Multi-modal emotion recognition-based social map construction and matching system

By collecting and generating multi-dimensional emotion vectors through a multimodal emotion recognition system, constructing an emotion map, and performing personalized matching, the problem of insufficient emotion capture and utilization in existing social platforms is solved. This enables the visualization and personalized matching of user emotions, thereby improving the social experience and emotional health management.

WO2026232648A1PCT designated stage Publication Date: 2026-11-12EMOTIVERSE TECH DEVELOPMENT (HANGZHOU) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/CN2025/097325
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-09
Filing Date
2025-05-27
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Existing social platforms lack the ability to capture and utilize users' real-time emotions, which means that traditional matching methods cannot meet users' deep needs for psychological resonance and personalized social interaction, and there is a lack of technical solutions for building dynamic social maps and refining matching.

Method used

The system uses a multimodal emotion recognition system to collect users' text input, voice expression, and facial expression data, generate multidimensional emotion vectors, construct an emotion map, and perform personalized matching recommendations based on the emotion vectors. This includes emotion intervention and guidance mechanisms, and the matching algorithm is updated and optimized in real time.

Benefits of technology

It enables the visualization and personalized matching of user emotions, effectively identifies and intervenes in negative emotions, enhances the emotional health index of social platforms, and provides an immersive virtual social experience.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN2025097325_12112026_PF_FP_ABST
    Figure CN2025097325_12112026_PF_FP_ABST
Patent Text Reader

Abstract

The present invention relates to the fields of artificial intelligence, social recommendation systems, and emotion computation, and in particular to a multi-modal emotion recognition-based social map construction and matching system. The system comprises a data collection module, an emotion recognition module, an emotion vector generation module, an emotion map construction module, an emotion matching recommendation module, and an emotion intervention and guidance module. The system collects text, voice, and expression data of a user, recognizes an emotion feature, generates a multi-dimensional emotion vector, and displays an emotional state of the user in an emotion map. The system performs emotion-driven user matching on the basis of the emotion vector, and generates personalized recommendation. The emotion intervention and guidance module is used for monitoring negative emotions and providing emotional support. In the present invention, by means of an emotion visualization map, a real social path is provided, and negative emotions are recognized and intervened, thereby enhancing the emotional health index in platforms.
Need to check novelty before this filing date? Find Prior Art

Description

A Social Mapping and Matching System Based on Multimodal Emotion Recognition Technical Field

[0001] This invention relates to the fields of artificial intelligence, social recommendation systems, and emotion computing, and in particular to a method and system for recognizing user emotions based on multimodal input and using emotional states to construct social maps and personalize matching recommendations. Background Technology

[0002] Currently, most mainstream social platforms use interest tags, geographical location, or behavioral preferences for user matching, lacking the ability to capture and utilize users' "real-time emotions" and their fluctuations. In an era where emotional expression is increasingly becoming the core of social interaction, traditional static matching methods can no longer meet users' deep needs for psychological resonance, emotional release, and personalized social interaction.

[0003] In recent years, although individual technologies such as natural language processing, speech recognition, and facial expression recognition have made some progress, there is still no complete solution to integrate multimodal emotion data for dynamic social map construction and refined matching. How to deeply integrate users' current emotional dynamics with social relationship recommendations is a pressing technological challenge that needs to be addressed. Summary of the Invention

[0004] This invention aims to address the technical problem of the lack of emotional dimension matching in existing social platforms, and provides a social matching system capable of capturing, analyzing, and utilizing users' emotional states in real time. Through multimodal emotion recognition and emotion-driven matching mechanisms, this invention can provide users with more personalized social connections that meet their psychological needs.

[0005] This invention proposes a social map construction and matching system based on multimodal emotion recognition, comprising:

[0006] The data acquisition module is used for:

[0007] Collect user's text input, voice expression, and facial expression data;

[0008] The collected multimodal data is transmitted to the emotion recognition engine;

[0009] The emotion recognition engine, which is communicatively connected to the data acquisition module, is used for:

[0010] Receive multimodal data transmitted by the data acquisition module;

[0011] Based on the multimodal data, identify the user's emotional characteristics;

[0012] The emotion vector generation module, which is communicatively connected to the emotion recognition engine, is used for:

[0013] Receive the emotional features identified by the emotion recognition engine;

[0014] Based on the emotional features, a multi-dimensional emotion vector is generated, which includes positive and negative emotion values, activity level, dominance, and confidence level.

[0015] The emotion map construction module, which is communicatively connected to the emotion vector generation module, is used for:

[0016] Receive the multidimensional emotion vector generated by the emotion vector generation module;

[0017] The multidimensional emotion vectors are mapped to an emotion coordinate system to form an emotion map;

[0018] The emotion map displays the distribution of users' emotional states.

[0019] The emotion matching recommendation system, communicatively connected to the emotion map construction module and the emotion vector generation module, is used for:

[0020] Receive the multidimensional emotion vector generated by the emotion vector generation module;

[0021] Receive emotion map data provided by the emotion map construction module;

[0022] Based on the multidimensional emotion vector and the emotion map data, emotion-driven user matching is performed.

[0023] Generate and output matching recommendation results;

[0024] The emotion intervention and guidance module, which is communicatively connected to the emotion matching recommendation system, is used for:

[0025] Receive user emotional state data provided by the emotion matching and recommendation system;

[0026] Monitor negative emotional states;

[0027] When a concentration of negative emotions is detected, emotional support services are provided.

[0028] Preferably, the emotion matching recommendation system includes:

[0029] The multidimensional emotion vector similarity calculation unit is used for:

[0030] Receive multidimensional emotion vectors from two users;

[0031] Calculate the similarity between the two multidimensional emotion vectors;

[0032] A matching threshold is set based on calculated similarity.

[0033] The emotional complementarity and resonance matching unit is used for:

[0034] Determine user matching pattern preferences based on users' historical interaction data;

[0035] Based on the matching pattern preference, select either the emotional resonance matching strategy or the emotional complementarity matching strategy;

[0036] Calculate the matching score between users based on the selected matching strategy;

[0037] The temporal emotion trajectory prediction unit is used for:

[0038] Collect the sequence of emotion vectors within a user's continuous time window to form an emotion trajectory;

[0039] Identify patterns of user emotional fluctuations;

[0040] Predicting users' short-term emotional state;

[0041] Multi-factor matching scoring unit, used for:

[0042] It considers multiple factors, including emotional similarity, timeliness, interaction history, social compatibility, and trend matching.

[0043] Generate the final match score;

[0044] Feedback optimization unit, used for:

[0045] Collect feedback data after user matching;

[0046] Adjust the matching algorithm parameters based on the feedback data;

[0047] Optimize subsequent matching and recommendation performance.

[0048] Preferably, the multidimensional emotion vector similarity calculation unit is further used for:

[0049] Different weighting coefficients are assigned to each dimension of the multidimensional emotion vector;

[0050] Calculate the weighted cosine similarity based on the weighting coefficients;

[0051] Calculate the Euclidean distance between emotion vectors and divide the matching levels based on the Euclidean distance.

[0052] Preferably, the emotion complementarity and resonance matching unit is further used for:

[0053] Emotional complementarity is calculated based on the difference between positive and negative emotion values ​​and the similarity of activity levels.

[0054] Build a user matching preference profile based on the distribution of successful matching patterns in the user's history, matching satisfaction, and explicit settings;

[0055] By combining emotional state characteristics, a context-aware matching mode switching is achieved.

