Negative emotion characterization analysis system for social contact in colleges and universities

Abstract

The invention, which relates to the technical field of social emotion analysis, discloses a university social negative emotion characterization analysis system comprising a data acquisition and preprocessing module, a feature extraction and anomaly detection module, an intelligent perception and negative emotion characterization module, and an anomaly risk assessment and response module. Emotional feature information is acquired in real time based on a multi-dimensional data source, and the acquired emotional feature information is cleaned and preprocessed. According to the method, by calculating the data consistency offset coefficient and the strong correlation coefficient, the multi-collinearity problem among multi-dimensional data features is solved, and instability of the model regression coefficient is avoided. And the system identifies abnormal correlation in real time and combines risk threshold classification processing to ensure robust model output. An analysis set is established by continuously collecting data, long-term emotion trend analysis is supported, personalized responses such as reminding, intervention or emergency measures are provided according to risk levels, and accurate management and timely adjustment of student emotional states are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of social emotion analysis, and in particular to a negative emotion representation analysis system for university social networks. Background Art

[0002] The analysis of negative emotional representation in college social interaction refers to the reverse analysis and modeling of college students' emotional state in social situations by using big data technology and psychology and sociology theories. Its core goal is to deduce students' deep psychological state and emotional fluctuations, especially those emotions that are not directly expressed on the surface, through information such as language, behavior and emotional expression in social platforms or interactions inside and outside the campus. For example, students may show positive emotions (such as optimism and happiness) in public interactions, but reverse representation analysis can reveal the negative emotions behind (such as anxiety and loneliness), so as to understand their true emotional state more comprehensively and deeply. This analysis is of great significance in mental health intervention and campus management - schools can identify potential psychological problems in a timely manner by identifying signals of long-term depression or social pressure, and then develop targeted psychological counseling programs to optimize interactions between students. At the same time, as college students increasingly rely on social platforms to express themselves, this analysis provides data support for education administrators to help them grasp the trend of emotional changes, formulate preventive measures, and create a healthy and harmonious campus environment.

[0003] In specific operations, the system uses sentiment analysis algorithms to identify explicit emotions (such as happiness, anger, and sadness), and combines psychological models (such as depression tendency detection and loneliness index) to infer potential implicit emotions and inner states. Through natural language processing technology (NLP), the system extracts emotional features such as emotional vocabulary, semantic structure, and emotional change trajectory from text and behavioral data. At the same time, behavioral data (such as frequency of likes and participation in activities) are combined with time series models to reveal the dynamic fluctuations of emotional states. Then, the system uses a multimodal fusion method to integrate features such as language, behavior, and emotional trajectory into a unified feature vector, and analyze their mutual relationships and changing trends. Combined with psychological models, the system can identify potential psychological risks and give early warnings, such as inferring whether a student may have a tendency to depression or loneliness. Ultimately, the analysis results provide schools with accurate data support, helping them take effective intervention measures before problems arise and improve the efficiency of students' mental health management.

[0004] The prior art has the following deficiencies:

[0005] In the characterization and analysis of negative emotions in college social interaction, multi-dimensional data (such as language, behavior, and emotional trajectory) are input into the fusion model after feature extraction to form a unified feature vector. However, when there is a strong correlation between the features of different data sources (such as the emotional trajectory is highly correlated with the language sentiment score), multicollinearity problems will arise. Multicollinearity can cause the regression coefficients or weights of the model to become unstable, making it difficult for the model to accurately determine the independent contribution of each feature to emotional judgment. Even if the data input changes slightly, the parameters may fluctuate dramatically. This instability can cause the model to output completely different emotional state predictions in different time periods or data sets, resulting in uncontrollable results and difficulty in long-term monitoring and trend analysis.

[0006] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not constitute the prior art that is already known to one of ordinary skill in the art. Summary of the invention

[0007] The purpose of the present invention is to provide a negative emotion characterization and analysis system for social interaction in colleges and universities. By calculating the data consistency offset coefficient and the strong correlation coefficient, the multicollinearity problem caused by the strong correlation between multi-dimensional data features is effectively solved, and the risk of unstable model regression coefficients and weights is avoided. The system identifies abnormal correlations in real time and processes emotional feature information in combination with risk threshold classification to ensure the robustness and accuracy of the model output. By continuously collecting data to establish an analysis set, the system supports long-term emotional trend analysis and provides personalized responses based on risk levels, such as reminders, interventions or emergency measures. This dynamic monitoring and response mechanism ensures the accurate management and timely adjustment of students' emotional states to solve the problems in the above-mentioned background technology.

[0008] In order to achieve the above-mentioned purpose, the present invention provides the following technical solutions: a negative emotion representation analysis system for social networking in colleges and universities, including a data collection and preprocessing module, a feature extraction and anomaly detection module, an intelligent perception and negative emotion representation module, and an abnormal risk assessment and response module:

[0009] The data collection and preprocessing module obtains emotional feature information in real time based on multi-dimensional data sources, cleans and preprocesses the acquired emotional feature information, and ensures that different emotional feature information is analyzed on the same dimension;

[0010] The feature extraction and anomaly detection module extracts key features from the preprocessed emotional feature information, then performs an anomaly analysis on the extracted feature information, and detects and marks atypical associations between emotional feature information;

[0011] The intelligent perception and negative emotion representation module deploys the pre-trained machine learning model in the emotion monitoring system. Through the intelligent perception ability of the machine learning model, it dynamically perceives subtle changes in feature data, identifies potential abnormal fluctuations, and maps them to negative emotion representation results;

[0012] The abnormal risk assessment and response module further analyzes the identified potential abnormal fluctuations, determines the risk level of the potential abnormal fluctuations, and formulates different response measures for different potential abnormal fluctuation risk levels.

[0013] Preferably, the emotional characteristic information includes language emotional polarity information and diurnal emotional variation information. Language emotional polarity information refers to judging the emotional direction and degree of emotional intensity by analyzing the vocabulary, grammar and sentence patterns in language text or voice expression. Diurnal emotional variation information is used to measure the magnitude and pattern of changes in an individual's emotional state in different time periods of a day.

[0014] Preferably, after obtaining the language emotion polarity information and diurnal emotion variation information in the emotion feature information, the language emotion polarity information is analyzed for abnormalities to generate a language emotion polarity fluctuation index, and the diurnal emotion variation information is analyzed for abnormalities to generate a diurnal emotion variation index. The generated language emotion polarity fluctuation index and diurnal emotion variation index are used to identify atypical correlations between the obtained multi-dimensional data.

