; In the formula, For the first The roles in the dialogue (user / assistant); For the first The content text of the dialogue; For the first The moment the dialogue occurred; For the number of dialogue memory turns loaded, It consists of 5 to 10 rounds.

[0074] Dialogue memory ensures dialogue coherence, allowing responses to follow the flow of the conversation. A time decay weight is introduced for each record. ; In the formula, For the first Time decay weighting for each dialogue record; the more recent the time, the higher the weight. This is the memory decay rate factor; This refers to the current system time. For the first The moment the dialogue occurred.

[0075] (4) User profile: basic demographic attributes, social style tags, interest tags, and social behavior tags; structured profile information, including basic attributes and social style tags: ; In the formula, Create structured user profile feature vectors; It includes basic demographic attributes, such as gender, age, region, and education level. Social style tags represent a user's communication style and social tendencies; Interest tags represent a user's areas of interest and hobbies; These are social behavior tags, including system-labeled behavioral characteristics such as social activity level and preference stability assessment.

[0076] A dynamic attention mechanism is used to calculate the weights of each dimension, and the weights are dynamically adjusted according to the scenario: the weight of behavioral data increases in passively triggered scenarios, the weight of dialogue memory increases in emotional expression scenarios, and the weight of preference analysis increases in explicit preference expression scenarios.

[0077] Intent triage and chat interaction also include context fusion and weight allocation. The information across the four dimensions is heterogeneous and its importance varies with the scenario; a dynamic weight fusion strategy is adopted. Specifically, the fusion context is defined. for: ; In the formula, This is the context vector after multi-dimensional fusion; For the first Dynamic weights for each information dimension; This is a feature encoding function that maps the original information into a semantic vector; For the first Information sources in each information dimension (Behavior / Preference / Memory / Profile), Weighting Calculated through scene-aware attention mechanisms: ; In the formula, A scene importance scoring function, based on behavioral data. For query conditions, behavioral data As a query condition, the importance of each dimension in the current scenario is dynamically evaluated; , Information dimensions (behavior / preference / memory / profile).

[0078] Passive triggering scenarios ( ): Behavioral data weights Increasing the size of the response affects the direction of the topic. User emotional expression scenarios: Dialogue memory weighting Enlarge it, and maintain emotional continuity; Clearly define the scenario for expressing preferences: Preference analysis weights Increased size facilitates referral conversion.

[0079] When generating chat messages, the merged context is input into a large language model to generate responses: In the formula, For the generated chat reply text; Generate functions for large language models; Input for the fusion context; To generate a set of constraints (style, coherence, length, safety). The generated constraints are shown in Table 1 below: Table 1. Generate Constraints

[0080] The output structure is as follows:

[0081] In the formula, This serves as the main body of the chat reply text; A list of interactive command buttons, such as "Recommend one for me" or "Let's chat about something else," provides users with quick access to operations.

[0082] During the chat, the system continuously calculates the confidence level of the friend request for each round of user messages. Specifically, when extracting the feature vector of the request signal, a signal feature vector is defined. Extract from the current message and conversation context: ; In the formula, The intensity of loneliness signals, such as "nobody's with me" or "I'm so bored"; The strength of the signal indicating interest in making friends, such as "want to get to know" or "do you have any recommendations"; The strength of the signal indicating a need for companionship, such as "wanting to find someone to chat with"; For explicit recommendation request signals, such as "help me find one" or "recommend one"; Confidence calculation of dating requests: ; In the formula, The confidence level for making friends is set to a value between 0 and 1. The Sigmoid activation function maps the input to the 0-1 range. For the first Weights of signal types; For the first The strength of social appeal signals; The weight of signals accumulated for historical demands; Accumulate the average value of appeal signals in historical messages, and capture progressive expressions across messages through dialogue memory calculation.

[0083] In this embodiment, the continuous monitoring of dating request signals specifically refers to: Extract four types of signal features from the current message and conversation context: loneliness, friendship intention, companionship need, and explicit recommendation request; and calculate the confidence level of friendship request. Based on the confidence threshold, the guidance strategy is divided into three levels: Confidence level below the low threshold: Pure companionship chat, without recommendation guidance; Confidence level between low and high thresholds: Social topics are subtly incorporated into responses to provide guidance; Confidence level higher than or equal to the high threshold: Explicitly provide recommendations and direct guidance; Once the user replies with confirmation after being directly guided, the recommendation and matching process officially begins.

[0084] Hierarchical guidance decision-making is represented as: ; In the formula, Output for the bootstrap level; To assess the credibility of dating requests; The high confidence threshold; The threshold is set to low confidence. Table 2 shows the performance of each level of guidance at the copywriting generation level. Table 2. Guidance at all levels in copywriting generation

[0085] When the guidance level is Level 3 and the user's next round of reply indicates confirmation, such as "OK" or "Help me find", the system will combine the previous guidance context in the next round of S20 intent recognition, determine the intent category as recommend, and formally enter the S40-S70 recommendation process. This progressive transformation mechanism ensures that the transition from chat to recommendation is natural and smooth, rather than a mechanical mode switch.

[0086] The LLM generation result needs to be verified by the security detection model. The detection content is the concatenation of speech_text and all command_text: ; In the formula, is the content security detection result; an output of pass indicates that the verification is passed, otherwise it is not passed; is the content security detection model function; is the string concatenation function, which concatenates multiple segments of text into a complete text in order; is the main body of the chat reply copy (speech_text); is all interactive instruction button text (command_text).

[0087] When the risk control fails in the passive trigger scenario, silence is selected to avoid disturbing the user by sending worthless fallback content.

[0088] When rendering the copy with highlights, mark labels are used in the generated copy to specify the keywords that need visual emphasis, and the extraction method is as follows: scan the <strong> tag pairs in the copy with regular expressions; extract the text inside the tags as highlighted words; replace the tags with the placeholder %s; generate a highlighted rendering description array: , where, is the th keyword text that needs to be rendered with highlights; is the th display color corresponding to the highlighted keyword; is the th identifier indicating whether the highlighted keyword is bolded; is the th position index / routing information of the highlighted keyword in the copy; the client renders the highlight effect at the corresponding position according to the description array.

