Infectious disease social risk perception system
By constructing a social risk perception system for infectious diseases, and employing multidimensional feature fusion and a weighted linear model, the system addresses the shortcomings of existing technologies in terms of real-time performance and multidimensional analysis. It achieves efficient and accurate risk perception and early warning of social media data, supporting public health decision-making.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TIANJIN UNIV
- Filing Date
- 2025-12-19
- Publication Date
- 2026-05-01
AI Technical Summary
Existing methods for social risk perception are insufficient in terms of real-time, dynamic and multi-dimensional analysis capabilities. They cannot reflect fluctuations in public sentiment in a timely manner, affecting decision-making efficiency. Furthermore, they have high requirements for data cleaning and processing and are easily affected by noisy data.
A social risk perception system for infectious diseases is constructed, including modules for data acquisition, preprocessing, feature analysis, and risk early warning. It adopts multi-dimensional feature fusion and weighted linear models, combined with topic screening, keyword library, noise removal, and exponential smoothing techniques, to achieve automated processing and real-time risk assessment of social media data.
It improves the accuracy of risk perception and real-time monitoring capabilities, can automatically process large-scale unstructured data, reduce the impact of short-term noise, provide timely and accurate risk warning information, and support rapid response and long-term social governance in public health management.
Smart Images
Figure CN121964181A_ABST
Abstract
Description
A social risk perception system for infectious diseases Technical Field
[0001] This invention relates to data processing technology, and more specifically, to a social risk perception system for infectious diseases. Background Technology
[0002] In recent years, the outbreak of emerging infectious diseases worldwide has brought enormous challenges to society. Sudden infectious diseases not only put pressure on public health systems but also trigger widespread social panic and anxiety, affecting people's daily lives, economic activities, and government public management decisions. The spread of an epidemic and the implementation of government response measures depend heavily on public cooperation and support. However, public behavior is often influenced by their perception of risk, especially in the early stages of epidemic prevention and control, where information asymmetry and extreme public opinion can easily lead to social instability. Therefore, timely and accurate capture of changes in public perception of social risk has become crucial for effectively responding to public health emergencies.
[0003] Social media, as a crucial platform for public expression of emotions and dissemination of information, reflects social sentiment and public risk perception. Especially during emergencies, social media provides a low-cost, high-efficiency channel for information dissemination. Public emotional fluctuations, reactions to events, and the formation and evolution of public opinion can all be presented in real time through social platforms. Social media platforms such as Weibo and WeChat not only allow users to express personal opinions but also showcase the collective emotions of social groups through interactive behaviors such as likes, reposts, and comments. In recent years, researchers and government agencies have increasingly emphasized the use of social media data analysis to understand society's emotional responses and risk perceptions to public health emergencies, thereby providing decision support and optimizing public health management and social governance.
[0004] Social media contains a wealth of user sentiment information, and sentiment analysis, as an important task in natural language processing, is widely used in fields such as public opinion monitoring and sentiment marketing. Most existing Weibo sentiment classification models rely on static word embeddings and extract text features through a single convolutional layer, which is not ideal in some scenarios, especially when dealing with multimodal data and complex sentiment classification.
[0005] With the development of artificial intelligence and big data technologies, social media-based public opinion analysis has matured and is widely used in research on social risk perception. Currently, techniques such as sentiment analysis, emotion mining, and social network analysis are widely applied in the processing and analysis of social media data. Sentiment analysis primarily utilizes natural language processing and machine learning to analyze the sentiment tendencies of social media texts, identifying negative or positive emotions to help predict trends in public opinion. Social network analysis focuses on studying the propagation paths and structures of information on social platforms, assessing the spread and influence of public opinion by analyzing user interaction behavior on social networks. Furthermore, with the rise of deep learning technology, sentiment analysis based on pre-trained models such as BERT and LSTM has significantly improved the accuracy and real-time performance of public opinion analysis.
[0006] However, existing methods for social risk perception mainly rely on traditional questionnaires or statistical data analysis, which are insufficient in terms of real-time, dynamic, and multi-dimensional analytical capabilities. Traditional social risk perception methods are usually based on static data or limited samples, and cannot dynamically capture the changing trends of social sentiment and public risk perception. Especially in the early stages of public health events, information dissemination is rapid and complex, and traditional methods often fail to reflect fluctuations in public sentiment in a timely manner, affecting decision-making efficiency. At the same time, existing research focuses on the analysis of single data sources, such as sentiment scores or topic popularity, neglecting the comprehensive analysis of social interaction characteristics and multi-dimensional information such as time and location, making it difficult to provide comprehensive and accurate risk assessments. In addition, traditional methods have high requirements for data cleaning and processing, and are highly sensitive to noisy data, which can easily affect the stability and reliability of the analysis results.
