Network public opinion hotspot prediction method and device based on integrated network classification model
By integrating network classification models and semantic feature extraction, the problem of insufficient accuracy and generalization ability in predicting online public opinion hotspots is solved, and more efficient prediction of public opinion hotspots is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NO 30 INST OF CHINA ELECTRONIC TECH GRP CORP
- Filing Date
- 2026-03-31
- Publication Date
- 2026-07-21
AI Technical Summary
Existing methods for predicting online public opinion hotspots are insufficient in accuracy and generalization when faced with the large volume and timeliness of online public opinion information in the era of intelligence, making it difficult to effectively cope with complex public opinion information.
We employ an ensemble network classification model approach, including SVM, XGBoost, and LightGBM models, combined with a sentence-transformer for semantic feature extraction and weight learning. We optimize model parameters and weights through cross-validation to achieve multi-model weighted prediction.
It improves the accuracy and generalization ability of predicting online public opinion hotspots, reduces errors caused by word segmentation mistakes, and enhances the flexibility and predictive performance of the model.
Smart Images
Figure CN122432335A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of public opinion prediction technology, and in particular to a method and apparatus for predicting online public opinion hotspots based on an integrated network classification model. Background Technology
[0002] Online public opinion refers to the collection of opinions, attitudes, emotions, and behavioral tendencies expressed by the public on the internet regarding specific events, topics, policies, or entities. It primarily spreads through platforms such as social media, forums, and blogs, mainly reflecting the collective psychology and public opinion dynamics of social groups. With the development of internet technology and the explosive growth of news information, online public opinion exhibits characteristics such as instant dissemination, rampant dissemination of false information, and emotional polarization.
[0003] Currently, methods for predicting online public opinion primarily rely on data-driven approaches, such as using models like ARIMA and Prophet to analyze historical data and then predict future public opinion trends. However, traditional methods for predicting online public opinion hotspots are mostly based on single models. Faced with the challenges of the large volume and time-sensitive nature of online public opinion information in the intelligent era, traditional methods have significant limitations in accuracy and generalization ability. Therefore, it is necessary to propose an innovative and efficient prediction method specifically for predicting online public opinion hotspots. Summary of the Invention
[0004] The technical solution adopted in this invention is to improve the accuracy and generalizability of online public opinion hotspot prediction results. In view of this, this invention provides a method and apparatus for predicting online public opinion hotspots based on an integrated network classification model.
[0005] The technical solution of this invention proposes a method for predicting online public opinion hotspots based on an ensemble network classification model, including: Step 1: Obtain news data within a preset time range to construct a dataset, and preprocess the dataset; Step 2: Use the sentence-transformer model to extract semantic features from the preprocessed news text and obtain the corresponding news semantic vector; Step 3: Use the news semantic vector as a feature and input it into a set of ensemble network classification models for training; Step 4: Based on the acquired training data, obtain the weights of each classification model in the ensemble network classification model through cross-validation learning. The weights will be used to weight the prediction results of each classification model in the subsequent prediction process. Step 5: Obtain the news data to be processed, use the trained ensemble classification model to make predictions, and output the final hot topic prediction results through the weight combination.
[0006] In one implementation, the integrated network classification model includes three types: SVM, XGBoost, and LightGBM.
[0007] In one implementation, the SVM uses the news semantic vector as input features, uses radial basis function kernels to map the data to an infinite-dimensional space to make it linearly separable, and sets the model parameters according to the characteristics of the model and the characteristics of the actual dataset. The XGBoost is used to construct a gradient boosting tree model based on the news semantic vector. The complexity of the classification model is controlled by adding and adjusting the number of leaf nodes and the weight of leaf nodes in the objective function. The LightGBM employs gradient-based one-sided sampling and mutually exclusive feature binding techniques, along with a leaf-wise gradient boosting framework and histogram algorithm to accelerate training, thereby reducing computational load and memory consumption.