[0089] When outputting, the output is: Chat intent path: send the generated chat reply copy (including interactive instruction buttons) to the user, and update the dialogue memory (enter S70); Recommendation intent path: hand over the control flow to S40, and start memory loading and feature fusion.

[0090] Although the two paths have different processing logics, they share the same memory update mechanism (S70) to ensure that the dialogue content of either path is remembered by the system.

[0091] S40. Memory Loading and Feature Fusion: Parallel loading of three heterogeneous data sources: dialogue memory data, user self-profile data, and current preference features. After processing through three stages—dimensional alignment and standardization, priority weighted fusion, and conflict detection and resolution—enhanced query conditions are constructed, including hard constraints, strong preference conditions, weak preference conditions, and exclusion conditions.

[0092] In this step, the dialogue memory data is the historical preference record of the current user loaded from the memory storage by the dialogue memory module. The memory storage comprises two layers: a short-term memory layer and a long-term memory layer. The short-term memory layer stores the complete context of the most recent K rounds of dialogue, including the user's original input, system response, entity tags extracted from each round of interaction, and user feedback signals for each recommendation result (acceptance, ignore, explicit rejection). In this embodiment, K = 5 to 10 rounds. Short-term memory is used to capture the user's preference trends and immediate needs within the current interaction cycle.

[0093] The long-term memory layer stores persistent preference summaries accumulated across sessions, including the dimensions of preferences explicitly expressed by the user in the past and their frequency statistics, common characteristics of recommendations accepted by the user (i.e., positive feedback induction), and common characteristics of recommendations explicitly rejected by the user (i.e., negative feedback induction). Long-term memory employs an exponential time decay weighting mechanism, which applies to user preferences in the past... The preference characteristics expressed in the next session, at the current moment The memory weights are: ; In the formula, For the first time in history Memory weights of secondary session preference features; In this embodiment, the decay rate factor is used. The time unit is the day; The time interval between the current moment and the moment when this preference was generated; For the current moment, This represents the historical moment in which this preference feature was generated. The decay mechanism ensures that recent interactions are weighted higher than earlier ones, allowing the preference summary to adaptively reflect the user's latest aesthetic and social tendencies.

[0094] In this step, the user's self-profile data is structured attribute data obtained from the user profiling service, including: basic demographic attributes (gender, age, region, education level, etc.), authentication information (real-name authentication status, professional authentication status, etc.), social behavior profile (active time periods, chat style preferences, historical interaction frequency, etc.), and behavioral tags assigned to the user by the system (such as social activity level, preference stability assessment, etc.). The role of the user's self-profile data is not to directly serve as a filtering condition, but rather to provide a "self-awareness context" for feature fusion. For example, for a user labeled as "low social activity," the system will appropriately reduce the weight of the requirement of "high activity level of the other party" during the fusion process, because such users may prefer communication partners with similar frequencies.

[0095] In this step, the current preference feature is the preference feature vector extracted from the intent recognition output for this interaction, including explicit preferences (filtering conditions directly expressed by the user) and implicit preferences (potential tendencies obtained through image analysis or semantic reasoning). The current preference feature represents the user's most explicit need signal at this moment and has the highest priority in the fusion process.

[0096] In this embodiment, the priority weighted fusion is performed according to the priority of current preference > short-term memory > long-term memory > user profile inference, with corresponding weight coefficients of 0.5, 0.25, 0.15, and 0.10, respectively; the conflict detection and resolution adopts the principle of proximity in time: current preference covers short-term memory, and short-term memory covers long-term memory; the feature marking of user explicit negative feedback is used as an exclusion condition.

[0097] In this embodiment, the three heterogeneous data streams are specifically: Dialogue memory data is divided into short-term memory and long-term memory layers. Short-term memory stores the complete context of the most recent 5 to 10 rounds of dialogue and recommendation feedback signals. Long-term memory stores cross-conversation preference frequency statistics, positive feedback inductive features, and negative feedback inductive features, using exponential time decay weighting. User profile data includes basic demographic attributes, authentication information, social behavior profiles, and system behavior tags, which are used to provide context for self-awareness. Current preference features: explicit and implicit preferences obtained from the intent recognition output have the highest priority in the fusion process.

[0098] After the three sets of information are loaded, the system performs a structured feature fusion process, mapping the heterogeneous information into enhanced query conditions that can be retrieved by the recommendation engine. The fusion process includes the following three stages: Phase 1: Dimension Alignment and Standardization. Information from the three data sources involving the same semantic dimension is aligned. For example, the current preference for "young," the long-term memory preference for "22-26 years old," and the short-term memory record of the previous recipient's age being 24 all point to the unified dimension of "age." The system converts fuzzy preferences described in natural language, such as "young" and "mature," into quantifiable intervals or label sets using a predefined semantic mapping table, allowing them to be compared with structured data within the same representation space.

[0099] Phase Two: Priority-Weighted Fusion. Information from different data sources on the same dimension is fused in descending order of priority: Current Preference > Short-Term Memory > Long-Term Memory > User Profile Inference. For the first... Each preference dimension is used to calculate the fused preference strength value. : ; In the formula, For the first The preference intensity value after fusion of all dimensions; The weight of the current preference; For current preferences in dimensions The value on (0 when there is no expression); Weights for short-term memory; Dimensions in short-term memory The normalized strength; Weights for long-term memory; Dimensions in long-term memory The mean preference weighted by the time decay factor; Weights inferred for user profiling; Creating user profiles in dimensions The inferred value is based on the weighting coefficients. and In this embodiment, the values ​​are taken as 0.5, 0.25, 0.15, and 0.10, respectively. When a data source is in dimension When there is no valid information, its weight is redistributed to the other data sources.