[0056] Preferably, the temporal emotion trajectory prediction unit is further used for:

[0057] The sliding window mechanism is used to balance the weights of historical data and the current state.

[0058] Define four basic patterns of emotional fluctuations: rising, falling, fluctuating, and stable.

[0059] Improve prediction accuracy by incorporating contextual factors such as time, environment, and social events;

[0060] The prediction results are attached with a confidence score to guide the matching decision.

[0061] Preferably, the multi-factor matching scoring unit is further used for:

[0062] Design a time-effect decay factor to ensure that the match is based on the current emotional state;

[0063] Create a unique matching parameter configuration profile for each user;

[0064] The matching configuration is switched according to different social scenarios to optimize the matching effect.

[0065] Preferably, the feedback optimization unit is further used for:

[0066] Distinguish between active and passive feedback data;

[0067] Construct a comprehensive evaluation model that weighs the importance of different feedback indicators;

[0068] Implement a two-tier optimization architecture for global and individual parameters;

[0069] The influence ratio of the two-layer parameters is dynamically adjusted to balance universality and personalization.

[0070] Preferably, the emotion vector generation module is further used for:

[0071] Generate sentiment semantic features based on text analysis;

[0072] Based on speech analysis, features of pitch, speech rate, and emotional intensity are extracted;

[0073] Based on facial expression image analysis, emotion category features are identified;

[0074] By integrating features from various modalities, a unified multidimensional emotion vector representation is formed;

[0075] The generated emotion vector is labeled with the collection time and validity period attributes.

[0076] Preferably, the emotion map construction module is further used for:

[0077] A two-dimensional emotion space is constructed by mapping positive and negative emotion values ​​to the horizontal axis and activity levels to the vertical axis.

[0078] Based on emotion vectors, user emotion points are drawn, and emotional states are represented by color, shape, or dynamic particles.

[0079] Aggregate sentiment data from multiple users to construct a sentiment distribution view in the form of a heatmap;

[0080] The sentiment map is updated in real time to reflect the dynamic changes in group sentiment.

[0081] Preferably, the emotion intervention and guidance module is further used for:

[0082] Detect areas of concentrated negative emotions in the emotion map;

[0083] Analyze the risk of negative emotions spreading and set intervention priorities;

[0084] Provide users experiencing negative emotions with soothing chat rooms, AI companionship services, or mood-regulating content;

[0085] Provide support to guide users with complementary emotional states into the negative emotional zone;

[0086] Record the trajectory of users' emotional changes and evaluate the effectiveness of interventions.

[0087] The beneficial effects of this invention are mainly reflected in:

[0088] 1. Create a "visualized map of user emotions," breaking through the traditional static labeling method in social media;

[0089] 2. Provide a more personalized and authentic matching social path;

[0090] 3. Effectively identify and intervene in the spread of negative emotions, and enhance the platform's emotional health index;

[0091] 4. Provide emotion tracking and intervention support technologies for the field of mental health;

[0092] 5. It helps to create an emotion-driven, immersive virtual social ecosystem. Attached Figure Description

[0093] Figure 1 is a diagram of the overall architecture of the system of the present invention;

[0094] Figure 2 is a structural diagram of the emotion map system of the present invention;

[0095] Figure 3 is a block diagram of the emotion matching mechanism of the present invention;

[0096] Figure 4 is a closed-loop diagram of user emotion evolution and system feedback in this invention. Detailed Implementation

[0097] Please refer to Figures 1-4. The present invention will be further described in detail below with reference to the figures and specific embodiments. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0098] As shown in Figure 1, the social map construction and matching system based on multimodal emotion recognition provided by this invention includes: a data acquisition module 1, an emotion recognition engine 2, an emotion vector generation module 3, an emotion map construction module 4, an emotion matching recommendation system 5, and an emotion intervention and guidance module 6. The modules form a complete system functional chain through data flow and information exchange.

[0099] The data acquisition module 1 is primarily responsible for collecting users' multimodal emotion expression data. In a preferred embodiment of the present invention, this module collects three types of key data: user text input, voice expression, and facial expression behavior data.

[0100] Specifically, for text data, the system collects text content, private messages, and comments posted by users on social media platforms; for voice data, the system captures the user's voice characteristics in voice chats, including parameters such as pitch, volume, and speech rate; and for facial expression data, the system monitors the frequency of emojis used by users and facial expressions in selfies or live video streams.

[0101] In practical applications, the data acquisition module 1 can also be configured with a sampling frequency control mechanism to dynamically adjust the sampling frequency based on system load and user activity. For example, for users with high-frequency interaction, a higher sampling frequency (e.g., 5 seconds / time) can be set; while for users with low-frequency interaction, the sampling frequency can be reduced (e.g., 30 seconds / time) to balance the relationship between system resource consumption and the real-time nature of emotion capture.

[0102] After the data collection is completed, the data collection module 1 transmits the multimodal data to the emotion recognition engine 2 through a secure encrypted channel for further processing.

[0103] The emotion recognition engine 2 is connected to the data acquisition module 1. Its main function is to receive and analyze multimodal data and identify the user's current emotional characteristics. This invention employs a modality-specific recognition method, applying specialized recognition algorithms to different types of input data.

[0104] For text data, the emotion recognition engine 2 applies a semantic analysis model based on the Transformer architecture to extract the emotional tendencies contained in the text. This model evaluates the emotional polarity and intensity of the text by analyzing word choice, sentence structure, and semantic relationships.

[0105] For voice data, the engine employs acoustic feature extraction and analysis methods, including the calculation and synthesis of parameters such as pitch variation, speech rate, and volume. For example, higher pitch fluctuations and faster speech rate may indicate that the user is excited or nervous, while a steady, low tone may indicate calm or depressed mood.

[0106] For facial expression data, the engine uses computer vision algorithms to analyze facial expressions, extracting key feature point displacements and expression change patterns. The system also analyzes the types and frequency of emojis used by users as an auxiliary basis for judging emotional state.

[0107] Preferably, the emotion recognition engine 2 employs an ensemble learning method to improve recognition accuracy, assigning different weights to the recognition results of different modalities to form a preliminary fused emotion feature representation. This fusion considers the reliability of each modality; for example, in noisy environments, the system automatically reduces the weight of the speech modality and increases the weight of text and facial expression data. Practice shows that this dynamic weight adjustment mechanism can improve the accuracy of emotion recognition by 15% to 20%.

[0108] After the recognition is completed, the emotion recognition engine 2 passes the recognition result to the emotion vector generation module 3 to generate a standardized emotion vector representation.

[0109] The emotion vector generation module 3 communicates with the emotion recognition engine 2. Its core function is to convert unstructured emotion recognition results into structured multidimensional emotion vector representations. This vectorized representation allows for mathematical calculations and spatial mapping of emotion states, laying the foundation for subsequent map construction and matching recommendations.

[0110] In one embodiment of the present invention, the emotion vector is defined as a four-dimensional vector:

[0111] V(e)=[Valence,Arousal,Dominance,Confidence],

[0112] in:

[0113] Valence is a positive or negative value for emotion, ranging from [0,1]. The larger the value, the more positive the emotion.

[0114] Arousal represents the level of activity, ranging from [0,1]. A larger value indicates a higher level of emotional activity.

[0115] Dominance is the degree of dominance, ranging from [0,1]. The larger the value, the stronger the dominance of the emotion expression.

[0116] Confidence is the level of certainty, ranging from [0,1]. A higher value indicates a higher level of confidence in the emotion recognition result.