[0008] In one implementation, after configuring the initial weights, a loss function is defined to adjust the weights, specifically using the cross-entropy loss function to represent the difference between the prediction result and the actual situation: in, This indicates the actual tag, using 1 and 0 to represent historical hot topics and non-hot topics, respectively. Indicates the first The model predicts the first... The probability that a sample is a trending news item. Indicates the first The weights of each model, Indicates the number of samples.
[0009] In one embodiment, the preprocessing includes: noise removal and text cleaning of the acquired news data.
[0010] In one embodiment, the method further includes: Based on the prediction results and the actual situation corresponding to the news data, the accuracy, recall, and F-score are calculated to evaluate the actual performance of the ensemble network classification model. Based on the performance evaluation results, adjust the parameters of the ensemble network classification model.
[0011] In one implementation, during model training, k-fold cross-validation is used to randomly and evenly divide the training set into k subsets of similar size, and to perform k training and validation cycles to evaluate the stability and generalization ability of the current model on different subsets.
[0012] Another aspect of the present invention provides a device for predicting online public opinion hotspots based on an integrated network classification model, comprising: The acquisition unit is used to acquire news data within a preset time range, construct a dataset, and preprocess the dataset. The extraction unit is used to extract semantic features from the preprocessed news text using the sentence-transformer model and obtain the corresponding news semantic vector. The training unit is used to input the news semantic vector as a feature into a set of ensemble network classification models for training. The weighting unit is used to obtain the weights of each classification model in the ensemble network classification model through cross-validation learning based on the acquired training data. The weights will be used to weight the prediction results of each classification model in the subsequent prediction process. The prediction unit is used to acquire news data to be processed, make predictions using the trained ensemble classification model, and output the final hot topic prediction result through the weight combination.
[0013] Another aspect of the present invention provides an electronic device, comprising: a memory and a processor, wherein the memory stores a computer program, and the processor is configured to execute the computer program to implement the network public opinion hotspot prediction method based on the integrated network classification model as described in any of the preceding claims.
[0014] Another aspect of the present invention provides a computer storage medium storing a computer program that is executed to implement the online public opinion hotspot prediction method based on an integrated network classification model as described in any of the preceding claims.
[0015] By adopting the above technical solution, the present invention has at least the following advantages: The method provided by this invention can leverage the advantages of multiple classification models, improving the model's generalization ability and prediction accuracy. At the same time, the adaptive weights obtained through learning enable the ensemble classification model to automatically adjust the contribution of each classification model, improving the flexibility of prediction. The weights are dynamically adjusted according to the characteristics of different data and the performance of the model to achieve optimal prediction results. Attached Figure Description
[0016] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a schematic diagram of the online public opinion hotspot prediction method based on an integrated network classification model according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the integrated network classification model structure according to an embodiment of the present invention. Detailed Implementation
[0017] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments.
[0018] While exemplary embodiments of the invention are shown in the accompanying drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the invention and to fully convey its scope to those skilled in the art. The invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0019] This invention provides a method for predicting online public opinion hotspots based on an ensemble network classification model, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain news data within a preset time range to construct a dataset, and preprocess the dataset; Step 2: Use the sentence-transformer model to extract semantic features from the preprocessed news text and obtain the corresponding news semantic vector; Step 3: Use the news semantic vector as a feature and input it into a set of ensemble network classification models for training; Step 4: Based on the acquired training data, obtain the weights of each classification model in the ensemble network classification model through cross-validation learning. The weights will be used to weight the prediction results of each classification model in the subsequent prediction process. Step 5: Obtain the news data to be processed, use the trained ensemble classification model to make predictions, and output the final hot topic prediction results through the weight combination.
[0020] Specifically, Figure 1 This is a flowchart from an embodiment of the present invention. For example... Figure 1As shown, a method for predicting online public opinion hotspots based on an ensemble network classification model includes data preprocessing, feature extraction, ensemble classification model training, weight learning, comprehensive judgment, performance evaluation, and model optimization. This invention provides a news data processing method that helps improve the accuracy and generalizability of online public opinion hotspot prediction.