[0007] Therefore, there is an urgent need for an efficient, accurate, and real-time method for social risk perception and early warning to help governments and related agencies optimize public health decision-making and improve social management efficiency. By combining multi-dimensional analysis of social media data, it is possible to more comprehensively capture the trajectory of changing social sentiment and monitor the dynamic changes in risk perception in real time, thereby providing decision-makers with more timely and accurate risk warning information. This early warning method can not only help governments respond quickly in the early stages of public health emergencies and reduce social panic, but also provide data support for long-term social governance and improve society's ability to respond to emergencies. Summary of the Invention
[0008] In view of this, the purpose of this invention is to provide a social risk perception system for infectious diseases.
[0009] To address the aforementioned technical problems, the technical solution of this invention is: a social risk perception system for infectious diseases, comprising: a data acquisition module for acquiring interactive data from an interactive platform; a data preprocessing module for preprocessing the interactive data to obtain an interactive dataset; a feature analysis module configured with several feature analysis items and a feature analysis unit for each feature analysis item, wherein the feature analysis unit is used to analyze the feature item data of the interactive data to obtain the feature item information of the interactive data; and a risk warning module configured with a risk analysis algorithm and risk triggering conditions, wherein the risk analysis algorithm is used to calculate the confidence risk value of each interactive data and output risk warning information according to the risk triggering conditions satisfied by the confidence risk value.
[0010] Furthermore, the data acquisition module includes a data identification unit and a data crawling unit. The data identification unit is configured with a topic filtering model and an epidemic keyword library. The topic filtering model is used to extract topic words from the interactive data. The epidemic keyword library is configured with epidemic keywords. The data identification unit combines the epidemic keywords in the interactive data with the interactive data. The data crawling unit determines whether the crawling conditions are met based on the topic words and matching results. If the conditions are met, the interactive data is crawled.
[0011] Furthermore, the data preprocessing module includes a data denoising unit, a missing information compensation unit, and a data standardization unit. The data denoising unit is used to identify and delete redundant information in the interactive data. The missing information compensation unit is used to identify and supplement missing information in the interactive data. The data standardization unit is used to adjust the format of the interactive data according to the type of the interactive data so that the interactive data meets the baseline format corresponding to that type.
[0012] Further: The feature analysis module includes a counting feature unit, an emotion feature unit, a semantic feature unit, an interaction feature unit, and a spatiotemporal feature unit. The counting feature unit is used to calculate the counting features in the interaction data; the emotion feature unit is used to calculate the emotion features in the interaction data; the semantic feature unit is used to calculate the semantic features in the interaction data; the interaction feature unit is used to calculate the interaction features in the interaction data; the spatiotemporal feature unit is used to calculate the spatiotemporal features in the interaction data; the aggregation feature unit is used to calculate the aggregation features in the interaction data; and the sequence derivation feature unit is used to calculate the sequence derivation features in the interaction data.
[0013] Furthermore, the counting features include the total amount of text, the amount of text related to the epidemic, and the proportion of text related to the epidemic.
[0014] Furthermore: the emotion feature unit is configured with an emotion classification model, which is used to analyze the interaction data to obtain an emotion intensity index, and the emotion feature includes the emotion intensity index.
[0015] Furthermore: the semantic feature unit is configured with a topic analysis model, which is used to analyze the interaction data to obtain a semantic topic index, and the semantic feature includes the semantic topic index.
[0016] Furthermore, the spatiotemporal features include time periodic features and geographical distribution features, and the time series features include rate of change features and moving average features.
[0017] Furthermore: the aggregation feature unit is used to calculate an aggregation index based on the feature data of other feature analysis items of the interaction data, and the aggregation feature includes the aggregation index.