[0117] The emotion vector generation module 3 first generates emotion semantic features based on text analysis. For example, when analyzing text such as "I'm so happy!", the system may generate a preliminary semantic feature vector [0.9, 0.8, 0.7, 0.85], representing high positivity, high activity, moderate dominance, and high confidence.

[0118] Meanwhile, based on speech analysis, the system extracts acoustic features. For example, if the system detects that the user speaks quickly (120 words / minute), has frequent pitch changes (standard deviation > 30Hz), and has an upward intonation, it may generate an acoustic feature vector [0.85, 0.9, 0.75, 0.8].

[0119] In addition, based on facial expression image analysis, the system identifies facial expression categories (such as smiling and frowning) and intensity, generating facial expression feature vectors [0.8, 0.7, 0.6, 0.75].

[0120] The emotion vector generation module 3 uses a weighted fusion algorithm to integrate the feature vectors of these three modalities: V(e) = α·V text +β·V voice +γ·V expression ,

[0121] Where: V(e) is the final fused emotion vector, which is a four-dimensional vector; V text V is the text feature vector, with the same dimension as V(e); voice V is the speech feature vector, with the same dimension as V(e); expression The expression feature vector has the same dimension as V(e); α, β, and γ are weight coefficients, representing the importance weights of text, speech, and expression modalities, respectively, and α+β+γ=1 to ensure weight normalization.

[0122] The weighting coefficients are dynamically adjusted based on the quality and reliability of each modality of data. For example, in cases where the text content is rich but the speech quality is poor, α = 0.6, β = 0.15, and γ = 0.25 might be set.

[0123] On a social media platform, when user A posts a status update saying "Today is such a terrible day, nothing is going right," the system may generate a sentiment vector [0.2, 0.6, 0.5, 0.9] based on text analysis, representing a relatively negative, moderately active emotional state, and the recognition result has a high degree of confidence. This vectorized representation will then be used for localization and matching calculations in the sentiment map.

[0124] In addition to generating emotion vectors, this module also tags each vector with a timestamp and an expiration date. Generally, the default expiration date for emotion vectors is set to 30 minutes, but the system dynamically adjusts this based on user activity and the frequency of emotion changes. For example, for users with frequent emotion fluctuations, the system may shorten the expiration date to 10 minutes to ensure the real-time accuracy of the emotion map.

[0125] The generated multidimensional emotion vectors are transmitted to the emotion map construction module 4 and the emotion matching recommendation system 5, respectively, for map construction and matching calculation.

[0126] The emotion map construction module 4 communicates with the emotion vector generation module 3, and is responsible for mapping abstract emotion vectors to a visualized emotion space to form an intuitive emotion map. This visualization makes complex emotion data intuitive and easy to understand, making it easier for users to understand the distribution of their own and others' emotional states.

[0127] In a preferred embodiment, the emotion map construction module 4 uses a two-dimensional coordinate system as the basic mapping framework: the horizontal axis represents the positive or negative value of emotion (Valence), and the vertical axis represents the level of activity (Arousal). This two-dimensional representation is derived from the widely accepted Circumplex Model of Affect in psychology, which is both concise and expressive.

[0128] Specifically, the emotion map building module 4 performs the following mapping operations:

[0129] First, the Valence value in the emotion vector is mapped to the horizontal axis coordinate using the formula: X = (Valence - 0.5) × Scale.

[0130] Where: X is the horizontal axis coordinate value, representing the horizontal position of the user's emotion on the emotion map; Valence is the positive or negative value of the emotion, ranging from [0,1]; Scale is the scaling factor, used to control the map size, usually set to 100, so that the coordinate range is between [-50,50], which is convenient for visualization.

[0131] Then, the Arousal value is mapped to the vertical axis using the formula: Y = (Arousal - 0.5) × Scale.

[0132] Where: Y is the vertical axis coordinate value, representing the vertical position of the user's emotion on the emotion map; Arousal is the activity level, ranging from [0,1]; Scale is the same scaling factor as the horizontal axis, ensuring that the horizontal and vertical axes are consistent in scale.

[0133] In this way, each user's emotional state is mapped to a point on the emotion map with coordinates (X,Y). The origin (0,0) of the emotion map represents a neutral emotional state, the upper right quadrant represents "positive excitement", the lower right quadrant represents "positive calm", the upper left quadrant represents "negative excitement", and the lower left quadrant represents "negative calm".

[0134] For example, if a user has just received an acceptance letter from their ideal university, and their emotion vector is [0.9, 0.8, 0.7, 0.95], then their coordinates on the emotion map are X = (0.9 - 0.5) × 100 = 40 and Y = (0.8 - 0.5) × 100 = 30, which are located in the upper right quadrant, indicating a positive and excited emotional state.

[0135] To enhance the expressiveness of the map, the emotion map construction module 4 also features a rich set of visual encoding methods:

[0136] User emotion points are represented by dynamic particles, and the particle color is determined according to the positive or negative value of the emotion. For example, a Valence value of 0-0.3 is mapped to cool colors (blue, purple), a value of 0.7-1 is mapped to warm colors (red, orange), and a value of 0.3-0.7 is mapped to neutral colors (green, yellow).

[0137] Particle size is related to the Dominance value; the stronger the dominance, the larger the particle. Particle transparency is related to the Confidence value; the higher the confidence, the less transparent the particle. The specific calculation methods are: Size = BaseSize + (Dominance × SizeRange), Opacity = MinOpacity + (Confidence × OpacityRange).

[0138] Where: Size is the particle display size (pixels); BaseSize is the base size, usually set to 5 pixels; SizeRange is the size range, usually set to 10 pixels; Opacity is the particle opacity, ranging from [0,1]; MinOpacity is the minimum opacity, usually set to 0.3; OpacityRange is the opacity range, usually set to 0.7.

[0139] In practical applications, emotion maps may simultaneously display the emotional states of hundreds or even thousands of users. For example, during a large social event (such as a music festival livestream), the system may display the real-time emotional distribution of thousands of viewers, presenting a visual effect of the overall "flow" of emotions. Viewers can see the location of their own emotional points on the map and also perceive the emotional state of the entire group, enhancing social connection and shared experience.

[0140] In addition, the map displays the movement trajectory of emotion points, reflecting the changing trends of users' emotions. For example, if a user's emotion gradually moves from the lower left quadrant to the upper right quadrant, it indicates that their emotion is changing from "negative calm" to "positive excitement." The trajectory is represented by semi-transparent lines, with the most recent changes being darker and the more distant changes gradually fading, visually showcasing the history of emotion changes.

[0141] In addition to individual user sentiment representations, the sentiment map building module 4 also aggregates multi-user data to form a heatmap-style sentiment distribution view. The formula for calculating the heatmap is:

[0142] Where: Intensity(x,y) is the thermal intensity at coordinate point (x,y); n is the total number of users; w i The weight of the i-th user may be related to user activity or influence; K is the kernel function, usually a Gaussian kernel. Let (x, y) be the distance from point (x, y) to the i-th user emotion point (X). i ,Y i The Euclidean distance is ; h is the bandwidth parameter, which controls the range of thermal diffusion, and is usually set to 15.

[0143] This heatmap representation allows the system to visually display emotional "hot spots." For example, during a live sports broadcast, the system might detect a significant heat concentration in the upper right corner of the emotion map (the positive and excited area) after a thrilling moment, reflecting the shared emotional response of the audience.

[0144] The emotion map is updated in real time every 30 seconds to ensure it reflects the user's current emotional state. The system also saves historical map data, supporting retrospective analysis and research into emotion trends.