[0015] Preferably, the language emotion polarity fluctuation index and the diurnal emotion variation index generated after abnormal analysis obtained through multi-dimensional data sources are input into a pre-trained machine learning model, and a data consistency offset coefficient is generated by the machine learning model. The acquired emotion feature information is intelligently evaluated by the data consistency offset coefficient.

[0016] Preferably, the generated data consistency deviation coefficient is compared and analyzed with a preset data consistency deviation coefficient reference threshold to identify a strong correlation between the acquired emotional characteristic information. The analysis results are as follows:

[0017] If the data consistency deviation coefficient is greater than or equal to the preset data consistency deviation coefficient reference threshold, a potential abnormal correlation is generated, indicating that there is a strong correlation between the acquired emotional characteristic information;

[0018] If the data consistency deviation coefficient is less than a preset data consistency deviation coefficient reference threshold, emotional characteristic independence is generated, indicating that there is no strong correlation between the acquired emotional characteristic information.

[0019] Preferably, when there is a strong correlation between the acquired emotional characteristic information and a potential abnormal correlation is generated, the data consistency deviation coefficient generated after the abnormal analysis of the emotional characteristic information is continuously obtained, an analysis set is established, and the data consistency deviation coefficient in the analysis set is compared with the first-level reference threshold and the second-level reference threshold and the data consistency deviation coefficient reference threshold, wherein the second-level reference threshold is greater than the first-level reference threshold, and the first-level reference threshold is greater than the data consistency deviation coefficient reference threshold, and the data consistency deviation coefficient is compared and analyzed with the second-level reference threshold, the first-level reference threshold and the data consistency deviation coefficient reference threshold, and the number of data consistency deviation coefficients that are less than the first-level reference threshold and greater than or equal to the data consistency deviation coefficient reference threshold is calibrated as Fa, the number of data consistency deviation coefficients that are less than the second-level reference threshold and greater than or equal to the first-level reference threshold is calibrated as Fb, and the number of data consistency deviation coefficients that are greater than or equal to the second-level reference threshold is calibrated as Fc;

[0020] Comprehensively analyze Fa, Fb, and Fc to generate a strong correlation coefficient SCC based on the following formula:

[0021]

[0022] , where h1, h2, and h3 are the preset proportional coefficients of Fa, Fb, and Fc, respectively, and Fa, Fb, and Fc are all greater than 0.

[0023] Preferably, the generated strong correlation coefficient is compared and analyzed with a preset first strong correlation coefficient reference threshold and a preset second strong correlation coefficient reference threshold, and the analysis results are as follows;

[0024] If the strong correlation coefficient is less than the first strong correlation coefficient reference threshold, the potential abnormal correlation is further classified as a low-risk strong correlation anomaly;

[0025] If the strong correlation coefficient is greater than or equal to the first strong correlation coefficient reference threshold and less than the second strong correlation coefficient reference threshold, the potential abnormal correlation is further classified as a medium-risk strong correlation anomaly:

[0026] If the strong correlation coefficient is greater than or equal to the second strong correlation coefficient reference threshold, the potential abnormal correlation is further classified as a high-risk strong correlation anomaly.

[0027] Preferably, the logic of performing abnormal analysis on the language sentiment polarity information and generating the language sentiment polarity fluctuation index is as follows:

[0028] Under the detection window, for a given language input, the sentiment polarity score S is extracted through the sentiment analysis model i, and convert it into time series data. The input language text time series data is as follows: T = {t i}={t1, t2, t3,...,t n}, where T represents the language text sequence collected in the detection window, t i represents the i-th language content at time point t, n represents the total number of texts collected in the detection window, and the output sentiment polarity sequence data is as follows: S = {S i}={S1,S2,……,S m}, where S represents the sentiment polarity sequence, S i represents the sentiment polarity score of the ith language content at time point t, m represents the total number of language texts in the detection window, that is, how many language text data are analyzed in total, reflecting how many sentiment records are processed in the detection window;

[0029] The emotion polarity change rate at each time point is calculated on the emotion polarity sequence S. The emotion polarity change trend is captured by high-order difference and the sensitivity to mutation is enhanced. The calculation expression is as follows:

[0030]

[0031] , where G t represents the gradient of sentiment polarity change at time point t, is the second-order difference, which measures the acceleration of the change in sentiment polarity and captures the sharp fluctuations. is the first-order difference, capturing the rate of change of sentiment polarity at consecutive time points, and α is the balance coefficient, which is used to adjust the influence of the first-order and second-order gradients;

[0032] A nonlinear weight function is introduced to emphasize the gradient of each sentiment polarity change G t The sentiment polarity of is abnormal, and the expression is as follows: In the formula, W(G t ) is a nonlinear weight parameter, which gives a higher influence to dramatic changes in emotional polarity and reduces the interference of small fluctuations. β is a weight amplification coefficient, which is used to control the degree of nonlinear amplification. e represents the natural base number.

[0033] The nonlinear weights of all time points are accumulated to generate the language sentiment polarity fluctuation index, which is expressed as follows:

[0034]

[0035] , where SPFI represents the language sentiment polarity fluctuation index, γ is the nonlinear scaling coefficient used to control the influence of the polarity score in the fluctuation index, and max(|S|) is the maximum absolute value in the sentiment polarity sequence S.

[0036] Preferably, the logic of performing abnormal analysis on the diurnal mood variation information and generating the diurnal mood variation index is as follows:

[0037] Under the detection window, the emotional data of the target individual is continuously collected and the collected emotional data is marked as E t , E t It is the emotional state value of the detected individual at time point t;

[0038] Emotional states fluctuate over time, so we need to calculate the instantaneous rate of change of the time series to capture the magnitude of emotional changes. Using the derivative form, we define the emotional change rate function, and the calculation expression is as follows:

[0039]

[0040] , where ΔE(t) represents the instantaneous rate of change of emotion, reflecting the amplitude of emotion fluctuation at time point t, and f circ (t) is the circadian rhythm weight function, which is used to consider the periodic influence of day and night.

[0041] in, 2π is the factor used to map the period of a day to the period of trigonometric functions;

[0042] In the instantaneous change rate of emotion ΔE(t), a nonlinear fluctuation intensity function is introduced to detect abnormal emotion fluctuations, which is used to quantify the abnormal fluctuation amplitude of emotion. The expression is as follows:

[0043]

[0044] , where F anom (t) represents the abnormal emotion fluctuation intensity function, t0 represents the starting time of emotion monitoring, t1 represents the end time of emotion monitoring, θ is a nonlinear adjustment parameter used to control the sensitivity of emotion fluctuation, λ is an adjustment weight used to control the influence of emotion state on abnormal fluctuation, g(E(t)) is an emotion state adjustment function, where g(E(t)) = exp(-σE(t)), where σ is a parameter that controls the decay rate;

[0045] The time-weighted result based on the abnormal mood fluctuation intensity function is combined with the circadian rhythm to generate the circadian mood variability index. The calculation expression is as follows:

[0046]

[0047] , where DEVI represents the circadian mood variability index and k is the circadian rhythm weight adjustment factor.