[0021] 1. Data preparation 1) First, news data within a certain timeframe and scope needs to be collected to build a dataset, providing a sufficient data foundation for model training. Specifically, the collected news data includes news titles and content, which can be obtained from news websites and social media platforms to ensure data diversity and representativeness. In addition, based on the collected news data, it is necessary to determine whether the news is trending based on metrics such as readership, number of comments, and number of reposts, combined with some manual quality control, and to label it accordingly. In this embodiment, 1 represents trending news and 0 represents non-trending news.
[0022] 2) To ensure the purity and quality of the data, and thus the correctness of feature extraction and model training, the collected news data needs to be preprocessed according to the actual situation. First, natural language processing (NLP) technology is used to identify and mark useless information in the collected data, such as advertisements, and remove them. Then, regular expressions are used to match the news content, identify and remove HTML tags from the text, retaining only the plain text content. Next, special characters in the text, such as "@", "#", and "&", need to be uniformly removed or replaced to reduce noise from special characters and their impact on model training results. Finally, to make the text data in the dataset more standardized, this invention standardizes the text encoding format in the collected news data and handles issues such as capitalization and spaces in the text.
[0023] 2. Feature Extraction Traditional text semantic extraction methods require pre-segmentation of the text, and the effectiveness of semantic feature extraction is largely affected by the segmentation process. To address this issue, this invention uses a sentence-transformer model to extract semantic features from pre-processed news text, obtaining corresponding news semantic vectors. This process eliminates the need for segmentation, thus effectively avoiding error propagation caused by segmentation errors compared to traditional methods. This improves the accuracy and efficiency of feature extraction and reduces semantic information loss during the extraction process to some extent.
[0024] Specifically, the feature extraction step first requires loading a pre-trained sentence-transformer model. This invention uses the all-MiniLM-L6-v2 model, which has 6 layers of transformer modules and is pre-trained on a sufficiently large corpus, thus possessing good semantic understanding capabilities. Simultaneously, this model utilizes knowledge distillation techniques to maintain high performance while maintaining a small size and fast inference speed.
[0025] Next, the preprocessed text string needs to be converted into a numerical format supported by the model through a text encoding process. Specifically, the text needs to be tokenized, that is, the text needs to be broken down into smaller semantic units, converted into a token sequence, and special markers need to be added at the beginning and end of the text sequence.
[0026] Subsequently, the encoded text tensor is input into the model's transformer encoder. The multi-head attention mechanism of the transformer encoder captures and learns the semantic features of the text. Then, the output of the self-attention layer is used as the input of the feedforward neural network for further nonlinear transformation and feature integration to obtain the semantic vector of each token.
[0027] Finally, to ensure the consistency of the semantic vector's dimensions, the generated semantic vector also needs to be processed using an average pooling strategy to obtain a single, fixed-dimensional vector. This vector serves as the semantic representation of the entire sentence, aggregating information from all tokens and comprehensively capturing the semantic features of the text, including information from multiple levels such as vocabulary, syntax, and semantics.
[0028] 3. Model Training After feature extraction is complete, the extracted semantic vectors need to be used as features and input into a set of integrated classification models for training. For this invention, to fully leverage the advantages of multiple classification models and learn from the speech vectors of historical hot and non-hot news, ultimately providing high prediction performance, the set of classification models used in this invention includes SVM, XGBoost, and LightGBM.
[0029] Meanwhile, for the complete dataset, this invention divides it into three categories: training set, validation set, and test set according to a ratio of 6:2:2 or 7:1:2. The three categories of data do not overlap to ensure that the model has enough data to learn during the training process, while reserving some data for validation and testing of the model's performance.
[0030] In addition, this invention employs k-fold cross-validation, randomly dividing the training set into k equally sized subsets and performing k training and validation iterations to evaluate the model's stability and generalization ability across different subsets, thus avoiding overfitting and underfitting. In this invention, k is typically taken as 5 or 10, depending on the specific data.