[0100] The specific rules are as follows: If the current preference provides explicit conditions in a certain dimension (such as "same city"), then the value of the current preference is directly used as a hard constraint for that dimension. If the current preference does not involve a certain dimension, but short-term memory shows that the user has repeatedly emphasized that dimension in recent interactions (such as mentioning "tall" three times in a row), then that dimension will be included as a strong preference soft constraint in the query conditions. If neither the current preference nor short-term memory involves a certain dimension, but statistics in long-term memory show that the dimension is a persistent user preference (e.g., more than 80% of the objects accepted in the past have a certain feature), then the dimension is included as a weak preference soft constraint. The information inferred from user profiles is only used to fill in dimensions not covered by the first three methods, and is included only with the lowest weight.

[0101] In this embodiment, the enhanced query conditions also include preference confidence metadata, which labels the source and confidence rating of each dimension of the conditions, for use in differentiated weighting during the recommendation and ranking stage; When there are too few effective dimensions, an active completion strategy is initiated, which generates inferred preferences based on the statistical preferences of users in the same group, and fills the blank dimensions with the lowest confidence level.

[0102] Phase Three: Conflict Detection and Resolution. When different data sources provide contradictory information on the same dimension—for example, long-term memory shows a preference for "same city," but the current preference explicitly states "no location restrictions"—the system executes conflict resolution logic: the data source with the highest time proximity takes precedence; that is, the current preference overrides short-term memory, and short-term memory overrides long-term memory. Furthermore, when a user provides explicit negative feedback on a recommendation result, such as "I don't like this type," the system marks that feature dimension as a negative constraint in real time, which then takes effect as an exclusion condition in this query.

[0103] After feature fusion, the system outputs structured enhanced query conditions, which include the following four types of fields: Hard constraints: Screening conditions that must be strictly met; if they are violated, the candidate will be directly excluded. These include core dimensions such as gender, upper and lower bounds of age range, and geographical scope. Strong preference conditions: high-priority soft constraints. When these conditions are not met, the matching score of the candidate decreases significantly but the candidate is not directly excluded. Examples include height preference and education preference. Weak preference conditions: soft constraints with low to medium priority, which slightly reduce the matching score when not met, such as interest tags, aesthetic styles, etc. Exclusion criteria: Exclusion rules generated by negative feedback. If the criteria are met, the candidate object is directly excluded, such as feature types that the user explicitly states they do not accept.

[0104] In addition, the enhanced query conditions also include a set of preference confidence metadata, which annotates the source of the preference conditions (current expression / memory inference) and confidence rating (high / medium / low) for each dimension, so that the recommendation engine in step S50 can perform differentiated weight allocation during the ranking stage.

[0105] When feature fusion reveals that the effective dimensions in the enhanced query conditions are insufficient—meaning the user hasn't expressed enough preference information and lacks usable historical data—the system initiates a proactive completion strategy: based on the current user profile and the statistical preference distribution of users in the same group, a set of "inferred preferences" is generated as default soft constraints to fill in the blank dimensions. These inferred preferences are labeled with the lowest confidence level to ensure that the recommendation range is not overly constrained, while providing the recommendation engine with sufficient signals to avoid completely random recommendations. The generation of inferred preferences is based on: high-frequency preference statistics of users of the same gender and age group in the user's region, and complementary inferences from the user's own profile attributes. For example, a 25-year-old male user is statistically most likely to prefer female users aged 22-28.

[0106] S50, Multi-level Funnel Recommendation Matching: The entire pool of candidate users is subjected to three levels of processing in sequence: hard constraint coarse screening, multi-dimensional weighted fine ranking, and diversity rearrangement. In the fine ranking stage, a weighted multiplicative comprehensive scoring model is adopted, and the two-way matching degree score is used as a multiplication factor to participate in the final score calculation, and the final list of candidate target objects is output.

[0107] In this step, a three-stage processing pipeline of "coarse screening, fine ranking, and diversity rearrangement" transforms the enhanced query conditions output in step S40 into a final list of candidate target objects presented to the user. This multi-stage funnel-shaped recommendation matching architecture solves the engineering challenge of efficiently retrieving highly matching candidate objects from a large-scale user pool while ensuring the diversity and freshness of the recommendation results.

[0108] In this embodiment, the multi-level funnel recommendation matching specifically includes the following three-level funnel recommendation matching: Level 1: Hard Constraint Coarse Screening. After receiving the enhanced query conditions, the recommendation engine first performs a hard constraint coarse screening stage. This stage quickly filters the entire pool of candidate users based on the hard constraints and exclusion conditions in the enhanced query conditions, eliminating objects that do not meet the basic conditions. The specific execution logic is as follows: Hard constraint filtering: Perform exact matching filtering based on hard constraints such as gender, age range, and geographical scope, and retain only candidate objects that meet all hard constraints; Exclusion rule filtering: Remove features based on exclusion conditions generated by negative user feedback to ensure that feature types that users have explicitly rejected do not reappear in the recommendation results; Basic compliance filtering: Perform platform-level compliance checks to exclude accounts that have been banned, deleted, or are inactive (have not logged in for more than a preset number of days), as well as accounts that have been blocked / blacklisted by the current user; Historical deduplication filtering: Excludes objects that have been recommended to the user within the most recent preset time window (72 hours in the example) to avoid repeatedly recommending the same person in a short period of time, which would degrade the user experience.

[0109] The design goal of the coarse screening stage is to reduce the candidate set from the full user pool (usually in the millions to tens of millions) to a coarse screening candidate set in the thousands to tens of thousands within milliseconds, thereby reducing the computational pressure on the fine ranking stage.

[0110] Level Two: Multi-Dimensional Weighted Refinement. After the initial candidate set enters the refinement stage, the system calculates a comprehensive matching score for each candidate. The refinement scoring model integrates signals from the following multiple dimensions: Dimension 1: Preference Matching Score. For both strong and weak preference conditions in the enhanced query, the matching degree between candidate objects and user preferences is calculated dimension by dimension. The calculation method for matching degree varies depending on the dimension type, and specifically includes: For continuous dimensions (such as age and height): a Gaussian decay function is used, with the user preference center value set to 1. Candidate objects in dimension The actual value on Then the matching score for this dimension is: ; In the formula, Candidates In dimensions Continuous matching score; Candidates In dimensions The actual value on; For users in dimensions The preference center value on; For dimension The tolerance width parameter, in this embodiment, is the age dimension. Age, height dimensions The closer a candidate value is to the preference center, the higher its score; the greater the deviation, the lower the score. The score approaches zero if the range is too wide.