[0048] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0049] The present invention effectively solves the multicollinearity problem caused by the strong correlation between multi-dimensional data features by calculating the data consistency offset coefficient and the strong correlation coefficient. This collinearity problem often leads to instability of the regression coefficient and weight of the model, affecting the accurate judgment of the independent contribution of emotional features to the emotional state. By identifying potential abnormal correlations in real time and combining with the set risk threshold, the system can accurately classify and process emotional feature information when multicollinearity occurs, ensuring the output stability of the model. This mechanism improves the accuracy of emotional state prediction, avoids the volatility of prediction results due to data correlation, ensures the robustness of the model in different time periods and data sets, and thus improves the accuracy and reliability of emotion monitoring.

[0050] The present invention continuously collects emotional feature information, dynamically generates data consistency deviation coefficients, establishes an analysis set, and supports long-term emotional change trend analysis. By comparing the data in the analysis set with the set threshold, the system can identify the risk level of the emotional feature information and provide personalized response strategies based on this. For example, low-risk anomalies may only require routine emotional management reminders, medium risks will trigger psychological counseling interventions, and high risks will activate emergency response mechanisms. This multi-level risk analysis and response strategy enables the system to have the ability to continuously and dynamically monitor emotional states, support long-term trend analysis and refined management of emotional characteristics, and ensure that students' emotional states can be managed and adjusted in a timely and effective manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0052] Figure 1 It is a module schematic diagram of the university social negative emotion representation and analysis system of the present invention. DETAILED DESCRIPTION

[0053] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of the present disclosure will be more comprehensive and complete, and the concept of the example embodiments will be fully conveyed to those skilled in the art.

[0054] The present invention provides Figure 1 The university social negative emotion representation analysis system shown in the figure includes a data collection and preprocessing module, a feature extraction and anomaly detection module, an intelligent perception and negative emotion representation module, and an abnormal risk assessment and response module:

[0055] The data collection and preprocessing module obtains emotional feature information in real time based on multi-dimensional data sources, cleans and preprocesses the acquired emotional feature information, and ensures that different emotional feature information is analyzed on the same dimension;

[0056] Emotional feature information includes language emotion polarity information and diurnal emotion variation information. Language emotion polarity information refers to the degree of emotion direction (positive, neutral or negative) and emotion intensity determined by analyzing the vocabulary, grammar and sentence structure in language text or speech expression. Diurnal emotion variation information is used to measure the magnitude and pattern of changes in an individual's emotional state in different time periods of the day (such as morning, afternoon and night).

[0057] Multidimensional data sources cover a variety of characteristic information of individuals in social situations, mainly including the following categories:

[0058] Language feature data: such as language sentiment polarity information, speaking speed, intonation, and sentiment word frequency in conversation texts, comments, and speech transcripts.

[0059] Behavioral characteristic data: including micro-expressions, facial expression changes, body movements and interactive behavior activity.

[0060] Physiological characteristic data: collected through wearable devices, such as heart rate variability, galvanic skin response, breathing rate and sleep quality.

[0061] Time series emotion data: such as diurnal emotion variation information, emotion stability, and behavioral rhythm characteristics.

[0062] These data sources reflect the individual's psychological state and social behavior from different perspectives, providing comprehensive support for negative emotion representation analysis.

[0063] The feature extraction and anomaly detection module extracts key features from the preprocessed emotional feature information, then performs an anomaly analysis on the extracted feature information, and detects and marks atypical associations between emotional feature information;

[0064] After obtaining the language emotion polarity information and diurnal emotion variation information in the emotion feature information, the language emotion polarity information is analyzed for abnormalities to generate a language emotion polarity fluctuation index, and the diurnal emotion variation information is analyzed for abnormalities to generate a diurnal emotion variation index. The generated language emotion polarity fluctuation index and diurnal emotion variation index are used to identify atypical associations between the obtained multi-dimensional data;

[0065] Atypical correlations between emotional feature information refer to abnormal or unexpected relationships between different data types that do not conform to conventional patterns or statistical laws. For example, in sentiment analysis, we usually expect positive language emotional expressions to be associated with high social activity, but if positive language coexists with low interactive behavior, this abnormal combination is an atypical correlation. Atypical correlations often reveal complex psychological states of individuals, such as emotional masking, behavioral inhibition, or a disharmony between physiological state and psychological expression.

[0066] The atypical correlations between emotional feature information are as follows:

[0067] 1. The association between positive language emotion polarity and low social interaction: Individuals express themselves very positively in language, but show social withdrawal in behavior, which may indicate emotional concealment or potential social anxiety.

[0068] 2. The association between circadian mood variability and adequate sleep: Although the sleep duration is sufficient, the mood variability is large, which may indicate that biological rhythm disorders or psychological stress lead to emotional instability.

[0069] 3. The association between high speech rate and low emotional intensity: Faster speech rate but lack of emotional expression in language may indicate an anxious state or emotional numbness under psychological pressure.

[0070] 4. The association between high heart rate and stable emotional expression: A significant increase in heart rate but calm outward emotional expression indicates that the individual may be suppressing emotions or experiencing a high-stress situation.

[0071] These atypical associations reveal the complex relationships among emotions, behaviors, and physiological states, which require comprehensive analysis to avoid misjudgments caused by simple linear inferences.

[0072] During the data fusion process, if the language sentiment polarity changes significantly, it means that there may be a strong correlation between the preprocessed emotional feature information, which is indeed easy to cause multicollinearity problems. The reason for multicollinearity is that there is a high linear correlation between the features of the model input, which makes the model unable to effectively distinguish the independent contribution of each feature. For example, there may be a strong intrinsic correlation between language sentiment polarity and features such as diurnal emotion variation and speech rate variation: when an individual's emotions fluctuate greatly over time, the emotional expression in the language will also tend to be extreme, or the speech rate will increase under high anxiety, and the emotional polarity of the language will also be significantly enhanced. This strong correlation will cause the model to have coefficient instability in regression analysis, and even the model output will fluctuate violently when the data changes slightly. In addition, if the model assigns a higher weight to these redundant features, the model's interpretability will decrease, and it will be difficult to clarify the independent contribution of each feature to the final emotion judgment. Especially in the application scenario of emotion monitoring, this instability will bring about prediction bias, leading to incorrect trend judgment or monitoring misjudgment.