[0031] The SVM model uses semantic vectors as input features. Its core idea is to find a hyperplane for given input data and learning objectives, such that data points of different categories can be distinguished as much as possible. For linearly inseparable data in real-world scenarios, this invention uses a Radial Basis Function (RBF) kernel to map the data to an infinite-dimensional space, making it linearly separable. Furthermore, based on the characteristics of the model and the properties of the actual dataset, the model parameters need to be set appropriately, such as the penalty coefficient C and the RBF kernel parameter gamma. These parameters help control the model's tolerance for misclassified samples and the impact range of a single sample. The specific parameter settings need to be adjusted based on the cross-validation results.
[0032] In the specific training process, the first step is to standardize all semantic vectors in the training set and use the same mean and variance for both the validation and test sets to mitigate the impact of feature scaling on the SVM. Next, in k-fold cross-validation, a grid search is performed on different training sets to obtain the parameter combination with the highest cross-validation score. This parameter combination is then used to retrain the SVM model on the entire training set. Finally, after training, the model performance is evaluated to avoid overfitting. Through these training steps, the final SVM model is obtained.
[0033] The core idea of the XGBoost model is Gradient Boosting, which involves sequentially training a series of decision trees. Each new tree learns by fitting the residual of the predictions made by the previous tree, thus correcting prediction errors through the continuous stacking of decision trees. This invention constructs a gradient boosting tree model based on semantic vectors. By adding and adjusting regularization terms such as the number of leaf nodes and their weights in the objective function, it effectively controls model complexity and prevents overfitting. Simultaneously, it utilizes parallel computing and feature subsampling to improve the training rate, enabling the model to learn useful patterns from the data.
[0034] The LightGBM model, also based on the Gradient Boosting concept, differs from other methods in that it employs gradient-based one-sided sampling and mutually exclusive feature binding techniques, along with a leaf-wise gradient boosting framework and histogram algorithms to accelerate training. This significantly reduces computational and memory consumption, making it particularly suitable for processing large-scale data and efficiently learning hotspot patterns from historical news. Depending on the characteristics of the actual dataset and the cross-validation process, adjustments to parameters such as the learning rate, tree depth, and leaf depth are necessary during the implementation of this invention.
[0035] The training process of XGBoost and LightGBM models is similar. They require sampling in the parameter space using a random search method. Each set of parameters is cross-validated with k folds on the training set. During the cross-validation process, an early stopping mechanism is used to stop training when no improvement is achieved in multiple consecutive rounds. The number of decision trees and other parameter combinations are determined by the average number of iterations in the cross-validation and the average score of each set of parameters. The model is then retrained on the entire training set, and the final model is saved.
[0036] 4. Weight Learning Since this invention uses multiple classification models for prediction, in order to ensure the accuracy of the final prediction results, it is necessary to learn the weights of each classification model based on the training data through methods such as cross-validation. These weights will be used to weight the prediction results of each model in the subsequent prediction process, so that the model can automatically adjust the contribution of each classification model, thereby improving the flexibility and accuracy of the prediction.
[0037] To facilitate the weight learning calculation process, it is necessary to first set the initial weights for each classification model. These initial weights can be obtained through methods such as equal or random settings. In this invention, the initial weights for SVM, XGBoost, and LightGBM are set to 0.3, 0.3, and 0.4, respectively.
[0038] After setting the initial weights, a loss function needs to be defined to measure the difference between the prediction results of the ensemble classification model and the actual labels for a given combination of weights. This invention uses the cross-entropy loss function to represent the gap between the prediction results and the actual situation, as follows: in, This indicates the actual tag, using 1 and 0 to represent historical hot topics and non-hot topics, respectively. Indicates the first The model predicts the first... The probability that a sample is a trending news item. Indicates the first The weights of each model, Indicates the number of samples.
[0039] Subsequently, based on the determined loss function and training data, gradient descent is applied to the loss function. By calculating the gradient of the loss function with respect to the weights, the weights are iteratively adjusted with a certain learning rate to minimize the loss function and gradually approach the optimal solution, so that the prediction result of the combined model is as close as possible to the true result.
[0040] 5. Predicted Output and Model Optimization After completing the model training process, the trained ensemble classification model can be used to predict new news data, and the final hot topic prediction results can be output through the weight combination obtained from training, providing highly accurate online public opinion hot topic prediction information.