[0111] Discrete dimensions (such as region, education level): use exact matching or hierarchical matching (same city > same province > no limit), the higher the matching level, the higher the score.

[0112] Tag-based dimensions (such as interest tags, aesthetic styles): Set similarity calculation (Jaccard coefficient) is used, assuming the user preference tag set is... The candidate object tag set is Then the tag dimension matching score is In the formula, Candidates In dimensions Continuous matching score; A set of user preference tags; Candidates A collection of tags.

[0113] The matching scores for the continuous, discrete, and labeled dimensions mentioned above are weighted according to the preference confidence metadata attached in step S40: the weight coefficient of the high confidence dimension. Greater than medium confidence, and medium confidence is greater than low confidence. In this embodiment, high / medium / low confidence are respectively taken as... Dimension The weighted preference matching score is In the formula, Candidates The weighted preference matching total score; For dimension The confidence weight; Candidates In dimensions Match score; This represents the total number of preference dimensions; ensuring that explicitly expressed user preferences have a greater impact on the ranking.

[0114] Dimension Two: Object Quality Score. Independent of user preferences, the system maintains a quality baseline score for each candidate, reflecting their social activity and popularity on the platform. The quality score is calculated based on the following factors: completeness of profile (avatar, bio, photos, etc.), authentication level (real-name authentication, professional authentication, etc.), recent activity (last login time, average daily online time), and historical social performance (acceptance rate of advances, average response time, etc.). The quality score ensures that the recommendation results do not contain a large number of "zombie accounts" or long-term inactive users, improving the probability of effective conversion from recommendations.

[0115] Dimension Three: Two-Way Matching Score. Unlike traditional one-way recommendation (which only considers the current user's preference for candidate matches), this invention introduces a two-way matching evaluation mechanism. The system uses the candidate's historical behavioral data (such as the distribution of their tendency to accept advances from potential partners) to inversely evaluate whether the current user is likely to be accepted by the candidate. Let the current user be... Candidates are The two-way matching score is defined as: ; In the formula, Candidates For the current user The two-way matching degree (acceptance probability); Candidates A collection of users who have historically been approached by strangers. Candidates A collection of users who have historically refused advances. For the current user eigenvectors; Historical users eigenvectors; The cosine similarity function; To ensure smoothness and prevent the denominator from being zero, this score reflects the similarity between the current user and the candidate's historical preference demographics; a higher score indicates a greater probability of acceptance. The two-way matching score is used as a multiplicative factor in the final score calculation, achieving mutually possible matching optimization.

[0116] Dimension Four: Timeliness and Freshness Scoring. To avoid stagnation in recommendation results, the system introduces a timeliness factor. Newly registered users or users who have recently uploaded photos / updated information receive a timeliness bonus, making them more likely to appear in recommendation results; while objects that have existed on the platform for a long time but have not been touched by the current user receive a small "long-tail discovery" bonus, helping users discover potential matches that were previously overlooked.

[0117] In this embodiment, the multi-dimensional weighted ranking integrates four dimensions of scoring: Preference matching score: Gaussian decay function is used for continuous dimensions, hierarchical matching is used for discrete dimensions, and Jaccard coefficient is used for labeled dimensions; weighted according to preference confidence. Object quality score: calculated based on a combination of data completeness, authentication level, recent activity level, and historical social behavior; Two-way matching score: Based on the set of users who have accepted and rejected candidates in the past, calculate the cosine similarity between the current user and the users who have historical preferences, which represents the probability of acceptance by the other party; Timeliness and freshness score: Newly registered users and users with newly updated information receive bonus points, and long-tail untouched objects receive discovery bonus points.

[0118] In the comprehensive score calculation, the system uses a weighted multiplication model to calculate the candidate objects based on the scores of the above four dimensions. Overall match score: ; In the formula, Candidates The final overall score; Candidates Preference matching score; Candidates Object quality rating; Candidates Timeliness and freshness rating; Candidates Two-way matching score; These are the weighting coefficients for preference matching. The weighting coefficient for the quality weight of the object; The weighting coefficients for timeliness and freshness are set to 0.6, 0.25, and 0.15, respectively. The two-way matching degree is calculated using a multiplicative factor to ensure that even with an extremely high preference matching degree, the overall score will significantly decrease if the other party's acceptance probability is extremely low. The system sorts the initial candidate set in descending order of score, and then selects the top 50-100 candidates from the overall score in descending order for the re-ranking stage.

[0119] Level 3: Diversity Reordering and Result Truncation. While the Top-M candidate list obtained from the fine-grained ranking is arranged according to the best matching degree, it may suffer from homogenization issues, such as top-ranked candidates all coming from the same region or having highly similar appearances. To improve the user's browsing experience and exploration space, the system performs diversity reordering before the final output.

[0120] In this embodiment, the diversity rearrangement applies three types of constraints: Regional diversity constraint: The proportion of candidates from the same city shall not exceed 60%; in the final recommendation list, the proportion of candidates from the same city shall not exceed the preset threshold of the total number, which is 60% in this embodiment, to ensure that users can see different regional options.

[0121] Style diversity constraint: Include at least two visual style types; based on the visual style tags of the candidate objects, such as intellectual, sweet, sporty, etc., ensure that the recommendation list contains at least two different style types.

[0122] Interaction status diversity constraint: It includes both completely new and untouched objects and objects with whom there has been light interaction; the recommendation list includes both "completely new objects that have never been encountered" and "objects with whom there has been light interaction, such as browsing each other's homepages but not initiating a conversation", taking into account both the psychological needs of users for novelty exploration and relationship progress.

[0123] After diversity rearrangement, the system selects no more than a preset number N_result candidate target objects from the adjusted list as the output for this recommendation matching. In this embodiment, 3 to 5 candidate target objects are selected as the output after rearrangement. Each output candidate object includes: a unique object identifier, an object profile summary (used for copy generation in step S60), a matching score and its sub-dimensional score details (used for generating recommendation reasons in step S60), and a list of matching preference dimensions (indicating which dimensions of user preferences the object performs well in).