[0018] Furthermore: the risk analysis algorithm is configured as follows: Step A1, using a time window With the region Define the original feature set for granularity:
[0019] in, For counting characteristics, As an emotional characteristic, For semantic features, As an interactive feature, Spatiotemporal characteristics, As an aggregation feature, For time series features; Step A2, convert the feature vector obtained in step A1 into a time series feature vector. Input into the risk calculation model and perform a linear weighted combination:
[0020] The linear result is mapped to an interval using the Sigmoid function. :
[0021] in, , , , , , , These are the weights of the feature analysis terms, and their values are greater than or equal to 0. For bias terms, The activation function is Sigmoid; Step A3: Apply the risk index obtained in Step A2. Perform time smoothing to obtain the smoothed risk index. The calculation method is as follows:
[0022] in, The smoothing risk index represents the time window t and the region r. This represents the smoothed risk index for the previous time window; This represents the original risk index calculated in step A2; Represents the exponential smoothing coefficient, with a range of values. Step A4: Judge the result obtained in step A3. When the risk warning information is output, θ is the preset risk threshold.
[0023] The main technical effects of this invention are reflected in the following aspects: 1. Multi-dimensional fusion features improve the accuracy of risk perception: This invention constructs a multi-dimensional feature system of counting, emotion, semantics, interaction, spatiotemporal, aggregation and sequence. Compared with traditional methods that only use text emotion or dissemination volume, it can comprehensively depict social emotions, public opinion dissemination and spatial diffusion status, and greatly improve the reliability of risk index.
[0024] 2. Model-level integrated computation enhances real-time monitoring capabilities: This invention employs a weighted linear model and a Sigmoid mapping structure, which enables the model to have advantages such as high computational efficiency and the ability to generate risk indices in real time, meeting the needs of rapid response in public health emergencies.
[0025] 3. Exponential smoothing mechanism enhances the stability of risk signals: The introduction of an exponential smoothing mechanism into the risk index effectively weakens the short-term noise impact of social media data, enabling the risk index to maintain trend stability over continuous time windows and improving the accuracy of risk judgment.
[0026] 4. Automated data processing enhances system robustness: Through techniques such as topic filtering, keyword database matching, noise removal, and format standardization, this invention can automatically process large-scale, unstructured social media data, improving the system's adaptability to complex data environments. Attached Figure Description
[0027] Figure 1: System architecture diagram of the present invention; Figure 2: Emotion classification model diagram of the present invention; Figure 3: Flowchart of risk index calculation of the present invention. Detailed Implementation
[0028] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, so that the technical solution of the present invention can be more easily understood and mastered.
[0029] Referring to Figure 1, an infectious disease social risk perception system includes a data acquisition module for acquiring interactive data from an interactive platform. The data acquisition module comprises a data identification unit and a data crawling unit. The data acquisition module collects interactive data from the social platform in real time, including text, interactive behaviors (forwarding, commenting, liking), posting time, and geographic information. The data identification unit is equipped with a topic filtering model and an epidemic keyword database. The topic filtering model uses a word embedding model (such as BERT, word2vec) + multi-layer convolutional or multi-head attention structure to learn representations of the interactive data and identify its topic words, such as disease-related words like "fever," "vomiting," "cough," and "diarrhea."
[0030] The topic selection model is used to extract topic words from interactive data. The topic selection model adopts a combination of BERT model and multi-head attention mechanism (Transformer). This architecture can effectively extract topic words related to infectious diseases from social media data. The specific architecture and working principle are as follows: (1) Model architecture: 1. Input layer: The input of the model is preprocessed text data. The input data is first processed by BERTTokenizer and then the words are converted into 768-dimensional word vectors using BERT Embedding Layer.
[0031] 2. BERT Model Layer: A pre-trained BERT model is used for bidirectional text encoding, enabling the acquisition of a semantic representation of each word within its context. Internally, BERT employs a Transformer architecture, containing 12 encoder layers, each with 12 attention heads. The input to each layer is a combination of the representation output from the previous layer and the multi-head attention mechanism of the current layer.
[0032] 3. Output Layer: Through the output of the last layer of the BERT model, combined with the classifier, the model can predict disease-related topic words in the text based on the output feature vectors. The output format of the topic words is a probability distribution, representing the probability of each word being a topic word in a specific context.
[0033] (2) Training Process: Model Initialization and Fine-tuning: The model first initializes the word embedding layer using a pre-trained BERT model, pre-training on a large amount of unlabeled text data. The goal of pre-training is to enable BERT to learn rich language expression capabilities through the Masked Language Model and Next Sentence Prediction tasks. Based on this, the model is fine-tuned using labeled social media data to focus on extracting topic words related to infectious diseases. The purpose of fine-tuning is to enable the model to learn and identify specific disease-related topics based on the specific content and context of social media.