[0041] First, the new news data undergoes the same preprocessing operations as the training data, such as noise removal and text cleaning. Then, the same sentence-transformer model is used to extract semantic features from the data, and pooling is used to obtain semantic vectors of the same dimension as the training data. This ensures that the new news data is consistent with the training data in terms of format and quality, while also ensuring that the new data can be correctly understood and processed by the integrated classification model.
[0042] The generated semantic vectors are then input into the trained ensemble classification model to obtain the prediction results of each model. Based on the learned weights, the prediction results of each model are then weighted and fused. The specific process is as follows: Figure 2 As shown. Specifically, the probability of the fused hotspot prediction is expressed as follows: in, This represents the normalized weights of the learned models. This represents the prediction probability of each model. The weighted combined prediction probability is compared with a pre-set threshold to obtain the final output of the ensemble classification model.
[0043] To ensure the model's predictive accuracy, it is necessary to regularly collect the model's prediction results in real-world applications and compare them with actual trending news events. Metrics such as accuracy, recall, and F-score should be calculated to evaluate the model's actual performance. Subsequently, based on the performance evaluation results, any problems with the model should be analyzed, and the model's parameters and hyperparameters should be adjusted accordingly.
[0044] Compared to existing technologies, most online public opinion hotspot prediction methods use single classification models, which suffer from issues such as susceptibility to misinformation and difficulty in scaling to multiple news domains. To address the poor accuracy and generalization ability of existing online public opinion hotspot prediction methods, this invention proposes an ensemble network classification model-based method. This method integrates multiple classification models to predict online public opinion hotspots and weights them according to certain weights to obtain the final prediction result. Compared to traditional online public opinion hotspot prediction methods, the proposed method fully leverages the advantages of multiple classification models, improving the model's generalization ability and prediction accuracy. Furthermore, the adaptive weights obtained through learning allow the ensemble classification model to automatically adjust the contribution of each classification model, increasing prediction flexibility. By dynamically adjusting the weights based on the characteristics of different data and the model's performance, optimal prediction results can be achieved. Furthermore, traditional semantic feature extraction methods mostly require pre-segmentation of the text, so errors in segmentation can significantly impact the semantic feature extraction results. To address this issue, this invention employs a sentence-transformer for news semantic analysis, avoiding the propagation of errors caused by segmentation mistakes, improving the accuracy of semantic feature extraction, and more effectively capturing the deep semantic information of news texts, providing more reliable data support for predicting public opinion hotspots. The proposed online public opinion hotspot prediction method based on an ensemble network classification model is stable and easy to operate. Compared to traditional online public opinion hotspot prediction methods, the proposed method has higher prediction accuracy and generalization ability, and can better handle complex public opinion hotspot prediction tasks.
[0045] In a second embodiment of the present invention, a device for predicting online public opinion hotspots based on an integrated network classification model can be understood as a physical device for executing the method provided in the first embodiment. This device may specifically include: The acquisition unit is used to acquire news data within a preset time range, construct a dataset, and preprocess the dataset. The extraction unit is used to extract semantic features from the preprocessed news text using the sentence-transformer model and obtain the corresponding news semantic vector. The training unit is used to input the news semantic vector as a feature into a set of ensemble network classification models for training. The weighting unit is used to obtain the weights of each classification model in the ensemble network classification model through cross-validation learning based on the acquired training data. The weights will be used to weight the prediction results of each classification model in the subsequent prediction process. The prediction unit is used to acquire news data to be processed, make predictions using the trained ensemble classification model, and output the final hot topic prediction result through the weight combination.
[0046] According to a third embodiment of the present invention, an electronic device includes a memory and a processor, which are connected to each other via a communication interface. The memory stores a computer program, and the processor executes the computer program to implement the network public opinion hotspot prediction method based on the integrated network classification model as described in the first embodiment above.
[0047] In a fourth embodiment of the present invention, a computer storage medium is provided, wherein a computer program is stored on the medium, and the computer program is executed to implement the network public opinion hotspot prediction method based on the integrated network classification model as described in the first embodiment.
[0048] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims. All of these forms are within the protection scope of this application.