[0124] When a user provides feedback on the recommendation results (accepting a specific item, ignoring all, or explicitly expressing dissatisfaction), the feedback signal is sent back to this module in real time to dynamically adjust the recommendation strategy for positive feedback, ignoring feedback, or negative feedback: Positive feedback: After a user accepts a candidate, the features of that candidate are marked as "positive confirmation", and the confidence level of the corresponding preference dimension is strengthened in the memory update in step S70; Ignore feedback: If a user browses the recommendation list but does not take any action on any object, the system infers that the matching accuracy is insufficient and will appropriately expand the tolerance range of the preference conditions in the next recommendation to increase the exploratory nature; Negative feedback: If a user explicitly states that they "don't like these" or expresses dissatisfaction with the recommendation results, the system triggers a preference correction process. By asking follow-up questions, the system guides the user to supplement or correct their preference description, updates and enhances the query conditions, and then re-executes the recommendation matching.

[0125] Intent-based traffic segmentation, chat interaction, and multi-level funnel recommendation matching achieve a complete transformation from vague user preferences to a precise list of candidate objects. Through multi-source information fusion, the system overcomes the limitations of a single data source; through a three-level funnel architecture, the system efficiently identifies the most matching candidates from a massive user pool; and through bidirectional matching degree evaluation and diversity re-ranking, the system ensures that the recommendation results are both accurate and have practical conversion potential, providing a high-quality candidate object foundation for the contextualized copywriting generation in step S60.

[0126] S60. Recommended Scenario Copy Generation: Based on enhanced query conditions and candidate object features, perform dimensional matching degree correlation analysis, select the optimal matching dimension, and generate three sets of differentiated copy: chat guidance copy, recommendation reason copy, and icebreaker copy. After content security verification, these are assembled into a recommendation card sequence and sent to the user.

[0127] After preference fusion filtering in S40 and multi-level funnel-style recommendation matching in S50, the system has obtained no more than N high-quality candidate objects. This recommendation scenario copy generation step is used to generate three sets of differentiated copy for these candidate objects: chat copy (to guide the user and introduce the recommendation results), recommendation reasons (explaining why this person is recommended), and icebreaker copy (an opening line that the user can directly send to the candidate). In this embodiment, the generation constraints of the three sets of differentiated copy are as follows: Chat prompts should connect to recent conversation topics, convey a sense of companionship, and naturally lead to recommended results. Recommendation copy: Based on the Top-K optimal matching dimension, highlighting the commonalities or complementarities between the two parties, including specific details; Icebreaker copywriting: Focus on the candidate's interests, match the user's social style, and create a hook for conversation by combining common ground between the two parties.

[0128] When generating copy for recommended scenarios, the input comes from the cumulative output of previous steps, as shown in Table 3: Table 3 Input sources for recommended scenario copy generation

[0129] Before invoking LLM, the system performs structured preference-feature association analysis to provide matching criteria for copy generation. This involves analyzing each dimension of the preference conditions after S40 fusion. The matching degree is calculated with the corresponding features of the candidate object: ; In the formula, For dimension Preference-feature matching degree; For users in dimensions Preference values, For the candidate object in the first dimensional eigenvalues This represents the range of numerical normalization.

[0130] The top-K dimensions with the highest matching degree are selected as the core materials for the recommendation reasons: In the formula, The set of Top-K dimensions with the highest matching degree; To take the front The function whose maximum value corresponds to the subscript; For the first Dimensional matching degree; The optimal number of dimensions to select.

[0131] For example, if a user's preference is "travel + Shanghai" and the user's characteristics are "have been to 12 countries + located in Shanghai", then "travel interests" and "region" will be selected as matching highlights.

[0132] The association analysis results and multi-dimensional context are input into the LLM (Local Management Module), and three sets of content are generated through scenario-based constraints: chat text (introduction), recommendation reasons, and icebreaker text. The chat text (introduction) is used to naturally lead the user to the recommendation results, continuing the previous dialogue context, and its generation constraints are as follows: The prompts should follow up on recent topics from the conversation, such as when the user just said, "Help me find someone who likes to travel," and the prompts should echo this need. This reflects the assistant's role of "companionship" and "helping to introduce"; The tone was natural, like a friend saying, "I've found someone for you."

[0133] Example output: "Based on what you said, I found a girl who also loves to travel~ Let's see if she's to your liking."

[0134] The recommendation reason explains why this object is recommended, enhancing user trust and click intent. The generation constraints are: it must be based on Top-K matching dimensions (avoiding fabricated non-existent associations); it should highlight the commonalities or complementarities between the two parties; it should be specific and detailed, such as "has been to 12 countries" being better than "loves to travel"; the recommendation reasons for different candidates should have differentiated perspectives. For example, the output could be: "She has been to 12 countries, recently returned from a certain place, and like you, enjoys exploring new places."

[0135] Icebreaker text is used to generate the first sentence that users can directly send to their target audience, lowering the barrier to social interaction. The generation constraints are: it must revolve around the target's characteristics and interests (ensuring the target has something to reply to); it must match the social style in the user profile (extroverted users can be more direct, introverted users more subtle); it must feel natural, like a normal greeting, rather than a templated "Hello, nice to meet you"; it should appropriately incorporate common ground to create conversation hooks; the quality of the icebreaker text must simultaneously meet three constraints, and the quality score of the icebreaker text is calculated by weighting these three dimensions.

[0136] In the formula, The overall quality score for the icebreaker copywriting; The relevance of the icebreaker copy to the characteristics of the target audience; To ensure the icebreaker copy matches the user's communication style; To ensure the responsiveness of icebreaker copywriting; , , These are the weight coefficients for the corresponding dimensions. For example, the output might be: "I saw you went to [place name], do you have any recommendations for fun places to visit? I've always wanted to go but haven't planned my route yet."