[0145] Users can view emotion maps through a dedicated interface on the social platform, allowing them to browse the overall emotion distribution or view the emotion status of specific friends or interest groups. Furthermore, the system supports emotion map filtering, such as displaying only users in specific emotion zones (e.g., "positive and excited"), making it easier to find social contacts with similar emotions.

[0146] The completed emotion map data is transmitted to the emotion matching and recommendation system 5 for initial matching and filtering based on spatial location relationships. Simultaneously, the map is also directly displayed to users through the user interface, providing insights into self-awareness of emotions and the distribution of group emotions.

[0147] The emotion matching and recommendation system 5 is the core innovative module of this invention, and it communicates with the emotion map construction module 4 and the emotion vector generation module 3. This system is responsible for accurately matching and recommending suitable social partners based on the user's emotional state.

[0148] As shown in Figure 3, the emotion matching recommendation system 5 includes a multi-dimensional emotion vector similarity calculation unit 51, an emotion complementarity and resonance matching unit 52, a temporal emotion trajectory prediction unit 53, a multi-factor matching scoring unit 54, and a feedback optimization unit 55.

[0149] The multidimensional emotion vector similarity calculation unit 51 is responsible for calculating the degree of similarity between the emotion vectors of two users, which is the basis of emotion matching.

[0150] In a preferred embodiment of the invention, the unit first assigns different weight coefficients to each dimension of the emotion vector to reflect the importance of each dimension in the matching. For example, in scenarios seeking emotional resonance, the Valence and Arousal dimensions are usually more important and can be assigned weights of 0.4 and 0.3 respectively, while Dominance and Confidence are each assigned a weight of 0.15.

[0151] Based on these weighting coefficients, the unit uses weighted cosine similarity to calculate the similarity between two sentiment vectors:

[0152] Where: Sim(V i V j V represents the similarity between the emotion vectors of user i and user j, ranging from [-1, 1]. A value closer to 1 indicates a more similar emotional state. i [k] and V j [k] represents the k-th dimension value of the emotion vectors of user i and user i, respectively; w k is the weight coefficient of the k-th dimension, reflecting the importance of this dimension in the matching calculation; n is the vector dimension, which is 4 in this embodiment, corresponding to the four dimensions of the emotion vector (Valence, Arousal, Dominance, Confidence); ∑ represents the summation operation, which accumulates the weighted products of all dimensions from k=1 to n; √ represents the square root operation, used to calculate the magnitude of the weighted vector, realizing the normalization of cosine similarity.

[0153] For example, if the emotion vectors of user A and user B are [0.8, 0.7, 0.6, 0.9] and [0.85, 0.65, 0.5, 0.85] respectively, and the weights of each dimension are [0.4, 0.3, 0.15, 0.15], then the similarity calculation process is as follows:

[0154] Molecular calculation: 0.4×0.8×0.85+0.3×0.7×0.65+0.15×0.6×0.5+0.15×0.9×0.85=0.272+0.1365+0.045+0.11475=0.56825

[0155] Weighted modulus length of vector A: √(0.4 2 ×0.8 2 +0.3 2 ×0.7 2 +0.15 2 ×0.6 2 +0.15 2 ×0.9 2 )=√(0.1024+0.0441+0.0081+0.01822)=√0.17282=0.4158

[0156] Weighted modulus length of vector B: √(0.4 2 ×0.85 2 +0.3 2 ×0.65 2 +0.15 2 ×0.5 2 +0.15 2 ×0.85 2 )=√(0.1156+0.03802+0.00562+0.01628)=√0.17552=0.4189

[0157] Similarity: 0.56825 / (0.4158×0.4189)=0.56825 / 0.1742=0.9747

[0158] The result indicates that the emotional states of the two users are highly similar (the similarity is close to 1), and the system may recommend them to each other as potential social contacts.

[0159] In addition to calculating the similarity, this unit also calculates the Euclidean distance between the emotion vectors for dividing matching levels:

[0160] Where: Distance(V i ,V j ) is the weighted Euclidean distance of the emotion vectors of user i and user j; the meanings of other symbols are the same as those in the similarity calculation formula; √ represents the square root operation; ∑ represents the summation operation, which accumulates the weighted squared differences of each dimension from k=1 to n.

[0161] Based on the calculated distance value, the system sets thresholds to divide the matching into three levels: high resonance (Distance≤0.2), moderate resonance (0.2<Distance≤0.4) and low resonance (Distance>0.4). These thresholds are determined based on a large amount of user feedback data analysis, and can effectively distinguish emotion matching effects of different degrees.

[0162] In social media usage scenarios, when users are browsing the "Emotional Exploration" feature, the system will present different matching recommendations based on similarity and distance calculations. For example, users with high resonance will be labeled "emotionally resonant," users with moderate resonance will be labeled "emotionally similar," and users with low resonance but complementary qualities may be labeled "emotionally balanced," helping users choose different types of social connections according to their own needs.

[0163] The emotional complementarity and resonance matching unit 52 is responsible for implementing two matching modes—emotional resonance (similar) matching and emotional complementarity matching, and intelligently selects the appropriate mode based on user characteristics.

[0164] For calculating emotional complementarity, this unit designed a specific calculation method that focuses on the difference between positive and negative emotional values ​​and the similarity of activity levels: Comp(V i V j )=w1·|V i [Valence]-V j [Valence]|+w2·(1- |V i [Arousal]-V j [Arousal]|),

[0165] Where: Comp(V) i V j V represents the emotional complementarity between user i and user j, ranging from [0,1]. A larger value indicates stronger complementarity. i [Valence] and V j [Valence] represents the positive or negative value of the sentiment of user i and user i', respectively, ranging from [0,1]; V i [Arousal] and V j [Arousal] represents the activity level of user i and user i', ranging from [0,1]; w1 and w2 are weighting coefficients that control the importance of positive and negative emotion differences and activity similarity, respectively, and are usually set to 0.6 and 0.4; |x| represents taking the absolute value of x, used to calculate the difference between the two values.

[0166] The design concept of this formula is that emotional complementarity is mainly reflected in the difference between positive and negative emotions (for example, positive emotions can balance negative emotions), while maintaining similar activity levels (ensuring that the pace of communication matches).

[0167] For example, if user C is depressed (Valence = 0.2, Arousal = 0.4) and user D is positive (Valence = 0.8, Arousal = 0.5), then the complementarity is calculated as: 0.6 × |0.2 - 0.8| + 0.4 × (1 - |0.4 - 0.5|) = 0.6 × 0.6 + 0.4 × 0.9 = 0.36 + 0.36 = 0.72, indicating a high degree of complementarity. On a social platform, the system might recommend user D to user C who is depressed, helping to improve their mood.

[0168] To determine a user's matching pattern preference, this unit constructs a user matching preference profile, taking into account the following factors: userPreference = α·historicalMatchRate + β·psychologicalTraits + γ·explicitSettings.

[0169] Wherein: userPreference is the user preference score, ranging from [0,1]. A larger value indicates a greater inclination towards resonant matching, while a smaller value indicates a greater inclination towards complementary matching; historicalMatchRate is the historical match success rate data, based on the success rate of resonant and complementary matching in the user's past matching records, ranging from [0,1]; psychologicalTraits is the result of psychological characteristic analysis, based on the user's social tendencies derived from personality tests or behavioral analysis, ranging from [0,1]; explicitSettings is the user's explicitly set preferences, the matching preferences that the user directly expresses in the settings, ranging from [0,1]; α, β, γ are weight coefficients, which control the influence of each factor respectively, and α+β+γ=1 to ensure weight normalization.