[0034] Hyperparameter training: During training, the cross-entropy loss function was used as the training objective, and the Adam optimizer was employed to update the model parameters. The initial learning rate was set to 2e-5, and the training process was optimized by dynamically adjusting the learning rate. The batch size was 32, and the model was trained for 20 epochs. Cross-validation was used to evaluate its performance, and the evaluation metrics were precision, recall, and F1 score.
[0035] The epidemic keyword database is configured with epidemic keywords, which include the following categories: ① Symptom-related words: fever, diarrhea, nausea, etc.; ② Scenario-related words: hospital, crowd gathering, school, etc.; ③ Behavioral words: seeking medical treatment, isolation, queuing, etc.
[0036] The data extraction unit calculates a relevance score based on the matching degree between topic words and epidemic-related keywords:
[0037] like If the value is greater than or equal to a preset threshold T, then the data is determined to be epidemic-related data and is captured. The data identification unit identifies epidemic keywords and interactive data in the interactive data, and the data capture unit determines whether the capture conditions are met based on the topic words and matching results. If the conditions are met, the interactive data is captured.
[0038] The data preprocessing module is used to preprocess interactive data to obtain an interactive dataset. This module includes a data denoising unit, a missing data compensation unit, and a data standardization unit. The data denoising unit identifies and removes redundant information from the interactive data. It uses regularization rules combined with a noise dictionary to remove noise elements such as URLs, emojis, repeated characters (e.g., "ahhhhh"), and meaningless text (advertisements, spam comments). The missing data compensation unit identifies and supplements missing information in the interactive data. It compensates for missing fields: if a time field is missing, it fills it in using a timestamp returned by the platform; if geographic information is missing, it can be parsed using user attributes or GPS; if behavioral data (e.g., likes) is missing, it fills it with 0. The data standardization unit adjusts the format of the interactive data according to its type to ensure it conforms to the corresponding baseline format. The data standardization unit unifies all data into a structured format. This ensures that the data meets the input requirements of the feature analysis module.
[0039] The feature analysis module is configured with several feature analysis items and a feature analysis unit for each feature analysis item. The feature analysis unit is used to analyze the feature item data of the interaction data to obtain the feature item information of the interaction data. The feature analysis module includes a counting feature unit, an emotion feature unit, a semantic feature unit, an interaction feature unit, a spatiotemporal feature unit, an aggregation feature unit, and a time series feature unit, which are used to calculate seven categories of features to form a complete feature vector.
[0040] Where t represents the time window and r represents the region.
[0041] The feature analysis module includes a counting feature unit, an emotion feature unit, a semantic feature unit, an interaction feature unit, and a spatiotemporal feature unit; the counting feature unit is used to calculate the counting features in the interaction data; the counting features include the total amount of text, the amount of epidemic-related text, and the proportion of epidemic text. The counting features include the total amount of interaction, the amount of epidemic-related text, and the proportion of epidemic text: (1) Total amount of interaction:
[0042] in, The number of interactive data entries within the time window t and the region r.
[0043] (2) Volume of text related to the epidemic:
[0044] in, This indicates that the i-th data entry contains keywords related to the epidemic. This indicates that it does not include.
[0045] (3) Proportion of COVID-19 related texts:
[0046] in, Indicates counting characteristics; This represents the total amount of interactive data within the time window t and the region r. This indicates the number of COVID-19 related texts within the same window.
[0047] The emotion feature unit is used to calculate the emotion features in the interaction data; the emotion feature unit is configured with an emotion classification model, which is used to analyze the interaction data to obtain an emotion intensity index, and the emotion features include the emotion intensity index. Based on the emotion classification model (as shown in Figure 2), the emotion feature unit outputs an emotion probability distribution:
[0048] The emotion classification model is used to extract and classify emotion features from social media or other text data. The model is based on a convolutional neural network and combines multi-scale convolution operations and average pooling to efficiently extract emotion information from text. Its model architecture is shown in Figure 2. The specific architecture and working principle are as follows: (1) Model architecture: Input layer: The input of the model is text data, which first needs to be segmented. Through Word2Vec word embedding, each word is converted into a 300-dimensional word vector. After each text data is embedded, it forms a two-dimensional matrix, where each row represents the embedding representation of a word in the text. Therefore, the input dimension is .