[0137] The dialogue memory records previously recommended objects and the angles from which they were used in the recommendation process. During copy generation, constraints are applied to object, angle, and copy deduplication. Object deduplication: Previously recommended objects have already been excluded in the S50 funnel stage. Angle deduplication: If the same user triggers recommendations multiple times, LLM, combined with dialogue memory, avoids previously used recommendation angles; for example, if the "travel" angle was used last time, it can be switched to the "same city" angle this time. Copy deduplication: To avoid generating expressions highly similar to historical recommendation reasons.

[0138] When the recommendation reason or icebreaker text generated by the LLM is empty, the system executes a fallback mechanism based on field segmentation. In this embodiment, a fallback degradation mechanism is set up: when the generated text is empty, a fallback text is selected from a pre-configured fallback pool; the fallback pool is shuffled by Fisher-Yates and then allocated in order to ensure that the text is not repeated in scenarios with multiple candidate objects; when the content security verification fails, the fallback text is sent in the actively triggered scenario, and is not sent in the passively triggered scenario.

[0139] When the recommendation reason is empty, it is randomly selected from the pre-configured default recommendation reason fallback pool. The text in the pool is shuffled in advance and assigned in order to avoid duplication within the same batch. When the icebreaker text is empty, it is randomly selected from the default pre-input fallback pool; when the chat text is empty, a general guidance template is used. The fallback pool adopts a Fisher-Yates shuffling and sequential assignment strategy to ensure that the fallback texts for each object are not duplicated in scenarios with multiple candidates.

[0140] After the three sets of text are generated, they must undergo content security checks before being sent. All text is then concatenated and sent to the risk control interface. ; In the formula, This is the complete text to be tested, which is then assembled and sent to the risk control interface. Chat prompts (introductory text); , ...the recommendation text for the 1st, 2nd...th candidate; , ...icebreaker copy for 1, 2... candidate objects; This is a text concatenation character, indicating that multiple text segments are concatenated in sequence.

[0141] If the risk control fails, a backup message will be sent to replace the default message if the scenario is actively triggered; otherwise, no message will be sent.

[0142] After risk control is passed, a recommendation card message is assembled for each candidate. The card information structure of each card includes: Basic information about the target (avatar, nickname, age, height, city); Geographical distance (calculated using the Haversine formula between male and female users); Recommendation reason text; Icebreaker text (pre-filled into the input box); A unique identifier, greet_id (for tracking purposes).

[0143] In this embodiment, the rule for arranging the recommended card sequence is as follows: Image 1: Chat prompts, recommended card information, and keyword highlighting; assistant avatar displayed. Cards 2 through N-1: Purely recommended card information, assistant avatar hidden; Card N: Recommended card information and interactive operation buttons; assistant avatar hidden. Multiple cards are sent at fixed intervals with progressively increasing timestamps, and each icebreaker message is assigned a unique identifier for persistent storage.

[0144] The system assigns a unique greeting_id to each icebreaker message and writes it to the database in batches, allowing users to use it with one click or edit and send it. After sending the complete recommended display unit to the user, the system enters S70 to update the dialogue memory (introduction, recommendation card and operation button), and then enters S70 to update the dialogue memory again, writing the complete content of this recommendation (who was recommended, the reason for the recommendation, and the icebreaker message) into the memory to avoid repeated recommendations in the future.

[0145] S70, Dialogue Memory Update: Asynchronously write the user input, system response, entity label and user feedback signal of this interaction into the memory module, update the short-term dialogue memory and long-term preference memory respectively, forming a closed loop for continuous evolution of recommendation accuracy.

[0146] This step asynchronously writes the user input and system response of this interaction into the memory module, updates the short-term context window and long-term preference memory, and makes them available for loading in steps S30 and S40, forming a closed loop for continuous evolution of recommendation accuracy; among them, the dialogue content under the chat intent (S30) path is also written into the memory module, which is used to understand the user's communication style and potential preferences when identifying intent and making recommendations.

[0147] In this embodiment, the long-term preference memory update includes statistical preference dimension occurrence frequency, summarizing positive feedback features and summarizing negative feedback features. The updated memory data is used for intent recognition, feature fusion and recommendation matching steps.

[0148] See Figure 4 As shown, this embodiment of the invention provides a social recommendation icebreaker system with multimodal intent recognition and behavior awareness, comprising: Intelligent triggering module: used to generate trigger signals through active input paths and behavior-aware passive triggering paths. After the passive triggering signal is filtered by multi-level frequency control, cooling window constraints and pre-entry verification, it is integrated with the active triggering signal to output global triggering criteria. Multimodal intent recognition module: This module is used to split multimodal inputs, extract the preference features of each modality, and combine them with short dialogue memory with time decay for disambiguation, outputting intent category, preference feature vector and confidence score. Intent Triage and Chat Interaction Module: This module is used to triage and process messages based on intent categories. It integrates multi-dimensional information to generate chat reply text under each chat intent and continuously monitors signals of friendship requests, guiding users to naturally transition into the recommendation process in a tiered manner. The memory loading and feature fusion module is used to load heterogeneous data from three sources—dialogue memory, user profile, and current preferences—in parallel under the recommendation intent. After dimensional alignment, weighted fusion, and conflict resolution, it constructs enhanced query conditions. Multi-level funnel recommendation module: It is used to perform hard constraint coarse screening, multi-dimensional weighted fine ranking and diversity rearrangement on the candidate user pool in sequence under the recommendation intention. It uses a weighted multiplication model to calculate the comprehensive score by using the bidirectional matching degree as the multiplication factor and outputs a list of candidate target objects. Recommendation copy generation module: used to generate chat prompts, recommendation reasons and icebreakers based on preference and feature correlation analysis under the recommendation intent, and then assemble them into a recommendation card sequence for sending after security verification; Dialogue Memory Module: Used to asynchronously store interaction data, update short-term dialogue memory and long-term preference memory respectively, forming a closed loop of preference evolution.

[0149] In this embodiment, the intelligent triggering module monitors six types of user behavior through behavior perception passive triggering path monitoring: homepage browsing pauses for more than 8 seconds, no private messages viewed in the foreground for 30 seconds, no chat initiated for 30 consecutive seconds, no opening remarks sent in a new conversation for 120 seconds, no reply to 3 consecutive read private messages, and no input on the assistant page for 10 seconds; multi-level frequency control includes three levels of frequency control: global daily level, conversation cycle, and event granularity, with high-priority rules directly rejecting low-priority requests.