Claims
1. A method for predicting online public opinion hotspots based on an ensemble network classification model, characterized in that, include: Step 1: Obtain news data within a preset time range to construct a dataset, and preprocess the dataset; Step 2: Use the sentence-transformer model to extract semantic features from the preprocessed news text and obtain the corresponding news semantic vector; Step 3: Use the news semantic vector as a feature and input it into a set of ensemble network classification models for training; Step 4: Based on the acquired training data, obtain the weights of each classification model in the ensemble network classification model through cross-validation learning. The weights will be used to weight the prediction results of each classification model in the subsequent prediction process. Step 5: Obtain the news data to be processed, use the trained ensemble classification model to make predictions, and output the final hot topic prediction results through the weight combination.
2. The method for predicting online public opinion hotspots based on an ensemble network classification model according to claim 1, characterized in that, The integrated network classification model includes three types: SVM, XGBoost, and LightGBM.
3. The method for predicting online public opinion hotspots based on an ensemble network classification model according to claim 2, characterized in that, The SVM uses the news semantic vector as input features, uses radial basis function kernels to map the data to an infinite-dimensional space to make it linearly separable, and sets the model parameters according to the characteristics of the model and the characteristics of the actual dataset. The XGBoost is used to construct a gradient boosting tree model based on the news semantic vector. The complexity of the classification model is controlled by adding and adjusting the number of leaf nodes and the weight of leaf nodes in the objective function. The LightGBM employs gradient-based one-sided sampling and mutually exclusive feature binding techniques, along with a leaf-wise gradient boosting framework and histogram algorithm to accelerate training, thereby reducing computational load and memory consumption.
4. The method for predicting online public opinion hotspots based on an ensemble network classification model according to claim 3, characterized in that, After configuring the initial weights, a loss function is defined to adjust these weights. Specifically, the cross-entropy loss function is used to represent the difference between the predicted results and the actual situation. in, This indicates the actual tag, using 1 and 0 to represent historical hot topics and non-hot topics, respectively. Indicates the first The model predicts the first... The probability that a sample is a trending news item. Indicates the first The weights of each model, Indicates the number of samples.
5. The method for predicting online public opinion hotspots based on an ensemble network classification model according to claim 1, characterized in that, The preprocessing includes noise removal and text cleaning of the acquired news data.
6. The method for predicting online public opinion hotspots based on an ensemble network classification model according to claim 1, characterized in that, The method further includes: Based on the prediction results and the actual situation corresponding to the news data, the accuracy, recall, and F-score are calculated to evaluate the actual performance of the ensemble network classification model. Based on the performance evaluation results, adjust the parameters of the ensemble network classification model.
7. The method for predicting online public opinion hotspots based on an ensemble network classification model according to claim 1, characterized in that, During model training, k-fold cross-validation is used to randomly divide the training set into k similarly sized subsets, and k training and validation cycles are performed to evaluate the stability and generalization ability of the current model on different subsets.
8. A device for predicting online public opinion hotspots based on an ensemble network classification model, characterized in that, include: The acquisition unit is used to acquire news data within a preset time range, construct a dataset, and preprocess the dataset. The extraction unit is used to extract semantic features from the preprocessed news text using the sentence-transformer model and obtain the corresponding news semantic vector. The training unit is used to input the news semantic vector as a feature into a set of ensemble network classification models for training. The weighting unit is used to obtain the weights of each classification model in the ensemble network classification model through cross-validation learning based on the acquired training data. The weights will be used to weight the prediction results of each classification model in the subsequent prediction process. The prediction unit is used to acquire news data to be processed, make predictions using the trained ensemble classification model, and output the final hot topic prediction result through the weight combination.
9. An electronic device, characterized in that, include: The system includes a memory and a processor, wherein the memory stores a computer program and the processor executes the computer program to implement the method for predicting online public opinion hotspots based on an integrated network classification model as described in any one of claims 1 to 7.
10. A computer storage medium, characterized in that, The medium stores a computer program that is executed to implement the method for predicting online public opinion hotspots based on an integrated network classification model as described in any one of claims 1 to 7.