[0170] In practical applications, these weighting coefficients are typically set to α = 0.5, β = 0.3, and γ = 0.2, reflecting the importance of historical behavioral data while also taking into account users' psychological characteristics and subjective preferences. These parameter values ​​are determined through large-scale A / B testing and can accurately predict users' matching preferences in more than 75% of scenarios.

[0171] For example, a user who historically interacts better with people who have similar emotions (historicalMatchRate = 0.8), whose psychological test shows a tendency to seek resonance (psychologicalTraits = 0.7), but who indicates interest in both modes in the settings (explicitSettings = 0.5), would have a preference score calculated as: 0.5 × 0.8 + 0.3 × 0.7 + 0.2 × 0.5 = 0.4 + 0.21 + 0.1 = 0.71, indicating that this user is more suited to the resonance matching mode.

[0172] Based on calculated preference scores, the system intelligently selects the matching mode:

[0173] Where: matchScore is the final matching score used for sorting; Sim(V i V j ) represents the similarity of emotion vectors; Comp(V) i V j ) represents the emotional complementarity; userPreference represents the user matching preference score; the conditional judgment "if userPreference>0.6" means that when the user preference score is greater than the threshold of 0.6, the resonance matching mode is used, otherwise the complementary matching mode is used.

[0174] The threshold of 0.6 was determined based on user experience testing and can effectively balance the application ratio of the two matching modes. In social platform practice, approximately 60% of users tend to use resonant matching and 40% tend to use complementary matching, and this threshold setting is basically consistent with the natural distribution of users.

[0175] Furthermore, this unit also enables context-aware matching mode switching, automatically adjusting strategies based on specific emotional patterns. For example, when it detects that a user is in a persistently low mood (Valence < 0.3 for more than 24 hours), the system will automatically switch to a complementary matching mode, recommending users with positive emotions to communicate with, thus helping to improve their emotional state.

[0176] In a real-world application scenario, user E posted negative content for three consecutive days, with an average emotion vector of [0.25, 0.3, 0.4, 0.85]. Even though this user usually prefers resonance matching (userPreference = 0.75), the system will temporarily activate a context-aware mechanism and switch to a complementary matching mode, recommending users with positive emotions to communicate with. Practice shows that this intervention mechanism can enable about 35% of users with long-term low mood to show significant mood improvement within a week.

[0177] The temporal emotion trajectory prediction unit 53 is responsible for analyzing the temporal change patterns of users' emotions, predicting possible future emotional states, and achieving forward-looking matching.

[0178] This unit first collects the user's emotion vector sequence within a continuous time window, forming an emotion trajectory:

[0179] Where: EmotionTrajectory is the user's emotional trajectory, which consists of multiple emotional vectors over a time series; For time point t iThe emotion vector contains four dimensions: Valence, Arousal, Dominance, and Confidence; n is the number of emotion vectors in the trajectory, usually taken from the 30 most recent sampling points.

[0180] To balance the weights of historical data and the current state, the unit employs a sliding window mechanism, using the 30 most recent sentiment sampling points to form the current trajectory window, and assigning time decay weights to data at different time points:

[0181] Where: weight(t) i (t) represents time point t i The weights range from 0 to 1, with higher weights closer to the current time; t current The current time, in seconds; t i λ is the time of the historical sentiment sampling point, in seconds; λ is the decay coefficient, which controls the rate at which the weight decays over time, usually set to 0.05 (about a 20-minute half-life); e is the base of the natural logarithm, approximately equal to 2.71828.

[0182] For example, in social applications, when a user is watching an emotional movie, the system continuously collects their emotional state. If the user experiences intense emotional fluctuations during the movie's climax, the system will assign higher weight to emotional data from the most recent 10 minutes (e.g., data from 10 minutes ago has a weight of approximately 0.6), making the predictions more reflective of the current emotional trend and thus recommending more suitable social interactions for the moment.

[0183] Based on emotional trajectory data, this unit defines four basic emotional fluctuation patterns:

[0184] 1. Rising type: Positive changes in sentiment at multiple consecutive time points, satisfying the condition V(t+1)[Valence]-V(t)[Valence]>δ for most t;

[0185] 2. Declining type: Negative changes in sentiment at multiple consecutive time points, satisfying the condition V(t+1)[Valence]-V(t)[Valence]<-δ for most t;

[0186] 3. Fluctuating type: The direction of mood changes frequently, satisfying the condition variance(V[Valence])>θ;

[0187] 4. Stable type: Small fluctuations in mood, satisfying the condition variance(V[Valence])<θ.

[0188] Where: V(t)[Valence] represents the Valence value in the emotion vector at time point t; δ is the change threshold, which represents the minimum difference that is judged as a meaningful emotional change, usually set to 0.05; variance() is a function to calculate variance, which measures the dispersion of the data; θ is the fluctuation threshold, which represents the lower limit of variance for judging a fluctuating emotion, usually set to 0.03.

[0189] These thresholds are empirical values ​​derived from the analysis of a large amount of user sentiment data, and they can correctly classify sentiment fluctuation patterns in more than 90% of cases.

[0190] In practical applications, the system might detect a clear upward trend in user F's emotions during video chats with friends (Valence gradually increases from 0.5 to 0.8), indicating that the interaction has a positive impact on their mood. Based on this pattern recognition, the system might prioritize recommending more similar social activities to maintain and reinforce this positive emotional trend.

[0191] For predicting future emotional states, this unit uses an LSTM (Long Short-Term Memory) model, incorporating contextual factors such as time, environment, and social events: V(e) t+1 =LSTM(EmotionTrajectory,contextualFactors),

[0192] Where: V(e) t+1 The predicted emotion vector for the next time point; LSTM() is a long short-term memory neural network model function specifically designed for prediction of sequence data; EmotionTrajectory is historical emotion trajectory data; contextualFactors include time features (such as time of day, weekend / weekday), environmental factors (such as weather, location), and social events (such as receiving messages, receiving likes), which are encoded in vector form and input into the model.

[0193] To improve prediction reliability, the system assigns a confidence score to each prediction result, calculated using the formula: confidenceScore = f(dataQuality, patternStability, predictionError).

[0194] Where: confidenceScore is the prediction confidence level, ranging from [0,1], with a higher value indicating a more reliable prediction; f() is the scoring function, usually using a weighted average method; dataQuality reflects the quality and completeness of sentiment data, ranging from [0,1]; patternStability represents the stability of the sentiment pattern, ranging from [0,1]; predictionError is the historical prediction error, ranging from [0,1], with a smaller value indicating a more accurate prediction.

[0195] The confidence score ranges from [0,1], and only predictions with a score greater than 0.7 are used to guide matching decisions. This threshold ensures that the system relies only on relatively reliable predictions, avoiding a decline in matching quality caused by erroneous predictions.

[0196] In social media platforms, this predictive feature is particularly helpful for event planning. For example, if the system predicts that user G's mood typically shifts from work fatigue (low Valence) to relaxation (high Valence) on Friday evening, it can recommend suitable social activities or chat partners for Friday afternoon to help them relax, thus enhancing the user experience's forward-looking approach.

[0197] The multi-factor matching scoring unit 54 is responsible for comprehensively considering multiple factors and calculating the final matching score to achieve comprehensive and accurate matching recommendations.