[0073] The logic of abnormal analysis of language sentiment polarity information and generating language sentiment polarity fluctuation index is as follows:

[0074] Under the detection window, for a given language input (such as a conversation or text), the sentiment polarity score S is extracted through the sentiment analysis model i , and convert it into time series data. The input language text time series data is as follows: T = {t i}={t1, t2, t3,...,t n}, where T represents the language text sequence collected in the detection window, t i represents the i-th language content (such as a comment or a conversation sentence) at time point t, i from 1 to n represents the order of the text in the time series, and n represents the total number of texts collected in the detection window, that is, how many language content expressions in the time series are included in the analysis. Each language content corresponds to a time point, and the output sentiment polarity sequence data is as follows: S = {S i}={S1,S2,……,S m}, where S represents the sentiment polarity sequence, S i represents the sentiment polarity score of the ith language content at time point t, m represents the total number of language texts in the detection window, that is, how many language text data are analyzed in total, reflecting how many sentiment records are processed in the detection window;

[0075] In sentiment analysis, language text T consists of a series of text units (such as sentences, phrases, or dialogues). These texts cannot be used directly to calculate or analyze sentiment polarity fluctuations, but need to be processed by sentiment analysis models to extract sentiment-related numerical features, namely sentiment polarity scores S.i .

[0076] Through the sentiment analysis model, the sentiment-related numerical features are extracted, namely the sentiment polarity score S i The steps are as follows:

[0077] First, we need to analyze the original text T = {t i}={t1, t2, t3,...,t n} Clean and normalize to ensure that the sentiment analysis model can process it efficiently.

[0078] After text preprocessing, each text unit t i Passed as input to the sentiment analysis model. The model determines the sentiment tendency of the text based on vocabulary and context and generates a sentiment polarity score S i .

[0079] For each text unit t i The result of the analysis is a sentiment polarity score S i , indicating the emotional tendency and intensity of the text.

[0080] Finally, the entire language text T is converted into a sentiment polarity sequence S = {S i}={S1,S2,……,S m}, each S i Corresponding to a text unit t i The sentiment polarity score of .

[0081] In sentiment analysis and mood swing analysis, the given language input refers to the text or language data used for model analysis. These inputs come from various social, conversational or recorded environments and are used to extract emotional information to reflect changes in emotional states. Language input is the core data source of the sentiment analysis system. After processing, it will be converted into numerical features, such as the sentiment polarity score Si, for further analysis by the model.

[0082] The source of the given language input

[0083] 1. Social platform data: short text data such as user comments, status updates, Weibo posts, forum speeches, etc.

[0084] 2. Conversation records: chat records of instant messaging tools (such as WeChat, Slack), or customer service conversation texts.

[0085] 3. Speeches and interviews: Recordings are transcribed into text, such as meeting minutes and interview content.

[0086] 4. Written text: News reports, blog posts, email bodies, and text in research papers.

[0087] 5. Speech-to-text data: Conversation content generated by speech transcription technology (such as recording-to-text conversion in voice assistants).

[0088] The characteristics of the given language input

[0089] 1. Different levels of structuring:

[0090] Some language input is structured (such as a customer service conversation with a clear sequence of questions and answers), while some is unstructured (such as user comments on social platforms).

[0091] 2. Context Dependency:

[0092] Certain language inputs, such as chat conversations, require contextual information to accurately identify sentiment. For example, “that’s awesome” could be sarcastic or genuinely positive depending on the context.

[0093] 3. Diversified expression:

[0094] Language input may contain a variety of expressions such as emoticons, abbreviations, slang, irony, etc., which require the model to properly identify and understand them.

[0095] The role of language input in sentiment analysis

[0096] 1. Data source: Language input is the basis for the model to extract emotional features. By analyzing these inputs, a quantitative emotional polarity score S is generated. i .

[0097] 2. Capturing emotional fluctuations: Organizing these input data in time series can capture the dynamic changing trends of emotions.

[0098] 3. Quantitative analysis support: Through the extracted polarity score, further calculate indicators such as the emotional fluctuation index to reveal the changes in emotional states and the correlation between characteristics.

[0099] The given language input refers to the text or conversation data used by the model for sentiment analysis, which reflects the emotional expression of individuals in different scenarios. After preprocessing and sentiment extraction, the language input will be converted into sentiment polarity scores, providing a basis for further sentiment monitoring and data fusion.

[0100] The emotion polarity change rate at each time point is calculated on the emotion polarity sequence S. The emotion polarity change trend is captured by high-order difference and the sensitivity to mutation is enhanced. The calculation expression is as follows:

[0101]

[0102] , where G trepresents the gradient of sentiment polarity change at time point t, is the second-order difference, which measures the acceleration of the change in sentiment polarity and captures the sharp fluctuations. is the first-order difference, capturing the rate of change of sentiment polarity at consecutive time points, and α is the balance coefficient, which is used to adjust the influence of the first-order and second-order gradients;

[0103] A nonlinear weight function is introduced to emphasize the gradient of each sentiment polarity change G t The sentiment polarity of is abnormal, and the expression is as follows: In the formula, W(G t ) is a nonlinear weight parameter, which gives a higher influence to dramatic changes in emotional polarity and reduces the interference of small fluctuations. β is a weight amplification coefficient, which is used to control the degree of nonlinear amplification. e represents the natural base number.

[0104] Giving higher influence to dramatic changes in emotional polarity and reducing the interference of small fluctuations means that by introducing nonlinear weight functions, the model pays more attention to those significant and sudden emotional fluctuations when processing emotional data, while ignoring or weakening those frequent but small emotional changes. The core of this method is that in the process of sentiment analysis, not all emotional fluctuations contribute equally to the analysis results. Drastic changes in emotional polarity often mean potential emotional abnormalities (such as sudden outbursts of anger and anxiety), so higher weights need to be given to highlight their importance. Small emotional fluctuations may be noise or normal emotional fluctuations, and should not have too much impact on the overall analysis results. Through nonlinear weight functions (such as exponential functions), the model can amplify large polarity changes and automatically filter out the interference of small fluctuations, thereby improving the robustness of the analysis and the ability to identify emotional abnormalities.