[0049] Multi-scale convolutional layers: This model uses multi-scale convolutional operations, employing four convolutional kernels of different sizes (1, 2, 3, and 4) to capture local features of the text at different scales. The output of the convolutional operation is a feature map, which is the feature representation generated by each convolutional kernel.
[0050] Average pooling layer: After convolution, average pooling is used to pool each feature map, reducing feature dimensionality and extracting the most important emotional features. After pooling, the output feature vector is flattened and used as input to the next layer.
[0051] Fully connected layer: The pooled feature vectors pass through a fully connected layer, which combines the extracted features and ultimately maps them to the probability distribution of the class labels. The ReLU activation function is used to introduce a non-linear transformation, enhancing the model's expressive power.
[0052] Output layer: The output layer is a Softmax layer, which is used to transform the output of the fully connected layer into the probability distribution of each category, and finally outputs the sentiment category predicted by the model.
[0053] (2) Training process: Word2Vec word embeddings are used to convert words into 300-dimensional word vectors. For texts of different lengths, zero padding is performed to ensure that all input texts have the same length. The cross-entropy loss function is used to calculate the difference between the predicted class and the actual class, and the Adam optimizer is used to update the parameters. The batch size is set to 64, the initial learning rate is set to 1e-4, and it is dynamically adjusted according to the training process. The model is trained for 30 epochs, that is, the model will be trained once using the entire training dataset in each epoch. After each epoch, a validation evaluation is performed, and the decision to continue training is made based on the performance on the validation set. The model's performance on the validation set is monitored during training, and the parameters are adjusted as needed. To prevent overfitting, an early stopping strategy is adopted during training. If the validation set loss does not improve within 5 consecutive epochs, training is terminated to avoid unnecessary computation and potential overfitting.
[0054] in, The emotional intensity index represents the time window t; Number of emotion categories; This represents the probability that the interaction data output by the emotion classification model belongs to the i-th emotion category within the time window t; The value represents the emotion intensity weight corresponding to the i-th emotion category (e.g., +1 for positive, 0 for neutral, and -1 for negative). The above emotion classification model includes an embedding layer, a convolutional layer, or an attention layer structure, which is used to extract emotion-related features from the text.
[0055] The semantic feature unit is used to calculate the semantic features in the interactive data; the semantic feature unit is configured with a topic analysis model, which is used to analyze the interactive data to obtain a semantic topic index, and the semantic features include the semantic topic index. The semantic feature unit adopts a semantic coding model. The topic analysis model is a pre-trained semantic coding model. Its specific application principle is as follows: (1) Data preparation and preprocessing: a pre-trained BERT model is used to map each word to a 768-dimensional vector.
[0056] (2) Model Training: The model is initialized using a pre-trained BERT model and pre-trained on large-scale unsupervised data. Then, it is fine-tuned using labeled social media data to focus on extracting disease-related features and generating topic vector representations. The final layer of the model weights and sums the vectors from different topics to obtain a comprehensive semantic topic index. The formula is:
[0057] in, The semantic topic index represents the time window t; Indicates the number of topics; This represents the topic probability of the interactive data on the j-th topic; This represents the semantic weight of the j-th topic in relation to the risk of the epidemic (used to distinguish the importance of semantic categories such as "fever", "cough", and "diarrhea").
[0058] The training process uses the cross-entropy loss function to evaluate the difference between the topic distribution of the text and the actual labels. The Adam optimizer is used to adjust the model weights.
[0059] (3) Hyperparameter settings: The initial learning rate is set to 1e-4, and a learning rate decay strategy is adopted to gradually reduce the learning rate as the training progresses. The parameter is set to 32, and 32 samples are used for parameter updates in each training iteration. Training is performed for 30 rounds, with each round iterating through all training data. During training, cross-validation is used to evaluate the model's performance. If the validation set performance does not improve within 10 consecutive rounds, training will stop early to avoid overfitting.
[0060] The interaction feature unit is used to calculate the interaction features in the interaction data; the interaction feature unit is used to construct interaction intensity features based on user behavior fields, and its calculation method is as follows:
[0061] in, Indicates the interaction intensity index; This represents the number of forwards within the time window t; Indicates the number of comments; This indicates the number of likes; The interaction features are used to reflect the spread of epidemic-related content and the degree of user response.
[0062] The spatiotemporal feature unit is used to calculate the spatiotemporal features in the interactive data; the spatiotemporal features include time period features and geographical distribution features.