[0198] This unit is designed with a comprehensive scoring formula that integrates five key factors: Score ij =α·Sim(V i V j )+β·TimeDecay+γ·InteractionHistory+δ·SocialCompatibility+ε·TrendMatch,

[0199] Among them: Score ij Sim(V) represents the combined matching score between user i and user j, ranging from [0,1]. A higher value indicates a higher degree of matching. i V j ) represents the similarity (or complementarity, depending on the selected matching pattern) of the emotion vectors, ranging from [-1,1] to [0,1] after normalization; α is the timeliness decay factor, ranging from (0,1), reflecting the timeliness of sentiment data; InteractionHistory is the historical interaction success rate, ranging from [0,1], reflecting the quality of past interactions between two users; SocialCompatibility is the social compatibility index, ranging from [0,1], calculated based on interest tags, social styles, etc.; TrendMatch is the sentiment trend matching degree, ranging from [0,1], provided by the time-series sentiment trajectory prediction unit; α, β, γ, δ, and ε are weight coefficients, respectively controlling the importance of each factor in the comprehensive score, and α+β+γ+δ+ε=1 to ensure weight normalization.

[0200] In general, these weighting coefficients are set to α = 0.4, β = 0.15, γ = 0.2, δ = 0.15, and ε = 0.1, reflecting the core role of emotional similarity in matching while also taking into account the influence of other factors. These parameter values ​​are the optimal solutions derived from the analysis of a large amount of user matching data, and can produce matching results that satisfy users in most scenarios.

[0201] For the time-sensitivity decay factor, the decay parameter λ is usually set to 0.02, which means that the weight of sentiment data from 30 minutes ago will decay to about 55% of its original weight, and data from 1 hour ago will decay to about 30% of its original weight. This setting ensures that the system mainly matches based on the user's current emotional state, but does not completely ignore recent emotional history.

[0202] For example, on a social media platform, user H and user I have a current emotion vector similarity of 0.85. They had a pleasant group chat a week ago (InteractionHistory = 0.7), share a love for photography and travel (SocialCompatibility = 0.8), and have similar future emotion trends (TrendMatch = 0.75). However, user H's emotion data was collected 20 minutes ago (TimeDecay = 0.67). Therefore, the overall matching score is calculated as: 0.4 × 0.85 + 0.15 × 0.67 + 0.2 × 0.7 + 0.15 × 0.8 + 0.1 × 0.75 = 0.34 + 0.1005 + 0.14 + 0.12 + 0.075 = 0.7755. This indicates a high degree of compatibility, and the system might recommend them as "emotional confidants."

[0203] To achieve personalized matching, this unit creates a unique matching parameter configuration profile for each user, adjusting the weights of each factor based on the user's matching preferences and feedback history. For example, for users who value emotional resonance, the system might increase α to over 0.5; while for users who prioritize shared interests, it might increase the value of δ. This personalized configuration allows matching results to better align with users' individual preferences, improving user satisfaction.

[0204] Furthermore, this unit switches matching configurations based on different social scenarios. For example, in the scenario of "finding someone to confide in," the system emphasizes matching based on complementary emotions; while in the scenario of "finding partners for shared activities," it emphasizes matching based on social compatibility and activity level. This scenario-based configuration effectively improves the matching accuracy for different social needs.

[0205] In practical applications, users can select different social intentions on social platforms, such as "I want to confide in someone," "Looking for resonance," or "Participating in activities together." The system will automatically adjust the matching weight configuration based on the selected scenario, providing more targeted social recommendations. Data shows that this scenario-based matching improves user satisfaction and interaction depth by an average of 28% compared to general matching.

[0206] The feedback optimization unit 55 is a key component for the system's self-evolution. It is responsible for collecting feedback data after user matching and optimizing the matching algorithm parameters accordingly.

[0207] This unit is designed with a multi-dimensional feedback acquisition mechanism, taking into account both active and passive feedback:

[0208] Active feedback includes direct user ratings (1-5 stars), text reviews, and rematch requests; passive feedback includes behavioral data such as chat duration (minutes), message frequency (messages / minute), sentiment change vector (sentiment difference before and after matching), and interaction depth (such as adding friends, number of days of continuous chatting).

[0209] To reduce the user's workload, the system adopts a non-intrusive feedback collection design. For example, it naturally guides users to provide simple feedback at the end of a chat, or provides one-click emotional feedback through floating emoji buttons. Practice shows that this lightweight feedback mechanism can increase user feedback participation rate from the traditional 15% to over 65%, significantly increasing the amount of data available for optimization.

[0210] Based on the collected multidimensional feedback data, this unit constructs a comprehensive evaluation model to calculate the matching quality score: matchQuality=w1·explicitRating+w2·f duration (chatDwration)+ w3·f frequency (messageFrequency)+w4·ositiveEimotionShift-w5·reMatchRequest,

[0211] Where: matchQuality is the match quality rating, ranging from [0,1], with higher values ​​indicating better matching results; explicitRating is the user's explicit rating, ranging from [0,1], obtained by normalizing a 1-5 star rating; f duration(chatDuration) is a rating function for chat duration, mapping the duration to the interval [0,1]; f frequency `messageFrequency` is the scoring function for message frequency, mapping frequency values ​​to the [0,1] interval; `positiveEmotionShift` is the positive emotion shift, ranging from [0,1], calculated as the increase in the Valence value after matching compared to before matching (if any); `reMatchRequest` is the rematch request indicator, with values ​​[0,1], where 1 indicates a user request for rematching; `w1`, `w2`, `w3`, `w4`, and `w5` are the weighting coefficients for each indicator, controlling their importance in the scoring. The weights are determined based on the reliability and importance of the feedback type, typically set to `w1 = 0.3`, `w2 = 0.2`, `w3 = 0.15`, `w4 = 0.25`, and `w5 = 0.1`. The function `f`... duration and f frequency For example, a normalization function:

[0212] Where: t is the actual chat duration (minutes); t optimal The optimal chat duration threshold is typically set to 15 minutes; r represents the actual message frequency (messages / minute); r optimal For optimal message frequency, it is usually set to 3 messages per minute; min(a,b) means taking the smaller value between a and b;

[0213] The conditional statement indicates that when the frequency is below the optimal value, the score increases with the frequency; when the frequency is above the optimal value, the score decreases with the frequency (to avoid low-quality communication caused by excessively rapid messages).

[0214] These parameter values ​​and function forms are determined by analyzing successfully matched interaction patterns. For example, data shows that high-quality social interactions typically last more than 15 minutes, with a message frequency of 2-4 messages per minute; frequencies that are too high or too low may indicate poor communication.

[0215] For example, on a social platform, users J and K chat for 20 minutes after being matched (f duration =1), message frequency is 2.5 messages / minute (f frequency=0.83), user J's emotion Valence increased from 0.4 to 0.7 (positiveEmotionShift = 0.3), the user gave a 4-star rating (explicitRating = 0.8), and did not request a rematch (reMatchRequest = 0). Therefore, the match quality score is: 0.3×0.8+0.2×1+0.15×0.83+0.25×0.3-0.1×0=0.24+0.2+0.1245+0.075+0=0.6395, indicating a medium-to-high level of match quality.

[0216] This unit employs reinforcement learning to optimize matching parameters, using the matching quality score as a reward signal: rewardSignal = matchQuality - baseline parameterUpdate = learningRate × rewardSignal × eligibilityTrace.