[0105] The nonlinear weights of all time points are accumulated to generate the language sentiment polarity fluctuation index, which is expressed as follows:

[0106]

[0107] , where SPFI represents the language sentiment polarity fluctuation index, γ is a nonlinear scaling factor used to control the influence of the polarity score in the fluctuation index, and max(|S|) is the maximum absolute value in the sentiment polarity sequence S, which is used to normalize each sentiment polarity score S i ;

[0108] From the calculation expression of the language sentiment polarity fluctuation index, it can be seen that the larger the performance value of the language sentiment polarity fluctuation index generated after abnormal analysis of the language sentiment polarity information, the more dramatic the fluctuation of the language sentiment polarity information in the detection window, which means that the strong correlation between the emotional feature information (such as language emotion, behavioral characteristics, time series emotions, etc.) is high, thereby increasing the probability of multicollinearity problems. This is because when the fluctuation of a certain feature is highly synchronized with other features, it is difficult for the model to distinguish the independent contribution of each feature, resulting in unstable weights and regression coefficients. The smaller the performance value of the language sentiment polarity fluctuation index, the lower the correlation between the features, the higher the feature independence, and the relatively smaller the multicollinearity risk of the model.

[0109] During the data fusion process, changes in diurnal mood variability may indicate strong correlations between the preprocessed mood feature information, thus causing multicollinearity problems. The diurnal mood variability index reflects the magnitude of changes in individual mood over time and has a potential strong correlation with other mood features (such as language emotion polarity, anxiety level, and sleep quality). For example, when an individual's mood fluctuates frequently, his or her language emotion polarity also tends to show an extreme tendency, or a decrease in sleep quality may also directly affect diurnal mood variability. The high correlation between these mood features will make it difficult to distinguish their independent contributions in the fusion model, and the regression coefficients or weights of the model will become unstable, thus causing multicollinearity. Multicollinearity not only causes inconsistent performance of the model on different data sets, but also may cause unreliable parameter estimation and difficulty in interpreting the results. Especially when conducting long-term mood monitoring, it will make the model sensitive to small changes in the data and reduce its robustness.

[0110] The logic of abnormal analysis of diurnal mood variation information and generating diurnal mood variation index is as follows:

[0111] Under the detection window, the emotional data of the target individual is continuously collected and the collected emotional data is marked as E t , E t It is the emotional state value of the detected individual at time point t;

[0112] Emotional data refers to information used to reflect an individual's psychological state and emotional expression, usually derived from multiple dimensions such as language, behavior, physiology, and environment. This data can be structured or unstructured and plays a key role in emotion monitoring and analysis models. The following are the main categories and detailed explanations of emotional data:

[0113] 1. Language feature data

[0114] Source: Text or audio from social media, chat logs, speeches, interviews, etc.

[0115] Sentiment polarity score: Indicates the positive, negative, or neutral sentiment in the language expression (as calculated by an NLP model).

[0116] Emotion word frequency: the frequency of use of specific emotion words (such as "happy" and "angry").

[0117] Speed ​​and tone of voice: Increased speech rate or changes in tone of voice may point to anxiety or excitement.

[0118] Keyword extraction: Identify key emotional themes (e.g., stress, achievement) through semantic analysis.

[0119] Is it necessary to collect:

[0120] Capturing language data is important when language is the primary channel for expressing emotions, such as in social situations or conversation analysis.

[0121] However, in some non-language environments (such as motor behavior or sleep monitoring), data can be selectively not collected.

[0122] 2. Behavioral characteristic data

[0123] Source: micro-expression analysis, body language monitoring, and interactive behavior recording.

[0124] Facial expression analysis: Capture subtle emotional reactions through facial expressions, such as surprise, anger, sadness, etc.

[0125] Behavioral activity: reflects the degree of social participation and the frequency of behavior, such as conversation initiative or participation duration.

[0126] Behavioral inhibition or avoidance: Observing individuals for avoidant behaviors or social withdrawal, revealing underlying anxiety or depression.

[0127] Movement and posture data: Changes in body movements (such as gestures, walking frequency) can reflect emotional states.

[0128] Is it necessary to collect:

[0129] Behavioral data is particularly critical when studying non-verbal expressions, such as micro-expressions or body movements.

[0130] However, in scenarios such as remote interviews or text analysis, behavioral data may not be collected.

[0131] 3. Physiological characteristic data

[0132] Sources: Wearables, biosensors, ECG, etc.

[0133] Heart rate variability (HRV): High heart rate variability is often associated with a state of relaxation, while low variability may indicate a state of stress.

[0134] Electrodermal activity (EDA): Changes in skin conductivity can reflect levels of emotional arousal.

[0135] Respiratory rate: Rapid breathing may be associated with anxiety or tension.

[0136] Sleep Quality: Analyze emotional stability and health status through sleep duration and stages.

[0137] Is it necessary to collect:

[0138] Physiological data is very important in high-precision emotion monitoring, such as anxiety and stress detection. However, the collection of this data requires equipment support, so it can be collected selectively according to the research purpose.

[0139] 4. Time Series Sentiment Data

[0140] Source: Emotion monitoring log, emotion fluctuation data within a time period.

[0141] Circadian Emotional Variability Information (DEVI): reflects the fluctuation amplitude and regularity of emotions during a day.

[0142] Emotional state labeling: Record emotional states at different time points (e.g., scoring every hour).

[0143] Emotional trends and cyclical changes: Identify long-term and short-term trends and cyclical changes in emotions.

[0144] Is it necessary to collect:

[0145] Time series data is crucial when analyzing the patterns of emotional changes and long-term monitoring is required. If it is only for short-term emotional analysis, it is not necessary to collect it.

[0146] 5. Environmental data

[0147] Source: Environmental sensors or geolocation data.

[0148] Ambient volume and brightness: Loud noise or excessive light can cause emotional distress.

[0149] Temperature and humidity: Extreme weather can affect emotional states.

[0150] Geographic location: Changes in location may be associated with mood swings (eg, more relaxed at home, more stressed at the office).

[0151] Is it necessary to collect:

[0152] Environmental data has an indirect impact on emotions and should be collected if emotional changes are related to the environment. However, in situations where language or physiological data are the main analysis, environmental data can be ignored.

[0153] It is not necessary to collect all data. The collection of emotional data needs to be determined in combination with specific application scenarios and analysis goals. For example:

[0154] If the focus is on analyzing an individual's social behavior, language and behavioral data are particularly critical, and physiological or environmental data can be ignored.

[0155] If the study is of stress or anxiety states, physiological data are necessary, but language data may not be required.

[0156] If you want to study long-term mood trends, you need to collect time series mood data and circadian rhythm information.