[0063] The spatiotemporal feature unit constructs spatiotemporal features based on the time and location attributes of the interactive data, specifically including: (1) time period features:
[0064] Where t represents the time window (in hours). This represents time characteristics encoded in a 24-hour cycle, used to capture day-night activity patterns.
[0065] (2) Geographical distribution characteristics:
[0066] in, This represents the amount of interactive data within the time window t and the region r. This represents the total amount of interactive data across all regions within that time window; This represents the relative proportion of region r within the time window t, used to measure regional clustering.
[0067] therefore:
[0068] The aggregation feature unit is used to calculate aggregation features in the interactive data; the aggregation feature unit is used to calculate an aggregation index based on the feature data of other feature analysis items of the interactive data, and the aggregation features include the aggregation index. The aggregation feature unit is used to generate an aggregation index based on the weighted summation of multidimensional features, and calculates it according to the following formula:
[0069] in, The aggregate risk index represents the time window t; This represents the number of interactive data entries within the time window t; This represents the multidimensional feature vector (including emotion features, semantic features, interaction features, etc.) corresponding to the i-th interaction data. This represents a feature mapping function used to map multidimensional features into aggregateable risk scores; Indicates the credibility of the i-th interaction data; This is a confidence weighting function used to reduce the impact of noisy data on the aggregation results.
[0070] The sequence-derived feature unit is used to calculate the sequence-derived features in the interactive data. The time series features include rate of change features and moving average features. Time series features are used to characterize the trend of feature values over time, including rate of change and moving average: (1) Rate of change feature (trend change):
[0071] in, This represents the numerical value of a certain feature within a time window t and a geographical region r (such as a sentiment index, count features, etc.). This represents the feature value corresponding to the previous time window; This represents the rate of change, used to describe the increase or decrease trend of public opinion related to the epidemic over time.
[0072] (2) Moving average characteristics (smooth changes):
[0073] in, Indicates the preset size of the movable window; : The moving average of the most recent k time windows, used to smooth fluctuations and extract long-term trends.
[0074] final:
[0075] The risk warning module is equipped with a risk analysis algorithm and risk triggering conditions. The risk analysis algorithm is used to calculate the confidence risk value of each interactive data and output risk warning information according to the risk triggering conditions satisfied by the confidence risk value.
[0076] The risk analysis algorithm is configured as follows: Step A1, using a time window With the region Define the original feature set for granularity:
[0077] in, For counting characteristics, As an emotional characteristic, For semantic features, As an interactive feature, Spatiotemporal characteristics, As an aggregation feature, For time series features; Step A2, convert the feature vector obtained in step A1 into a time series feature vector. Input into the risk calculation model and perform a linear weighted combination:
[0078] The linear result is mapped to an interval using the Sigmoid function. :
[0079] in, , , , , , , These are the weights of the feature analysis terms, and their values are greater than or equal to 0. For bias terms, The activation function is Sigmoid; Step A3: Apply the risk index obtained in Step A2. Perform time smoothing to obtain the smoothed risk index. The calculation method is as follows:
[0080] in, The smoothing risk index represents the time window t and the region r. This represents the smoothed risk index for the previous time window; This represents the original risk index calculated in step A2; Represents the exponential smoothing coefficient, with a range of values. Step A4: Judge the result obtained in step A3. When the risk warning information is output, θ is the preset risk threshold.
[0081] Training and optimization of the risk calculation model: The dataset is divided into training and test sets in an 8:2 ratio. The training set is used for model training, and the test set is used to evaluate the model's performance. The risk calculation model is a linear regression model, using the mean squared error (MSE) as the optimization objective, and is trained based on the difference between the true label and the predicted risk index. Each iteration calculates the gradient of the loss function and updates the model parameters using the learning rate. The Adam optimizer is used to adjust the learning rate, with an initial value of 0.01. An L2 regularization term is added to avoid overfitting.
[0082] The model is trained using the training set, continuously optimizing the model parameters (regression coefficients) to minimize the loss function until it converges. After each training round, the model's performance on the validation set is evaluated to check for overfitting. During training, the model's performance is evaluated using the test set, and evaluation metrics such as MSE, RMSE, and MAE are calculated. .
[0083] Of course, the above are just typical examples of the present invention. In addition, the present invention may have many other specific embodiments. All technical solutions formed by equivalent substitution or equivalent transformation fall within the scope of protection claimed by the present invention.