[0217] Wherein: rewardSignal is the reward signal, indicating whether the matching quality has improved or decreased relative to the baseline; baseline is the benchmark, usually taken as the moving average of recent matching quality, reflecting the current average performance of the system; parameterUpdate is the parameter update amount, indicating the direction and magnitude of adjustment for each matching parameter; learningRate is the learning rate, controlling the step size of parameter updates, usually set to 0.05 to balance learning speed and stability; eligibilityTrace is the eligibility trace, representing the contribution of each parameter to the current matching result, with larger contributions resulting in more significant adjustments. This parameter optimization mechanism based on reinforcement learning enables the system to continuously adapt to changes in user preferences and emerging social patterns, maintaining the timeliness and accuracy of the matching algorithm.

[0218] To balance generality and personalization, this unit implements a two-tier optimization architecture of global parameters and personal parameters: finalParameters = λ × globalParameters + (1 - λ) × personalParameters,

[0219] Where: finalParameters is the final matching parameter configuration of the application; globalParameters is the global optimization parameter, reflecting the matching preferences of the overall user group; personalParameters is the personalized parameter, reflecting the personalized matching needs of a specific user; λ is the balancing factor, ranging from [0,1], which controls the mixing ratio of global parameters and personal parameters.

[0220] λ is dynamically adjusted based on the sufficiency of individual data. For example, for new users or users with little data (such as those who have been registered for less than a week or have had fewer than 5 matches), λ may be set to 0.8 or higher, mainly relying on global optimization parameters; while for users with long usage time and rich data (such as those who have been registered for more than three months and have had more than 30 matches), λ may be reduced to below 0.4, relying more on personalized parameters.

[0221] In practical applications, this parameter optimization mechanism effectively solves the "cold start" problem, and the matching recommendations become increasingly personalized as users spend more time using the system. Data shows that active users who have used the system for more than three months have an average matching satisfaction rate 22% higher than new users, mainly due to the system's deep learning of their matching preferences.

[0222] Through this closed-loop feedback mechanism, the system can continuously learn and adapt to user preferences, constantly improving matching quality. Actual testing shows that after using this mechanism, the system's matching satisfaction increased by an average of 23%, user activity increased by 17%, and user retention rate improved by 15%, significantly enhancing the core competitiveness of the social platform.

[0223] The emotion intervention and guidance module 6 communicates with the emotion matching and recommendation system 5, and is primarily responsible for monitoring the emotional health status of the platform and providing intervention and guidance services when necessary. This module reflects the system's focus on users' mental health and is an important supplement to traditional social platforms.

[0224] This module first receives user emotional state data from the emotion matching recommendation system 5 and monitors areas of negative emotion clustering in the emotion map. The system defines "negative emotion clustering" as follows: the user density in a certain emotion area (usually the left half of the emotion map) increases by more than 30% in a short period of time (e.g., 1 hour), and the average Valence value of users in that area is less than 0.3.

[0225] For detected areas of concentrated negative emotions, this module analyzes the risk of the negative emotions spreading and sets intervention priorities. The risk assessment considers the following factors: the number of users in the concentrated area, the intensity of the negative emotions, the connection density between users (reflecting the possibility of emotion contagion), and the development trend of similar situations in historical data.

[0226] Based on the risk assessment results, the module adopts a tiered intervention strategy:

[0227] For low-risk situations (such as small-scale, short-term negative emotional fluctuations), the system may only monitor without taking intervention measures;

[0228] For moderate-risk situations, the system will recommend "emotion-relieving content" to users in a negative emotional state, such as positive articles, relaxing videos, or calming music.

[0229] In high-risk situations (such as widespread and persistent negative emotions), the system will activate "emotional support services," including soothing chat rooms, AI companionship services, or recommendations for professional psychological counseling resources. Simultaneously, the system will guide users with complementary emotional traits (those with positive and stable emotions and a tendency to support others) into negative emotion zones to provide support, forming an "emotional mutual aid network."

[0230] While implementing intervention measures, this module records the user's emotional changes and evaluates the intervention's effectiveness. Evaluation indicators include: the magnitude of emotional improvement after intervention (the increase in Valence score), the duration of improvement (the duration of the improved emotional state), and the user's feedback on the intervention measures.

[0231] For example, following a global negative news event, the system detected a significant increase in the concentration of negative emotions in a certain city area, with the average Valence value rapidly dropping from 0.6 to 0.25. The system immediately initiated high-risk intervention measures, launching a "Mood Support Station" campaign and recommending "Mood Guide" users with positive emotions to users in the area. 72 hours after the intervention, the average Valence value in the area rebounded to 0.45, and user feedback showed that 82% of participants felt their emotions were effectively relieved, demonstrating the positive effect of the intervention.

[0232] In addition, this module also provides personalized guidance services for specific emotional patterns. For example, for users with frequent mood swings (Valence standard deviation > 0.25), the system may recommend emotion management techniques or meditation exercises; for users who are in a state of low activity for a long time (Arousal < 0.3 for more than a week), the system may recommend active communities or interest groups to encourage participation and interaction.

[0233] The design philosophy of Module 6, Emotional Intervention and Guidance, is to respect users' autonomy while providing necessary emotional support and guidance to maintain the overall emotional health ecosystem of the platform. Practice shows that this module can effectively reduce the spread of negative emotions within the platform (by an average of 35%) and help 20%–30% of users experiencing negative emotions achieve emotional improvement.

[0234] The system operation process of this invention can be roughly divided into the following stages:

[0235] Data flow stage:

[0236] 1. Users engage in activities on social platforms (posting text, voice chat, using emoticons, etc.), and data collection module 1 collects this multimodal data in real time;

[0237] 2. The collected data is transmitted to the emotion recognition engine 2 via a secure channel for modality-specific emotion analysis;

[0238] 3. The emotion recognition result is fed into the emotion vector generation module 3, where it is converted into a standardized multidimensional emotion vector;

[0239] 4. The emotion vector is simultaneously transmitted to the emotion map construction module 4 and the emotion matching recommendation system 5, which are used for map construction and matching calculation, respectively.

[0240] Emotion Mapping Phase:

[0241] 1. The emotion map construction module 4 receives multi-dimensional emotion vectors and performs coordinate mapping;

[0242] 2. The system generates user emotion points and designs visual codes (color, size, transparency, etc.);

[0243] 3. The system aggregates data from multiple users to construct a sentiment heatmap and distribution view;

[0244] 4. The emotion map is updated in real time at a fixed frequency (e.g., every 30 seconds) to reflect the current distribution of emotional states.

[0245] Matching and recommendation phase:

[0246] 1. The user initiates a matching request or the system actively triggers the matching process;

[0247] 2. The multidimensional emotion vector similarity calculation unit 51 calculates the emotion similarity or complementarity between users;

[0248] 3. The emotional complementarity and resonance matching unit 52 determines the matching mode based on user preferences;

[0249] 4. The temporal emotion trajectory prediction unit 53 analyzes the trend of emotion changes and predicts future states;

[0250] 5. Multi-factor matching scoring unit 54 calculates the final matching score by integrating various factors;

[0251] 6. The system generates matching recommendations and displays the best match to the user.

[0252] Interaction and feedback phase:

[0253] 1. Users engage in social interactions with recommended individuals (chatting, sharing content, etc.);

[0254] 2. Feedback optimization unit 55 collects active and passive feedback data during the interaction process;

[0255] 3. The system evaluates the matching quality and adjusts the matching algorithm parameters accordingly;

[0256] 4. The adjusted parameters are applied to the next round of matching calculations to form an optimized closed loop.