[0150] In this embodiment, the multimodal intent recognition module includes a text intent recognition unit, a visual intent recognition unit, and a memory fusion disambiguation unit. The visual intent recognition unit takes an image as input and extracts three types of visual features: body shape and clothing, scene atmosphere, and temperament / style, which are determined by default to be recommendation intent. The memory fusion disambiguation unit loads 5-10 rounds of short-term dialogue memory and uses exponential time decay weighted aggregation to resolve semantic ambiguity in a single round. In this embodiment, the intent triage and chat interaction module includes a multi-dimensional context fusion unit, a chat text generation unit, and a friend-making demand monitoring and guidance unit. The friend-making demand monitoring and guidance unit extracts four types of signals: loneliness, friend-making intention, companionship need, and explicit request. After calculating the confidence level, it executes a three-level guidance strategy: pure chat, topic hints, and direct recommendation.

[0151] In this embodiment, the multi-level funnel recommendation module includes a coarse screening unit, a fine ranking unit, and a diversity rearrangement unit. The fine ranking unit calculates scores for four dimensions: preference matching degree, object quality, bidirectional matching degree, timeliness, and freshness. A comprehensive score is calculated using a weighted multiplication model, with the bidirectional matching degree score used as a multiplication factor in the calculation.

[0152] In this embodiment, the recommended text generation module includes a matching degree correlation analysis unit, three sets of text generation units, a security verification unit, and a card assembly unit; the card assembly unit arranges the recommended card sequence according to the rule that the first card has a guiding text and a highlight, the middle cards are pure cards, and the last card has an operation button.

[0153] The multimodal intent recognition and behavior-aware social recommendation icebreaker method and system provided by this invention extracts intents through three modalities (text, image, and voice) and combines them with time-decayed dialogue memory for joint intent recognition. Users can directly express subtle preferences using natural language or images, which helps lower the threshold for preference expression and reduce preference loss due to insufficient tag dimensions. Through behavior-aware passive triggering paths, it identifies six types of implicit request behaviors and uses a three-level frequency control system (global daily level, application session level, and subdivided event level), exponential decay cooling, and pre-entry interlocking with existing users before triggering. This allows it to identify when users need help while constraining the frequency of contact, achieving a balance between increasing activity and avoiding disturbance. Furthermore, it generates specific matching parameters through dimensional correlation analysis between users and candidate objects. The invention includes detailed recommendation reasons and icebreaker texts that revolve around the target's interests, take into account the user's social style, and ensure reproducibility. These are pre-filled into the input box, reducing the burden on users to organize opening remarks and providing topics for follow-up between matched individuals. By using exponential decay weighting of short-term and long-term memories, conflict resolution based on temporal proximity, and asynchronous memory updates after interaction, recent preferences are given higher weight, and negative feedback features are promptly eliminated. This allows the recommendation results and text angles to be continuously adjusted as user preferences evolve. The invention introduces a two-way matching degree based on the candidate's historical acceptance and rejection groups in the ranking process, with a multiplicative factor participating in the comprehensive score. At the same time, it uses three diversity constraints—region, style, and interaction status—for re-ranking, which helps to balance the possibility of mutual acceptance and the exploration space of the recommendation list.

[0154] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art can still combine, add, delete, or otherwise adjust the features of the various embodiments of the present invention according to the circumstances without conflict or creative effort, thereby obtaining different technical solutions that do not fundamentally depart from the concept of the present invention. These technical solutions also fall within the scope of protection of the present invention.

Claims

1. A social recommendation icebreaker method based on multimodal intent recognition and behavior perception, characterized in that, Includes the following steps: Active trigger signals are generated through active input paths, and original passive trigger signals are generated through behavior-aware passive trigger paths. The original passive trigger signal is filtered by a triple interlocking system of multi-level frequency control, cooling window constraint and pre-access verification, and then integrated with the active trigger signal to output a global trigger criterion to trigger the subsequent interaction process. Modality splitting is performed on the input multimodal data, and preference features corresponding to each modality are extracted differentially. Load short-term dialogue memory with time decay weights for fusion and disambiguation, and output intent category, preference feature vector and intent confidence; Traffic is segmented based on the identified intent category; If the intent category is chat intent, then user behavior data, preference analysis results, dialogue memory context, and user profile information are integrated to generate chat response text that conforms to the scenario constraints. During the conversation, signals of a desire to make friends are continuously monitored, and the user is guided naturally into the recommendation process according to confidence level, followed by the dialogue memory update step. If the intent category is recommendation intent, then the following recommendation matching process is executed: The system loads three heterogeneous data sources in parallel: dialogue memory data, user profile data, and current preference features. After processing through three stages—dimensional alignment and standardization, priority weighted fusion, and conflict detection and resolution—enhanced query conditions are constructed, including hard constraints, strong preference conditions, weak preference conditions, and exclusion conditions. The entire pool of candidate users is subjected to three levels of processing: hard constraint coarse screening, multi-dimensional weighted fine ranking, and diversity rearrangement. The fine-ranking stage employs a weighted multiplication comprehensive scoring model, using the two-way matching score as a multiplication factor in the final score calculation, and outputting the final list of candidate target objects. Based on enhanced query conditions and candidate object features, a dimensional matching degree correlation analysis is performed to select the optimal matching dimension and generate three sets of differentiated copy: chat guidance copy, recommendation reason copy, and icebreaker copy. After content security verification, these are assembled into a recommendation card sequence and sent to the user. After the recommendation matching process is completed, the dialogue memory update step is executed: the user input, system response, entity label and user feedback signal of this interaction are asynchronously written into the memory module, and the short-term dialogue memory and long-term preference memory are updated respectively, forming a closed loop for continuous evolution of recommendation accuracy.

2. The social recommendation icebreaker method based on multimodal intent recognition and behavior perception according to claim 1, characterized in that, The active input path receives three types of multimodal input: text, images, and voice. Text carries fuzzy dating preferences, images convey appearance and aesthetic preferences, and voice is transcribed into text by the ASR engine and then incorporated into a unified processing pipeline. An active trigger signal is generated upon detecting any valid user input.