[0157] Emotional states fluctuate over time, so we need to calculate the instantaneous rate of change of the time series to capture the magnitude of emotional changes. Using the derivative form, we define the emotional change rate function, and the calculation expression is as follows:

[0158]

[0159] , where ΔE(t) represents the instantaneous rate of change of emotion, reflecting the amplitude of emotion fluctuation at time point t, and f circ (t) is the circadian rhythm weight function, which is used to consider the periodic effects of day and night (such as biological rhythms), where 2π is the factor used to map the period of a day to the period of trigonometric functions, which ensures that f circ (t) complete a full cycle (i.e., from 0 to 24 hours), which corresponds to the diurnal cycle in biological rhythms;

[0160] In the instantaneous change rate of emotion ΔE(t), a nonlinear fluctuation intensity function is introduced to detect abnormal emotion fluctuations, which is used to quantify the abnormal fluctuation amplitude of emotion. The expression is as follows:

[0161]

[0162] , where F anom (t) represents the abnormal emotion fluctuation intensity function, t0 represents the starting time of emotion monitoring, t1 represents the end time of emotion monitoring, θ is a nonlinear adjustment parameter used to control the sensitivity of emotion fluctuation, λ is an adjustment weight used to control the influence of emotion state on abnormal fluctuation, g(E(t)) is an emotion state adjustment function used to dynamically adjust the influence of emotion state on abnormal fluctuation detection during the calculation process, where g(E(t)) = exp(-σE(t)), where σ is a parameter that controls the attenuation rate;

[0163] The time-weighted result based on the abnormal mood fluctuation intensity function is combined with the circadian rhythm to generate the circadian mood variability index. The calculation expression is as follows:

[0164]

[0165] , where DEVI represents the circadian mood variability index, and κ is the circadian rhythm weight adjustment factor, which determines the contribution of the circadian rhythm to the variability index;

[0166] From the calculation expression of the diurnal mood variability index, it can be seen that the larger the performance value of the diurnal mood variability index generated after the abnormal analysis of the diurnal mood variability information, it usually indicates that the correlation between the emotional features within the detection window is stronger. In this case, the strong correlation between the diurnal mood variability information and other emotional features (such as language emotion polarity, behavioral inhibition, anxiety level, etc.) may be enhanced, increasing the probability of multicollinearity problems in the model. When multicollinearity exists, it is difficult for the model to accurately identify the independent contribution of each feature to the result, and the regression coefficient may be unstable, resulting in fluctuations in the prediction results. When the performance value of the diurnal mood variability index is smaller, it means that the correlation between the emotional features is weaker, and the contribution of each feature to the model is more independent, so the probability of causing multicollinearity problems is also lower.

[0167] The intelligent perception and negative emotion representation module deploys the pre-trained machine learning model in the emotion monitoring system. Through the intelligent perception ability of the machine learning model, it dynamically perceives subtle changes in feature data, identifies potential abnormal fluctuations, and maps them to negative emotion representation results;

[0168] The language emotion polarity fluctuation index SPFI and the diurnal emotion variation index DEVI generated after abnormal analysis obtained from multi-dimensional data sources are input into the pre-trained machine learning model, and the data consistency deviation coefficient DCDC is generated by the machine learning model. The obtained emotion feature information is intelligently evaluated by the data consistency deviation coefficient DCDC;

[0169] The machine learning model is not limited here, and any model that can realize the generation of data consistency deviation coefficient DCDC from language emotion polarity fluctuation index SPFI and diurnal emotion variation index DEVI is acceptable. In order to realize the technical solution of the present invention, the present invention provides a specific implementation method;

[0170] The calculation formula for generating the data consistency deviation coefficient DCDC is as follows:

[0171]

[0172] , where d1 and d2 are the preset proportional coefficients of the language emotion polarity fluctuation index SPEI and the diurnal emotion variability index DEVI, and both d1 and d2 are greater than 0.

[0173] From the calculation expression of the data consistency shift coefficient, it can be seen that the larger the expression value of the language emotion polarity fluctuation index generated after the abnormal analysis of the language emotion polarity information, the larger the expression value of the diurnal emotion variation index generated after the abnormal analysis of the diurnal emotion variation information, that is, the larger the expression value of the data consistency shift coefficient, the higher the strong correlation between the obtained emotion feature information, thereby increasing the probability of multicollinearity problems. On the contrary, it means that the strong correlation between the obtained emotion feature information is low, and the probability of multicollinearity problems is smaller.

[0174] The generated data consistency deviation coefficient is compared and analyzed with the preset data consistency deviation coefficient reference threshold to identify the strong correlation between the acquired emotional characteristic information. The analysis results are as follows:

[0175] If the data consistency deviation coefficient is greater than or equal to the preset data consistency deviation coefficient reference threshold, a potential abnormal correlation is generated, indicating that there is a strong correlation between the acquired emotional characteristic information;

[0176] If the data consistency deviation coefficient is less than the preset data consistency deviation coefficient reference threshold, the emotional characteristic independence is generated, indicating that there is no strong correlation between the acquired emotional characteristic information;

[0177] The abnormal risk assessment and response module further analyzes the identified potential abnormal fluctuations, determines the risk level of the potential abnormal fluctuations, and formulates different response measures for different potential abnormal fluctuation risk levels.

[0178] When there is a strong correlation between the acquired emotional characteristic information and a potential abnormal correlation is generated, the data consistency deviation coefficient generated after the abnormal analysis of the emotional characteristic information is continuously obtained, an analysis set is established, and the data consistency deviation coefficient in the analysis set is compared with the first-level reference threshold, the second-level reference threshold and the data consistency deviation coefficient reference threshold, wherein the second-level reference threshold is greater than the first-level reference threshold, and the first-level reference threshold is greater than the data consistency deviation coefficient reference threshold, and the data consistency deviation coefficient is compared and analyzed with the second-level reference threshold, the first-level reference threshold and the data consistency deviation coefficient reference threshold, and the number of data consistency deviation coefficients that are less than the first-level reference threshold and greater than or equal to the data consistency deviation coefficient reference threshold is calibrated as Fa, the number of data consistency deviation coefficients that are less than the second-level reference threshold and greater than or equal to the first-level reference threshold is calibrated as Fb, and the number of data consistency deviation coefficients that are greater than or equal to the second-level reference threshold is calibrated as Fc.

[0179] Comprehensively analyze Fa, Fb, and Fc to generate a strong correlation coefficient SCC based on the following formula:

[0180]

[0181] , where h1, h2, and h3 are the preset proportional coefficients of Fa, Fb, and Fc, respectively, and Fa, Fb, and Fc are all greater than 0.