Claims
1. A social risk perception system for infectious diseases, characterized in that: Includes a data acquisition module, used to acquire interactive data from the interactive platform; The data preprocessing module is used to preprocess the interactive data to obtain the interactive dataset; the feature analysis module is configured with several feature analysis items and a feature analysis unit for each feature analysis item. The feature analysis unit is used to analyze the feature item data of the interactive data to obtain the feature item information of the interactive data. The risk warning module is equipped with a risk analysis algorithm and risk triggering conditions. The risk analysis algorithm is used to calculate the confidence risk value of each interactive data and output risk warning information according to the risk triggering conditions satisfied by the confidence risk value.
2. The infectious disease social risk perception system as described in claim 1, characterized in that: The data acquisition module includes a data identification unit and a data crawling unit. The data identification unit is configured with a topic filtering model and an epidemic keyword library. The topic filtering model is used to extract topic words from the interactive data. The epidemic keyword library is configured with epidemic keywords. The data identification unit combines the epidemic keywords in the interactive data with the interactive data. The data crawling unit determines whether the crawling conditions are met based on the topic words and matching results. If the conditions are met, the interactive data is crawled.
3. The infectious disease social risk perception system as described in claim 1, characterized in that: The data preprocessing module includes a data denoising unit, a missing information compensation unit, and a data standardization unit. The data denoising unit is used to identify and delete redundant information in the interactive data. The missing information compensation unit is used to identify and supplement missing information in the interactive data. The data standardization unit is used to adjust the format of the interactive data according to the type of interactive data so that the interactive data meets the baseline format corresponding to that type.
4. The infectious disease social risk perception system as described in claim 1, characterized in that: The feature analysis module includes a counting feature unit, an emotion feature unit, a semantic feature unit, an interaction feature unit, and a spatiotemporal feature unit. The counting feature unit is used to calculate the counting features in the interaction data; the emotion feature unit is used to calculate the emotion features in the interaction data; the semantic feature unit is used to calculate the semantic features in the interaction data; the interaction feature unit is used to calculate the interaction features in the interaction data; the spatiotemporal feature unit is used to calculate the spatiotemporal features in the interaction data; the aggregation feature unit is used to calculate the aggregation features in the interaction data; and the sequence derivation feature unit is used to calculate the sequence derivation features in the interaction data.
5. The infectious disease social risk perception system as described in claim 1, characterized in that: The counting features include total text volume, epidemic-related text volume, and the proportion of epidemic-related text.
6. The infectious disease social risk perception system as described in claim 1, characterized in that: The emotion feature unit is configured with an emotion classification model, which is used to analyze interaction data to obtain an emotion intensity index. The emotion feature includes the emotion intensity index.
7. The infectious disease social risk perception system as described in claim 1, characterized in that: The semantic feature unit is configured with a topic analysis model, which is used to analyze interactive data to obtain a semantic topic index. The semantic feature includes the semantic topic index.
8. The infectious disease social risk perception system as described in claim 1, characterized in that: The spatiotemporal features include time periodic features and geographical distribution features, and the time series features include rate of change features and moving average features.
9. The infectious disease social risk perception system as described in claim 4, characterized in that: The aggregation feature unit is used to calculate the aggregation index based on the feature data of other feature analysis items of the interaction data, and the aggregation feature includes the aggregation index.
10. The infectious disease social risk perception system as described in claim 1, characterized in that: The risk analysis algorithm is configured as follows: Step A1, using a time window... With the region Define the original feature set for granularity: ;in, For counting characteristics, As an emotional characteristic, For semantic features, As an interactive feature, Spatiotemporal characteristics, As an aggregation feature, For time series features; Step A2, convert the feature vector obtained in step A1 into a time series feature vector. Input into the risk calculation model and perform a linear weighted combination: Map the linear result to an interval using the Sigmoid function. ： ;in, 、 、 、 、 、 、 These are the weights of the feature analysis terms, and their values are greater than or equal to 0. For bias terms, The activation function is Sigmoid; Step A3: Apply the risk index obtained in Step A2. Perform time smoothing to obtain the smoothed risk index. The calculation method is as follows: ;in, The smoothing risk index represents the time window t and the region r. This represents the smoothed risk index for the previous time window; This represents the original risk index calculated in step A2; Represents the exponential smoothing coefficient, with a range of values. Step A4: Judge the result obtained in step A3. When the risk warning information is output, θ is the preset risk threshold.