3. The social recommendation icebreaker method based on multimodal intent recognition and behavior perception according to claim 1, characterized in that, The behavior-aware passive triggering path collects user application behavior through client-side tracking points. An original passive triggering signal is generated when any one of the following six types of triggering behavior is met: homepage browsing pause time exceeds 8 seconds; no private messages are viewed in the foreground for 30 seconds. No initiation of a chat for 30 consecutive seconds; no opening remarks sent for 120 seconds after creating a new conversation; viewing private messages from 3 users consecutively without replying; no input for 10 seconds after entering the assistant page.

4. The social recommendation icebreaker method based on multimodal intent recognition and behavior perception according to claim 1, characterized in that, The multi-level frequency control adopts a top-down three-level frequency control mechanism, where high-priority rules directly reject low-priority triggering requests, including: Global daily frequency control: The maximum number of passive triggers per day is 10; Session lifecycle frequency control: The maximum number of times this function can be triggered within the lifecycle of a single application is 5. Event-level frequency control: Count the trigger frequency of each type of behavior at the hourly and daily levels, and block it if it exceeds the corresponding threshold; The total frequency control factor is taken as the minimum value of the three-level frequency control factors.

5. The social recommendation icebreaker method based on multimodal intent recognition and behavior perception according to claim 1, characterized in that, The cooling window constraint is set to a uniform cooling cycle of 2 minutes, and the cooling factor is calculated using an exponential decay formula. In the formula, This is the attenuation coefficient, with a value of 0.

5. ; To standardize the cooling cycle; The current moment; The time since the last trigger is used; if the interval between the last trigger and the last trigger is sufficient, the passage is allowed; if the interval is insufficient, the passage is blocked by exponential decay.

6. The social recommendation icebreaker method based on multimodal intent recognition and behavior perception according to claim 1, characterized in that, The pre-access verification is passively triggered when all three of the following conditions are met: The message interval between the user and the assistant is greater than or equal to the cooldown period; The message interval between the assistant and the user is greater than or equal to the cooldown period; The users are existing users who are interacting with the platform, not new users of the new features. The global trigger criterion is the maximum value of the active trigger signal and the filtered passive trigger signal. If the criterion is 1, the intent recognition process is started; otherwise, the system remains silent.

7. The social recommendation icebreaker method based on multimodal intent recognition and behavior perception according to claim 1, characterized in that, Modal splitting of the input multimodal data includes the following steps: Text input is fed into the LLM intent recognition model, which splits intent categories through cue engineering and few-shot learning. The image input is fed into the VLM visual model, which extracts three types of visual features: body shape and clothing, scene atmosphere, and temperament and style, and determines them as recommendation intent by default. After voice input is transcribed into text, the text processing link is reused. All modal features are uniformly mapped to the same semantic vector space.

8. The social recommendation icebreaker method based on multimodal intent recognition and behavior perception according to claim 1, characterized in that, The most recent 5-10 rounds of short-term dialogue memory are loaded for fusion and disambiguation, and the memory weights are calculated using an exponential time decay formula. In the formula, The decay rate factor is 0.1d. -1 ; This refers to the current system time. For the first The timing of the dialogue; the weight of long-term dialogues decreases, while the weight of short-term demands increases; After weighted aggregation of historical memory vectors, the current input semantic features are superimposed to resolve semantic ambiguity from three dimensions: topic continuity, emotional fluctuation, and preference accumulation.

9. The social recommendation icebreaker method based on multimodal intent recognition and behavior perception according to claim 1, characterized in that, The specific meaning of continuously monitoring dating request signals is as follows: Extract four types of signal features from the current message and conversation context: loneliness, friendship intention, companionship need, and explicit recommendation request; and calculate the confidence level of friendship request. Based on the confidence threshold, the guidance strategy is divided into three levels: Confidence level below the low threshold: Pure companionship chat, without recommendation guidance; Confidence level between low and high thresholds: Social topics are subtly incorporated into responses to provide guidance; Confidence level higher than or equal to the high threshold: Explicitly provide recommendations and direct guidance; Once the user replies with confirmation after being directly guided, the recommendation and matching process officially begins.

10. A social recommendation icebreaker system based on multimodal intent recognition and behavior perception, characterized in that, A social recommendation icebreaker method for performing multimodal intent recognition and behavior awareness according to any one of claims 1-9, comprising: Intelligent triggering module: used to generate trigger signals through active input paths and behavior-aware passive triggering paths. After the passive triggering signal is filtered by multi-level frequency control, cooling window constraints and pre-entry verification, it is integrated with the active triggering signal to output global triggering criteria. Multimodal intent recognition module: This module is used to split multimodal inputs, extract the preference features of each modality, and combine them with short dialogue memory with time decay for disambiguation, outputting intent category, preference feature vector and confidence score. Intent Triage and Chat Interaction Module: This module is used to triage and process messages based on intent categories. It integrates multi-dimensional information to generate chat reply text under each chat intent and continuously monitors signals of friendship requests, guiding users to naturally transition into the recommendation process in a tiered manner. The memory loading and feature fusion module is used to load heterogeneous data from three sources—dialogue memory, user profile, and current preferences—in parallel under the recommendation intent. After dimensional alignment, weighted fusion, and conflict resolution, it constructs enhanced query conditions. Multi-level funnel recommendation module: It is used to perform hard constraint coarse screening, multi-dimensional weighted fine ranking and diversity rearrangement on the candidate user pool in sequence under the recommendation intention. It uses a weighted multiplication model to calculate the comprehensive score by using the bidirectional matching degree as the multiplication factor and outputs a list of candidate target objects. Recommendation copy generation module: used to generate chat prompts, recommendation reasons and icebreakers based on preference and feature correlation analysis under the recommendation intent, and then assemble them into a recommendation card sequence for sending after security verification; Dialogue Memory Module: Used to asynchronously store interaction data, update short-term dialogue memory and long-term preference memory respectively, forming a closed loop of preference evolution.

Images