[0182] It can be seen from the strong correlation coefficient calculation expression that the larger the strong correlation coefficient expression value, the greater the probability that the collected emotional characteristic information has a strong correlation, and the smaller the strong correlation coefficient expression value, the smaller the probability that the collected emotional characteristic information has a strong correlation.

[0183] The generated strong correlation coefficient is compared and analyzed with the preset first strong correlation coefficient reference threshold and the second strong correlation coefficient reference threshold, and the analysis results are as follows;

[0184] If the strong correlation coefficient is less than the first strong correlation coefficient reference threshold, the potential abnormal correlation is further classified as a low-risk strong correlation anomaly;

[0185] If the strong correlation coefficient is greater than or equal to the first strong correlation coefficient reference threshold and less than the second strong correlation coefficient reference threshold, the potential abnormal correlation is further classified as a medium-risk strong correlation anomaly:

[0186] If the strong correlation coefficient is greater than or equal to the second strong correlation coefficient reference threshold, the potential abnormal correlation is further classified as a high-risk strong correlation anomaly;

[0187] When a low-risk strong correlation anomaly is detected, it indicates that the data consistency deviation coefficient exceeds the reference threshold but has little impact, and multicollinearity has limited interference with the stability of the model and the analysis results. At this time, the system will send a low-level warning prompt to remind data analysts to pay attention to potential risks, but no immediate intervention is required. The sampling frequency should be increased to continuously monitor the data, and the data should be re-analyzed when necessary to ensure that the feature selection of the model is reasonable.

[0188] When a medium-risk strong correlation anomaly is detected, it means that data offset and feature correlation have caused significant interference to the model, and the model needs to be further optimized. The system will issue an intermediate warning prompt, requiring the team to analyze the sentiment characteristic information and record the fluctuations. In addition, a periodic analysis report should be generated to track the changes in the strong correlation coefficient so that the strategy can be adjusted in time during the future monitoring cycle.

[0189] When high-risk strongly correlated anomalies are detected, it indicates that the model may be seriously disturbed, with systematic errors, data anomalies, or extreme fluctuations in emotional states. At this time, the system will issue an emergency red alert, requiring immediate inspection and correction of the source of the anomaly. Strong regularization methods (such as Lasso regression) or removal of highly correlated features should be used to avoid model inaccuracy. At the same time, other data sources can be introduced for cross-validation to ensure the accuracy of the analysis results, and temporary strategies can be implemented based on the monitoring results to ensure the reliability and long-term robustness of the model.

[0190] The present invention effectively solves the multicollinearity problem caused by the strong correlation between multi-dimensional data features by calculating the data consistency offset coefficient and the strong correlation coefficient. This collinearity problem often leads to instability of the regression coefficient and weight of the model, affecting the accurate judgment of the independent contribution of emotional features to the emotional state. By identifying potential abnormal correlations in real time and combining with the set risk threshold, the system can accurately classify and process emotional feature information when multicollinearity occurs, ensuring the output stability of the model. This mechanism improves the accuracy of emotional state prediction, avoids the volatility of prediction results due to data correlation, ensures the robustness of the model in different time periods and data sets, and thus improves the accuracy and reliability of emotion monitoring.

[0191] The present invention continuously collects emotional feature information, dynamically generates data consistency deviation coefficients, establishes an analysis set, and supports long-term emotional change trend analysis. By comparing the data in the analysis set with the set threshold, the system can identify the risk level of the emotional feature information and provide personalized response strategies based on this. For example, low-risk anomalies may only require routine emotional management reminders, medium risks will trigger psychological counseling interventions, and high risks will activate emergency response mechanisms. This multi-level risk analysis and response strategy enables the system to have the ability to continuously and dynamically monitor emotional states, support long-term trend analysis and refined management of emotional characteristics, and ensure that students' emotional states can be managed and adjusted in a timely and effective manner.

[0192] The above description is only by way of illustration of certain exemplary embodiments of the present invention. It is undoubted that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. The negative emotion representation analysis system for college social interaction is characterized by: It includes data collection and preprocessing module, feature extraction and anomaly detection module, intelligent perception and negative emotion representation module, and abnormal risk assessment and response module: The data collection and preprocessing module obtains emotional feature information in real time based on multi-dimensional data sources, cleans and preprocesses the acquired emotional feature information, and ensures that different emotional feature information is analyzed on the same dimension; The feature extraction and anomaly detection module extracts key features from the preprocessed emotional feature information, then performs an anomaly analysis on the extracted feature information, and detects and marks atypical associations between emotional feature information; The intelligent perception and negative emotion representation module deploys the pre-trained machine learning model in the emotion monitoring system. Through the intelligent perception ability of the machine learning model, it dynamically perceives subtle changes in feature data, identifies potential abnormal fluctuations, and maps them to negative emotion representation results; The abnormal risk assessment and response module further analyzes the identified potential abnormal fluctuations, determines the risk level of the potential abnormal fluctuations, and formulates different response measures for different potential abnormal fluctuation risk levels.

2. The negative emotion representation analysis system for social interaction in colleges and universities according to claim 1 is characterized by: Emotional feature information includes language emotion polarity information and diurnal emotion variation information. Language emotion polarity information refers to judging the emotional direction and intensity of emotion by analyzing the vocabulary, grammar and sentence patterns in language text or voice expression. Diurnal emotion variation information is used to measure the magnitude and pattern of changes in an individual's emotional state at different time periods throughout the day.

3. The negative emotion representation analysis system for college social interaction according to claim 2 is characterized by: After obtaining the language emotion polarity information and diurnal emotion variation information in the emotion feature information, the language emotion polarity information is analyzed for abnormalities to generate a language emotion polarity fluctuation index, and the diurnal emotion variation information is analyzed for abnormalities to generate a diurnal emotion variation index. The generated language emotion polarity fluctuation index and diurnal emotion variation index are used to identify atypical correlations between the acquired multi-dimensional data.

4. The negative emotion representation analysis system for college social interaction according to claim 3 is characterized by: The language sentiment polarity fluctuation index and diurnal emotion variation index generated after abnormal analysis obtained from multi-dimensional data sources are input into a pre-trained machine learning model, and the data consistency offset coefficient is generated by the machine learning model. The acquired emotional feature information is intelligently evaluated through the data consistency offset coefficient.