[0257] Emotional intervention phase:

[0258] 1. The Emotion Intervention and Guidance Module 6 continuously monitors the aggregation of negative emotions in the emotion map;

[0259] 2. Systematically analyze the risk of negative emotions spreading and determine intervention priorities;

[0260] 3. Implement tiered intervention strategies (monitoring, content recommendation, support services, etc.) based on risk levels;

[0261] 4. The system records the intervention effect and continuously optimizes the intervention strategy.

[0262] The social map construction and matching system based on multimodal emotion recognition of the present invention can provide users with accurate social matching services based on emotional state, significantly improving the quality of social experience.

[0263] Compared with traditional social platforms, this system has the following advantages:

[0264] 1. Real-time emotion perception: The system can capture the user's emotional state "at this moment", rather than relying solely on static labels or historical behavioral data;

[0265] 2. Multi-dimensional matching: Through multi-dimensional emotion vectors and multi-factor scores, a more comprehensive matching consideration is achieved than that of a single dimension;

[0266] 3. Dynamic adaptability: The system can continuously optimize the matching strategy based on user feedback to achieve a personalized matching experience;

[0267] 4. Emotional Health Focus: Proactively maintain the platform's emotional health ecosystem through emotional intervention and guidance mechanisms.

[0268] Actual test data shows that compared with traditional matching methods, this system improves user matching satisfaction by 35% to 50%, increases the average duration of social interaction by 45%, and increases message frequency by 60%. For users in a negative emotional state, through the matching and intervention services provided by the system, 40% of users achieved emotional improvement within a short period (within 24 hours).

[0269] In addition, the system's emotion map feature provides users with a novel social experience, allowing them to intuitively understand the distribution of their own and others' emotional states, enhancing social transparency and a sense of connection. User feedback shows that 75% of users found the emotion map feature "fun and helpful," and 65% said it helped them "better understand their own and others' emotional states."

[0270] Overall, this invention, through multimodal emotion recognition and emotion-driven matching mechanisms, successfully constructs a social platform technology solution that better meets users' psychological needs and promotes emotional health, providing a new technological path for the future development of social networks.

[0271] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A social map construction and matching system based on multimodal emotion recognition, characterized in that, include: The data acquisition module is used for: Collect user's text input, voice expression, and facial expression data; The collected multimodal data is transmitted to the emotion recognition engine; The emotion recognition engine, which is communicatively connected to the data acquisition module, is used for: Receive multimodal data transmitted by the data acquisition module; Based on the multimodal data, identify the user's emotional characteristics; The emotion vector generation module, which is communicatively connected to the emotion recognition engine, is used for: Receive the emotional features identified by the emotion recognition engine; Based on the emotional features, a multi-dimensional emotion vector is generated, which includes positive and negative emotion values, activity level, dominance, and confidence level. The emotion map construction module, which is communicatively connected to the emotion vector generation module, is used for: Receive the multidimensional emotion vector generated by the emotion vector generation module; The multidimensional emotion vectors are mapped to an emotion coordinate system to form an emotion map; The emotion map displays the distribution of users' emotional states. The emotion matching recommendation system, communicatively connected to the emotion map construction module and the emotion vector generation module, is used for: Receive the multidimensional emotion vector generated by the emotion vector generation module; Receive emotion map data provided by the emotion map construction module; Based on the multidimensional emotion vector and the emotion map data, emotion-driven user matching is performed. Generate and output matching recommendation results; The emotion intervention and guidance module, which is communicatively connected to the emotion matching recommendation system, is used for: Receive user emotional state data provided by the emotion matching and recommendation system; Monitor negative emotional states; When a concentration of negative emotions is detected, emotional support services are provided.

2. The system according to claim 1, characterized in that, The emotion matching recommendation system includes: The multidimensional emotion vector similarity calculation unit is used for: Receive multidimensional emotion vectors from two users; Calculate the similarity between the two multidimensional emotion vectors; A matching threshold is set based on calculated similarity. The emotional complementarity and resonance matching unit is used for: Determine user matching pattern preferences based on users' historical interaction data; Based on the matching pattern preference, select either the emotional resonance matching strategy or the emotional complementarity matching strategy; Calculate the matching score between users based on the selected matching strategy; The temporal emotion trajectory prediction unit is used for: Collect the sequence of emotion vectors within a user's continuous time window to form an emotion trajectory; Identify patterns of user emotional fluctuations; Predicting users' short-term emotional state; Multi-factor matching scoring unit, used for: It considers multiple factors, including emotional similarity, timeliness, interaction history, social compatibility, and trend matching. Generate the final match score; Feedback optimization unit, used for: Collect feedback data after user matching; Adjust the matching algorithm parameters based on the feedback data; Optimize subsequent matching and recommendation performance.

3. The system according to claim 2, characterized in that, The multidimensional emotion vector similarity calculation unit is further used for: Different weighting coefficients are assigned to each dimension of the multidimensional emotion vector; Calculate the weighted cosine similarity based on the weighting coefficients; Calculate the Euclidean distance between emotion vectors and divide the matching levels based on the Euclidean distance.

4. The system according to claim 2, characterized in that, The emotional complementarity and resonance matching unit is further used for: Emotional complementarity is calculated based on the difference between positive and negative emotion values ​​and the similarity of activity levels. Build a user matching preference profile based on the distribution of successful matching patterns in the user's history, matching satisfaction, and explicit settings; By combining emotional state characteristics, a context-aware matching mode switching is achieved.

5. The system according to claim 2, characterized in that, The temporal emotion trajectory prediction unit is further used for: The sliding window mechanism is used to balance the weights of historical data and the current state. Define four basic patterns of emotional fluctuations: rising, falling, fluctuating, and stable. Improve prediction accuracy by incorporating contextual factors such as time, environment, and social events; The prediction results are attached with a confidence score to guide the matching decision.

6. The system according to claim 2, characterized in that, The multi-factor matching scoring unit is further used for: Design a time-effect decay factor to ensure that the match is based on the current emotional state; Create a unique matching parameter configuration profile for each user; The matching configuration is switched according to different social scenarios to optimize the matching effect.

7. The system according to claim 2, characterized in that, The feedback optimization unit is further used for: Distinguish between active and passive feedback data; Construct a comprehensive evaluation model that weighs the importance of different feedback indicators; Implement a two-tier optimization architecture for global and individual parameters; The influence ratio of the two-layer parameters is dynamically adjusted to balance universality and personalization.

8. The system according to claim 1, characterized in that, The emotion vector generation module is further used for: Generate sentiment semantic features based on text analysis; Based on speech analysis, features of pitch, speech rate, and emotional intensity are extracted; Based on facial expression image analysis, emotion category features are identified; By integrating features from various modalities, a unified multidimensional emotion vector representation is formed; The generated emotion vector is labeled with the collection time and validity period attributes.

9. The system according to claim 1, characterized in that, The emotion map construction module is further used for: A two-dimensional emotion space is constructed by mapping positive and negative emotion values ​​to the horizontal axis and activity levels to the vertical axis. Based on emotion vectors, user emotion points are drawn, and emotional states are represented by color, shape, or dynamic particles. Aggregate sentiment data from multiple users to construct a sentiment distribution view in the form of a heatmap; The sentiment map is updated in real time to reflect the dynamic changes in group sentiment.

10. The system according to claim 1, characterized in that, The emotion intervention and guidance module is further used for: Detect areas of concentrated negative emotions in the emotion map; Analyze the risk of negative emotions spreading and set intervention priorities; Provide users experiencing negative emotions with soothing chat rooms, AI companionship services, or mood-regulating content; Provide support to guide users with complementary emotional states into the negative emotional zone; Record the trajectory of users' emotional changes and evaluate the effectiveness of interventions.