5. The negative emotion representation analysis system for college social interaction according to claim 4 is characterized by: The generated data consistency deviation coefficient is compared and analyzed with the pre-set data consistency deviation coefficient reference threshold to identify the strong correlation between the acquired emotional characteristic information. The analysis results are as follows: If the data consistency deviation coefficient is greater than or equal to the preset data consistency deviation coefficient reference threshold, a potential abnormal correlation is generated, indicating that there is a strong correlation between the acquired emotional characteristic information; If the data consistency deviation coefficient is less than a preset data consistency deviation coefficient reference threshold, emotional characteristic independence is generated, indicating that there is no strong correlation between the acquired emotional characteristic information.

6. The negative emotion representation analysis system for social interaction in colleges and universities according to claim 5 is characterized by: When there is a strong correlation between the acquired emotional characteristic information and a potential abnormal correlation is generated, the data consistency deviation coefficient generated after the abnormal analysis of the emotional characteristic information is continuously obtained, an analysis set is established, and the data consistency deviation coefficient in the analysis set is compared with the first-level reference threshold, the second-level reference threshold, and the data consistency deviation coefficient reference threshold, wherein the second-level reference threshold is greater than the first-level reference threshold, and the first-level reference threshold is greater than the data consistency deviation coefficient reference threshold, and the data consistency deviation coefficient is compared and analyzed with the second-level reference threshold, the first-level reference threshold, and the data consistency deviation coefficient reference threshold, and the number of data consistency deviation coefficients that are less than the first-level reference threshold and greater than or equal to the data consistency deviation coefficient reference threshold is calibrated as Fa, the number of data consistency deviation coefficients that are less than the second-level reference threshold and greater than or equal to the first-level reference threshold is calibrated as Fb, and the number of data consistency deviation coefficients that are greater than or equal to the second-level reference threshold is calibrated as Fc; Comprehensively analyze Fa, Fb, and Fc to generate a strong correlation coefficient SCC based on the following formula: ， Wherein, h1, h2, and h3 are preset proportional coefficients of Fa, Fb, and Fc, respectively, and Fa, Fb, and Fc are all greater than 0.

7. The negative emotion representation analysis system for college social interaction according to claim 6 is characterized by: The generated strong correlation coefficient is compared and analyzed with the preset first strong correlation coefficient reference threshold and the second strong correlation coefficient reference threshold, and the analysis results are as follows; If the strong correlation coefficient is less than the first strong correlation coefficient reference threshold, the potential abnormal correlation is further classified as a low-risk strong correlation anomaly; If the strong correlation coefficient is greater than or equal to the first strong correlation coefficient reference threshold and less than the second strong correlation coefficient reference threshold, the potential abnormal correlation is further classified as a medium-risk strong correlation anomaly: If the strong correlation coefficient is greater than or equal to the second strong correlation coefficient reference threshold, the potential abnormal correlation is further classified as a high-risk strong correlation anomaly.

8. The negative emotion representation analysis system for college social interaction according to claim 3 is characterized in that: The logic of abnormal analysis of language sentiment polarity information and generating language sentiment polarity fluctuation index is as follows: Under the detection window, for a given language input, the sentiment polarity score S is extracted through the sentiment analysis model i , and convert it into time series data. The input language text time series data is as follows: T = {t i }={t1, t2, t3,...,t n }, where T represents the language text sequence collected in the detection window, t i represents the i-th language content at time point t, n represents the total number of texts collected in the detection window, and the output sentiment polarity sequence data is as follows: S = {S i }={S1,S2,……,S m }, where S represents the sentiment polarity sequence, S i represents the sentiment polarity score of the ith language content at time point t, m represents the total number of language texts in the detection window, that is, how many language text data are analyzed in total, reflecting how many sentiment records are processed in the detection window; The emotion polarity change rate at each time point is calculated on the emotion polarity sequence S. The emotion polarity change trend is captured by high-order difference and the sensitivity to mutation is enhanced. The calculation expression is as follows: ， In the formula, G t represents the gradient of sentiment polarity change at time point t, is the second-order difference, which measures the acceleration of the change in sentiment polarity and captures the sharp fluctuations. is the first-order difference, capturing the rate of change of sentiment polarity at consecutive time points, and α is the balance coefficient, which is used to adjust the influence of the first-order and second-order gradients; A nonlinear weight function is introduced to emphasize the gradient of each sentiment polarity change G t The sentiment polarity of is abnormal, and the expression is as follows: In the formula, W(G t ) is a nonlinear weight parameter, which gives a higher influence to dramatic changes in emotional polarity and reduces the interference of small fluctuations. β is a weight amplification coefficient, which is used to control the degree of nonlinear amplification. e represents the natural base number. The nonlinear weights of all time points are accumulated to generate the language sentiment polarity fluctuation index, which is expressed as follows: ， Wherein, SPFI represents the language sentiment polarity fluctuation index, γ is the nonlinear scaling coefficient used to control the influence of the polarity score in the fluctuation index, and max(|S|) is the maximum absolute value in the sentiment polarity sequence S.

9. The negative emotion representation analysis system for college social interaction according to claim 3 is characterized by: The logic of abnormal analysis of diurnal mood variation information and generating diurnal mood variation index is as follows: Under the detection window, the emotional data of the target individual is continuously collected and the collected emotional data is marked as E t , E t It is the emotional state value of the detected individual at time point t; Emotional states fluctuate over time, so we need to calculate the instantaneous rate of change of the time series to capture the magnitude of emotional changes. Using the derivative form, we define the emotional change rate function, and the calculation expression is as follows: ， In the formula, ΔE(t) represents the instantaneous rate of change of emotion, reflecting the amplitude of emotion fluctuation at time point t, and f circ (t) is the circadian rhythm weight function, which is used to consider the periodic influence of day and night. in, 2π is the factor used to map the period of a day to the period of trigonometric functions; In the instantaneous change rate of emotion ΔE(t), a nonlinear fluctuation intensity function is introduced to detect abnormal emotion fluctuations, which is used to quantify the abnormal fluctuation amplitude of emotion. The expression is as follows: ， In the formula, F anom (t) represents the abnormal emotion fluctuation intensity function, t0 represents the starting time of emotion monitoring, t1 represents the end time of emotion monitoring, θ is a nonlinear adjustment parameter used to control the sensitivity of emotion fluctuation, λ is an adjustment weight used to control the influence of emotion state on abnormal fluctuation, g(E(t)) is an emotion state adjustment function, where g(E(t)) = exp(-σE(t)), where σ is a parameter that controls the decay rate; The time-weighted result based on the abnormal mood fluctuation intensity function is combined with the circadian rhythm to generate the circadian mood variability index. The calculation expression is as follows: ， Where DEVI represents the diurnal mood variability index, and k is the circadian rhythm weight adjustment factor.

Images